The Firm That Thinks: Company Brains, Agent Coworkers, and the Future of Organizational Autonomy

A theory of the cognitive enterprise: shared institutional knowledge, agent coworkers, economics, cybersecurity, governance, and the limits of autonomy.

The organizational question behind the agent revolution

Consider two companies that produce the same digital service, serve the same customers, and face the same contractual obligations. The first has investors, financial assets, software infrastructure, and no employees; processes are initiated and executed by artificial agents, with statutory human governance retained wherever the law requires it. The second employs hundreds of people, yet each can interrogate an organization-wide knowledge and reasoning infrastructure, and each can work alongside persistent agents that initiate, negotiate, execute, and verify operational activities. They differ profoundly in employment and social structure. They may nevertheless share the same underlying cognitive architecture.

The juxtaposition matters because debate about enterprise artificial intelligence often conflates two separate propositions. The first is that computational systems can perform work previously assigned to humans. The second is that the firm itself might acquire durable mechanisms for representing its condition, coordinating action, retaining experience, and learning across changes in personnel and software. Automating tasks would not establish the second proposition: isolated competent agents can remain collectively incoherent, whereas a company with predominantly human activity may achieve unusually effective collective reasoning through a common institutional memory. The relevant unit of analysis is therefore neither the model nor the individual job, but the organization as an information-processing and action-producing system.

The hypothesis investigated here is that organizational intelligence can increasingly be engineered as an institutional capability, while the identities of the actors exercising that capability—and the degree to which they act without intervention—vary independently. The term Company Brain designates the persistent, governed substrate through which knowledge, current state, decisions, and authorized actions become mutually intelligible. Agent coworkers are computational participants with persistent operational responsibilities, not merely interfaces that await human questions. These are analytical definitions, not claims that an organization is conscious or that software has human interests.

This distinction is timely but not yet empirically settled. In a 2026 study of production agents, 68% of surveyed deployments executed at most ten steps before human intervention; the same study identified reliability as the leading development problem.¹ Meanwhile, contemporary workplace experiments document material gains in some contexts and losses in others, without showing that an autonomous firm can sustain strategy, finance, compliance, and execution indefinitely.²³ The article accordingly treats the two extreme companies as controlled thought experiments, develops an explicit model of their common architecture, and separates evidence from extrapolation. Its claim is not that agent-native firms must emerge, but that existing theories of the firm can be sharpened by asking how cognition, delegated action, and legal responsibility are coupled—or deliberately kept apart.

What is a firm when employees are no longer definitional?

A registered company is a juridical arrangement; a productive firm is an organized system for combining resources and carrying out economic activity. The concepts overlap but are not interchangeable. A legal entity can own rights, owe duties, and survive changes in personnel, whereas an operating organization may coordinate resources across several entities, contractors, platforms, and external suppliers. A thought experiment about a firm with no employees must therefore specify whether it concerns legal constitution, internal labor, operational performance, or the absence of human contribution anywhere in its economic network. More importantly, it must acknowledge that the productive firm has never been just an aggregation of individuals: its structures, procedures, and technologies have long been designed to achieve coordination beyond the capacities of any individual participant.

The evolution of organizational structure can be understood, in part, as a succession of responses to human cognitive and coordination limits. Specialization distributes knowledge and attention; hierarchy concentrates certain decisions while delegating others; standardized procedures make recurring situations manageable without renewed deliberation; accounting and reporting systems preserve information beyond individual memory. These arrangements expand collective productive capacity, but they also create new costs of communication, escalation, interpretation, and control. Simon’s analysis of administrative decision-making makes bounded rationality a fundamental feature of organizational design: individuals cannot consider every alternative or consequence, so organizations supply decision premises, routines, and structures within which choices become tractable.⁴ The firm is thus already an engineered information-processing arrangement, even before computers enter the picture.

Coase’s foundational argument explains the firm partly through the cost of using market transactions: when coordinating an activity internally costs less than negotiating and monitoring successive market exchanges, an entrepreneur may prefer direction within an organization.⁵ Williamson’s later transaction-cost program added the importance of uncertainty, contractual incompleteness, and asset specificity. Neither account requires every coordinating function to be performed by a human employee; the economic question is how governance arrangements affect the combined costs of production, coordination, opportunism, and failure. An agent-mediated transaction might reduce the expense of acquiring information while increasing the cost of detecting manipulation or enforcing an ambiguous agreement. Lower search cost alone does not settle the boundary of the firm.

The information-processing tradition makes the historical trajectory more explicit. Galbraith did not describe organizations as fixed hierarchies awaiting better software: he identified alternative ways to cope with uncertainty, including reducing the need for information processing through slack or self-contained tasks, and increasing processing capacity through vertical information systems and lateral relationships.⁶ Technological integration therefore belongs to a broader repertoire of organizational redesign, rather than constituting a wholly new response to complexity. Hayek’s argument about dispersed knowledge adds a crucial qualification: the information needed for effective decisions is often local, contextual, and incomplete, and its availability to a central decision-making mechanism cannot simply be presumed.⁷ Increasing the speed at which information circulates does not, by itself, resolve whether it has been interpreted correctly.

Enterprise software progressively extended these organizational mechanisms into executable systems. Early transaction-processing and material-requirements-planning applications codified calculations and recurring administrative decisions; manufacturing resource planning broadened their operational scope; and enterprise resource planning (ERP) subsequently pursued integration across functional boundaries.⁸ The important development was not merely that firms stored more information digitally. Activities previously coordinated through manual reconciliation, individual memory, departmental correspondence, and repeated managerial intervention could increasingly be performed through shared records, configured rules, and software-controlled transactions. A replenishment calculation, an invoice validation, or a production-planning update could be initiated by an event or schedule and executed without a human reconstructing the procedure each time.

An ERP system is consequently better understood as an integrated transactional and procedural infrastructure than as a corporate database. It represents business objects and their states—orders, inventory, assets, invoices, accounts, and resources—while enforcing at least some of the rules governing their permitted transitions. Shared master data, posting logic, authorization controls, and cross-functional workflows can make distinct departments act upon a more consistent operational record. Davenport’s analysis of enterprise systems emphasized both the integration opportunity and its organizational price: adopting common software processes can require a firm to change how work is performed, not simply automate its existing practices.⁹ A software configuration is therefore also an organizational design decision, encoding assumptions about responsibilities, legitimate operations, and the meaning of business events.

ERP was, however, only one component of the broader computational transformation. Enterprise platforms, in the relevant architectural sense, are reusable digital environments that provide common data models, application services, execution mechanisms, and interfaces through which multiple organizational processes and actors interact. Their functions differ. ERP, customer relationship management, product lifecycle management, and related domain systems maintain authoritative records for particular activities; business process management and workflow engines coordinate tasks, transitions, and exceptions; integration middleware, APIs, and event infrastructures connect otherwise separate applications; collaboration and content-management systems preserve documents and conversations; and analytical platforms make operational data available for reporting, forecasting, optimization, or decision support. These are functional distinctions, not mutually exclusive product classes: a modern enterprise application may implement several of them. Collectively, they allow portions of organizational memory, coordination, and execution to persist outside the heads of employees and outside any single department.¹⁰

This transition was accompanied by changes in the organization itself. The business-process redesign literature treated information technology not merely as a means of making existing tasks faster, but as an opportunity to reorganize work across departmental boundaries.¹¹ Yet the relationship is reciprocal rather than technologically predetermined: systems constrain and enable practices, while their design, interpretation, and actual use are shaped by organizational choices. Orlikowski’s structurational analysis cautions against treating technology as an external force that automatically produces a particular organizational form.¹² A highly integrated ERP can support standardized control, but it can also coexist with fragmented responsibilities, informal workarounds, duplicated records, and decisions made outside the system. The presence of a shared technical platform does not establish the existence of shared organizational understanding.

The knowledge-based interpretation exposes why this distinction matters. Grant’s theory treats the integration of specialized know-how as a central organizational function, rather than equating knowledge with stored data.¹³ Nelson and Winter emphasize routines and historically accumulated capabilities; Walsh and Ungson explain organizational memory in terms of acquisition, retention, and retrieval, including the possibility of error and misuse.¹⁴¹⁵ An ERP may establish that a purchase order was approved, but the evidence explaining why it was approved, which alternatives were rejected, whether the decision rested on an exceptional commercial judgment, and whether its underlying assumptions remain valid may be dispersed across documents, messages, people, and external events. Conversely, not all such knowledge can be made explicit without loss. Capturing additional records is not equivalent to preserving the capacity to interpret and evaluate them.

The distinction is therefore not between an allegedly unintelligent past and an intelligent future. Traditional enterprise software already embodies substantial fragments of organizational intelligence in codified rules, process logic, transaction semantics, optimization routines, and decision-support models. What remains difficult is making those fragments interoperable at the level of meaning and justified action. A process engine knows which step follows another under its configured model; an ERP validates a transaction against its rules; an analytical system identifies a statistical anomaly; a collaboration platform retains an explanation. None of these functions, considered separately, guarantees that an organizational actor can reconstruct the changing situation across systems, distinguish evidence from inference, identify applicable authority, evaluate competing responses, and learn from the verified consequences of its actions. Even when such capabilities exist locally, they may not be durable or coherent at enterprise scale.

A minimal operating firm must repeatedly sense relevant conditions, interpret observations, preserve context, choose among alternatives, commit resources, carry out interventions, verify effects, and revise plans. Historically, organization design and enterprise computing have progressively redistributed those functions among people, procedures, and machines. Their distribution remains heterogeneous: a deterministic warehouse controller, a workflow engine, a human negotiator, an optimization model, and an AI agent may all operate in different parts of the same feedback system. It would nevertheless be invalid to infer that every function is equally automatable: interpretation may depend on unavailable context, allocating money can invoke legally restricted powers, and modifying fundamental objectives requires legitimate authority rather than merely computational capability.

The first result is therefore both ontological and historical. Human employment is not necessary for every productive or coordinating operation, but human legal responsibility and social dependence cannot be removed merely by substituting software for workers. More significantly, the prospect of an agent-operated firm is not a discontinuity emerging from an entirely non-computational organization. It extends a decades-long transfer of organizational functions into software, from records and calculations to rules, transactions, workflows, and increasingly autonomous decisions. The unresolved question is whether these distributed capabilities can be connected into a persistent, context-aware, provenance-preserving, and governable organizational cognitive architecture—one that does not merely execute predefined processes but can relate the firm’s knowledge to its changing circumstances and to authorized action. That is the problem to which the concept of a Company Brain is addressed.

From enterprise knowledge to a Company Brain

The history of enterprise computing established that a firm can delegate substantial parts of its memory, coordination, and execution to software. An ERP system maintains transactions and enforces configured business rules; a workflow engine coordinates activities; a data platform integrates observations; process mining reconstructs aspects of what actually occurred. Yet the persistence of these systems does not necessarily produce the persistence of the organizational understanding required to use them. A purchase order can remain perfectly recorded after the people who approved it have departed, while the reasoning that made its terms acceptable, the exceptions it embodied, and the assumptions on which it depended disappear. The question is therefore not how to assemble a larger repository, but whether a firm can construct a durable computational capacity to interpret its own situation, preserve the grounds of its decisions, and connect what it knows—or believes it knows—to what it is entitled to do.

This ambition has deeper antecedents than contemporary generative AI. Walsh and Ungson conceptualized organizational memory through the acquisition, retention, and retrieval of information, while warning that memory can be misused as well as used.¹⁶ Alavi and Leidner subsequently situated knowledge-management systems within the creation, transfer, and application of organizational knowledge, rather than the storage of documents alone.¹⁷ Wegner’s theory of transactive memory supplied a further insight: a group’s effective memory depends partly on knowing who knows what and how expertise can be located and coordinated, not on every participant holding the same information.¹⁸ What these traditions describe is not an information deficit cured by universal access, but a problem of maintaining useful relations among knowledge, context, expertise, and collective action.

Distributed cognition makes the unit of analysis more radical. Hutchins’s study of navigation showed how cognitive accomplishments arise through interactions among people, instruments, representations, and established practices; Hollan, Hutchins, and Kirsh extended this perspective to the design of computational environments.¹⁹²⁰ On this account, cognition need not be localized inside one individual in order to be studied as a property of an organized activity system. Nevertheless, Tsoukas’s description of the firm as a distributed knowledge system supplies a warning against the literal brain metaphor: relevant knowledge is dispersed, situated, continuously emerging, and not exhaustively available to any overseeing mind.²¹ These arguments are not contradictory. The first legitimizes analysis of organizational cognition; the second rejects the assumption that cognition requires, or could achieve, an omniscient organizational center. A Company Brain should be designed from that tension, not by ignoring it.

This article proposes the Company Brain as a persistent, governed, and potentially federated organizational cognitive infrastructure: a system of interoperable representations, evidentiary records, interpretive mechanisms, decision constraints, and feedback processes through which authorized human and artificial actors can establish the grounds for organizational judgments, coordinate action, and revise those judgments in light of outcomes. The definition is functional and architectural. It makes no claim that a company is conscious, possesses beliefs in the phenomenological sense, or can acquire legal personality through computation. Nor does it imply that an accurate representation of the entire enterprise is attainable. Its distinctive objective is to make organizational reasoning institutionally persistent and inspectable even when the people, models, and applications contributing to it change.

Two distinctions are decisive. First, organizational memory is not organizational knowledge: a retained record may be false, obsolete, incomplete, misclassified, or inapplicable to the present case. Second, organizational knowledge is not organizational agency: the ability to identify a justified response does not entail authority to implement it. A mature Company Brain would therefore need to mediate between an epistemic order—what the organization has observed, asserted, inferred, and disputed—and an institutional order—what its actors are permitted, required, and forbidden to do. It is the traceable connection between these orders, not conversational fluency or the volume of indexed data, that marks the proposed departure from conventional enterprise knowledge management.

The distinction becomes clearer once the supposedly homogeneous category of enterprise information is decomposed. Declarative representations describe entities and relations; procedural models specify ways of performing work; episodic records preserve events, decisions, and their circumstances; and causal models express assumptions about how interventions affect outcomes. Normative representations specify obligations, prohibitions, permissions, and decision rights; social and relational representations describe responsibilities, expertise, commitments, and interdependencies; operational-state models approximate changing business conditions; and metaknowledge describes the origin, temporal scope, quality, uncertainty, and limitations of the other representations. These categories are analytically different even when implemented by overlapping technologies. A contract clause is not merely another sentence resembling a policy; a forecast is not an observation; a supplier’s promise is not proof of delivery; and an inference from past correlation is not evidence that an intervention will cause the predicted result.

Computer science already provides components for respecting these differences. Gruber’s work on portable ontology specifications treated shared conceptualization as a prerequisite for exchanging formally represented knowledge across systems.²² Ontologies and knowledge graphs can represent entities, identities, constraints, and relationships across domains; their value lies in explicit semantics and reusable relations rather than in graph structure alone.²³ Operational records, document corpora, event streams, and process definitions supply different forms of evidence. BPMN offers a standardized notation for business-process models, while process mining can examine the relation between intended and observed execution.²⁴ The World Wide Web Consortium’s PROV-O provides a formal vocabulary for describing how entities were generated, used, or attributed to activities and agents, and SHACL provides means to validate RDF data against specified structural constraints.²⁵²⁶ None of these standards, separately or together, warrants the stronger conclusion that the resulting claims are true: syntactic conformance, semantic compatibility, and evidential justification are different achievements.

More difficult still, organizational concepts are neither always uniform nor always worth forcing into a single canonical definition. A customer in a sales system, a contracting legal entity in finance, and the beneficiary of a service in operations may refer to overlapping but non-identical sets of actors. Even the meaning of available inventory may depend on reservations, quality holds, ownership, physical accessibility, and the promised delivery date. Star and Griesemer’s work on boundary objects is relevant precisely because cooperation among specialized communities often depends on representations that remain recognizable across domains while accommodating different local interpretations.²⁷ The ambition should consequently be semantic interoperability without epistemic uniformity: different domains can preserve their legitimate conceptual distinctions, while mappings make agreements, disagreements, and translation assumptions explicit. A universal vocabulary that suppresses such differences can create the appearance of integration at the cost of incorrect decisions.

Time poses a related problem, because the enterprise does not inhabit a single, synchronized informational present. Temporal-database research distinguishes valid time—when a recorded proposition is claimed to hold in the modeled world—from transaction time—when that proposition is recorded as part of the database’s state.²⁸ An event’s physical occurrence, its detection, its communication, its accounting recognition, and its eventual correction may happen at different moments. A supplier may report a disruption after a production schedule has been approved; an invoice may be posted after the period to which it pertains; a sensor observation may arrive only after an automated decision was executed. The Company Brain must therefore be able to reconstruct not only an updated best view of a past situation, but also what information was available, to whom, and under which rules when a particular decision was made. Retrospective correctness and contemporaneous reasonableness are not the same standard.

This leads to a more demanding notion of the enterprise record. For a material decision, it should be possible to recover the relevant assertions, their source and version, their temporal applicability, the transformations that produced them, and the conditions under which they were accepted or challenged. The model should distinguish direct observations from reported claims, interpretations, forecasts, causal conjectures, and authoritative instructions. Contradictions should not be silently resolved by whichever document appears highest in a search ranking; some can be settled by a designated system of record, some require reconciliation, and others express legitimate competing interpretations. The absence of a record must not be mistaken for proof that an event did not occur. Even numerical confidence is meaningful only when its method of estimation, calibration, and domain of validity are understood. In this sense, the Brain’s most valuable output may sometimes be an explicit statement that no sufficiently supported conclusion is available.

Generative models can help translate between heterogeneous representations, extract candidate relations, formulate explanations, and expose evidence through natural language; those capabilities do not make a language model an authoritative store of organizational truth. Retrieval-augmented generation provides a way of consulting external information, but retrieval does not establish source validity, resolve conflicting claims, or preserve permissions automatically. Empirical work on long-context language models has demonstrated that even locating relevant material within a supplied context can be sensitive to its position, under the models and tasks studied.²⁹ Accordingly, an intelligent interface should be understood as one possible access and reasoning mechanism over the Company Brain, not as the Brain itself. Durable records, explicit semantics, executable controls, and independently verifiable outcomes cannot safely depend on the model’s transient context window.

The difference between explaining a historical outcome and predicting the consequences of an action is equally fundamental. A process-mining system may detect that delayed approvals correlate with missed deliveries; it cannot infer from that association alone that removing approvals would improve performance without creating other harms. Causal inference requires explicit assumptions about the mechanisms that connect interventions and outcomes, and some effects will remain unidentified from observational data.³⁰ Research on digital twins of organizations demonstrates an important step toward operationally coupled representations: Park and van der Aalst connected process-state views, constraints, and action mechanisms in a feasibility study involving an artificial order-handling information system.³¹ Its significance lies in closing part of the loop between process observation and controlled intervention; its experimental setting does not establish that a general organizational twin, still less a general Company Brain, has been realized. The proposed Brain extends this research trajectory from process-state monitoring toward situated, evidence-preserving organizational deliberation—but that extension remains an engineering and scientific hypothesis.

Organizational understanding also cannot be reduced to converting everything employees know into documents. Nonaka’s account of knowledge creation emphasizes interaction between tacit and explicit knowledge and the organizational conditions under which new interpretations are developed.³² Weick’s treatment of sensemaking similarly directs attention to how participants construct workable interpretations of ambiguous situations, often retrospectively and through social interaction.³³ A maintenance engineer’s judgment may depend on physical cues not captured by sensors; a supplier manager’s anticipation may reflect years of interactions that cannot be translated into reliable rules. The appropriate objective is not universal extraction of tacit knowledge, but the preservation of its discoverability, relevant context, practical consequences, and opportunities for contestation. A Company Brain should be able to locate expertise, preserve reasoning when it can be articulated, record exceptions, and learn which questions require renewed human or specialized investigation. The prospect of an enduring organizational memory does not eliminate situated expertise; it changes the conditions under which that expertise can be found, combined, and transmitted.

These requirements favor a logically coherent but physically federated system. Operational authorities should remain close to the systems and people that can establish their validity; sensitive knowledge may need to remain within defined legal or security boundaries; and different domains may require different models, update frequencies, and evidentiary standards. Simon’s account of nearly decomposable complex systems provides a theoretical basis for coordinating relatively autonomous subsystems through controlled interfaces rather than assuming complete centralization.³⁴ Federation, however, must not be an excuse for unresolved fragmentation. A coherent Company Brain needs reliable identification across domains, explicit semantic mappings, traceable provenance, mechanisms for resolving or preserving conflicts, and well-defined rules determining which sources and authorities govern particular decisions. Its unity is a property of interoperable reasoning and accountable coordination, not a demand that all information be copied into one database.

The architecture must also represent itself as a fallible institution. It should know, in an operationally meaningful sense, which domains it covers, which sources have failed, which models are stale, which conclusions depend on disputed assumptions, and which actions exceeded or approached the boundary of their authorization. It should preserve the history of its own interpretations and corrections. This is not self-consciousness; it is organizational epistemic reflexivity—a proposed capability to inspect and revise the evidentiary and procedural grounds on which the organization acts. Without it, deeper integration can amplify a common error across many departments and agents. A shared representation may make coordination faster while making independent dissent less likely; strong institutional cognition therefore requires not only shared context, but also opportunities for independent verification and legitimate disagreement.

Consider a seemingly elementary question: Can we promise a customer that one hundred units will ship next Friday? A transactional system might report sufficient stock; a planning system might show allocated capacity; a supplier notice might disclose a late component; a quality record might impose a hold; a contract might distinguish dispatch from delivery; and a salesperson might have authority to make only certain commitments. A conventional dashboard can juxtapose these observations, and a chatbot can summarize them. The proposed Company Brain would go further by preserving the relationships among the underlying claims, identifying which dates and meanings apply, exposing contradictions and missing evidence, reconstructing the reasons supporting alternative commitments, and determining which classes of action are permitted under the relevant policy. Whether an actor subsequently makes the promise is a separate question of agency and authorization. Whether the shipment actually arrives is a separate question of verification. Whether the organization changes future commitments after a failure is a separate question of learning. The point is not that one system already performs all of these functions, but that they jointly define a testable research target.

The proposed concept therefore requires a stronger criterion than the presence of a knowledge graph, an enterprise search interface, or an AI assistant. A Company Brain becomes meaningfully distinct when organizational representations are semantically interpretable, temporally situated, evidentially traceable, institutionally governable, usable in consequential decisions, and revisable through observed outcomes. These are design properties, not a claim that complete knowledge, flawless reasoning, or autonomous execution can be guaranteed. The underlying research proposition is that connecting them can produce a form of epistemic continuity: the organization may retain not only records of what it did, but inspectable grounds for what it concluded, why it acted, when those grounds ceased to hold, and how subsequent experience changed its understanding. This would be a new degree of institutional capability, built from a long history of organizational knowledge practices and enterprise computing rather than appearing ex nihilo with generative AI.

The final distinction prepares the next step of the argument. The Company Brain is not the organization’s agent; it is the governed cognitive infrastructure through which different actors can acquire context, examine reasons, and coordinate decisions. A firm populated entirely by humans and one relying heavily on artificial agents could, in principle, use the same infrastructure. Their divergence begins when a representation of what might be done is coupled to a participant capable of initiating and carrying out action. That is why the transition from chatbots to agent coworkers must be analyzed separately from the construction of the Company Brain itself.

From conversational assistants to bounded agent coworkers

The Company Brain, as defined in the preceding section, is not itself an actor. It maintains an inspectable relationship among organizational evidence, interpretations, rules, and decisions; someone or something must still initiate an intervention, assume an operational role, and remain within a legitimate allocation of authority. The transition from conversational AI to organizational agents is therefore more consequential than a change in user interface. It connects the firm’s emerging cognitive infrastructure to the mechanisms through which the firm actually operates. Once that connection exists, an additional transformation becomes possible: more of the organization can be represented not only as information to be consulted, but as an explicit and selectively executable structure of roles, processes, obligations, and decision rights.

A conversational assistant is conventionally activated by a human request and returns information, analysis, or proposed artifacts. The same underlying model can instead be placed inside a runtime that responds to business events, retains task state across interruptions, invokes tools, coordinates work with other actors, and examines whether requested changes occurred. This is not a categorical division between chatbots and agents: a conversation can initiate an agentic workflow, a workflow can call a language model, and a persistent agent may communicate primarily through dialogue. Nor does every agent need to be based on a language model. The relevant distinction is the scope of delegated action and the architecture governing its persistence, decisions, interfaces, and consequences, rather than whether its interface resembles a colleague’s.

An agent coworker, in the deliberately functional sense used here, is a computational participant assigned a recurrent organizational role, supplied with access to relevant context and capabilities, and authorized to initiate or advance specified work under enforceable constraints. It need not reproduce human intentions, consciousness, or general intelligence. Its organizational identity is not legal personality; being responsible for a process queue is not the same as owning assets or owing duties in law. Coworker is useful because it invites analysis of sustained collaboration, handoffs, expertise boundaries, and shared outcomes. It becomes misleading if it erases the difference between software execution, human accountability, employment, and corporate representation. The deployer and other legally relevant persons or entities remain answerable under applicable law; the exact allocation depends on the action, jurisdiction, contracts, and governance arrangements.

The deeper novelty lies in what this form of participation demands of organizational design. In organization theory, formalization already has a precise historical meaning: the extent to which organizational instructions, roles, procedures, and communications are explicitly documented and preserved. The Aston research tradition treated it as a dimension distinct from specialization, standardization, and centralization.³⁵ It would therefore be historically incorrect to claim that AI agents introduce formalization into an otherwise informal enterprise. Hierarchies, accounting controls, operating procedures, ERP configurations, and workflow engines have progressively codified organizational activity for decades. What changes when artificial participants can act across processes is the required depth and operational consequence of formalization: descriptions of how the firm is supposed to function increasingly have to be translated into models that a computational actor can interpret, that an execution environment can enforce, and that an independent observer can inspect.

Two meanings of formal must nevertheless remain separate. In organizational sociology, a formal rule may be a written instruction whose application still depends on human interpretation; in computer science, a formal specification has an explicitly defined syntax and semantics permitting particular forms of mechanical analysis. A policy written in natural language is more explicit than an unwritten convention but is not thereby executable or mathematically verifiable. Conversely, a machine-checkable rule can be perfectly implemented while encoding a mistaken business assumption. The movement from documented organization to computationally formalized organization is therefore neither automatic nor equivalent to technological determinism. It involves additional acts of semantic definition, institutional authorization, software engineering, validation, and revision.

One can distinguish several progressively stronger forms of organizational explicitness. A procedure may first be documented so that participants can discover it; then structured as defined entities, conditions, decisions, and roles; then operationalized as a workflow, decision service, or authorization rule; then constrained and tested against selected invariants; and finally observed and revised using evidence from actual execution. This is a proposed analytical progression, not an established maturity standard, and its stages need not apply uniformly to every activity. The critical conceptual advance is that the organization begins to acquire an executable institutional model: a versioned representation of selected relationships among organizational purposes, knowledge, commitments, roles, decision rights, and permissible transitions. The model is neither the entire company nor a digital copy of all social reality. It is the part of the institution that can be made sufficiently explicit to govern consequential interactions between people, agents, and enterprise systems.

Important pieces of this apparatus already exist, although they address different problems. The Object Management Group’s Semantics of Business Vocabulary and Business Rules (SBVR) specifies a structured approach to business concepts and rules; BPMN represents process coordination; and Decision Model and Notation (DMN) provides a way to express decisions and business rules, including decision tables that may be validated or executed.³⁶³⁷³⁸ These standards help separate what terms mean, which activities can follow which others, and how particular decisions are derived. They do not automatically establish the truth of inputs, the legal sufficiency of a rule, or the legitimacy of the authority that adopted it. In an agent-native setting their significance increases because agents may invoke a decision service or follow a formally specified process without needing to reconstruct the rule from prose at every step.

Long before the present generation of language-model agents, research on organizational multi-agent systems had already investigated how computational agents could occupy roles within structures governed by goals and norms. The MOISE+ model, for example, distinguishes structural, functional, and deontic specifications: which roles and groups exist, how collective purposes are organized, and which obligations or permissions connect roles to activities.³⁹ Work on social commitments likewise examined how promises, obligations, and interactions can be represented as relations between agents within a social context.⁴⁰ These contributions establish important intellectual precedents for an agent organization: a collection of individually capable programs does not become a firm merely because they exchange messages. It requires a recognizable institutional order specifying how their actions relate to shared objectives and to one another. A computationally represented commitment, however, must not be confused with a legally enforceable undertaking; mapping between the two requires separate legal and organizational analysis.

The Company Brain and the executable institutional model would have complementary, but not identical, functions. The Brain supplies context-sensitive evidence, historical explanations, semantics, and competing interpretations; the institutional model specifies the conditions under which an actor may turn a proposed decision into an operative commitment. Agents use the former to deliberate within the limits of available information and the latter to determine what they may actually request or perform. In a credible architecture, neither a plausible answer from a model nor a favorable prediction is sufficient authorization. The boundary between inference and action is a first-class organizational boundary, enforced by transaction systems, independent policy checks, and appropriately authorized participants rather than by a sentence embedded in a model prompt.

This separation suggests a more demanding agent role contract than a job title or a list of accessible APIs. For each organizational role, the runtime would need a persistent identity and an authorized purpose; a definition of permitted inputs and the evidentiary quality required for consequential decisions; allowed tools and action classes; financial, operational, temporal, and jurisdictional limits; dependencies and segregation-of-duties requirements; escalation and exception conditions; and explicit rules for credential issuance, state retention, audit, suspension, and termination. These constraints should be versioned, so a later reviewer can reconstruct which mandate and policy were in force when an action occurred. A task may be delegated to another agent, but that delegation must not silently expand privileges or dissolve the originating actor’s organizational responsibility. Nothing in the possession of a general-purpose tool should imply unlimited permission to use its functions.

Consider a procurement agent responding to an anticipated component shortage. The Company Brain can connect a production forecast to a supplier’s reliability history, an open contractual obligation, inventory reservations, a liquidity projection, and the provenance of an incoming disruption notice. A decision service may determine that replenishment is justified under specified assumptions. But the agent’s authority still depends on whether the material falls within its mandate, whether spending and exposure ceilings are respected, whether a supplier must be screened, and whether approving the purchase would violate separation-of-duties rules. It might be permitted to prepare a purchase requisition, automatically issue a low-value catalog order, or seek approval for an exceptional commitment; the same evidence does not authorize all three. What makes the arrangement agentic is the ability to initiate and coordinate these activities without a new human prompt for every step. What makes it governable is that the permissible consequences are determined outside the agent’s discretionary interpretation of its instructions.

Interoperability standards assist, but do not themselves provide this institutional framework. The Model Context Protocol (MCP) supplies mechanisms through which clients access tools and contextual resources; its 28 July 2026 revision introduced a stateless protocol core and an extensible framework, including mechanisms relevant to long-running tasks and enterprise authorization.⁴¹ Agent2Agent (A2A) addresses discovery, communication, and task coordination among heterogeneous agent systems.⁴² A stateless protocol is not evidence that an organizational process is stateless; persistent task history belongs in an application or workflow layer. Equally, being able to call a tool or communicate with a peer does not entail authority to create liabilities, transfer money, disclose information, or alter infrastructure. Transport interoperability, task orchestration, identity, authorization, and legal capacity answer different questions and must remain separate layers.

This distinction also clarifies why persistent agents need more than a long conversation history. Multi-step organizational work crosses failures, deadlines, personnel changes, model upgrades, and transactions with irreversible external effects. Durable workflow systems such as Temporal illustrate how execution state can be reconstructed from recorded events after a worker failure, rather than depending on uninterrupted process memory.⁴³ In an enterprise setting, an agent’s durable record should identify the task, its originating authorization, relevant versions of evidence and policy, the actual tool invocations and their responses, and the status of external commitments. Retry behavior requires special care: repeating a payment, a purchase order, or a control command is not equivalent to retrying a read-only query. Idempotency keys, transactional boundaries, compensating actions, reconciliation, and explicit acknowledgment of uncertain outcomes become parts of organizational reliability, not merely infrastructure implementation details.

There is a further distinction between permission and institutional validity. Languages such as Cedar show that application authorization can be separated from application logic, evaluated against explicit principal–action–resource–context relationships, and subjected to bounded formal analysis. Research on Cedar has used formal modeling and verification-guided development to establish properties of the policy language and its implementation.⁴⁴ This is significant: selected rules can be made inspectable and testable rather than left as informal expectations of model behavior. Yet a formally verified authorizer proves properties about its specified rules and implementation, not that the rules express the right corporate policy, that the factual context is correct, or that a technically authorized act is lawful. Even well-formed permission sets may allow a harmful sequence of individually permitted transactions. A credible enterprise architecture must therefore combine local authorization with process-level invariants, aggregate exposure limits, independent checks, and the ability to halt or reverse actions where feasible.

A concrete architecture for an executable institutional role

The institutional distinction can be realized without postulating an autonomous enterprise operating system. Consider a manufacturer with an existing ERP—such as SAP S/4HANA or Dynamics 365 Finance and Supply Chain Management—a CRM, supplier portals, document repositories, an identity provider, and operational event sources. The design objective is to introduce a persistent procurement agent that can investigate shortages, propose replenishment, and issue only those purchase orders that fall within a delegated mandate. In architectural terms, this is a controlled extension of the existing enterprise rather than its replacement. The ERP remains authoritative for purchase orders and financial commitments; identity and authorization services remain authoritative for access; and the Company Brain supplies contextual representations whose claims can be traced to their sources. A human or artificial actor may reason over the same governed evidence without thereby acquiring the same right to act.

The concrete deployment separates five concerns. An observation and integration plane extracts authenticated business events and source records through supported APIs, change-data capture, or transactional event publication. CloudEvents can normalize event envelopes across producers, while a transactional outbox can reduce inconsistencies between a database transaction and the event announcing it; neither technique makes the business content of an event trustworthy by itself.⁴⁵⁴⁶ A federated knowledge plane resolves entity identities, preserves temporal qualifications and provenance, and assembles authorized context from relational stores, document indexes, knowledge graphs, and source systems. A workflow and agency plane owns durable cases, deadlines, retries, task assignments, model/tool invocations, and human handoffs. A governance and transaction plane maintains versioned organizational mandates, decision rules, limits, and approval requirements, applying them independently of model-generated reasoning. Finally, an assurance plane correlates business events, authorization decisions, committed transactions, and independently observed outcomes.

These planes need not correspond one-to-one with products. A plausible reference deployment could use an event broker with change-data capture, a relational and graph-backed semantic service with document retrieval, a durable workflow engine such as Temporal or a BPMN implementation, isolated agent workers, and a policy decision service such as Open Policy Agent or Cedar. The institutional model would combine a role-and-delegation registry, process definitions, decision tables where suitable, and signed or otherwise authenticated approvals. Workload identity can be issued independently of user identity; SPIFFE, for example, defines verifiable workload identifiers, but such an identifier authenticates software rather than granting it business authority.⁴⁷ NIST’s zero-trust model distinguishes policy decision from enforcement, and OPA documents a corresponding separation between a policy decision point and the applications acting as enforcement points.⁴⁸⁴⁹ The essential requirement is not the choice of framework but the fact that there is no direct mutation path from the reasoning runtime to the system of record.

Figure 1 depicts this role-level deployment. Solid arrows represent actual transfers, requests, or execution paths; dotted arrows represent rules and identity constraints. The gateway is the enforcement boundary: read and write privileges are differentiated, and an agent’s ability to invoke a tool is not sufficient to invoke its privileged operations.

%%{init: {"theme": "neo", "look": "handDrawn", "layout": "elk"}}%%
flowchart TB
  subgraph S[Systems of record and observation]
    ERP[(ERP and finance ledger)]
    SRC[(CRM, suppliers, documents, operations)]
    EVT[(Business event stream)]
    ERP -->|publish committed events| EVT
    SRC -->|publish observations| EVT
  end
  subgraph K[Federated Company Brain]
    INT(Semantic and temporal integration)
    KB[(Evidence, provenance and domain models)]
    CTX(Scoped context service)
    EVT -->|deliver events| INT
    INT -->|qualify and link records| KB
    KB -->|provide relevant evidence| CTX
  end
  subgraph A[Workflow and agency]
    WF(Durable case and workflow engine)
    AG[Isolated agent runtime]
    EVT -->|initiate cases| WF
    WF -->|assign bounded tasks| AG
    CTX -->|deliver authorized context| AG
  end
  subgraph P[Authority and transaction boundary]
    M[(Versioned mandates and decision rules)]
    GW(Action gateway and policy enforcement)
    DEC{Policy and risk decision}
    H[Authorized human approver]
    EX(Restricted transaction adapter)
    AG -->|submit typed proposal| GW
    GW -->|request decision| DEC
    M -.->|governs| DEC
    DEC -->|allow decision| GW
    DEC -->|needs approval| H
    DEC -->|deny or defer| J
    H -->|signed decision and recheck| GW
    GW -->|enforce approved operation| EX
    EX -->|execute approved operation| ERP
  end
  subgraph V[Independent assurance]
    J[(Workflow, decision and action journal)]
    RC(Outcome reconciliation)
    WF -->|record state transitions| J
    DEC -->|record policy evaluation| J
    EX -->|record execution receipt| J
    ERP -->|provide authoritative readback| RC
    RC -->|submit verified result| INT
  end
  ID[(Identity provider and delegation registry)] -.->|authenticates and constrains| GW
  ID -->|issues workload identity| AG

Figure 1: A bounded agent role is implemented by separate evidence, workflow, authority, transaction, and assurance functions; the agent proposes actions but cannot write directly to systems of record.

The application boundary can be specified more precisely through a typed action proposal, rather than a natural-language instruction sent to a privileged connector. For an illustrative procurement operation, the proposal would include a durable case identifier; the agent’s workload identity and its originating organizational principal; the requested business operation and target resource; the amount, currency, supplier, and item; references to evidence and its observation times; the role mandate and policy versions; an expiration time; the ERP record version against which preconditions must be checked; and an idempotency key. This is a proposed application contract, not a pre-existing interoperability standard. A minimal envelope might look as follows:

{
  "case_id": "CASE-1842",
  "proposer": "agent:procurement-eu",
  "mandate_id": "procurement-routine-v7",
  "action": "purchase-order.create",
  "resource": "erp:company-A",
  "parameters": {"supplier": "SUP-142", "item": "COMP-18", "quantity": 40, "amount_eur": 7200},
  "evidence_refs": ["supplier-confirmation:903", "inventory-snapshot:455"],
  "expected_state_version": "inventory:455",
  "policy_version": "finance-policy:31",
  "expires_at": "2026-11-15T12:00:00Z",
  "idempotency_key": "CASE-1842-PO-1"
}

The examples are hypothetical, and the identifiers express the distinctions that matter rather than an endorsed schema. Before acting, the gateway authenticates the workload and, where applicable, the human or corporate principal on whose behalf it operates; checks the mandate, supplier eligibility, cumulative exposure, spending ceilings, segregation of duties, and evidence validity; and re-reads mutable source state to reject stale proposals. It must treat policy_version as a reference to be validated, not as permission for the agent to select an obsolete policy. A missing or contradictory input produces a denial, deferral, or explicit escalation, not an invitation to improvise authority. Approval by a natural person authorizes a defined transaction under defined conditions; it is not an unrestricted credential that the agent can reuse for subsequent actions.

After authorization, a constrained adapter invokes an approved ERP operation using narrowly scoped credentials. It records the response and subsequently reads the ERP’s authoritative state, because a network timeout does not reveal whether the remote transaction committed. The workflow stores the outcome as proposed, authorized, submitted, committed, or reconciled, rather than reducing all states to completed. Durable workflow systems preserve ordered event histories for recovery, but they do not by themselves guarantee exactly-once effects in external systems.⁵⁰ Idempotency, concurrency checks, reconciliation, and carefully designed compensation are therefore first-order organizational controls. If an order is placed but the supplier rejects it, the system must preserve that distinction, reopen the case, and update the Brain’s evidence rather than treating the creation of an ERP record as proof that the procurement objective was achieved.

The architecture yields a further, testable distinction between machine-readable organization and machine-governed organization. Codifying an organizational rule in a workflow definition, decision table, or policy language makes its current form executable; connecting that rule to traceable evidence, delegated authority, and independent verification makes its application accountable. The former is already achievable in limited domains with established technologies. The latter requires difficult integration across changing enterprise semantics, distributed transactions, conflicting policies, and real-world outcomes. This role-level deployment gives a concrete interpretation to the broader organizational architecture considered later in the article; it is not evidence that a complete autonomous firm is already available.

The emergence of a more explicitly modeled enterprise should not be confused with the elimination of informal judgment. Feldman and Pentland’s analysis of organizational routines distinguishes the abstract pattern of a routine from its situated performance, showing how enactment both reproduces and changes the routine.⁵¹ Written policies and executable workflows are incomplete descriptions of changing social and commercial reality. An experienced employee may recognize that an apparently routine order conceals a customer dispute; an agent may encounter an unprecedented contractual exception that defeats the assumptions under which its decision table was designed. Increasing formalization can make some activities reliable and auditable while making others rigid, opaque, or vulnerable to the gaming of metrics. A cognitive enterprise must therefore formalize the management of exceptions, disagreement, and revision without pretending that every exceptional judgment can itself be specified in advance.

The most far-reaching possibility is that organizational structure becomes partly reprogrammable. If roles, capabilities, process dependencies, approval paths, and decision constraints have explicit representations, changes to an organization need not be implemented solely through revised job descriptions, meetings, or ERP configuration tickets. Certain modifications could be proposed as changes to the institutional model, checked for conflicts, simulated against historical cases, approved by designated authorities, deployed in controlled versions, and monitored for unintended consequences. Agents could identify dysfunctional handoffs or propose a new division of labor; they should not acquire the unilateral power to enlarge their own mandate, rewrite overriding policies, or define what counts as acceptable oversight. The design objective is not unconstrained self-organization, but controlled institutional adaptation in which the authority to alter organizational rules remains distinct from the ability to execute work under them.

That prospect also changes the political meaning of enterprise software. More extensive formalization makes some decisions more inspectable and contestable, but it can concentrate control in the parties that define organizational categories, performance measures, and permissions. Research on algorithmic management has documented how computational systems can direct, evaluate, and discipline workers, with consequences for power and resistance.⁵² The presence of human employees does not guarantee that machine-interpretable policy empowers them; a Company Brain may distribute relevant knowledge widely while an executable institutional model narrows their practical discretion. The design problem includes who can challenge a rule, correct an incorrect representation, appeal an automated decision, and influence revisions to the governing model. Formal transparency without such rights may make authority easier to observe but no easier to contest.

Finally, every coupling between organizational knowledge and tool-enabled action introduces a security boundary. An adversary who contaminates an apparently credible supplier message, document, tool response, or peer-agent communication may attempt to convert untrusted content into operational instructions. The danger is not limited to incorrect text: it arises when an actor with valid credentials uses them for actions outside the legitimate intent of a task. OWASP’s treatment of excessive agency identifies the combination of excessive functionality, permissions, or autonomy with unreliable or manipulated model behavior as a route to damaging action.⁵³ Defenses must therefore be structural: constrain authority at the execution boundary, preserve the distinction between data and instructions, validate relevant facts independently, monitor sequences of actions, and retain the practical ability to suspend a role or revoke its capabilities. No amount of natural-language prompting can substitute for controls that remain effective when the model is mistaken or the environment is adversarial.

The resulting picture is more radical than a firm with better assistants, but also more precise than the slogan of an autonomous company. The historical enterprise has increasingly formalized its operations through hierarchy, rules, records, and software; the Company Brain proposes to make the grounds of organizational reasoning more continuous and inspectable; persistent agents make selected portions of that reasoning operational without requiring a fresh human instruction at every step. Their convergence points toward a firm whose knowledge, division of labor, commitments, and permissible actions are increasingly explicit, machine-interpretable, executable, and revisable—while its legitimacy and accountability remain questions that cannot be settled by computation alone. The two organizational extremes considered next differ less in whether they need this architecture than in who participates in it, which powers are delegated, and where human judgment remains institutionally indispensable.

Extreme one: the company with shareholders but no employees

The preceding sections described a possible transition from enterprise software that records and regulates transactions, through a Company Brain that preserves organizational evidence and reasons, toward agent coworkers acting within an executable institutional model. The most demanding test of this architecture is not whether an agent can perform a job previously assigned to an employee. It is whether the organization can continue to exist as a functioning economic and institutional arrangement when it has no employees at all. An employee-free enterprise still needs to discover opportunities, obtain resources, make commitments, deliver value, account for its activities, respond to failure, and change course. Eliminating employment does not eliminate these functions; it changes the mechanisms through which they are performed, coordinated, governed, and made answerable to others.

Consider a corporation that owns intellectual property and a digital service, while shareholders supply capital and exercise the rights attached to ownership. A collection of agents develops and maintains the product, markets it, contracts for infrastructure, acquires external services, handles customer interactions, performs financial administration, and monitors commercial performance. Transactions flow through authenticated enterprise systems; a Company Brain retains the temporally qualified evidence, process context, and decision history needed by these agents; and an executable institutional model defines roles, spending limits, commitments, exceptions, and approval rights. This is a coherent candidate operating model, not evidence that an unrestricted autonomous firm has already been built. The thought experiment becomes scientifically useful only after its organizational boundary, legal constitution, operating processes, and permitted degree of autonomy are specified.

Four boundaries of the employee-free hypothesis

At least four interpretations must be distinguished. In the purely theoretical limit, the only directly participating humans are shareholders, and neither corporate governance nor routine operations depend on another natural person occupying an internal role; this is an analytical construction, not a claim of legal feasibility. In a legally constituted employee-free corporation, no workers are employed, but required natural-person directors, officers, or other functionaries continue to exercise statutory powers and duties, perhaps as shareholders or independently appointed persons. In an agent-operated digital firm, the internal operating core is computational while cloud providers, banks, auditors, lawyers, contractors, and other counterparties supply capabilities that may depend extensively on human labor. A robotically extended firm moves some execution into the physical world while inheriting additional dependencies on energy, maintenance, logistics, safety systems, and regulation. These models are not interchangeable: a company with no employees can be economically dependent on thousands of people, and a company that operates with negligible routine human intervention can remain subject to intensive human governance.

The boundary is also temporal. An agent-operated firm might be incorporated, capitalized, authorized, and initially configured by humans, then left to conduct routine transactions. If its original purposes, products, suppliers, and regulatory conditions subsequently change, a claim of organizational autonomy must specify whether humans may intervene to revise objectives, resolve disputes, recapitalize the firm, or wind it down. An organization that operates unaided for ninety days in a stable digital environment demonstrates a different property from one that can survive unanticipated contractual, technological, and market changes for years. Autonomy is always relative to a class of disturbances, a duration, a set of delegated powers, and an environment of continuing external services; it is not equivalent to isolation from human society.

Organization without an employee hierarchy

Organizational theory offers a more exact starting point than imagining a boardroom full of software personas. Mintzberg distinguishes an operating core, strategic apex, middle line, technostructure, and support staff, alongside different coordination mechanisms such as direct supervision, mutual adjustment, and the standardization of processes, outputs, or skills.⁵⁴ These are analytically separable functions rather than necessarily separate groups of employees. Their persistence in an agent-operated firm is the first structural constraint: work still requires integration, supervision, standardization, support, and direction, even if software performs many of the corresponding activities. The organization chart cannot simply be replaced by a list of models with occupational titles.

The operating core could consist of agents and deterministic services organized into enduring product, customer, financial, and supplier domains. A coordination layer would manage dependencies among these domains, reconcile competing resource demands, and allocate work according to explicit service and risk constraints. A computational technostructure would maintain process models, decision rules, domain vocabularies, model evaluation procedures, and delegated mandates. Support functions would operate identity, security, infrastructure, legal-document preparation, accounting, and vendor interfaces, although regulated professional services and other reserved activities might have to be obtained from qualified external persons. The strategic apex would remain divided: agents could analyze markets and propose plans, but decisions legally reserved to directors, shareholders, or other authorized persons would remain at the applicable human or corporate governance level. An artificial CEO may be a useful operational coordinator, but its title supplies neither statutory office nor additional authority.

This is not necessarily a conventional pyramid recreated in software. Malone and Crowston define coordination in terms of managing dependencies among activities, providing a more fundamental explanation for why an organization requires coordinating mechanisms at all.⁵⁵ Where work can be decomposed into independently verifiable services with explicit interfaces, autonomous agent teams may operate with little centralized direction. Where actions share scarce resources, create contractual liabilities, alter common data, or expose the entire enterprise to risk, integration mechanisms become unavoidable. The organizational design problem is therefore to decide which dependencies should be resolved by protocols, which by shared plans, which by delegated managerial agents, and which by an independent or human authority. The cheapest architecture for completing tasks need not be the safest architecture for maintaining the firm.

A useful design is a federation of bounded operational domains rather than a homogeneous population of free-ranging agents. Each domain would have an explicit purpose, an authoritative set of records, a budget or resource envelope, a process mandate, measurable service obligations, and defined interfaces with other domains. Agents could be created, upgraded, replaced, or combined within a domain without changing the legal identity of the firm or automatically expanding its authority. The Company Brain would preserve the semantics of cross-domain coordination and the reasons for significant decisions; the executable institutional model would preserve the operative allocation of decision rights. This separates organizational roles, which may persist for years, from agent implementations, which may change weekly. If the two are confused, replacing a model can unintentionally alter what the company is permitted to do.

The minimal organizational design would accordingly require an explicit organizational constitution at several levels. Corporate constitutional documents and applicable law define legal powers and reserved matters. A governing mandate translates those powers into corporate purposes, risk appetite, capital constraints, and delegation rules. A machine-interpretable operating model defines process ownership, dependencies, agent roles, authorities, and escalation triggers. Technical controls implement and verify selected aspects of those rules. These levels must remain distinguishable: a successful workflow execution cannot amend corporate articles, a process owner cannot promote an operating exception into a new corporate objective, and an agent cannot grant itself permission to rewrite the policy by which its permissions are evaluated.

This is a proposed architecture for constitutional separation within an executable organization, not an existing legal category.

Figure 2 represents this organizational division of labor without treating the agents as artificial corporate officers. Human participation is concentrated in ownership and legally recognized governance, while the internal operating core consists of specialized agent domains coordinating through shared institutional evidence; external counterparties and professional providers remain outside the employee-free boundary. The authority gate separates an agent’s proposed business action from a valid corporate commitment, and verified outcomes return to the Company Brain without allowing the agents to approve changes to their own mandates.

%%{init: {"theme": "neo", "look": "handDrawn", "layout": "elk"}}%%
flowchart TB
    subgraph GOV["Ownership and lawful human governance"]
        SH[Shareholders]
        DIR[Directors and authorized corporate organs]
        M[(Approved mandates and reserved matters)]
        SH -.->|ownership rights| DIR
        DIR -->|establishes delegation| M
    end

    subgraph CORE["Employee-free operating organization"]
        B[(Federated Company Brain)]
        W(Durable process coordination)
        A1[Product and engineering agents]
        A2[Commercial and customer agents]
        A3[Procurement and finance agents]
        P{Independent authority gate}
        S[Enterprise systems of record]
        V(Outcome verification)

        B -->|governed context| W
        W -->|assigns work| A1
        W -->|assigns work| A2
        W -->|assigns work| A3
        A1 -->|proposed actions| P
        A2 -->|proposed actions| P
        A3 -->|proposed actions| P
        P -->|authorized transactions| S
        S -->|authoritative results| V
        V -->|qualified evidence| B
    end

    EXT[External customers, suppliers and service providers]
    M -.->|constrains| P
    P -->|reserved decisions| DIR
    DIR -->|authorized decisions| P
    S -->|transactions and requests| EXT
    EXT -->|responses and services| S

Figure 2: The employee-free operating firm: specialized agent domains execute ordinary work through a federated Company Brain, while legal governance retains reserved powers and an independent authority gate controls consequential commitments.

Processes, not simulated departments, become the basic units of operation

If the organization has no employees, it cannot depend on a manager noticing that work has stalled, an experienced accountant remembering an exception, or an informal conversation reconciling inconsistent departmental reports. Its operating model must make the life cycle of consequential work explicit. Business process management distinguishes identification, discovery, analysis, redesign, automation, and monitoring; a process is not fully governed merely because its expected sequence appears in a diagram.⁵⁶ The agentic firm therefore needs an enterprise process architecture connecting customer value creation to resource management, reporting, control, and institutional adaptation. At a minimum, each material process needs an accountable corporate sponsor or legal authority where required, a durable operational owner, authorized inputs and outputs, defined exception routes, and evidence of completed outcomes.

The following table is a proposed functional process map, not an assertion that these responsibilities can all be delegated legally or performed reliably by present systems. It avoids organizing the firm around personified agents such as AI CFO or AI lawyer and instead asks which processes must remain viable and which commitments require stronger authority.

Process family	Agent-executable work in a bounded design	Critical governance boundary
Strategy-to-portfolio	Market sensing, scenario development, resource proposals, performance analysis	Change of corporate purpose, material investments, capital structure and reserved strategic decisions
Concept-to-product and change-to-release	Product specification, coding, testing, deployment preparation, monitored releases	Product safety, intellectual-property rights, material risk acceptance, independent release criteria
Market-to-contract and order-to-cash	Qualification, quotation, contract preparation, service delivery, invoicing and collection	Representational authority, price and credit exposure, unfair or unlawful terms, contested customer rights
Source-to-contract and procure-to-pay	Supplier evaluation, comparison, requisitions, purchase orders, invoice matching	Commitment limits, supplier due diligence, segregation of duties, payment authorization
Record-to-report and treasury	Ledger reconciliation, forecasts, tax-data preparation, cash monitoring, management reports	Financial statement approval, statutory filings, banking controls, audit independence and insolvency triggers
Risk-to-response and incident-to-recovery	Continuous monitoring, anomaly triage, containment proposals, recovery coordination	Independent challenge, external notification duties, safety and business-continuity decisions
Evidence-to-policy and change-to-organization	Process mining, exception analysis, proposed mandate and workflow changes	Approval of new authorities, validation of new models, constitutional and legal constraints

Table 1: Proposed enterprise process architecture for an employee-free operating model.

The table shows why the right abstraction is end-to-end obligations rather than job replacement. A sale has not succeeded because the sales agent issued an accepted quotation: the customer must receive the agreed service, the company must satisfy its duties, the related financial records must reconcile, and disputes must have a legitimate route to resolution. Likewise, procurement is not complete because a purchase-order API returned success; the delivery, receipt, quality status, invoice, payment, and contractual consequences must be accounted for. In an organization lacking human employees, these inter-process obligations cannot be left to unnamed informal practices. They require explicit handoffs, evidence requirements, reconciliation points, and defined authority over exceptions.

A business process should consequently be represented as more than a prescribed sequence of tasks. Its control contract must say which event initiates it, which evidence is admissible, how actors are identified, which decisions are reversible, which resources may be committed, which external obligations are created, and what constitutes final settlement. It must also define what happens when two systems disagree or an expected reply never arrives. A resilient design allows uncertain work to remain visibly unresolved rather than fabricating a conclusion to keep the process moving. This distinction is particularly important because agent decisions may be probabilistic while corporate ledgers, legal obligations, and payment systems require reproducible states and enforceable controls.

Process ownership has two separate meanings in such a firm. A computational process steward can monitor progress, maintain workflow state, investigate exceptions, and recommend remediation. The legally or institutionally accountable party remains the corporation acting through duly authorized governance arrangements, or another person or body where law assigns a specific duty. An internal RACI-like matrix that labels an agent accountable does not move statutory obligations to the model. Moreover, control functions must not collapse into the same optimization objective: an agent evaluated for sales growth should not be the sole judge of the legality of its contract terms, the reliability of its sales recognition, or the adequacy of its own exceptions.

The distinction between processes and organizational change becomes acute when the firm tries to reorganize itself. A group of agents may detect that supplier qualification is the bottleneck and propose a new purchasing role or routing rule. Under the executable institutional model described earlier, that proposal could be checked for conflicting permissions, simulated against historical cases, evaluated for its effects on cost and risk, and promoted through a versioned approval process. But changing how work is organized must remain a different category of action from performing work under an existing organization. The power to create another agent is not automatically the power to create another budget; the power to alter an internal workflow is not the power to relax a corporate control. A firm that fails to separate ordinary execution from constitutional change would be capable of erasing the boundaries that purportedly make its autonomy safe.

Who governs a company that operates itself?

The shareholders-only phrase is legally and institutionally misleading unless ownership is separated from governance. Shareholders ordinarily exercise rights through the company’s applicable constitutional and statutory framework; they do not automatically conduct its day-to-day management. Under the UK Companies Act 2006, sections 154–155 require at least one director of a private company and at least one natural-person director; section 171 requires directors to act within their powers, while sections 172–174 address the company’s success, independent judgment, and reasonable care, skill, and diligence.⁵⁷⁵⁸ Delaware’s General Corporation Law likewise establishes management by or under the direction of a board, subject to statutory and charter provisions, and requires corporate directors to be natural persons under section 141(b).⁵⁹ Other jurisdictions and legal forms require separate analysis. These examples rule out treating software as the automatic replacement for every statutory organ; they do not rule out extensive automation of business operations under human legal oversight.

The UN Commission on International Trade Law adopted its Model Law on Automated Contracting in 2024 to provide a framework for recognizing the formation and performance of contracts using automated systems, including AI.⁶⁰ It is a model for legislators, not a globally operative statute, and it does not endow software with legal personhood. A validly attributed automated transaction may bind the parties under applicable law, but attribution, contractual validity, authority, error allocation, and remedies still have to be established for the relevant jurisdiction and transaction. The legal problem is thus not simply whether a digital signature can be generated. It is whether a company has lawfully authorized the underlying class of commitments, whether the system acted within that mandate, and who bears the consequences when it does not.

Jensen and Meckling’s agency-cost analysis remains relevant because separating ownership, decision-making, and control creates problems of divergent interests and imperfect monitoring.⁶¹ In an artificial-agent organization these problems mutate rather than disappear. Software does not need bonuses or suffer human career incentives in the ordinary sense, but its objectives may be misspecified, its training or retrieval data misleading, its optimization criteria incomplete, and its provider’s interests different from the firm’s. Holmström and Milgrom’s multitask analysis warns more generally that rewarding a measurable dimension can distort effort away from less measurable responsibilities.⁶² That mechanism has an analogue in machine-governed processes: maximizing collected revenue, response speed, or conversion rates may sacrifice customer fairness, service reliability, or long-term resilience when those objectives are poorly represented. The analogy is informative, not an assertion that human incentive theory transfers mechanically to algorithms.

Ownership also concerns residual control rights. Hart and Moore analyze the distinction between ownership and contracting partly through the ability to determine uses of assets when agreements are incomplete.⁶³ Agent systems may generate and execute far more detailed operating rules, but they cannot eliminate incomplete contracts or foresee every condition under which assets should be redeployed. The decisive governance question is therefore who can decide outside the rulebook. If an unexpected market shock requires selling the company’s core intellectual property, terminating a strategically important relationship, or entering an unanticipated line of business, the agent’s ability to calculate an attractive outcome does not by itself confer the authority to make that decision.

A plausible employee-free corporation would need at least three institutionally distinct governance functions, even if many preparatory tasks are automated. The first is direction: authorized human directors or other lawful organs set purposes, risk tolerance, capital constraints, reserved matters, and the scope of delegation. The second is management and risk control: operational agents conduct business while separately empowered controls validate actions, monitor aggregate exposures, and challenge proposed decisions. The third is assurance: independent reviews determine whether reported outcomes, financial statements, controls, and compliance claims can be trusted. The OECD’s 2023 corporate-governance principles identify board responsibilities for strategic guidance, risk oversight, internal control, reporting integrity, and accountability; the IIA’s Three Lines Model distinguishes management activity and risk expertise from independent assurance.⁶⁴⁶⁵ Neither framework implies that every firm must reproduce a large corporate bureaucracy, but both clarify why a single agent hierarchy reporting its own successes would provide weak governance.

The separation must be substantive rather than cosmetic. A risk-review agent using the same corrupted knowledge base, action logs, model configuration, and business objective as the agent it evaluates may repeat the same error while giving the appearance of independence. Stronger assurance could involve independently sourced records, separately administered validation systems, protected control infrastructure, tamper-evident evidence, external audits, and human review of matters that cannot be discharged by software or are reserved by law; different model families can contribute only where their relevant failure modes are sufficiently independent. The governing body must be able to obtain information without depending exclusively on the agents whose conduct is under scrutiny. A nominal human director who cannot understand, contest, suspend, or replace the operational system would satisfy neither the substantive aspiration of oversight nor necessarily the duties imposed by applicable law.

Governance should therefore be modeled as a set of reserved, delegated, and prohibited decisions. Routine low-exposure acts may be delegated under clear controls; material capital commitments, new business purposes, significant related-party transactions, unusual financing, and selected legal or safety decisions may require explicit authorization. Some actions remain forbidden irrespective of expected profitability. Others depend on conditions such as verified evidence, independent approval, the absence of conflicts, or demonstrated reversibility. This is not merely role-based access control: permissions must be evaluated in the context of an ongoing business process and against cumulative exposure. Ten individually authorized purchases can exceed a daily liquidity limit; multiple permissible supplier substitutions can collectively undermine a safety or quality obligation. A mature governance system must reason over portfolios and sequences of actions, not just isolated API calls.

The regulatory perimeter is similarly conditional rather than universal. In the European Union, the AI Act establishes human-oversight and deployer obligations for systems classified as high-risk within its scope; it does not make every enterprise agent a high-risk system or require approval of every automated transaction.⁶⁶ Article 22 GDPR restricts certain decisions based solely on automated processing that produce legal or similarly significant effects concerning individuals, subject to specified exceptions and safeguards.⁶⁷ Sectoral rules, licensing, financial reporting, cybersecurity duties, consumer protection, and professional reservations may impose additional requirements depending on the activity. A firm can outsource execution, but the allocation of legal duties under applicable law does not necessarily follow the location of the software or the absence of an employment contract.

The final governance test is the power to stop being a firm. Continued operation is not always the lawful or economically rational objective: insolvency risk, an exhausted capital mandate, the loss of a critical license, or unacceptable harm may require suspension, restructuring, sale, or winding-up. An autonomous operating system optimized for continuity could otherwise resist the very intervention that corporate governance requires. Its constitutional model should therefore reserve the authority to terminate activities, revoke credentials, preserve records, protect creditors and other stakeholders, and initiate legally prescribed proceedings. Agents may detect distress, model alternatives, assemble evidence, and execute approved steps, but they cannot treat their own persistence as a governing end. The ability to be suspended and ultimately dissolved is not an incidental safety mechanism; it is part of what makes an enterprise accountable to an institutional order beyond itself.

A processual stress test: when autonomy meets an exception

Suppose the agent-operated firm sells a subscription service whose availability depends on an external provider. Monitoring identifies an outage while a provider message proposes an emergency migration to an unfamiliar infrastructure vendor. The Company Brain records the observed loss of service, the provider’s assertions, the affected contracts, cash exposure, alternative vendors, and uncertainty about restoration time. An incident-management agent activates a durable workflow; supplier agents produce alternatives; financial agents evaluate available liquidity; and customer-service agents prepare communications. This chain contains reasoning, coordination, and execution, but it is not yet evidence of a viable institution.

The crucial decisions concern authority and conflicting objectives. If the approved migration option remains within tested configurations, spending limits, and contractual terms, an authorized action gateway may commit the order, record the transaction, and require subsequent verification of service restoration. If the cheaper option transfers regulated customer data to an unsuitable jurisdiction or needs an unapproved security exception, the existence of a technically workable plan should not authorize it. If every available response breaches a hard control or imposes a material loss, the system needs a clearly defined route to an authorized decision-maker rather than an instruction to do the best thing. In the employee-free legal model, that decision-maker might be a natural-person director or lawfully empowered external professional, not an employee. In the purely theoretical model, leaving the matter to agents creates an explicit and unresolved problem of legitimacy, not an invisible assumption that the software somehow inherits corporate sovereignty.

The same event tests organizational learning and control over change. The incident creates new evidence: which supplier claims proved accurate, which decisions were effective, which controls prevented harm, and which responses were unnecessarily slow. Agents may propose different contracting standards, reserve capacity, or a revised exception policy. The Company Brain should preserve why those changes are recommended and which episodes support them; a change-control process should test their effects and distinguish ordinary model updates from modifications of delegated authority. If the agents that suffered the failure can also redefine the metric by which success is assessed, suppress adverse records, and approve their own wider powers, the firm may exhibit rapid adaptation while losing accountability. The substantive test is whether it can correct itself without controlling away the evidence of its own mistakes.

Reliability, adaptability, and external dependence

At the technical level, the weakest inference is to extrapolate from high individual-task accuracy to reliable corporate continuity. If a workflow contains n sequential steps, each with independent and identical success probability p, and every step must succeed without repair, then

P(\text{all steps succeed}) = p^n. \tag{1}

With p=0.99 and n=100, the result is approximately 0.366. This is a sensitivity illustration, not a measured probability for any enterprise or agent. Independence, identical difficulty, and absence of recovery are unrealistic for most business processes. In practice, failures can be correlated through shared models and data, but also contained by checkpoints, deterministic execution, redundant evidence, transactional reconciliation, and bounded escalation. A realistic reliability claim must therefore concern the entire human–machine–institutional process, including the costs and effects of recovery, not an isolated language-model success rate.

Contemporary empirical studies justify this restraint. TheAgentCompany introduced tasks in a simulated software-firm environment and originally reported 24% autonomous task completion for its strongest evaluated baseline, not 24% of a whole firm’s work sustained over time.⁶⁸ EnterpriseBench’s 2026 interactive business environments likewise found that evaluated methods did not yet demonstrate stable, comprehensive reliability across enterprise decision tasks.⁶⁹ The 2026 READY framework addresses a more relevant deployment question: what reliability can an agent achieve on a specified workflow, with which oversight policy and at what operating cost? Its evidence remains workflow-bounded rather than a demonstration of organizational self-sufficiency.⁷⁰ These studies measure distinct constructs and must not be compressed into a single estimate of the proportion of a company that agents can run.

The economic feasibility boundary follows from the same reasoning. Digitally delivered products with testable outputs, modular dependencies, reversible transactions, modest capital exposure, and established legal arrangements are more plausible candidates for highly agentic operations than safety-critical infrastructure, regulated fiduciary functions, or activities requiring continuous physical intervention. But apparent independence can also be an accounting illusion: an employee-free firm may import significant human work through contractors, model providers, banks, cloud services, auditors, and legal representatives. Hart and Moore’s emphasis on ownership and contractual boundaries is relevant here because substituting suppliers for employees changes rights of control and exposure to hold-up, not merely personnel counts.⁷¹ A meaningful comparison must measure total coordination costs, external dependencies, risk concentration, and the distribution of human labor across the firm’s production network.

A further structural vulnerability is institutional monoculture. In a human organization, conflicting expertise, departments, and informal channels can cause delay and inconsistency, but they can also preserve independent warnings. An agent-only organization coordinated through one Company Brain and one family of models might reduce friction while eliminating useful diversity of interpretation. The capability to act quickly on a shared understanding becomes dangerous when that understanding is wrong. The architectural answer is not to reproduce every inefficient hierarchy of the past, but to make dissent, alternative evidence, adversarial review, limits on common-mode failure, and independent controls part of organizational design rather than optional model prompts.

The radical possibility is therefore not a business with zero human interactions, nor a legal person that has somehow ceased to require lawful governance. It is a firm whose routine operating organization is largely implemented as a persistent, inspectable, and partially self-adapting network of computational actors. Its products, processes, and internal divisions of labor could change without continuous employee intervention, but its legal authority, external dependencies, and rights of institutional redesign would remain subject to explicit constraints. Whether such an enterprise constitutes a new type of firm depends on more than reducing labor costs: it depends on demonstrating durable economic production, reliable cross-process coordination, independent accountability, lawful external commitments, and adaptation to circumstances not anticipated when its mandates were written. The employee-free firm is thus the limiting experiment for distinguishing an organization that automates work from one that can sustain and govern its own operational existence. The next limiting case retains human employees and asks whether the very same cognitive infrastructure can transform not the absence of human work, but the capabilities and distribution of intelligence within it.

Extreme two: the human company with universal cognitive access

The second limiting case preserves the human organization but transforms the informational and cognitive conditions under which its members operate. Employees, managers, specialists, executives, and statutory corporate organs remain present; their expertise, responsibilities, relationships, and legitimate authority do not disappear. What changes is the extent to which organizational knowledge can be accessed, interpreted, connected, and applied independently of the hierarchical position, personal networks, or departmental systems through which it was historically mediated. A product engineer can investigate how a design decision affects procurement commitments, production capacity, contractual obligations, and customer outcomes; a salesperson can examine whether a promised delivery date is supported by actual operational conditions; a finance analyst can reconstruct the processes and decisions responsible for an economic deviation without waiting for several departments to reconcile their separate explanations. The Company Brain makes these forms of inquiry possible through a common but governed cognitive infrastructure, while autonomous agents participate in the same organizational processes as bounded operational actors.

This scenario is not simply the opposite of a company without employees. In the first extreme, the central problem is whether an artificial operating organization can sustain economic activity under legitimate human governance. In the second, the problem is whether the integration of human and computational cognition can create organizational capabilities that neither conventional firms nor autonomous-agent systems could realize independently. The stronger proposition is that universal cognitive access may alter the fundamental relationship between specialization, hierarchy, coordination, and collective intelligence, even if no employee is replaced. Such a transformation would involve more than distributing an intelligent assistant to every worker. It would require reexamining why organizational structures exist, which functions they perform, and how those functions change when relevant knowledge can be reconstructed and acted upon across traditional boundaries.

From hierarchical information distribution to cognitively connected organizations

Modern organizations have historically addressed cognitive limitations through division of labor, specialization, standardized procedures, hierarchical escalation, and coordinating mechanisms. Galbraith’s information-processing theory explains organizational structure partly as a response to uncertainty and to the volume of information that tasks require participants to exchange.⁷² Hierarchy reduces the need for every participant to process every issue; specialized departments preserve expertise; managerial roles integrate decisions whose consequences cross organizational boundaries. These arrangements are not merely historical accidents or instruments of managerial control. They can be economically rational responses to finite attention, costly communication, limited expertise, and the difficulty of maintaining a sufficiently coherent view of organizational activity.

The problem is that specialization simultaneously creates the need for coordination. Malone and Crowston define coordination in terms of managing dependencies among activities, including shared resources, producer–consumer relationships, and task dependencies.⁷³ A manufacturing company illustrates the consequence: a commercial promise depends on engineering specifications, inventory availability, procurement commitments, production scheduling, quality assurance, and logistics. Each specialized function can be locally efficient while the enterprise fails to deliver the promised outcome. Traditional organizations address these dependencies through meetings, management hierarchies, liaison roles, project teams, planning procedures, and increasingly integrated information systems. Yet the cost of reconstructing the relevant context remains substantial, especially when the information is fragmented or its interpretation depends on specialized knowledge.

The Company Brain offers a different response to this problem. Instead of requiring information to travel repeatedly through people and organizational layers until it reaches someone who can assemble the necessary context, relevant representations may become computationally available at the point where the decision originates. This does not mean every employee should access every record, nor that a central model should determine what everybody ought to think. It means that an authorized participant could interrogate a logically connected representation of the enterprise, recover the evidence supporting a proposition, understand its operational consequences, identify relevant expertise, and initiate a structured decision process without reconstructing the entire informational path through personal contacts and departmental intermediaries.

The resulting transformation is best understood as a partial decoupling of access to organizational knowledge from position in the organizational hierarchy. Historically, hierarchy has often determined not only who may decide but also who can obtain the information needed to formulate a decision. A cognitively integrated enterprise could reduce this second dependency while preserving the first wherever it remains necessary. The distinction is fundamental: a maintenance engineer may acquire an accurate cross-site diagnosis without receiving authority to approve a capital investment; a salesperson may identify a profitable pricing opportunity without possessing the right to alter contractual limits. The organization can distribute the ability to understand a situation more widely than the authority to commit resources or accept risk.

A consequence follows that is less obvious than the prospect of flatter organizations. The Company Brain could reduce certain informational justifications for managerial intermediation without eliminating the need for management itself. Roles devoted predominantly to collecting updates, reconciling reports, forwarding information, or translating between departmental vocabularies may diminish in relative importance. Roles concerned with resolving conflicting objectives, allocating scarce resources, exercising judgment under uncertainty, legitimizing contested decisions, and accepting accountability may remain necessary or acquire greater significance. This is a hypothesis about changing organizational functions, not a prediction that hierarchy will uniformly disappear.

The same distinction applies to specialization. Universal cognitive access need not make everyone a generalist. An engineer who can understand the financial consequences of a technical decision does not thereby possess the expertise of a financial controller; a controller who can interrogate production constraints does not become a production engineer. Rather, specialized knowledge becomes more composable: individuals can bring expertise to bear on problems spanning several domains without having to acquire the complete professional competence of every other participant. Faraj and Sproull’s research on software development teams found that the coordination of expertise—knowing where expertise exists, where it is needed, and how to mobilize it—was associated with team performance beyond the mere presence of specialized knowledge.⁷⁴ A Company Brain could strengthen this capability by functioning not only as a repository of knowledge, but also as an organizational mechanism for locating, contextualizing, and connecting expertise.

The deeper transformation would therefore not necessarily be the abolition of departments, but the emergence of a more densely interconnected organizational cognition superimposed upon continuing specialization. Functional boundaries would remain meaningful for competence, accountability, and resource stewardship, while becoming less restrictive as boundaries of inquiry and collaborative problem-solving.

Whether this produces an improvement depends on the quality of the underlying representations, the cost of interpretation, the reliability of agents, and the organization’s ability to manage conflicts that greater visibility may expose rather than resolve.

Figure 3 uses the same visual grammar to show what changes when human specialists remain inside the operating organization. Engineering, commercial, and financial participants can interrogate the Company Brain directly under governed access, while agents contribute to the same durable processes; neither shared knowledge nor participation in a workflow equalizes the participants’ authority. The diagram preserves separate corporate governance, authorization, and outcome-verification mechanisms, making the structural contrast with Figure 2 explicit rather than presenting the human-centered firm as a less automated stage of the same hierarchy.

%%{init: {"theme": "neo", "look": "handDrawn", "layout": "elk"}}%%
flowchart TB
    subgraph GOV["Ownership and lawful human governance"]
        SH[Shareholders]
        DIR[Directors and authorized corporate organs]
        M[(Approved mandates and reserved matters)]
        SH -.->|ownership rights| DIR
        DIR -->|establishes delegation| M
    end

    subgraph CORE["Human–agent operating organization"]
        B[(Federated Company Brain)]
        H1[Engineering and operations specialists]
        H2[Commercial and procurement specialists]
        H3[Finance and delegated managers]
        A[Agent coworkers]
        W(Durable process coordination)
        P{Independent authority gate}
        S[Enterprise systems of record]
        V(Outcome verification)

        B -->|governed context| H1
        B -->|governed context| H2
        B -->|governed context| H3
        B -->|governed context| A
        H1 -->|evidence and proposals| W
        H2 -->|evidence and proposals| W
        H3 -->|decisions and proposals| W
        A -->|agent proposals| W
        W -->|typed action proposals| P
        P -->|delegated exceptions| H3
        H3 -->|authorized review| P
        P -->|authorized transactions| S
        S -->|authoritative results| V
        V -->|qualified evidence| B
    end

    EXT[External customers, suppliers and service providers]
    M -.->|constrains| P
    P -->|reserved decisions| DIR
    DIR -->|authorized decisions| P
    S -->|transactions and requests| EXT
    EXT -->|responses and services| S

Figure 3: The human–agent cognitive enterprise: specialized employees, delegated managers, and agent coworkers share governed organizational knowledge, while decisions and transactions remain subject to differentiated authority and independently verified outcomes.

The reorganization of work and the emergence of mixed human–agent processes

Once employees and agents operate through the same Company Brain, the process rather than the individual job becomes an increasingly important unit of organizational design. Traditional job descriptions allocate bundles of responsibilities to roles; business-process models define activities, dependencies, controls, and expected outcomes. The two representations have always coexisted, but persistent agents make their relationship more explicit because work can be reassigned among human and artificial actors without necessarily changing the business objective or the institutional responsibility for its achievement.

Consider an order-to-cash process. A commercial employee identifies an opportunity and negotiates the commercial relationship; an agent reconstructs customer history, contractual constraints, and available-to-promise quantities; a planning service evaluates capacity; another agent prepares a proposed delivery commitment; the authorized employee or manager approves any exceptional terms; and enterprise systems execute the transaction, monitor fulfillment, and reconcile the financial result. Some activities may be automated entirely, some may require human approval, and others may be performed through iterative collaboration. The process remains a coherent institutional object even as its executing actors change. Its effectiveness depends on the integrity of the handoffs, the meaning of shared data, and the conditions under which proposed actions become valid commitments.

The organizational novelty is that processes may increasingly be designed around complementary cognitive and operational capabilities rather than around the historical boundaries of occupations. A software engineer might delegate reproducible testing and defect localization while retaining architectural review; a procurement specialist might use agents to assemble evidence and monitor supplier obligations while managing negotiations involving ambiguous commercial interests; a maintenance professional might rely on an agent to correlate alarms, previous incidents, technical documentation, and spare-part availability while retaining responsibility for judgments involving physical safety or uncertain plant conditions. These divisions of labor are not fixed. They should depend on the observed reliability of the agent, the nature of the task, its consequences, and the institutional authority required to act.

The crucial design question is therefore not whether a process is performed by a person or an agent, but which activities require which combinations of evidence, expertise, discretion, authority, and verification. A deterministic decision rule may be preferable to generative reasoning when a condition is explicit and stable; an optimization model may allocate resources more reliably than a language model; a human professional may provide essential contextual judgment; and an AI agent may coordinate the interaction among them. Treating every organizational activity as a problem to be solved by a language-model agent would be as conceptually inadequate as assuming that every activity must remain attached to a fixed human position.

The executable institutional model introduced in the previous section supplies the necessary structure. Human and artificial participants would operate under explicit roles, mandates, decision rights, process obligations, and escalation procedures. Their capabilities would differ, but their interactions could be represented through common organizational objects: tasks, commitments, evidence, approvals, exceptions, and outcomes. The Company Brain supplies relevant context and institutional memory; workflow systems preserve process state; authorization mechanisms determine permitted actions; and independent systems of record confirm what actually happened. Humans remain part of the process not as generic emergency substitutes for unreliable software, but as participants assigned functions justified by their expertise, legitimate authority, or the particular consequences of a decision.

An important implication is that organizational processes could become more dynamically configurable without becoming institutionally unstructured. A cross-functional team might be assembled around a detected operational problem, with temporary agent roles performing investigation, modeling, scheduling, documentation, and monitoring. Once the problem is resolved, the temporary configuration may dissolve while its evidence, decisions, and lessons remain part of the organizational memory. This resembles project-based and adaptive organizational forms, but with a potentially lower cost of assembling shared context and coordinating specialized contributions. The limits remain substantial: processes cannot be reconfigured safely merely because an agent can generate a new workflow. Resource constraints, collective agreements, safety responsibilities, segregation of duties, and the right to alter organizational rules must still be respected.

What happens to management and organizational structure?

The implications for management may be more significant than the automation of individual administrative tasks. A large part of managerial activity has traditionally involved acquiring information, distributing assignments, monitoring execution, reconciling different accounts of reality, and resolving exceptions. Where these activities rely on repetitive information processing, they are natural candidates for computational support or partial automation. Where they require choosing between competing interests, determining legitimate priorities, managing interpersonal conflict, accepting responsibility, or exercising judgment under ambiguous conditions, the possibility of automation is more contested.

Recent empirical research provides evidence that this distinction is already economically relevant, although it does not establish a general theory of the agent-native firm. In a study published in Management Science in July 2026, Devigili, Yilmaz, Gaba, and Greve examined changes in hiring demand across 1,820 publicly listed US firms around the introduction of ChatGPT. Their findings indicate selective reductions in demand for skills associated with task division, information provision, monitoring, and certain operational exceptions, while task allocation and conflict resolution exhibited greater stability.⁷⁵ These results concern changes in firms’ demand for skills as expressed in job postings, not direct observation of fully reorganized enterprises. Nevertheless, they support the hypothesis that AI changes the economics of organizing work, not only the productivity of performing isolated tasks.

Theoretically, a Company Brain might allow a manager to supervise a larger organizational scope because relevant information is continuously accessible, agents monitor routine processes, and exceptions can be classified before escalation. It might also enable more decentralized structures because operational participants can obtain context without requesting it from higher hierarchical levels. Neither outcome is inevitable. If decision rights remain centralized, improved information infrastructure can strengthen central control rather than reduce hierarchy. If agents generate more exceptions, recommendations, and proposed changes than humans can evaluate, managerial workload may increase rather than decrease. The relationship between cognitive integration and managerial span of control is mediated by the rate, complexity, and consequences of decisions requiring human attention.

An additional transformation concerns the nature of managerial authority. In a traditional hierarchy, a manager may derive influence from positional rights, control over information, expertise, relationships, and the ability to coordinate access to resources. Universal cognitive access weakens some informational asymmetries, but it may increase the importance of whoever controls the rules governing the Company Brain. The selection of authoritative data sources, the classification of exceptions, the design of decision thresholds, and the criteria used to evaluate performance become consequential organizational choices. A new form of managerial power may therefore emerge at the boundary between business governance, enterprise architecture, and computational policy.

Management could become increasingly concerned with governing the conditions under which distributed actors reason and act, rather than directing every action through a hierarchical chain. This would involve defining objectives, resource constraints, delegated authority, acceptable risks, evidence requirements, escalation boundaries, and mechanisms for institutional learning. Managers would still make substantive decisions, but a greater share of their distinctive contribution might lie in ensuring that the organization can make many other decisions coherently without their immediate intervention.

The possibility is not purely decentralizing. Shared cognitive infrastructure may also enable unusually detailed central supervision, automated assessment, and standardized control of otherwise autonomous teams. The same technology that allows a technician to understand corporate strategy could allow senior management to inspect an unprecedented level of operational detail. Whether this produces a more participatory institution or a more comprehensive managerial apparatus depends on governance arrangements, not on the availability of AI alone.

Collective intelligence: beyond the sum of augmented individuals

A company in which every employee has an effective AI assistant is not necessarily a collectively intelligent organization. Individual productivity improvements may coexist with defective coordination, contradictory local decisions, duplicated work, or the inability to preserve and reuse organizational learning. The distinctive hypothesis behind the Company Brain is stronger: common cognitive infrastructure could enable the organization to integrate contributions, evidence, and experience across time and organizational boundaries, creating capabilities that cannot be inferred merely by summing the performance gains of individual workers.

This distinction has roots in distributed cognition, transactive memory, and knowledge-based theories of the firm. The organizational contribution of a specialist does not consist exclusively of what the specialist knows, but also of how that knowledge becomes available when other actors need it.⁷⁶⁷⁷ A shared cognitive infrastructure may reduce the cost of locating expertise, reconstructing prior decisions, discovering dependencies, and identifying relevant precedents. More importantly, it may preserve connections among contributions that individual employees would otherwise struggle to maintain. A technical decision can be connected to its commercial consequences; a customer complaint to an engineering modification; an operational failure to a policy change; and the outcome of that change to subsequent decisions.

This possibility should be distinguished from a simplistic assumption that more information always improves collective judgment. Organizational cognition depends on interpretation, attention, incentives, communication, and the willingness of participants to challenge prevailing assumptions. A Company Brain could make a false interpretation available everywhere just as efficiently as a correct one. If a single model or semantic layer becomes the dominant source of organizational explanations, the resulting system may reduce the diversity of interpretations needed to recognize unexpected situations. The organization could become more coordinated while becoming less capable of detecting shared mistakes.

James March’s analysis of exploration and exploitation is especially relevant. Organizational learning requires the exploitation of established knowledge and the exploration of alternatives, but successful exploitation can progressively suppress the variation from which future adaptation depends.⁷⁸ A Company Brain optimized exclusively for retrieving established procedures, reinforcing historically successful decisions, and aligning employees with common recommendations could intensify this tendency. Its greatest contribution to organizational intelligence may therefore depend not only on improving access to accumulated knowledge, but also on making the limits of that knowledge visible and preserving opportunities for independent inquiry.

This suggests a principle of epistemic pluralism within institutional coherence. The organization needs shared identities, authoritative records, and enforceable rules, but it need not require every participant to adopt the same interpretation of uncertain evidence. Domain experts should be able to challenge causal assumptions, document competing hypotheses, introduce local observations, and request independent evaluation. The Company Brain should preserve both the accepted institutional position and the reasons for unresolved disagreement. Authority to execute a decision does not imply epistemic infallibility; similarly, an employee who lacks decision authority may possess evidence that invalidates the premises of an authorized decision.

Amy Edmondson’s research on psychological safety and team learning reinforces the importance of the social conditions under which knowledge is challenged. Her study of work teams found an association between psychological safety and learning behavior, indicating that access to information alone is insufficient if people perceive interpersonal risk in raising concerns or exposing errors.⁷⁹ A computational infrastructure can make dissent technically possible without making it institutionally acceptable. The cognitive enterprise must therefore protect the organizational processes through which inconvenient evidence, minority interpretations, and failures become visible.

A further question concerns whether the introduction of agents can improve collective intelligence without weakening human knowledge networks. An employee who routinely asks an agent rather than a colleague may gain speed but lose opportunities to exchange tacit knowledge, develop relationships, or discover contextual information that was never documented. Conversely, a well-integrated agent may identify the relevant expert, reconstruct a problem sufficiently to make consultation productive, and enable interactions that previously would not have occurred. These effects cannot be determined from model capability alone.

A 2026 working paper by Büchsenschuss, Koch-Bayram, Biemann, and Puranam offers unusually relevant evidence. In a randomized field experiment involving 316 employees across 42 teams in a European technology-services company, access to a generative-AI assistant customized with organizational knowledge was associated with increased connectivity in collaboration and knowledge-sharing networks over three months. Specialists gained more in knowledge-network centrality, while generalists recorded larger gains in the measured number of projects handled.⁸⁰ The study’s setting, duration, and partly self-reported measures limit generalization, but its contribution is conceptually important: AI adoption can alter relationships among employees rather than simply replacing interpersonal interaction.

The findings complicate the familiar prediction that AI necessarily isolates workers or makes human collaboration redundant. The actual effect depends on how the technology changes the relative costs of independent problem-solving and interpersonal coordination. An effective Company Brain may make human expertise more valuable precisely because it allows that expertise to be discovered and applied more widely. Alternatively, poorly designed systems may encourage employees to substitute standardized AI answers for the difficult conversations through which organizations identify ambiguous problems and revise their assumptions.

Empirical evidence: augmentation is not the same as organizational transformation

The evidence for AI-assisted productivity is increasingly substantial, but it must not be interpreted as proof of the broader Company Brain hypothesis. Brynjolfsson, Li, and Raymond studied the staggered deployment of a generative-AI assistant among 5,172 customer-support workers and reported a 15% average increase in issues resolved per hour, with larger gains among less experienced or lower-performing workers.⁸¹ This demonstrates a meaningful productivity effect in the studied environment, including changes consistent with knowledge transfer to less experienced employees. It does not establish that an integrated enterprise-wide cognitive infrastructure would produce similar improvements across occupations or business processes.

A more directly organizational experiment was published in Organization Science in June 2026. Dell’Acqua and colleagues studied 791 professionals at Procter & Gamble, randomly assigning participants to work individually or in two-person teams, with or without generative AI, on product innovation tasks. Individuals using AI achieved performance comparable to conventional two-person teams within that experimental setting, while AI also helped participants generate proposals spanning technical and commercial perspectives.⁸² The results suggest that AI can reproduce some benefits associated with cross-functional collaboration, but the experimental tasks did not involve the persistent execution of enterprise processes or shared institutional memory. They provide evidence about the possibilities of cognitive augmentation and expertise integration, not the existence of a Company Brain.

Countervailing results reinforce the importance of context. METR’s randomized study of experienced open-source developers working on familiar repositories found that access to early-2025 AI coding tools increased task completion time by approximately 19%.⁸³ In February 2026, the researchers reported that subsequent experiments with newer tools suffered from selection and measurement problems, preventing a reliable estimate of their updated productivity effects.⁸⁴ This is not evidence that AI generally reduces software productivity; it demonstrates that even technically capable systems can impose integration, verification, and correction costs that outweigh their immediate assistance in particular environments.

At a broader economic level, a January 2026 European Investment Bank working paper using matched firm-level data reported an estimated 4% increase in labor productivity associated with AI adoption among European firms under its identification strategy, with the estimated gains linked to capital deepening rather than immediate employment reductions.⁸⁵ This evidence is relevant to the economics of complementary investment, but it concerns AI adoption broadly, not firms operating through a shared Company Brain or persistent agent coworkers.

Taken together, these studies support three restricted conclusions. AI can improve certain individual and collaborative outcomes; the magnitude and direction of its effects depend on the task, expertise, system design, and organizational context; and technology adoption can alter patterns of collaboration and skill demand. They do not establish whether the integration of shared organizational cognition and persistent agents produces an additional enterprise-level effect beyond deploying these technologies independently.

That distinction is essential for the article’s central hypothesis. The proper empirical unit for assessing a cognitive enterprise is not merely the individual task, but the end-to-end organizational process and the institution’s capacity to improve that process over time. A faster report is not necessarily a better decision; a better decision is not necessarily an authorized one; an authorized action is not necessarily a successful outcome; and a successful outcome does not prove that the organization has learned how to reproduce it.

A concrete operating scenario: the cognitively integrated manufacturing firm

A supply-chain disruption illustrates why the human–agent model should be evaluated at the level of a complete organizational process rather than as individual access to assistants. A supplier notice may affect purchasing, inventory, production, customer commitments, product qualifications, and liquidity at once, while the relevant evidence remains distributed among authoritative systems and the expertise of specialists. A Company Brain could correlate the observations, retain their provenance and temporal validity, and make competing consequences available to procurement, engineering, commercial management, and finance; agents could prepare scenarios and advance approved workflows, but neither shared access nor a plausible recommendation would confer authority to modify contracts or execute consequential purchases.

The distinctive contribution of the human participants is not simply an approval at the end of the workflow. An engineer may identify a qualification constraint absent from the data model, a procurement specialist may challenge a supplier assertion, and finance may reject an economically attractive but imprudent exposure; their observations can revise the common evidentiary representation without dissolving the institutionally defined decision rights. The later section A supply-chain disruption, executed two ways examines this case in detail under both operating extremes, including its security boundaries, transaction enforcement, reconciliation, and the conditions for legitimate institutional learning.

Governance: from universal access to distributed but accountable authority

The architecture raises a governance problem that is distinct from that of the employee-free company. In the first extreme, the main question concerns how owners and statutory organs can govern a largely autonomous operating system. In the second, corporate governance must additionally account for employees whose knowledge, professional discretion, working conditions, and relationships are mediated by the same cognitive infrastructure used by autonomous agents.

The expression universal cognitive access should therefore be interpreted as universal eligibility to use the organization’s governed cognitive capabilities, not unrestricted access to all information or unrestricted authority to act. Employees should be able to discover and interrogate relevant organizational knowledge according to their legitimate purposes, roles, and applicable restrictions. The Company Brain must enforce confidentiality, personal-data protection, commercial secrecy, contractual limitations, and operational security boundaries. In highly regulated or safety-critical environments, a common logical representation may need to be implemented through federated systems with different trust zones, access policies, and restrictions on information movement.

A coherent governance model needs to distinguish at least four institutional functions. Epistemic governance concerns what the organization accepts as evidence, which sources are authoritative, how contradictory claims are handled, and who can correct an institutional representation. Operational governance concerns process responsibility, task allocation, execution limits, and the coordination of people and agents. Corporate governance determines objectives, delegated powers, risk acceptance, reserved decisions, and accountability to shareholders and other legally relevant stakeholders. Participatory governance concerns the ability of employees and their representatives to understand, challenge, and influence decisions about the systems that shape their work. These functions are related, but their rights and responsibilities should not be conflated.

The distinction between the right to know, the right to challenge, and the right to act is especially important. Access to a conclusion does not establish its validity. The ability to demonstrate that a conclusion is incorrect does not automatically grant authority to reverse the associated business decision. Conversely, a formal decision-maker may be authorized to act without being the most knowledgeable participant in the matter under consideration. A well-governed cognitive enterprise must establish pathways through which evidence can challenge authority without dissolving the legitimate allocation of decision rights.

This structure resembles a form of polycentric organizational cognition: expertise and interpretation remain distributed across multiple domains, while overlapping coordination and governance mechanisms allow collective action. Ostrom’s research on polycentric governance demonstrates why complex systems need not depend exclusively on a single centralized authority, although her institutional findings cannot be mechanically transferred to corporate AI architectures.⁸⁶ The relevant design implication is more limited and precise: a Company Brain can support shared context and coordinated decisions without requiring a single actor, model, or department to monopolize every valid interpretation.

Governance must also address changes to the executable institutional model itself. Human employees and agents may discover dysfunctional processes, obsolete decision rules, contradictory responsibilities, or opportunities for more effective coordination. They can propose revisions, support them with evidence, and simulate possible effects. But the power to adopt a new rule depends on the organization’s authorized change procedures. An agent that identifies an inefficient approval threshold should not unilaterally increase its own spending authority; a manager should not be able to alter an employee’s access rights or professional obligations merely by instructing a generative model to reinterpret policy.

The governing institutional model must consequently be versioned, reviewable, and capable of distinguishing proposed, approved, deployed, and superseded arrangements. This is a direct continuation of the increasing formalization established in the preceding sections. As more of the organization becomes machine-interpretable, organizational governance itself becomes an object of explicit modeling. The company does not become a self-legitimating computational institution; rather, its rules, rights, and changes can become more transparent, testable, and enforceable.

The risks of surveillance, deskilling, and cognitive concentration

The same architecture that makes organizational knowledge more accessible can make employees more observable. Process histories, task execution traces, agent interactions, document access, communication patterns, and performance measures may reveal detailed information about the behavior of individual workers. Such information can be useful for detecting errors, understanding operational bottlenecks, and improving organizational processes; it can also be repurposed for intrusive monitoring, individual ranking, behavioral manipulation, or automated disciplinary decisions.

Kellogg, Valentine, and Christin’s analysis of algorithmic management identifies mechanisms through which software can direct, evaluate, and discipline workers, demonstrating that computational coordination is inseparable from questions of organizational power.⁸⁷ A Company Brain creates additional possibilities for such control because it connects information that would otherwise remain distributed among systems and social contexts. A system designed to explain why a production process failed could also be used to infer which employee communicated with whom, how quickly a decision was taken, or whether a worker deviated from an expected behavioral pattern. Technical capability does not establish that such uses are legitimate.

For European companies, these risks intersect with binding legal obligations. The GDPR imposes principles of lawfulness, fairness, transparency, purpose limitation, data minimization, and other safeguards; Article 22 addresses certain decisions based solely on automated processing that produce legal or similarly significant effects.⁸⁸ The EU AI Act establishes additional requirements for systems falling within its scope, including categories of high-risk employment and worker-management applications, subject to the regulation’s classification rules, exceptions, transitional arrangements, and applicable dates.⁸⁹ Existing employment law, information and consultation obligations, and national worker-protection rules also remain relevant. These instruments do not prohibit all organizational analytics or AI-supported management, but they preclude treating enterprise-wide cognitive visibility as an unlimited managerial entitlement.

An institutional distinction is therefore needed between process observability and personal surveillance. A company may require reliable evidence that a procurement process respected its controls without granting every supervisor unrestricted access to the conversations, working notes, or personal behavior of each participant. Different purposes require different data granularity, retention periods, access rights, and review procedures. Auditability should be designed around accountable processes and consequential actions, not assumed to justify continuous observation of every employee.

A second risk is cognitive dependency. If employees increasingly rely on agents to identify problems, formulate alternatives, explain decisions, and evaluate outcomes, they may become less practiced in activities that previously sustained their professional competence. This risk does not imply that retaining every manual task is beneficial; automation has always changed which skills are economically valuable. The concern is whether organizations preserve the capacity to detect incorrect recommendations, respond to unfamiliar situations, and operate when the computational infrastructure is unavailable or compromised.

Skill erosion may be particularly consequential when employees are nominally retained as approvers but are no longer sufficiently involved in the underlying work to exercise meaningful judgment. A human approval step can become ceremonial if the reviewer lacks time, expertise, access to counterevidence, or practical authority to refuse. Governance must therefore address the actual competence and decision conditions of human participants, rather than counting the presence of a person in a workflow as proof of effective oversight.

This concern introduces a tension between short-term efficiency and long-term organizational learning. Delegating routine analysis to agents may free specialists for more complex work, but it can also remove the experiences through which junior employees acquire expertise. Automating operational exceptions may improve response time while reducing opportunities for professionals to develop practical judgment. A human-centered cognitive enterprise may consequently require deliberately designed learning pathways, opportunities for independent problem-solving, and procedures that keep critical expertise exercised rather than merely documented.

The third risk is cognitive concentration. Even where the Company Brain is physically federated, dependence on a limited set of reasoning models, semantic definitions, retrieval mechanisms, or policy configurations can generate common-mode errors. Employees may receive consistent but mistaken explanations; managers may mistake computational agreement for independently corroborated evidence; and agents may reinforce decisions derived from the same flawed organizational assumptions. Such risks are not eliminated by giving every employee access to the same system. Indeed, universal adoption can increase their consequences.

The architectural response is to preserve meaningful independence where it is required: access to original evidence, provenance and uncertainty, competing models, domain-specific interpretation, independent review, and mechanisms for correcting institutional records. This should be combined with the procedural right to raise objections and the organizational capacity to evaluate them without retaliation. A Company Brain that makes every employee more dependent on a single uncontestable interpretation would represent greater computational integration, but not necessarily greater organizational intelligence.

The emerging form: a cognitively integrated human institution

The second extreme reveals a more complex evolutionary possibility than the gradual replacement of employees by software. The firm could become an institution in which human intelligence is increasingly amplified by persistent organizational cognition, while computational agents extend the range of coordinated action available to human participants. Organizational knowledge would no longer be confined primarily to individuals, informal networks, departmental repositories, and application-specific rules. It would become a more continuous institutional resource, capable of supporting inquiry, coordination, action, and learning across successive teams and generations of employees.

Such an arrangement could substantially change the economics of specialization and organizational scale. A smaller team might coordinate processes previously requiring a larger administrative structure; a specialist might contribute effectively to decisions spanning unfamiliar domains; and employees might be able to exercise greater initiative because relevant context is available without repeated hierarchical mediation. At the same time, larger organizations might gain new advantages by integrating accumulated operational experience, specialized capabilities, and complex processes through a shared cognitive substrate. The direction of change is therefore not predetermined: cognitive integration could support flatter and more decentralized organizations, or enable more extensive coordination and control within large ones.

The decisive organizational design variable is not simply how much information becomes accessible, but how the enterprise distributes cognitive capability, execution authority, and the power to shape the institutional rules under which both operate. A company can make information universally discoverable while keeping decision-making highly centralized. It can delegate operational authority while preserving fragmented knowledge. It can give every employee an agent while failing to integrate their contributions into a persistent organizational memory. None of these arrangements, individually, constitutes the fully developed cognitive enterprise proposed here.

The strongest version of the human-centered model would combine several capabilities: employees can interrogate the relevant organizational context; their expertise and judgments can contribute to an evolving institutional memory; agents can coordinate and execute bounded activities; decisions can be reconstructed from evidence and authority; and organizational rules can be challenged and revised through legitimate procedures. Its distinguishing quality would not be unrestricted transparency or maximal automation, but the institutional integration of distributed understanding and accountable action.

This configuration also offers a potential advantage over the employee-free extreme. Human participants do not merely intervene when agents fail; they contribute ongoing interpretation, situated expertise, social negotiation, criticism, and legitimacy to the operating organization. Their presence can increase cognitive diversity, but it does not guarantee better outcomes. People can share biases, reinforce organizational routines, suppress dissent, and make consequential errors. The hypothesis that human–agent organizations outperform predominantly agent-operated firms must therefore be evaluated under specified conditions, not assumed from the intrinsic superiority of human judgment.

An appropriate empirical test would compare organizations or comparable process units operating with conventional enterprise systems, individually deployed AI assistants, coordinated agents without shared institutional memory, and a governed Company Brain supporting both human and agent participation. Evaluation should examine not only task time and immediate productivity, but also end-to-end process performance, cross-functional coordination, error propagation, knowledge retention, decision quality, worker autonomy, resilience, and the organization’s ability to correct mistaken assumptions. Without such comparisons, improvements attributed to organizational cognition may simply reflect faster retrieval, greater computational expenditure, or conventional process redesign.

The broader theoretical implication is that the relationship between people and the firm may change without the disappearance of either. Historically, organizations accumulated knowledge through people, routines, and records, while enterprise software progressively formalized parts of their coordination and execution. A Company Brain could extend that process by making knowledge, evidence, and decision context more persistent and accessible; agent coworkers could connect that cognitive infrastructure to operational action. The human-centered firm would then become neither a traditional hierarchy equipped with AI tools nor a computational enterprise retaining people as ceremonial supervisors, but a hybrid institution in which cognitive work, operational responsibilities, and decision rights are consciously redistributed.

The central possibility is not that every employee becomes universally knowledgeable, but that the organization becomes capable of mobilizing relevant knowledge wherever it is needed without requiring each participant to possess it individually. That possibility is different from the employee-free company, yet both depend on the same underlying architecture of persistent cognition, explicit institutional structure, and bounded agency. The question that follows is whether these two extremes represent distinct organizational species or configurations within a common design space—and which combinations of human participation, collective intelligence, autonomy, and accountability are actually viable.

A four-property theory of the cognitive enterprise

The two organizational extremes reveal something more fundamental than the possibility of replacing employees with artificial agents or augmenting them with a shared intelligence. They expose a limitation in the conceptual apparatus through which enterprises are ordinarily understood. Economic theories explain why firms exist, organizational theories explain how work and authority are distributed, information-processing theories explain how organizations cope with uncertainty, and computer science explains how representations and executable systems can coordinate increasingly complex activities. Yet the convergence of persistent organizational knowledge, computational agency, and machine-interpretable institutional structures creates a new analytical problem: what constitutes the intelligence of an organization when the capacity to know, decide, and act is no longer primarily attached to the individuals who formally occupy its positions?

The question cannot be answered by treating the firm as a large artificial intelligence, nor by assuming that an organization becomes autonomous as the proportion of automated tasks increases. A company may possess extensive computational knowledge without allowing software to make consequential decisions; it may execute most operations automatically while retaining little understanding of why those operations are appropriate; and it may delegate substantial authority to agents while remaining unable to explain, challenge, or reliably correct their actions. Conversely, a company that retains human employees may develop a form of integrated organizational cognition that substantially exceeds the capabilities of its individual members. These configurations are not successive stages of a single technological progression. They are different arrangements of knowledge, action, authority, and institutional responsibility.

The preceding analysis has established the historical ingredients of this transformation. Traditional organizational structures compensated for the limited knowledge and attention of their members. Enterprise software progressively externalized records, calculations, procedures, and transactions. The Company Brain proposes to make organizational knowledge, reasoning context, and institutional memory more persistent, interoperable, and inspectable. Agent coworkers connect this cognitive infrastructure to sustained operational action. The increasing formalization of organizational structures, processes, and decision rights makes it possible to specify the conditions under which these actors operate. Taken together, these developments point toward a firm whose cognitive and operational capabilities are increasingly embedded in its institutional architecture rather than exclusively in the people who participate in it.

The central theoretical proposition of this article is that the emerging cognitive enterprise must be understood through four analytically separable properties: organizational cognition, organizational agency, organizational autonomy, and organizational accountability. Their interaction determines what the enterprise can understand, what it can do, how independently it can operate, and how its decisions remain subject to legitimate authority. None of these properties, taken alone, is sufficient to characterize an intelligent, viable, or responsibly governed organization.

From organizational cybernetics to a computational theory of the firm

The idea that an enterprise might possess something resembling a brain predates contemporary artificial intelligence. Stafford Beer’s Brain of the Firm, published in 1972, developed an organizational-cybernetic account of control, coordination, adaptation, and viability inspired partly by the nervous system.⁹⁰ Beer did not propose a company-wide language model or a universal repository of corporate knowledge. His concern was how an organization could maintain coherent behavior while facing environmental complexity, distributing operational responsibilities, and preserving the capacity to adapt. His Viable System Model remains important because it treats organizational intelligence as a problem of distributed regulation and recursive coordination rather than as the possession of a central computational device.

A related foundation appears in W. Ross Ashby’s cybernetics. His law of requisite variety concerns the relationship between the diversity of disturbances a system encounters and the regulatory distinctions available to control their effects.⁹¹ Conant and Ashby’s 1970 good-regulator theorem established, under its mathematical assumptions, a relationship between effective, economical regulation and the existence of a model of the regulated system.⁹² Neither result proves that an organization must build a literal digital brain, nor that accumulating more data necessarily improves control. Their relevance is more precise: effective organizational action depends on the capacity to distinguish consequential states, anticipate relevant relationships, and select responses appropriate to the disturbances and objectives of the system.

The contemporary Company Brain can be interpreted as an attempt to make portions of these regulatory capabilities computationally persistent. It connects representations of organizational conditions to the processes through which decisions are formulated, authorized, executed, and evaluated. Yet the representation is necessarily partial. Tsoukas’s conception of the firm as a distributed knowledge system challenges the assumption that organizational knowledge can be completely assembled within a single overseeing mind.⁹³ Relevant knowledge is continually produced through local practice, social relationships, and circumstances that cannot be comprehensively specified in advance. A Company Brain should therefore be conceived not as the elimination of distributed knowledge, but as an infrastructure that enables distributed knowledge to participate more effectively in collective reasoning and action.

This historical connection leads to a crucial distinction between representing an organization and regulating an organization. An enterprise architecture repository may represent systems, responsibilities, and processes without influencing their execution. An ERP system may regulate many transactions without representing the wider reasons that justify them. A Company Brain may improve the interpretation of organizational conditions without acquiring authority to change those conditions. An autonomous agent may act upon a valid representation without understanding the institutional consequences of its actions. The cognitive enterprise emerges from the controlled integration of these capabilities, not from their mere coexistence.

The proposed theory must therefore account for four different questions: what can the organization know; what can it cause to happen; under which conditions can it act without further intervention; and who possesses the authority and bears the consequences of those actions?

Four constitutive properties of the cognitive enterprise

Organizational cognition is the institutionally available capacity to construct, maintain, interrogate, and revise representations of the enterprise and its environment, including relevant facts, relationships, operational conditions, historical experience, causal hypotheses, decision contexts, and acknowledged uncertainty. It encompasses more than information retrieval or statistical prediction. Its defining feature is the ability to connect evidence to justified organizational judgments while preserving the provenance and limitations of those judgments. Cognition in this sense is an operational property of the organization, not a claim that the organization possesses subjective experience or consciousness.

The relevant distinction is between the existence of information and the organization’s capacity to use it correctly. A company may retain millions of transactions while being unable to reconstruct the circumstances under which a critical decision was made. Another may possess advanced analytical models but lack the means to determine whether their assumptions remain applicable. Organizational cognition becomes more substantial when relevant knowledge remains accessible, interpretable, contestable, and revisable across changes of personnel, software, and organizational structure. Its quality is measured by the correctness, usefulness, and durability of organizational judgments—not by the quantity of stored information or the fluency of an interface.

Organizational agency is the capacity of an institutionally organized system of actors to initiate, coordinate, and execute actions directed toward defined organizational purposes. Agency belongs here to the organized system, not necessarily to a single individual or software process. A production organization exercises agency through the combined actions of employees, controllers, workflow engines, planning systems, and authorized managers. An agent-native firm might implement many of the same functions through computational actors. The analytical issue is whether the organization can transform relevant conditions and objectives into coordinated interventions, including the allocation of resources, the creation of commitments, and the modification of its operating environment.

Agency must be distinguished from intelligence. A deterministic control system can exercise extensive operational agency within a defined domain while possessing little capacity for contextual interpretation or novel reasoning. Conversely, an exceptionally capable reasoning system may remain purely advisory. The existence of a Company Brain therefore does not entail organizational agency, just as the existence of automated execution does not establish sophisticated organizational cognition.

Organizational autonomy is the extent to which the enterprise can sustain specified decision and action processes without additional intervention from actors outside the delegated operating arrangement, within defined environmental conditions, temporal horizons, and authority boundaries. Autonomy is relational and conditional. A procurement process may be autonomous for routine catalog purchases while requiring human intervention for strategic supplier changes. A production system may operate independently under normal conditions while transferring control when unexpected disturbances exceed its validated capabilities.

A company may also possess substantial operational autonomy without possessing the authority to modify its fundamental objectives, corporate constitution, or reserved decision rights. Such restrictions do not make its operational autonomy fictitious. They define its institutional boundaries. Autonomy should therefore be measured in relation to a specified activity, operating environment, and intervention threshold; it is not an absolute property that either exists or does not exist.

Organizational accountability is the institutional and legally meaningful arrangement through which decision rights, duties, controls, review mechanisms, explanations, and consequences are allocated among recognized actors. It determines who may authorize an action, who must supervise a process, who can challenge a decision, who bears relevant responsibilities, and through which mechanisms an affected party may obtain correction or remedy. Accountability includes the ability to attribute actions to their origins, but attribution alone is insufficient. A perfect audit trail does not establish that an action was legitimate, that appropriate oversight existed, or that an injured party has an effective remedy.

Unlike the other three properties, accountability cannot be understood as a computational capability alone. Its institutional legitimacy ultimately depends on legal and social arrangements that exist outside the software executing organizational processes. Jensen and Meckling’s agency-cost framework explains why the separation of ownership, decision-making, and control creates economic problems even in entirely human organizations.⁹⁴ Introducing artificial actors changes how decisions may be delegated and monitored, but it does not abolish the underlying question of whose interests are served, who can exercise control, and who remains responsible for outcomes.

The four definitions must be interpreted as analytically separable rather than empirically or statistically independent. In practice, a change in one property may influence the others. A better Company Brain may enable broader delegation; broader delegation may require more elaborate accountability mechanisms; and stronger controls may constrain certain dimensions of autonomy. Their separation means that no such relationship follows automatically from the definitions. Each proposed causal connection must be established through an identifiable mechanism and, ultimately, empirical evidence.

The framework also distinguishes capability from institutional legitimacy. An organization may technically be capable of executing a transaction it has no right to perform; it may possess legal authority to make a decision for which its information is inadequate; and it may maintain extensive records while denying affected parties any meaningful possibility of challenge. These are different failures, requiring different remedies.

A multidimensional organizational design space

The four properties imply that the cognitive enterprise should not be represented by a single AI maturity scale. The traditional image of technological progress—moving from digitization to copilots, from copilots to agents, and finally to full autonomy—confuses advances in computational capability with decisions about organizational structure and authority. It also implies, without justification, that increasing autonomy is necessarily desirable.

Consider four illustrative configurations. A research-intensive company may maintain a sophisticated Company Brain supporting executives, engineers, and analysts while reserving nearly all consequential decisions to human decision-makers. Its organizational cognition may be highly developed, but its agent autonomy deliberately limited. An automated industrial facility may execute complex sequences through deterministic controls while lacking an integrated representation of the reasons, exceptions, and institutional commitments behind those sequences. Its operational agency may be substantial even where its broader cognitive integration remains limited. A poorly governed agent-native startup might delegate extensive decisions without maintaining adequate provenance, independent controls, or effective escalation; it would exhibit substantial autonomy without satisfactory accountability. Finally, a human–agent enterprise may combine extensive shared cognition with selective automation, distributed expertise, and carefully bounded delegated authority.

Because none is necessarily more advanced in every relevant respect, the article proposes an organizational configuration space whose coordinates correspond to the four properties. Cognition can be assessed through the availability, quality, integration, contextual relevance, and correction of organizational knowledge. Agency can be assessed through the range of activities the organization can initiate and coordinate, and through the reliability of their execution. Autonomy can be assessed through the scope of action possible without additional intervention. Accountability must be examined through the structure and effectiveness of authority, oversight, contestation, and remedy rather than simply through the existence of technical logs.

These assessments should not be collapsed into one unvalidated numerical index. In particular, accountability cannot be ranked solely by counting approvals, and cognition cannot be assessed solely by measuring the accuracy of answers to questions. A highly bureaucratic organization may require numerous approvals while exercising ineffective control; a model may retrieve correct facts while contributing to poor decisions because it fails to recognize their context.

The relevant unit of analysis is consequently not always the company as a whole. Organizations can contain domains with very different configurations. A treasury function may require strict authority controls, a software-development environment may permit broader experimentation, and an industrial safety system may operate under automatic protective logic that deliberately overrides ordinary managerial discretion. The enterprise-level configuration is a structured combination of these domain-specific arrangements.

This produces a stronger conception of organizational maturity: not the maximum degree of AI deployment, but the ability to deliberately configure cognition, agency, autonomy, and accountability according to the purposes, uncertainties, dependencies, and consequences of each organizational function.

Formalizing autonomy without confusing discretion with performance

Autonomy requires particularly careful treatment because its ordinary use conflates several different properties. A system that initiates its own tasks is not necessarily authorized to spend money. A system that plans for long periods is not necessarily capable of responding to unforeseen circumstances. A system that executes many transactions is not necessarily authorized to change its objectives. These distinctions are fundamental when moving from tool-using agents to organizational systems.

For a specified organizational domain and operating envelope, define an autonomy descriptor

\boldsymbol{\alpha}=(i,p,e,r,d,h,v,o,g), \qquad \alpha_j\in[0,1], \tag{2}

where i denotes the scope of independent initiation; p, planning discretion; e, authorized execution scope; r, control over organizational resources; d, the criticality of decisions within the delegated mandate; h, the operating horizon without required intervention; v, authority to undertake actions with irreversible consequences; o, the degree of freedom from routine external oversight; and g, the authority to revise objectives or governing decision criteria.

The interval is a proposed normalization convention, not a natural physical scale, and the components should be reported alongside their underlying units, operating envelopes, and risk constraints. Each dimension requires an explicit operational definition and assessment protocol. For example, temporal autonomy might be measured through a specified maximum unattended operating interval, while resource authority could be expressed through bounded expenditure categories and financial exposure. Normalization should not obscure these underlying units or permit comparisons across incompatible environments.

More importantly, the components are not all desirable attributes to maximize. Greater authority over irreversible operations may increase institutional risk; the ability to revise governing objectives may be impermissible; and the absence of oversight can represent a deficiency rather than an achievement. The vector records the configuration of delegated discretion, not the quality of the resulting organization.

An enterprise may therefore be highly autonomous in routine execution and deliberately non-autonomous in constitutional matters. That separation is a feature of mature institutional design. The ability to revise a production schedule within approved constraints is different from the authority to alter the criteria defining acceptable production risk. The first may be delegated operationally; the second may belong to an authorized corporate or regulatory decision process.

This distinction introduces two analytically separate forms of organizational adaptation. Operational adaptation changes plans, resources, and process execution within an established mandate. Constitutional adaptation changes the mandates, roles, objectives, decision rights, or rules governing the organization itself. Agents may be allowed to propose both, but their authority to enact them need not—and generally should not—be equivalent.

A useful general principle follows: the capacity to execute organizational decisions should not automatically entail authority to redefine the institutional conditions under which those decisions become legitimate. This is not a claim that constitutional rules must remain permanently immutable. It is a requirement that changes to those rules follow a separately authorized procedure.

A computational model of cognition, action, and institutional control

The four-property theory can be expressed more precisely by modeling the enterprise as a partially observable, institutionally constrained dynamical system.

Let s_t denote the relevant state of the enterprise and its environment at time t, including operational resources, contractual conditions, dependencies, and external circumstances. The full state is generally unavailable to any participant or computational system. Instead, the organization receives observations o_t from applications, measurements, communications, and external events. Its Company Brain maintains an evolving representation K_t constructed from those observations, retained records, semantic models, and evidentiary provenance.

Let m_t denote the currently applicable organizational purposes and delegated mandates, \mathcal P_t the operative policy and authority rules, u_t the authenticated acting role, and c_t the relevant execution context. A joint human–agent policy \pi may generate candidate actions, but candidates are not equivalent to executable commands. A simplified model is

\begin{aligned} o_t &\sim O(\cdot\mid s_t),\\ K_t &= U(K_{t-1},o_t,\nu_t),\\ \hat a_t &\sim \pi(\cdot\mid K_t,m_t),\\ a_t &= G(\hat a_t,u_t,c_t,\mathcal P_t),\\ s_{t+1} &\sim T(\cdot\mid s_t,a_t,\xi_t). \end{aligned} \tag{3}

Here, O describes the observation process; U updates the organizational representation using new observations and additional evidentiary context \nu_t, including provenance and verification results; \hat a_t is a proposed action; and G is an independently enforced institutional gate that either permits execution or produces a rejection or review outcome. To make the transition well typed, define a_t as the permitted action when authorization succeeds and as a distinguished no-command action \bot otherwise; the transition function T then allows the business state to change through external disturbances \xi_t even when a_t=\bot. This remains a conceptual abstraction, not a specification of the transactional implementation.

The equations are a conceptual abstraction rather than an empirically validated universal model of firms. Their purpose is to make several distinctions explicit. The true state is different from the organization’s representation of that state. The production of a candidate action is different from permission to execute it. Execution is different from achievement of the intended outcome. Finally, an observation of an outcome is not automatically a justified revision of organizational knowledge.

The distinction between s_t and K_t is particularly consequential. An organization may execute an entirely authorized action based on an incorrect representation of inventory, contractual obligations, or external conditions. It may also infer the situation correctly but propose an action outside the acting role’s mandate. These failure modes cannot be managed through the same control alone. Better retrieval or reasoning addresses aspects of the former; authorization mechanisms address aspects of the latter. Independent outcome verification is required to establish whether an action actually produced its intended effects.

For actions that reach an authoritative execution system, an admissibility requirement can be expressed as

\operatorname{Dispatch}(u,a,c,t)=1 \Longrightarrow \operatorname{Permit}_{\mathcal P_t}(u,a,c)=1 \;\land\; \operatorname{ValidPreconditions}(a,c)=1. \tag{4}

The implication specifies a design invariant: an execution gateway must not dispatch a consequential operation unless the applicable policy permits it and the required transaction preconditions hold. It does not prove that an authorized action is economically wise, physically safe, or legally effective, nor that the implementation is free of bypasses or defects. Those stronger conclusions require additional evidence and controls. In particular, resource constraints must account for cumulative exposure, concurrent actions, and distributed commitments; otherwise, several individually permitted transactions may collectively violate an organizational limit.

The significance of this model is that the Company Brain cannot authorize actions simply by producing convincing explanations. Nor may the agent runtime treat a proposed action as a valid commitment merely because it is consistent with an objective. The enterprise must preserve an enforceable separation between evidentiary interpretation, candidate decisions, institutional authority, and actual execution.

This separation creates the conditions for higher autonomy without requiring unconstrained self-government. Agents can independently investigate, propose, coordinate, and execute within explicit boundaries; authorized actors can review exceptional decisions; and organizational learning can occur through verified outcomes. In principle, the arrangement can accommodate both extreme companies. They differ in the composition of the actors implementing \pi and in the distribution of authority, but not in the need to distinguish knowledge, action, and legitimacy.

The executable institutional constitution

The most consequential implication of the model is that organizational formalization may extend beyond the digital representation of transactions and workflows to the representation of the institution’s own operating rules.

Historically, companies have formalized responsibilities through corporate statutes, organizational charts, delegated authorities, policies, procedures, contracts, and application configurations. These artifacts constrain behavior, but they are often distributed across documents and systems whose meanings must be reconstructed by people. The combination of a Company Brain and persistent agents creates an incentive to make a larger portion of this institutional structure explicit and computationally usable.

The article proposes the concept of an executable institutional constitution: a versioned, machine-interpretable representation of selected organizational roles, mandates, authority relationships, commitments, prohibitions, decision procedures, escalation mechanisms, and controlled change rules. The term constitution is used analytically, not to imply that a corporate software model replaces the legally binding constitution, articles of association, applicable law, or the authority of recognized corporate organs.

Its function is to connect the organizational model of what exists and what is believed with the institutional model of what may legitimately be done. The Company Brain supplies evidence, context, and competing interpretations. The executable institutional constitution supplies operative constraints and the procedure through which organizational decisions become authorized. Human and artificial actors participate through these interconnected structures, while systems of record establish actual transactions and their consequences.

The distinction can be expressed through three different kinds of organizational change.

The first changes knowledge: new observations, corrected records, revised hypotheses, and additional interpretations alter what the organization believes about its circumstances. The second changes operations: actors execute transactions, allocate resources, or coordinate activities under existing authority. The third changes institutional rules: authorized decision-makers modify mandates, responsibilities, objectives, controls, or decision procedures.

These changes are not interchangeable. A supplier disruption may justify revising an operational plan, but it does not authorize a procurement agent to increase its own financial limits. Discovery of a flawed business rule may justify proposing its amendment, but it does not establish that the agent discovering the defect may deploy a replacement without approval. Conversely, an authorized policy change must be propagated to operating systems and processes; otherwise, the institution may formally adopt a rule while continuing to act under an obsolete implementation.

The organization thus becomes partially self-describing, partially executable, and potentially self-adapting within explicitly governed limits. Its ability to modify its internal procedures may increase, but the legitimacy of those modifications remains anchored in the institutional authority through which they were adopted.

This does not imply that every organizational rule can be formalized completely. Contracts are incomplete; values can conflict; social expectations change; and practical expertise often relies on contextual judgments that are not expressible as reliable executable predicates. The executable institutional constitution is therefore necessarily partial. Its purpose is to make the formalizable portion of authority more explicit and enforceable, while identifying where interpretation, discretion, and legitimate human or institutional decision-making remain necessary.

The architecture is compatible with a company that retains many employees and with one whose operational roles are predominantly implemented by agents. In both cases, organizational roles can persist independently of the particular human or computational actors occupying them. The firm increasingly preserves not merely its records and procedures, but also a structured representation of how authority is distributed and exercised.

The architecture of a cognitive institution

Figure 4 brings the theoretical properties together in a single institutional architecture. It distinguishes the flow of evidence, the proposal and execution of actions, the application of authority, the verification of outcomes, and the controlled modification of governing rules.

%%{init: {"theme": "neo", "look": "handDrawn", "layout": "elk"}}%%
flowchart TB
    X["Environment and business systems"]
    B[("Federated Company Brain")]
    H["Human operational actors"]
    A["Artificial agent actors"]
    P{"Authority and execution gate"}
    E["Enterprise execution systems"]
    V("Independent outcome verification")
    G["Corporate governance and authorized reviewers"]
    C("Institutional change control")
    M[("Versioned mandates, roles and policies")]

    X -->|"observations and records"| B
    B -->|"governed evidence and context"| H
    B -->|"governed evidence and context"| A

    H -->|"proposed actions"| P
    A -->|"proposed actions"| P

    M -.->|"constrains"| P

    P -->|"permitted operations"| E
    P -->|"exceptions and escalation"| G
    G -->|"authorized decisions"| P

    E -->|"execution results"| V
    V -->|"qualified outcome evidence"| B
    V -->|"audit and incident evidence"| G

    G -->|"approves rule changes"| C
    C -->|"publishes governed revisions"| M

Figure 4: The cognitive enterprise integrates evidence, distributed agency, bounded execution, and institutional learning while preserving independent authority over organizational rules and consequential actions.

The diagram describes a conceptual architecture rather than a single deployable product. Its central point is that organizational cognition, agency, and authority interact but are implemented through distinguishable mechanisms. The Company Brain is responsible for making evidence and context available; the actors formulate and initiate interventions; the policy gate constrains execution; and outcome verification supplies evidence for learning and governance.

The governance branch is particularly important. The path by which the organization changes its rules is separate from the path by which actors execute business activities under those rules. An agent may identify the need for a new process, propose a change, and support the proposal with evidence, but the deployment of that change must follow the applicable institutional authority. This prevents the general capability to generate organizational modifications from becoming unrestricted permission to enact them.

The architecture also reveals why accountability cannot be implemented merely by assigning a role to an AI agent. A role may be the immediate source of an action, but a complete account must establish the mandate under which it operated, the policies and evidence relevant at the time, the execution mechanisms involved, the outcome observed, and the legally recognized persons or entities responsible for the relevant decision and its consequences.

The result is an institution that can become computationally more autonomous without becoming normatively self-sovereign. The distinction is essential: software can execute a policy, but technical execution does not itself establish the legitimacy of the policy, the authority that adopted it, or the acceptability of its consequences.

Four requirements for a viable cognitive enterprise

The model suggests four architectural conditions that must be examined together when assessing the viability of a cognitive enterprise.

The first is epistemic continuity. The organization must be able to preserve, reconstruct, and revise the grounds of its judgments across time, personnel changes, system replacements, and model upgrades. The relevant object of continuity is not the original language-model instance or its internal activations, but the institutionally significant evidence, assumptions, decisions, and interpretations. This is the strongest distinction between a Company Brain and an enterprise search system.

The second is operational coherence. The organization must coordinate its actions in ways that respect dependencies, resource constraints, and shared objectives. Individually effective agents can still create collectively dysfunctional behavior when they compete for the same resources, produce incompatible commitments, or optimize local performance measures at the expense of the enterprise. Coordination must therefore be evaluated at process and organizational levels, not inferred from the success of isolated tasks.

The third is constitutional integrity. The enterprise must maintain an effective distinction between executing authorized activities and changing the rules under which those activities are authorized. This involves controlled delegation, separation of duties, versioned policies, institutional change management, and the capacity to challenge or suspend actors whose behavior deviates from their mandates. Constitutional integrity does not require immutable rules; it requires legitimate procedures for changing them.

The fourth is adaptive viability. The organization must detect when its representations, plans, or operating assumptions cease to be adequate and must possess mechanisms for correcting them. A system that executes an established process with exceptional efficiency but cannot recognize that the process has become inappropriate is operationally capable but adaptively fragile. Learning therefore requires a connection between intended outcomes, observed consequences, independent evaluation, and justified revision.

These four conditions are proposed architectural requirements, not a theorem that establishes the sufficiency of any particular implementation for corporate viability. Their presence does not guarantee profitability, legality, social legitimacy, or survival. Nevertheless, their absence identifies distinct mechanisms through which a purported cognitive enterprise may fail.

A company can possess extensive memory but lack epistemic continuity if it cannot distinguish current evidence from obsolete assertions. It can deploy thousands of agents but lack operational coherence if their actions are not coordinated. It can automate nearly all routine operations but lack constitutional integrity if agents can silently expand their mandates. And it can execute approved plans faithfully while lacking adaptive viability if its feedback mechanisms cannot recognize an obsolete strategy.

This makes the quality of integration at least as important as the capabilities of individual components.

The possibility of failure also reveals an important tension. Greater cognitive integration can increase the speed and consistency of organizational action while exposing the enterprise to common-mode errors. A flawed representation, misinterpreted policy, or corrupted knowledge source may influence many agents and employees simultaneously. Formalization can improve enforceability while hardening mistaken assumptions into executable rules. More complete observation can improve accountability while intensifying surveillance. Greater autonomy can increase responsiveness while allowing damaging actions to propagate before intervention.

Nor are human overseers an automatic remedy. Bainbridge’s analysis of the ironies of automation showed why increasing automation can leave human operators responsible for exceptional conditions precisely when their reduced involvement makes those conditions harder to understand and manage.⁹⁵ A cognitive enterprise must therefore preserve meaningful intervention capabilities, relevant expertise, and independently available evidence, rather than treating the mere presence of a human approval step as proof of effective control.

The optimization target is consequently not maximal cognition, maximal agency, or maximal autonomy considered independently. It is reliable and legitimate organizational performance under uncertainty, with the capacity to detect error, contain its consequences, and revise the institutional response.

The two extremes as configurations of one institutional architecture

The two thought experiments can now be compared without assuming that either represents a more advanced evolutionary stage. They employ the same underlying categories but distribute cognitive participation, operational authority, and institutional responsibility differently.

In the employee-free operating model, the organization attempts to achieve economic continuity through predominantly artificial execution. Its central challenge is whether a computational population can maintain coordinated, reliable action over extended periods while remaining subject to legitimate corporate governance and external obligations. In the human–agent model, the organization uses the same cognitive infrastructure to connect specialized human understanding with computational capabilities. Its central challenge is whether this integration improves collective performance without undermining expertise, professional autonomy, accountability, or the legitimacy of organizational decisions.

Table 2 summarizes the distinction.

Dimension	Shareholder-governed, employee-free operating model	Human–agent cognitive enterprise
Principal operating actors	Persistent agents, deterministic services, and external providers	Employees, specialists, managers, agents, and deterministic services
Organizational cognition	Shared evidence and institutional memory support computational coordination	Shared evidence and institutional memory support human and artificial cognition
Organizational agency	Predominantly computational within delegated mandates	Distributed across human and computational participants
Operational autonomy	Potentially extensive for bounded processes	Variable by process, task, authority, and risk
Human participation	Ownership, legally required governance, reserved decisions, and external dependencies	Ownership, governance, operational expertise, judgment, collaboration, and oversight
Organizational structure	Persistent roles implemented primarily by computational services	Hybrid roles, teams, and processes with variable human–agent participation
Process coordination	Agent orchestration and executable institutional rules	Mixed human–agent workflows supported by common institutional representations
Institutional adaptation	Agents may propose changes; authorized governance must control their adoption	Humans and agents may propose changes through established organizational governance
Principal cognitive risk	Correlated errors, incomplete context, and weak detection of unexpected conditions	Automation bias, suppression of independent judgment, and cognitive dependency
Principal governance risk	Weak effective oversight, mandate drift, and unclear accountability across external dependencies	Surveillance, concentration of interpretive power, contested decision rights, and inadequate worker participation
Principal economic hypothesis	Lower internal coordination and labor costs for suitable operations	Complementarity between human expertise, shared cognition, and bounded automation
Primary empirical uncertainty	Long-horizon whole-firm reliability and institutional adaptability	Whether cognitive integration creates enterprise-level benefits beyond individual AI assistance
Demonstrated status as of November 2026	A theoretical limiting case; bounded component capabilities do not establish general-purpose firm autonomy	Relevant components and local benefits demonstrated; general enterprise-wide cognitive integration remains unproven

Table 2: The two organizational extremes as alternative configurations of cognition, agency, autonomy, and accountability, rather than successive stages of AI maturity.

The comparison suggests that the two extreme configurations may occupy different regions of the same organizational design space rather than constitute fundamentally incompatible forms. A company could rely on agents for some complete business processes while retaining human specialists in others. It could make organizational knowledge broadly accessible while centralizing certain decisions. It could delegate operational planning extensively while reserving institutional change to corporate organs. The resulting organization need not correspond to any familiar classification based on employee count, hierarchy, or technology deployment.

More importantly, the presence or absence of employees may become less informative about an organization’s cognitive and operational architecture than the distribution of its knowledge, decision rights, and execution capabilities. Two firms with the same number of employees may differ radically in their degree of institutional cognition and agency. Conversely, an employee-free digital company and a large industrial organization with thousands of workers may employ remarkably similar mechanisms for knowledge representation, workflow coordination, authorization, and verified learning.

The theoretical object is therefore not the automated company or the AI-augmented workforce, but the cognitive enterprise as an institutional configuration.

Empirical implications of the four-property framework

The framework entails testable, but distinct, claims about cognitive integration, delegated authority, and organizational performance. Shared memory may create an additional benefit when it preserves the context needed for decisions across functional boundaries; broader autonomy may create value only when executable mandates and independent controls support reliable end-to-end processes; and the distribution of human expertise may matter particularly under unfamiliar or poorly observable conditions. These are conditional relationships rather than stages of a universal maturity ladder, and their comparative performance cannot be inferred from the number of models, agents, or employees.

The later research program formalizes these claims as competing hypotheses, defines interventions separating shared memory from agent execution, and specifies causal comparisons, failure tests, and longitudinal measures. At this point the conceptual result is sufficient: the four properties identify different organizational capabilities whose economic value and mutual dependence must be established empirically rather than inferred from technical possibility.

Toward a new conception of the firm

The four-property framework leads to a reinterpretation of the institution itself. The traditional firm has never been reducible to the people it employs: its productive capabilities emerge from the organization of capital, expertise, routines, assets, knowledge, contracts, authority, and coordinated activity. What the convergence of the Company Brain and autonomous agents changes is the possible implementation of those capabilities. More of the institution’s knowledge can persist independently of its current participants; more of its processes can be represented and executed computationally; and more of its operating structure can be governed through explicit, versioned rules.

The result is not necessarily an artificial organization in the sense of a single machine possessing a unified will. On the contrary, the emerging enterprise may become increasingly distributed, with multiple computational and human actors contributing different kinds of intelligence within a shared institutional framework. Its apparent unity comes not from a central omniscient agent, but from the capacity to preserve relevant relationships among evidence, decisions, authority, commitments, actions, and outcomes.

This distinction reconciles two apparently conflicting perspectives. The firm remains a distributed knowledge system in which no participant can possess all relevant information. Yet the firm can increasingly maintain a common computational infrastructure through which distributed knowledge becomes accessible, decisions become traceable, and actions become coordinated. The Company Brain does not abolish distribution; it changes the mechanisms through which distributed cognition is integrated.

Equally, the executable institutional constitution does not eliminate discretion or governance. It makes selected powers and obligations more explicit, allowing the organization to delegate action while preserving the distinction between operational execution and institutional authority. Computational agents do not need to become legal persons for their actions to be organizationally meaningful; nor does the organization’s increasing ability to act independently imply that it can legitimately determine all its own purposes.

The conceptual advance is therefore not the disappearance of human employees, the universal deployment of agents, or the construction of an all-knowing corporate model. It is the possibility that organizational cognition and organizational agency become persistent, interoperable properties of the institution, while autonomy and accountability are deliberately configured rather than treated as automatic consequences of technological capability.

This may change the theory of the firm because it changes the costs and feasible forms of organizing knowledge and action. When relevant knowledge can be made available without repeatedly transferring it through people, some reasons for existing hierarchies and organizational boundaries may weaken. When sophisticated agents can coordinate activities across enterprise systems, some functions historically requiring substantial administrative labor may be implemented through comparatively small operating structures. When shared cognitive infrastructure increases the value of accumulated experience and cross-functional integration, other reasons for maintaining larger organizations may strengthen. The economic direction remains open; it depends on the costs of infrastructure, coordination, assurance, adaptation, and the governance of interdependent actors.

The more radical implication concerns organizational continuity. A firm has traditionally retained knowledge imperfectly through documents, procedures, databases, and employees whose experience may be lost when they depart. A mature cognitive enterprise could preserve a richer, independently inspectable history of what the institution understood, what it attempted, why its actions were authorized, what happened, and how subsequent evidence changed its judgments. Organizational learning would remain fallible, but parts of it could become less dependent on the uninterrupted presence of particular individuals.

Such an enterprise would not necessarily think as a human thinks. It would not possess consciousness merely because it can reason over institutional records, generate plans, or modify approved procedures. The term cognitive designates a functional and organizational property: the capacity to maintain and revise actionable representations through a distributed system of actors, artifacts, and controlled processes. The term enterprise designates the continuing institutional arrangement within which those capabilities are directed toward economic activity and subjected to legitimate authority.

The two extremes thereby become tests of the same foundational proposition. The company without employees asks how far organizational agency and autonomy can extend when human operational participation approaches zero. The human–agent company asks how far organizational cognition can extend when relevant knowledge is made accessible across the workforce and computational agents participate in the same processes. Neither scenario resolves the other, and neither is inherently the endpoint of organizational evolution.

The cognitive enterprise is not defined by who performs its work, but by the architecture through which its knowledge becomes judgment, its judgment becomes authorized action, and the consequences of that action become institutional learning.

That formulation is the theoretical center of the article. It implies that the decisive competition may not be between firms with more or fewer employees, nor even between firms using more or less advanced AI models. It may increasingly concern their ability to construct reliable systems of collective cognition and action whose capabilities persist, evolve, and remain governable despite continuous changes in their constituent human and computational actors.

The ultimate research question is no longer simply whether artificial intelligence can perform the tasks of a company. It is whether a company can become a durable cognitive institution: an organization whose capacity to understand and act is progressively embedded in its architecture, while the legitimacy of its purposes and the accountability for its consequences remain institutionally intelligible.

The firm that thinks, if such a form can be realized, will not be the firm that has eliminated human intelligence. It will be the firm that has learned to make intelligence an enduring organizational capability without confusing intelligence with authority, autonomy with legitimacy, or computational coherence with truth.

The economics of a firm with cheaper cognition

The four-property theory changes the economic question from whether artificial intelligence makes individual workers more productive to whether it changes the comparative advantages of alternative ways of organizing economic activity. The distinction is fundamental. A firm exists not merely because particular tasks must be performed, but because tasks, knowledge, resources, commitments, and decisions must be coordinated under uncertainty. If a significant portion of that coordination can be performed by computational agents operating through persistent organizational knowledge and explicit institutional rules, the economics of organization may change even when the underlying products, physical assets, and markets remain substantially the same.

The conventional intuition is that cheaper intelligence will produce smaller firms because fewer employees will be required to perform administrative, analytical, and coordinating functions. The opposite intuition is equally plausible: organizations that integrate their accumulated knowledge into a Company Brain may achieve increasing advantages from scale, making large enterprises more capable of exploiting the information, relationships, and operating experience they already possess. A third possibility is that the boundary between firm and market becomes more permeable, with networks of specialized enterprises coordinating through interoperable agents and machine-executable processes. None follows necessarily from the technology itself. Their relative plausibility depends on how the cost of cognition interacts with the costs of coordination, verification, governance, adaptation, and ownership.

The preceding sections provide the basis for a more precise economic argument. The Company Brain is not simply another information technology; it is a proposed mechanism for preserving and applying organizational knowledge across decisions and time. Agent coworkers are not merely cheaper substitutes for employees; they are computational actors capable of initiating and coordinating selected business activities. The executable institutional model formalizes portions of the rules under which those actors can make commitments and exercise authority. Together, these technologies could alter the costs of organizing production, but their economic value depends on whether they reduce the total cost of achieving verified, legitimate outcomes—not merely the cost of producing recommendations or executing individual tasks.

The central economic hypothesis is that a cognitive enterprise changes the boundary and scale of the firm by changing the relative costs of acquiring, preserving, interpreting, coordinating, and acting upon organizational knowledge. Its most important effect may therefore concern the economics of collective cognition, not just the automation of labor.

The theory of the firm revisited: from transaction costs to cognitive coordination costs

Coase’s original explanation of the firm begins with a deceptively simple observation: organizing production through markets is not free.⁹⁶ Discovering prices, identifying counterparties, negotiating agreements, supervising performance, and adapting contractual relationships require resources. A firm can sometimes perform these coordinating functions more economically through internal direction than through repeated market transactions. Yet internal organization also has costs, including managerial coordination, bureaucracy, information loss, and the possibility that resources are allocated inefficiently.

Williamson’s transaction-cost economics refined this comparison by examining how uncertainty, opportunism, contractual incompleteness, and asset specificity affect the relative efficiency of markets, hierarchies, and intermediate governance arrangements. Hart and Moore’s property-rights approach supplied another mechanism: when contracts cannot specify every future contingency, ownership matters because it affects residual control rights and incentives to make relationship-specific investments.⁹⁷ These theories remain relevant in an agent-mediated economy. Artificial intelligence does not eliminate contractual incompleteness merely because an agent can generate a detailed contract or execute a formalized workflow. Unknown future states, ambiguous obligations, contested interpretations, and the possibility of strategic behavior remain economic facts.

However, the cost of obtaining and using knowledge deserves more explicit treatment within this framework. Garicano’s theory of knowledge-based hierarchies explains how organizations allocate the acquisition and use of knowledge among production workers and specialized problem-solvers. When communication and problem-solving are costly, hierarchies allow routine problems to be solved locally while exceptional problems are referred to specialists.⁹⁸ A Company Brain could change precisely these costs. Problems that previously required a sequence of communications across organizational levels may become resolvable through shared representations, automated retrieval, and coordinated reasoning. Exceptions may be identified earlier, and expertise may be applied without requiring specialists to reconstruct the entire background of every case.

But reducing the cost of obtaining an answer is not the same as reducing the cost of a correct decision. A language model can produce a proposed interpretation at low marginal cost while leaving substantial work to establish whether the underlying data are current, whether the inferred relationship is valid, whether the proposed action is authorized, and whether the actual outcome satisfies the business objective. The economic unit of analysis should therefore be the verified and institutionally valid decision or completed process, rather than the individual inference or model invocation.

This distinction extends the economics of digital technology identified by Goldfarb and Tucker, who organized the analysis of digitization around reductions in search, replication, transportation, tracking, and verification costs.⁹⁹ Agentic systems potentially reduce additional costs associated with the interpretation and execution of interdependent decisions. Yet verification itself may become more demanding when decisions are generated by probabilistic systems, supported by multiple uncertain sources, or executed through chains of independently operating agents. The cost structure is not simply lower; it is redistributed among activities.

A firm may save on departmental reporting and administrative coordination while spending more on semantic integration, policy engineering, identity management, model evaluation, audit, cybersecurity, and recovery from unexpected actions. An outsourced agent may appear less expensive than an internal team while imposing substantial costs for ensuring contractual compliance, inspecting its behavior, and retaining the capacity to intervene. Conversely, a well-governed internal agent system may reduce these costs through access to authoritative data and enforceable organizational controls; the appropriate comparison must therefore include the entire institutional cost of producing a reliable outcome.

A comparative economic model of cognitive production

Consider a business activity or process x that the enterprise could perform internally, procure through an external market, or organize through a hybrid arrangement. Let g denote the chosen governance arrangement, with g belonging to the set of internal, market, and hybrid alternatives. For a specified output quality, risk tolerance, and operating horizon, define the expected total cost as

C_g(x)=P_g(x)+I_g(x)+V_g(x)+G_g(x)+R_g(x), \tag{5}

where P_g represents direct production and execution costs, including labor, computation, infrastructure, and external service charges; I_g represents the costs of integrating information, knowledge, processes, and counterparties; V_g represents verification and assurance costs; G_g represents contractual, organizational, regulatory, and authority-governance costs; and R_g represents expected residual losses from failures, disputes, interruption, and other risks not already included in the preceding categories.

The decomposition is an analytical accounting framework, not an empirical law. The components must be defined without double counting, and expected losses must be evaluated over a stated distribution of outcomes. Where severe tail risks matter, an expected-cost comparison alone is insufficient and must be supplemented by explicit risk constraints. Alternatives must also be compared for equivalent service quality and legally feasible operating arrangements.

The enterprise’s sourcing problem becomes the selection of the feasible governance arrangement with the lowest appropriately risk-adjusted total cost, subject to the desired outcomes and institutional constraints. The original make-or-buy comparison is therefore expanded, not abandoned. Internal provision is preferred when its relevant cost is lower than the cost of external provision; a hybrid arrangement becomes attractive when shared governance, coordinated specialization, or selective ownership offers an advantage over either extreme.

The novel feature is that a Company Brain and an agent infrastructure can change several components simultaneously. Internalizing a process may become more economical because an agent can reuse organizational knowledge without repeated contractual disclosure or interpretation. Externalizing the same process may become more economical because standardized agent interfaces reduce coordination costs and make supplier performance easier to observe. A hybrid network may become more effective because parties can exchange structured commitments and evidence while preserving ownership of their specialized assets.

Research predating modern AI already demonstrates why such choices cannot be reduced to production cost alone. Poppo and Zenger’s empirical comparison of alternative theories of the firm, using information services across 152 companies, found that transaction-cost, knowledge-based, and measurement explanations illuminate different aspects of make-or-buy decisions.¹⁰⁰ The emergence of computational agents reinforces the need for this integrated perspective: the economic value of an agent depends not only on whether it can perform a task, but on where the relevant knowledge resides, how its performance can be evaluated, and which governance structure can manage the consequences.

An important implication follows. Cheaper cognition does not necessarily make markets more efficient than firms, or firms more efficient than markets. It changes the relative cost curves of different institutional arrangements. The direction of organizational evolution depends on which costs fall, which new costs emerge, and how the resulting gains are distributed among participants.

The first possible equilibrium: smaller firms with disproportionately large capabilities

The first scenario is the emergence of enterprises whose productive and coordinating capabilities become unusually large relative to their internal employment. A small software company, for example, might combine human founders or directors with agents responsible for development, testing, customer operations, sales support, financial reconciliation, and routine administration. Functions previously requiring separate teams could be partially implemented through computational actors sharing a persistent organizational representation.

The economic mechanism involves a reduction in the minimum efficient organizational scale for certain activities. Historically, a firm needed enough employees and administrative capacity to sustain recurring functions even when the volume of productive transactions was relatively modest. If portions of accounting, customer support, procurement, product maintenance, and coordination can be provided as scalable computational services, the fixed organizational capacity required to operate a business may decrease.

The most favorable conditions are modular products, standardized interfaces, low physical dependence, predictable business processes, independently testable outputs, and bounded liability exposure. A digital product with automated tests and reproducible deployment provides a more suitable environment than a complex industrial enterprise whose activities depend on physical safety, asset maintenance, uncertain environmental conditions, and consequential human relationships.

The limiting possibility is an enterprise in which ownership and formal governance remain human, but a large share of the productive and administrative operating core is computational. Such a company could generate significant revenue with very few employees. Yet low employee count does not demonstrate low total resource consumption. Model inference, software subscriptions, cloud infrastructure, outsourced professional services, payment systems, external suppliers, and human labor embedded in those services remain economically necessary inputs.

A firm with no employees may therefore be a heavily outsourced firm rather than a firm without human-dependent production. The correct productivity measure must include purchased services and capital inputs rather than treating the absence of employees as evidence of extraordinary total factor productivity.

Moreover, the smallest firm is not necessarily the most competitive firm. A company with low administrative overhead may still lack distribution, reputation, finance, proprietary information, bargaining power, and the capacity to absorb losses. If its competitors can access the same agent services, the productivity improvement may become widely available and cease to provide a durable advantage. Lower entry costs may intensify competition and reduce margins rather than create exceptional profits for every entrant.

There is nevertheless a potentially transformative implication for entrepreneurship. Activities previously constrained by the need to assemble a large team may become feasible for founders with capital, specialized knowledge, and access to reliable agent infrastructure. The boundary between an individual professional practice, a small enterprise, and a larger operating organization could become more fluid. This would not abolish the theory of the firm; it would change the range of organizational configurations that can satisfy its economic conditions.

The second possible equilibrium: larger firms with increasing cognitive returns

The opposite scenario begins with the observation that a Company Brain may become more valuable as the scope of the organization’s relevant, verified experience increases.

Large firms possess extensive records of transactions, supplier behavior, product performance, customer relationships, operational incidents, contracts, and institutional decisions. Much of this material is presently fragmented, inaccessible, inconsistent, or poorly connected to the decisions for which it might be useful. A Company Brain that increases the proportion of organizational experience that can be interpreted and reused may unlock economic value unavailable through isolated applications.

This possibility connects to the economic literature on intangible capital. Brynjolfsson, Hitt, and Yang argued that investment in information technology is associated with complementary organizational assets whose value cannot be understood solely through the price of computers and software.¹⁰¹ Bresnahan, Brynjolfsson, and Hitt found evidence that information technology, workplace reorganization, and skilled labor are complementary in firm-level data.¹⁰² The significance is that technological investment produces value through changes in organizational capabilities and practices, rather than through computational equipment alone.

A Company Brain could extend this logic by turning portions of accumulated institutional knowledge into a more reusable productive resource. A previous supplier failure could inform future contracting decisions; the consequences of a product modification could inform engineering and commercial choices; an operational incident could update maintenance policies; and repeated interactions between departments could reveal structural inefficiencies that were previously understood only locally.

The resulting advantage would not arise from data volume alone. It would arise from the density, relevance, quality, and causal usefulness of connections among organizational experiences.

The strongest hypothesis is that knowledge acquired in one process may become useful in multiple others without requiring proportional duplication of the original acquisition effort. A verified engineering lesson can influence procurement, quality assurance, production, and customer commitments. An accurate supplier-risk model can improve planning across several plants. A well-specified contractual exception can guide future transactions involving similar obligations.

This creates a potential economy of cognitive scope: the value of an institutional representation may increase when it can support decisions in several interconnected activities. The effect resembles economies of scope in production, but its mechanism lies in the reuse and recombination of organizational knowledge.

However, the returns are conditional. More records can increase ambiguity, raise validation costs, and expose the enterprise to broader correlated errors. Integration across heterogeneous businesses may require costly semantic reconciliation. Different subsidiaries may use superficially similar concepts with materially different meanings. Regulatory restrictions can prevent some forms of data sharing. A large enterprise may consequently incur greater costs precisely because its organizational knowledge is more diverse and difficult to formalize.

The economic proposition is therefore not that the largest Company Brain is necessarily the most valuable. It is that the capacity to integrate relevant institutional knowledge may become a distinctive source of scale and scope economies, subject to increasing costs of semantic complexity, assurance, and governance.

This raises a further competitive possibility. If a firm combines proprietary institutional knowledge with reliable agents and integrated business processes, its advantage may be difficult to reproduce simply by purchasing access to the same underlying language models. Competitors may obtain comparable computational capabilities but lack the accumulated evidence, process semantics, validated decision rules, and institutional experience through which those capabilities generate value.

The durable asset would be the organization-specific cognitive system, not necessarily the foundation model providing part of its reasoning capability.

The third possible equilibrium: firms become more permeable through agent-mediated markets

A third scenario involves neither smaller independent firms nor more integrated corporate hierarchies, but a reconfiguration of economic activity into dynamically coordinated networks of specialized organizations.

Markets already support extensive interfirm coordination through contracts, logistics networks, payment services, outsourcing agreements, enterprise platforms, and digital marketplaces. Yet many exchanges continue to require substantial human effort to interpret specifications, resolve exceptions, verify commitments, reconcile transactions, and adapt to unexpected conditions. Autonomous agents could reduce some of these costs by conducting structured negotiations, coordinating schedules, exchanging evidence, and initiating authorized transactions.

A manufacturer might maintain its core design and customer relationships internally while contracting specialized production, testing, logistics, financing, and maintenance services through agent-mediated interfaces. An unexpected demand change could trigger a coordinated search for alternative capacity, evaluation of contractual constraints, and proposals for reallocating work among qualified partners.

The Company Brain would preserve the firm’s own operational context and institutional commitments, while interoperable external services would expose the specific information and actions required for authorized coordination. Each participating enterprise could retain its own cognitive infrastructure, creating a federation of organizational systems rather than one encompassing corporate brain.

This arrangement suggests that the firm’s boundary becomes increasingly distinguishable along three dimensions: ownership of productive assets, authority over decisions, and access to relevant knowledge. An organization may coordinate an activity without owning all the resources required to perform it; it may possess information relevant to an external supplier without acquiring authority over that supplier; and it may delegate execution while retaining responsibility for contractual outcomes.

The economics of modularity provides an important precedent. Baldwin and Clark’s work on modular design explains how explicit interfaces can permit independent experimentation and substitution within a larger system.¹⁰³ The same principle could apply to organizational capabilities when business processes, responsibilities, commitments, and technical interfaces are sufficiently specified for components to be substituted without reconstructing the entire operating arrangement.

But firms are not interchangeable software modules. Their contracts are incomplete, their incentives differ, their operating conditions evolve, and their commitments can generate legal or physical consequences. Standardized agent interfaces reduce particular coordination costs; they do not eliminate differences in competence, reliability, bargaining power, or institutional authority.

The third scenario is therefore best characterized as modular economic coordination under distributed ownership and governance. It could support firms that dynamically assemble capabilities from external providers, but the reliability of those arrangements would depend on interoperable business semantics, independently verifiable performance, enforceable commitments, and workable mechanisms for resolving disputes.

The deepest economic transformation may occur where the traditional distinction between internal organization and external transaction becomes less useful for describing how production is coordinated. The legal boundary of ownership remains, but the operational boundary of coordinated activity can extend across multiple firms.

Organizational cognitive capital: a new category of productive capability

The three scenarios suggest the need for a more explicit theory of the productive asset created by a Company Brain.

The term organizational cognitive capital is proposed here to describe the durable, institutionally controlled capability arising from the integration of relevant organizational knowledge, semantic representations, historical evidence, decision procedures, executable process models, and mechanisms for validating and improving decisions.

It is not synonymous with data, software, intellectual property, or employee expertise, although it depends on all of them. Nor is it proposed as a new accounting asset category automatically recognizable under financial reporting standards. It is an analytical concept intended to explain how accumulated organizational understanding can become a reusable input into production and coordination.

The distinction matters because the economic value of organizational knowledge has historically depended heavily on the people and routines through which it is exercised. A firm may possess an extensive documentation repository while losing operational capability when experienced employees leave. Conversely, a business may function effectively through tacit expertise that is difficult to reproduce or transfer.

A Company Brain could alter this relationship by preserving parts of the context necessary to reuse knowledge across different people, applications, and agent implementations. This would not eliminate the importance of tacit knowledge or human capabilities, but it could reduce the cost of recovering and applying selected forms of institutional experience.

Organizational cognitive capital has several economically relevant characteristics, beginning with its reuse potential can be high: the same verified representation may support many decisions without being exhausted by each use. Its complementarity is also substantial: a representation becomes more valuable when connected to reliable execution systems, specialized expertise, and appropriate decision authority. Its specificity may be high when it depends on proprietary processes, organizational vocabulary, or historical relationships. Its maintenance burden is continuous because relevant facts, regulations, operating conditions, and business priorities change. Finally, its appropriability depends on ownership, contractual rights, access controls, vendor arrangements, and the ability to prevent competitors from reproducing the relevant capabilities.

This combination resembles earlier forms of intangible organizational capital, but the proposed Company Brain adds an important possibility: portions of institutional knowledge may become directly available to computational actors capable of using that knowledge in authorized operational decisions.

The productive asset is therefore not simply knowledge that employees can consult. It is knowledge embedded in a governed system through which organizational actors can interpret, coordinate, and act.

The economic importance of this asset depends on its marginal contribution to decisions. A document that is retrieved frequently but never changes an action may have little operational value. A rarely consulted engineering constraint that prevents a catastrophic failure may have considerable value. Similarly, a historical decision that is reused automatically may improve efficiency only if its assumptions remain valid.

A useful formal interpretation comes from the expected value of information. Let s denote an uncertain state relevant to a decision, a an available action, U(a,s) the resulting organizational payoff, and z an additional observation or body of knowledge. Under risk-neutral expected-value decision-making and a fixed feasible action set \mathcal A, the net expected value of acquiring and using z can be expressed as

V(z)= \mathbb E_z\!\left[ \max_{a\in\mathcal A} \mathbb E[U(a,s)\mid z] \right] - \max_{a\in\mathcal A} \mathbb E[U(a,s)] - C(z), \tag{6}

where C(z) includes the incremental cost of obtaining, validating, integrating, and using the information.

The expression is a standard value-of-information construction adapted to the organizational setting. It does not establish that the value of the Company Brain can be calculated from readily observable data. In practice, the action set, utility consequences, and information-generating process may be uncertain or endogenous; multiple decisions may also share information, creating dependencies not captured by the single-decision model.

Nevertheless, the equation clarifies the underlying economic principle. Information has value because it can improve the selection of actions under uncertainty, not because its quantity increases.

The same reasoning applies to organizational memory. The value of retaining a past decision depends on whether the preserved evidence helps improve future decisions, prevent errors, reduce the cost of reconstruction, or support legitimate accountability. Some historical information can have negative net value when it is obsolete, misleading, unnecessarily sensitive, or expensive to maintain. A mature Company Brain must therefore manage knowledge acquisition, validation, retention, revision, and deletion as economic as well as epistemic problems.

The relevant productive asset is not accumulated information, but the durable capacity to transform relevant evidence into better coordinated and accountable action.

The paradox of cognitive integration: economies of scope and new asset specificity

The accumulation of organizational cognitive capital creates a paradox. The more effectively a Company Brain integrates a company’s knowledge and processes, the more valuable it may become internally—and the more costly it may become to replace, divide, or transfer.

Traditional transaction-cost economics recognizes that relationship-specific investments can alter bargaining positions and create dependence. A similar mechanism can arise when organizational decision processes become deeply embedded in proprietary knowledge representations, agent runtimes, workflow definitions, authorization policies, and external model services.

An enterprise may gradually encode years of operating experience into a cognitive architecture whose semantics are not easily portable. Process definitions may depend on proprietary orchestration mechanisms; historical reasoning records may be stored in vendor-specific formats; business rules may rely on implicit assumptions embedded in prompts or configurations; and critical integrations may depend on APIs that are difficult to replace.

Even if the underlying transactional data remain exportable, the organization may be unable to reconstruct the relationships among decisions, evidence, authority, and outcomes without substantial migration effort.

This represents a potential form of cognitive asset specificity: investments whose productive value depends on their incorporation into a particular organizational and technical arrangement.

Klemperer’s analysis of switching costs explains why such dependence can affect competition even when competing products offer comparable functionality.¹⁰⁴ The important question is whether the Company Brain becomes an asset owned and controlled by the enterprise or a capability effectively controlled by external infrastructure and software providers.

Consider a firm that relies on one supplier for model inference, another for orchestration, a third for enterprise search, and a fourth for agent identity and authorization. The apparent flexibility of a modular architecture may conceal dependencies that become visible only during migration, disruption, or contractual renegotiation. Conversely, an architecture based on explicit business semantics, versioned evidence, portable process models, replaceable agent implementations, and independent authorization services may preserve organizational bargaining power despite continued dependence on external technologies.

The distinction is not between purchasing software and developing everything internally. Even internally developed systems depend on hardware, standards, expertise, and external services. It is between forms of dependence that can be governed and substituted at acceptable cost, and forms that create excessive concentration of operational and institutional control.

The most important architectural consequence is that the enterprise must preserve the portability of its organizational cognition independently of the particular models and platforms through which that cognition is accessed. This is an economic argument for the separation established earlier between Company Brain, agent runtime, executable institutional model, and enterprise transaction systems. Their separation can reduce the cost of replacing individual components while preserving the continuity of knowledge, authority, and business operations.

The distribution of economic power: labor, capital, and the ownership of cognition

If organizational cognitive capital becomes an increasingly important productive resource, its ownership and governance will influence how economic returns are distributed.

The employee-free firm represents one extreme of this question. Shareholders may supply financial capital and acquire computational services capable of performing activities previously requiring substantial employment. If the resulting organization remains commercially viable, a larger proportion of its value added might accrue to owners of capital, proprietary knowledge, intellectual property, and computational infrastructure.

Yet this outcome is not automatic. Lower labor requirements can reduce costs without increasing profits when competition passes the gains to customers. Purchased AI services may transfer much of the surplus to infrastructure providers. Scarce human expertise may become more valuable when agents complement rather than replace it. And new types of firms and services may create demand for labor in activities not previously performed.

Acemoglu and Restrepo’s task-based framework distinguishes the displacement effect of automation from productivity improvements and the creation of new tasks that can restore demand for human labor.¹⁰⁵ The relevance to cognitive enterprises is that the consequences depend on which activities are automated, which capabilities are complemented, and how production and demand adjust. It is insufficient to infer aggregate employment effects from the number of tasks that agents can technically perform.

The Company Brain also introduces a more difficult question about the ownership of organizational knowledge. Much of an enterprise’s accumulated intelligence originates in the experience, decisions, observations, innovations, and cooperation of its employees and external partners. When these contributions are transformed into durable computational representations, the firm may preserve and reuse them after the original contributors leave.

This may improve organizational continuity and reduce dependence on particular individuals, but it can also change bargaining relationships. Employees whose expertise was historically scarce because it was difficult to codify may find that parts of its application become reproducible through shared infrastructure. Other employees may become more productive because they can access organizational knowledge previously unavailable to them. The distributional outcome depends on the substitutability and complementarity of their skills, the allocation of intellectual-property and data rights, and institutional bargaining arrangements.

The human-centered cognitive enterprise therefore does not guarantee a more equitable distribution of value than the employee-free model. Broad access to knowledge may empower employees operationally while the resulting intellectual and computational capital remains controlled by a narrow ownership structure. Conversely, low-cost agent services may allow independent professionals and small businesses to capture value previously concentrated in larger firms.

The economics of the cognitive enterprise must consequently distinguish the distribution of intelligence from the distribution of economic rights. Universal access to a Company Brain does not imply shared ownership of the Brain, equal authority over its development, or equal participation in the economic returns generated by its use.

Competition and industrial structure: concentration or decentralization?

The three organizational equilibria create opposing forces in market structure: lower costs of organizing a small enterprise may reduce barriers to entry. If AI services, reliable workflows, and common business interfaces are widely available, small firms may acquire capabilities once affordable only to large organizations. Entrepreneurial experimentation could become less costly, and specialized firms could enter markets without assembling extensive administrative infrastructures.

At the same time, large incumbents may exploit complementary assets that smaller rivals cannot easily replicate: trusted distribution channels, proprietary data, regulatory capabilities, finance, existing customer relationships, and accumulated organizational cognitive capital. The fixed costs of constructing and governing a sophisticated Company Brain may reinforce these advantages.

Competition may also become concentrated at the infrastructure layer even while applications become more decentralized. Thousands of small agent-operated enterprises could depend on the same limited number of model providers, cloud platforms, identity systems, payment networks, or digital marketplaces. Their apparent operational independence would then coexist with substantial dependence on common upstream infrastructure.

The existing literature on superstar firms provides a relevant but incomplete analogy. Autor and colleagues document patterns consistent with increasingly concentrated sales among highly productive firms and associated declines in the labor share in several industries.¹⁰⁶ These findings precede the present generation of agents and do not establish that cognitive enterprises will produce the same outcome. They demonstrate, however, that productivity-enhancing technological change can coincide with increased concentration rather than automatically diffusing economic advantage.

Recent AI-specific evidence remains mixed. An OECD study published in July 2026 found no significant association between non-generative AI adoption and increased market power in its examined data, while identifying both opportunities for smaller firms and advantages associated with established capabilities and concentrated AI innovation.¹⁰⁷ This is consistent with the proposition that the effects differ between firms that develop AI technology, those that adopt it, and those that control complementary assets. It does not settle the longer-term consequences of organizational agents or Company Brains, whose widespread deployment remains comparatively recent.

The economic implication is that competition may increasingly occur across several layers at once: the provision of general-purpose models, the infrastructure required to operate them, the ownership of organizational data and cognitive capital, the ability to coordinate business processes, and access to customers and markets.

A company may compete vigorously in its product market while depending on an oligopolistic market for computational capabilities. Another may control valuable organizational knowledge while purchasing interchangeable AI services. These are different configurations of economic power, even if both appear to be highly automated enterprises.

The question is thus not simply whether AI favors large or small companies. It is which layers of the cognitive enterprise exhibit economies of scale, which exhibit economies of scope, which depend on scarce assets, and which can be made competitive through interoperability and substitution.

What the evidence establishes as of November 2026

Current empirical evidence supports the importance of organizational complementarities but does not yet establish the proposed transformation of firm boundaries.

Research on enterprise information technology has long indicated that its value depends on associated organizational investments. Bresnahan, Brynjolfsson, and Hitt found complementarities between information technology, workplace organization, and skilled labor.¹⁰⁸ Bloom, Garicano, Sadun, and Van Reenen subsequently found that different information and communication technologies were associated with opposing changes in organizational autonomy: technologies improving local access to information tended to support decentralization, while technologies facilitating communication with superiors tended to strengthen centralization in their studied manufacturing firms.¹⁰⁹ This is particularly relevant to a Company Brain, whose effects may depend on whether it gives participants greater independent decision capacity or makes them more directly subject to centralized observation and instruction.

The evidence available in 2026 reinforces the importance of firm heterogeneity. A European Investment Bank study based on matched firm-level data estimated a positive labor-productivity effect associated with AI adoption, approximately 4% under its identification strategy, while finding that gains were concentrated among medium-sized and large firms and depended on complementary investments.¹¹⁰ The estimate concerns AI adoption broadly rather than integrated Company Brains, and the research does not establish an employee-free organizational model.

An OECD study covering 15 member countries similarly found that larger firms exhibited greater adoption of advanced technologies and that human and technological capital were important enablers. It also showed that some apparent productivity advantages of AI adopters weakened after accounting for complementary digital capabilities.¹¹¹ These findings caution against attributing productivity gains to AI alone when the adopting firms may already possess superior management, data infrastructure, skills, or financial resources.

More detailed US evidence is provided by Bonney and colleagues’ April 2026 analysis of the Census Bureau’s AI supplement. During the November 2025–January 2026 observation window, 18% of surveyed firms reported using AI in at least one business function, with usage concentrated in a limited number of functions and tasks; employment reductions associated with AI use were uncommon in the reported period.¹¹² Those findings demonstrate that AI diffusion is uneven and that broad enterprise integration cannot be inferred from the availability of advanced models. They are not a direct measure of how many firms operate persistent agents, and their estimates should not be conflated with surveys using broader definitions of AI use.

There is also reason to be cautious about macroeconomic extrapolation. Acemoglu’s analysis of the macroeconomics of AI shows how aggregate productivity effects depend on the fraction of economically relevant tasks affected and the magnitude of task-level cost savings; his bounded estimates under specific assumptions are much more modest than some forecasts derived from general model capabilities.¹¹³ Although the eventual effects of agentic organizations could differ materially from those modeled assumptions, the methodological warning remains applicable: local technical feasibility does not automatically translate into economy-wide productivity or institutional transformation.

None of these studies directly tests the proposition that a shared Company Brain combined with autonomous agents changes the efficient boundary of the firm. Current surveys usually measure AI use, investment, skills, or selected performance outcomes. They rarely distinguish transient assistance from persistent agents, or conventional enterprise integration from institutionally governed shared cognition. More importantly, they generally do not measure the complete costs of autonomous operation, including assurance, governance, exception handling, recovery, and the preservation of organizational knowledge.

As of November 2026, the economic case for cognitive enterprises is therefore theoretically consequential and supported by several adjacent bodies of evidence, but not empirically established as a general transformation of organizational form.

Measuring the real economic value of a Company Brain

A research program capable of evaluating the theory must measure more than employee productivity, agent utilization, or the number of automated tasks.

The primary economic unit should be a completed business process delivering an independently verified outcome. The relevant cost must include direct labor, computational resources, software and infrastructure, knowledge maintenance, process integration, governance, verification, exception handling, and losses from unsuccessful execution. Where the process crosses organizational boundaries, outsourced costs must remain visible.

The second measurement concerns decision-relevant knowledge reuse. A Company Brain should be evaluated according to whether previously acquired knowledge changes subsequent decisions, improves their quality, or reduces the cost of reaching an equivalent result. It is insufficient to count document retrievals or responses generated by an AI system. A meaningful estimate requires an appropriate counterfactual: what would have happened if the decision-maker lacked the shared institutional representation but possessed the same basic applications and general-purpose model?

The third concerns organizational coordination. If a Company Brain actually changes the economics of the firm, its benefits should be visible in reduced delays between dependent activities, fewer contradictory commitments, lower reconciliation costs, improved exception resolution, or more effective allocation of specialized expertise. Gains must be distinguished from conventional workflow automation and enterprise application integration.

The fourth concerns organizational resilience. Lower routine costs may be offset by an increased frequency or severity of common-mode failures. An enterprise that completes thousands of ordinary transactions automatically but cannot recover from an unexpected supplier disruption may have lower average processing costs and higher economic fragility. Evaluating autonomy therefore requires stress testing, measures of recovery capability, and analysis of rare but consequential failures.

Finally, the measurement must consider institutional flexibility. An architecture that becomes increasingly economical to operate but prohibitively expensive to modify may generate short-term productivity gains while diminishing the organization’s capacity to adapt. Relevant indicators include the cost of changing decision rules, replacing models or providers, migrating knowledge representations, restructuring processes, and introducing new organizational capabilities without losing historical evidence and control.

Together, these measurements would make it possible to estimate whether a Company Brain constitutes a distinct productive capability or merely repackages benefits already obtainable through existing enterprise software, data integration, and AI assistance.

The deeper economic consequence: the boundaries of knowledge and the boundaries of ownership diverge

The analysis leads to a conclusion extending beyond firm size, because historically organizational boundaries have partly reflected the difficulty of coordinating knowledge and activity across independent actors. Specialized knowledge could be expensive to transfer; complex processes required established routines; and internal authority sometimes provided a more efficient means of coordinating interdependent resources than market contracting. The Company Brain and autonomous agents could reduce some of these frictions by making organizational knowledge more persistent, interfaces more explicit, and coordination less dependent on repeated human communication.

Yet the same technologies may increase the value of proprietary institutional knowledge, create new forms of asset specificity, and reinforce the advantages of organizations capable of integrating complex activities. They may facilitate decentralized production while concentrating control over computational infrastructure. They may reduce internal employment while increasing dependence on external providers. They may democratize access to cognitive capabilities without democratizing ownership of the productive assets into which organizational knowledge is transformed.

The resulting economic landscape is therefore not captured by a simple opposition between hierarchy and market. It is better understood as a changing relationship among ownership, institutional knowledge, operational coordination, and delegated authority. These dimensions may become increasingly separable in technical implementation even while remaining tightly connected in economic consequences.

An enterprise might retain ownership of its intellectual property and customer relationships while delegating most operations to agents. Another might preserve a large human workforce while integrating its accumulated knowledge into a shared cognitive system. A third might orchestrate a network of independent suppliers through interoperable institutional interfaces. Each arrangement could be economically efficient under different conditions, and each would expose different forms of risk, dependence, and bargaining power.

This is why the economics of a cognitive enterprise cannot be reduced to the price of intelligence. The price of generating an answer may fall dramatically while the value of knowing whether the answer is applicable, whether the resulting action is permitted, and whether the outcome can be independently verified remains high. Organizational knowledge may become easier to access without becoming easier to interpret. Coordinating activities may become cheaper without making the consequences of coordination less uncertain.

The decisive economic variable is therefore not the cost of cognition in isolation, but the cost of transforming cognition into reliable, coordinated, and legitimate economic action.

That proposition also clarifies the relationship between the two extreme firms. The employee-free enterprise seeks to minimize the human labor required for organizational execution; the human–agent enterprise seeks to increase the productive value of human expertise through shared organizational cognition. Their economics may converge where the same infrastructure reduces coordination and knowledge-reconstruction costs, but diverge where human participation contributes distinctive expertise, institutional legitimacy, relationships, or adaptive capacity.

The firm of the future may consequently be smaller in employment, larger in productive reach, more dependent on accumulated cognitive capital, and more permeable to external providers—without necessarily becoming smaller in its economic footprint or less concentrated in its ownership.

The most radical possibility is not that intelligent machines eliminate the economic rationale for organizations. It is that the organization becomes a more important productive technology in its own right: an architecture that determines how human and artificial intelligence, institutional knowledge, authority, and resources are combined into economic outcomes.

Whether that technology favors entrepreneurial decentralization, corporate concentration, or federated economic networks remains an empirical question. The answer will depend on which forms of organizational cognition create genuine economies, how their costs and risks evolve, and who controls the accumulated knowledge and institutional mechanisms through which intelligence becomes productive.

One final tension follows from this analysis and leads directly to the next section. If a Company Brain creates economic value by making organizational cognition more integrated, the same integration may also create systemic dependence on shared representations, assumptions, and decision procedures. An institution may become more efficient at coordinating its activities while becoming less capable of recognizing when the common interpretation guiding those activities is wrong.

The economics of integration must therefore be examined alongside its epistemic consequences: the possibility that making an organization think more coherently may also make it fail more coherently.

Collective cognition and the paradox of a common mind

The economic argument developed in the preceding section leads to a problem that may ultimately determine whether the cognitive enterprise becomes a more intelligent organizational form or merely a more efficient mechanism for reproducing mistakes. If the Company Brain creates value by making organizational knowledge persistent, accessible, and reusable across decisions, its benefits depend on the integration of representations that were previously dispersed among individuals, departments, applications, and institutions. Yet integration introduces a fundamental tension. The same mechanisms that reduce informational fragmentation can also reduce the independence of judgment, amplify erroneous assumptions, and establish a common interpretation of reality that becomes increasingly difficult to challenge.

A company in which every employee and artificial agent consults a shared cognitive infrastructure may coordinate its activities more effectively than an organization dependent on disconnected information systems and informal communication. But it may also become more vulnerable to systemic epistemic failure: an incorrect representation can influence many decisions simultaneously; a mistaken causal explanation can acquire institutional authority through repeated reuse; and a conclusion generated by one agent can return through another as apparently independent corroboration. The organization may become more coherent without becoming more correct.

This is not merely a technical risk associated with unreliable language models. It is a problem at the intersection of distributed cognition, social epistemology, organizational learning, software reliability, and corporate governance. Human organizations have always confronted the tension between shared understanding and independent judgment. What changes when cognition becomes computationally integrated is the potential speed, scale, persistence, and operational reach of that tension. A mistaken belief may cease to be confined to the individuals who hold it and become embedded in the infrastructure through which the organization interprets subsequent evidence and authorizes action.

The central paradox is that the Company Brain may increase the productive value of collective knowledge precisely by introducing mechanisms that can reduce the epistemic independence on which reliable collective intelligence depends. The critical design challenge is therefore not simply to make organizational knowledge shared, but to make it collectively useful without making organizational judgment epistemically uniform.

Collective intelligence is not the accumulation of individual intelligence

The notion of an organization possessing cognitive capabilities independent of any particular member has a substantial theoretical foundation. Hutchins’s analysis of navigation demonstrated that cognitive processes can be distributed across people, instruments, representations, procedures, and coordinated practices.¹¹⁴ A ship’s navigational capability cannot be adequately understood by examining the isolated reasoning of one participant, because the relevant computations are performed through a system of interacting actors and artifacts. The unit of analysis is the organized activity through which information is transformed into coordinated action.

The same principle applies to enterprises. A company designs products, manages supply chains, evaluates risks, and allocates resources through interactions among specialists, documents, software, measurements, routines, and institutional rules. No individual needs to possess the complete knowledge required for these activities. Indeed, the economic rationale for specialization depends partly on the possibility of coordinating knowledge that remains distributed.

The distinction between individual and collective capability also has empirical support. Woolley and colleagues investigated the performance of small human groups across different cognitive tasks and identified a statistical factor associated with collective performance that was not strongly explained by the maximum or average individual intelligence of group members. In their experimental settings, social sensitivity and more balanced participation in conversation were associated with this factor.¹¹⁵ The results do not establish a universal organizational intelligence quotient, nor can they be extrapolated directly from small human groups to computational enterprises. They nevertheless challenge the assumption that collective performance can be predicted by aggregating the capabilities of individual participants.

An organization of exceptionally capable agents may perform poorly if its members cannot identify dependencies, exchange relevant context, resolve conflicting commitments, or recognize common errors. Conversely, an organization containing less capable individual participants may achieve superior outcomes through more effective coordination, specialization, and complementary reasoning.

Cohen and Levinthal’s concept of absorptive capacity supplies another important connection. They describe the organizational capacity to recognize the value of external information, assimilate it, and apply it to commercial ends, emphasizing the role of prior related knowledge and the diversity of expertise within the organization.¹¹⁶ The ability to access information is therefore different from the ability to understand its significance and integrate it into productive activity.

A Company Brain could increase organizational absorptive capacity by making prior experience and relevant expertise more discoverable, preserving connections among otherwise isolated knowledge domains, and helping participants interpret unfamiliar information through existing organizational context. However, it could also reproduce and reinforce the interpretive assumptions already embedded in that context.

This distinction suggests that collective cognition requires more than shared memory and computational reasoning. It requires mechanisms through which differently situated actors contribute observations, expose assumptions, evaluate interpretations, and coordinate consequential decisions. Its quality depends on both the integration of knowledge and the preservation of the differences that make integration informative.

The proposed Company Brain should therefore not be understood as a device that combines many partial minds into one complete mind. Its stronger interpretation is as an institutional infrastructure for coordinating distributed cognition while preserving the provenance, limitations, and contestability of its constituent knowledge.

The difference between collective knowledge and collective agreement

Organizations frequently seek agreement because coordinated action requires participants to adopt compatible plans, responsibilities, and commitments. A production schedule cannot be executed if procurement, manufacturing, logistics, and sales operate under mutually incompatible assumptions. Common business definitions and authoritative operational records are consequently necessary for many forms of coordination.

Yet the need for coordinated action does not imply that every participant must share the same interpretation of uncertain evidence. An organization may legitimately adopt one operational plan while preserving several competing explanations of the circumstances that motivated it. It may designate an authoritative inventory record without claiming that the record is infallible. It may assign a decision to a responsible manager while preserving the objections and alternative hypotheses raised by specialists.

The distinction between agreement required for action and agreement about what is true becomes especially important when the Company Brain supplies a common interpretive interface.

Research on human group decision-making reveals why consensus can be misleading. Stasser and Titus demonstrated that groups may fail to combine information effectively when some facts are shared among participants while others are known only to particular individuals. In their experimental setting, discussion tended to reinforce initially preferred alternatives rather than reliably integrate the information needed to identify the better choice.¹¹⁷ The result illustrates a structural problem: information that is widely shared may receive disproportionate attention even when less widely distributed evidence is more important.

The Company Brain could reduce this problem by making previously isolated information discoverable. But it could also transform the same mechanism into an institutional-scale failure if its retrieval, ranking, or summarization processes systematically prioritize already established interpretations.

Social influence introduces a related difficulty. Lorenz and colleagues experimentally demonstrated that even relatively mild information about others’ estimates could reduce the diversity of judgments and undermine the accuracy benefits associated with aggregating independent estimates in the tasks they studied.¹¹⁸ Participants became more similar in their responses without necessarily becoming more accurate.

A cognitive enterprise may generate an analogous effect when employees and agents repeatedly consult the same model-generated interpretation before forming their own judgments. Their subsequent agreement may reflect the common influence of the system rather than independent evaluation of the underlying evidence.

The economics of informational cascades provides a further explanation. Bikhchandani, Hirshleifer, and Welch showed theoretically how individuals who observe the behavior of others may rationally follow an emerging pattern even when their private information points elsewhere.¹¹⁹ Once enough participants rely on earlier decisions, collective behavior can become detached from the private evidence initially available to them.

In a Company Brain, this mechanism could operate through computational as well as social pathways. One agent recommends a supplier based on an incomplete evaluation. A second agent incorporates the recommendation into a procurement plan. A third agent interprets the plan as evidence that the supplier has already been validated. A fourth agent retrieves the accumulated record and identifies an apparent pattern of institutional preference. If the transformations are not distinguished from independent observations, the organization can construct increasingly convincing support for a conclusion whose evidentiary foundation has not improved.

The resulting failure is more subtle than an ordinary hallucination. An isolated model error may be detected and corrected. An error that becomes part of a shared institutional representation can influence subsequent decisions while acquiring an appearance of credibility through repetition and organizational adoption.

The epistemic danger is not that the Company Brain produces an incorrect answer. It is that the organization gradually loses the ability to distinguish independent evidence from the accumulated consequences of its own earlier interpretations.

Epistemic monoculture and the mathematics of correlated errors

The risk of a common cognitive infrastructure can be analyzed through the dependence among judgments: suppose several agents evaluate the same uncertain quantity. Their outputs may appear independent because the agents have different roles, prompts, or model implementations. Yet their estimates may depend on the same underlying records, semantic definitions, historical examples, or generated summaries. Different computational processes can therefore produce correlated errors even when they execute separately.

A simple statistical model makes the consequence explicit by assuming that n estimators produce unbiased estimates of the same quantity, each with error variance \sigma^2, and that every pair of estimation errors has the same correlation coefficient \rho. For their equally weighted average \bar X, the variance is

\operatorname{Var}(\bar X) = \frac{\sigma^2}{n} \left[1+(n-1)\rho\right]. \tag{7}

When the errors are independent, \rho=0, and the variance decreases proportionally to 1/n. When they exhibit positive correlation, the benefit of aggregation is smaller. For a fixed positive \rho, increasing the number of estimators cannot reduce the variance below the limiting value \rho\sigma^2.

Under these assumptions, an illustrative effective number of independent estimators can be defined as

n_{\mathrm{eff}} = \frac{n}{1+(n-1)\rho}. \tag{8}

For ten equally precise estimators with pairwise error correlation of 0.2, the variance of their average is 0.28\sigma^2, compared with 0.10\sigma^2 under independence. Their effective number of independent estimators is approximately 3.57 rather than ten.

These equations describe an idealized model of numerical estimation. Actual agent judgments may be biased, heterogeneous, categorical, or dependent through mechanisms that cannot be represented by a single correlation coefficient. The equations should not be used as a ready-made metric for organizational intelligence. Their value is conceptual: multiplying the number of judgments does not necessarily multiply the amount of independent evidence supporting them.

The problem has a direct precedent in software reliability. Knight and Leveson investigated the independence assumption underlying multiversion programming, in which independently developed implementations are executed to provide fault tolerance. Their experiment, involving 27 software versions and one million test inputs, found coincident failures occurring more frequently than independence assumptions would predict.¹²⁰ The result does not imply that implementation diversity is useless. It establishes that separate development does not automatically guarantee independent failure behavior.

The analogy with organizational agents is consequential. Several agents may use different models yet share the same erroneous specification. They may have different prompts yet retrieve the same contaminated document. They may operate in separate environments yet depend on a common authorization rule or semantic mapping. Their apparent agreement provides little additional assurance when the underlying cause of error is shared.

A cognitive enterprise therefore needs to distinguish implementation diversity from epistemic independence. Different models can produce similar errors; the same model can sometimes reason from genuinely different evidence; and independent verification may be more valuable than adding another agent to an existing consensus.

The relevant analytical object is therefore the dependency structure of the evidence and reasoning supporting a decision: a Company Brain should be able to determine whether multiple claims ultimately derive from one source, whether independent observations support them, and whether the transformations applied to those observations introduce common assumptions. What appears to be five corroborating assessments may be five descendants of a single unverified assertion; the number of agreeing actors must therefore not be confused with the number or strength of independent grounds for belief.

From a knowledge graph to a graph of epistemic dependencies

This observation suggests a further development of the Company Brain architecture: an ordinary enterprise knowledge graph can represent relationships among suppliers, products, processes, assets, contracts, and organizational actors. A provenance graph can record which entities, activities, and agents participated in producing information. The W3C PROV-O ontology supplies an interoperable vocabulary for expressing such relationships.¹²¹ However, establishing provenance is not equivalent to establishing truth, and documenting a computational transformation does not prove that its conclusion is justified.

A more demanding Company Brain would maintain an explicit relationship among claims, underlying observations, supporting assumptions, inferential transformations, competing evidence, temporal applicability, and institutional acceptance.

This is closely related to an earlier tradition of artificial intelligence that predates modern language models. Jon Doyle’s work on truth-maintenance systems addressed the need for reasoning programs to preserve the justifications supporting their beliefs and revise those beliefs when assumptions change or contradictions emerge.¹²² The central insight is that a reasoning system cannot be adequately characterized by its current conclusions alone. It must also maintain information about why those conclusions are held and which other conclusions depend on them.

Applied to organizational cognition, this idea becomes particularly powerful: consider an enterprise that believes a supplier can deliver a critical component within ten days. That conclusion may depend on a contractual lead-time clause, a recent supplier message, historical delivery performance, and an assumption that an identified transport route remains available. A planning agent uses the conclusion to generate a production schedule; a commercial agent uses the schedule to confirm customer delivery dates; a financial agent incorporates the expected revenue into a cash-flow projection.

If the transport route becomes unavailable, updating a document repository or recording the new event is insufficient. The organization must identify which earlier conclusions depended on the invalidated assumption, determine which commitments may have been affected, and initiate the appropriate reassessment.

A credible epistemic infrastructure would preserve the dependency chain rather than requiring each actor to rediscover it independently, while also distinguishing the status of the propositions it represents. An independently observed event, a statement made by a supplier, a forecast generated by a model, an inference accepted by a specialist, and a policy approved by a competent authority belong to different epistemic and institutional categories. Their evidentiary significance cannot be determined by textual similarity or retrieval rank.

In particular, institutional acceptance does not convert a proposition into an empirical fact. A company can authorize a decision based on the best information available while recognizing that the underlying evidence remains uncertain. A policy can be validly adopted while its anticipated effects remain unproven. A historical outcome can be accurately recorded without its causes being identified.

The Company Brain should consequently preserve at least two different relationships: epistemic support, describing why a proposition is considered plausible or established within defined limits, and institutional authority, describing why a particular decision or action is permitted or binding.

The distinction is a continuation of the four-property theory: organizational cognition concerns what can be justified, organizational agency what can be initiated, autonomy the conditions under which action proceeds without additional intervention, and accountability the legitimate allocation of authority and consequences. A sophisticated cognitive enterprise must connect these relationships without confusing them.

The recursive danger: when institutional memory learns from itself

The risk of epistemic monoculture becomes more severe when the Company Brain does not merely retrieve historical knowledge but continuously incorporates the outputs of its own agents.

Imagine an agent that produces a concise explanation of a supplier failure. The explanation is plausible but relies partly on an unverified causal assumption. Another agent later retrieves the summary as evidence of the supplier’s historical behavior. A third agent incorporates the reconstructed history into a risk assessment. Eventually, the organization may treat the original conjecture as an established pattern.

No external observation has necessarily strengthened the initial conclusion. The apparent accumulation of knowledge consists partly of repeated transformations of the same uncertain interpretation.

This is a form of epistemic feedback contamination: information generated by the organization’s own reasoning mechanisms re-enters its evidence environment without sufficiently preserving the distinction between the original observations and subsequent inferences.

An adjacent problem has been demonstrated in research on generative-model training. Shumailov and colleagues investigated the consequences of training generative models recursively on model-generated data, showing that certain training regimes can progressively lose information about the original data distribution.¹²³ Their results concern statistical learning under specified conditions; they do not prove that an enterprise retrieval system will experience the same form of model collapse. The organizational analogy is nevertheless informative because both situations involve the possibility that generated representations displace or distort the evidence from which they were originally derived.

In a Company Brain, the danger is often not that the language model’s parameters change. It is that the institutional knowledge base changes in ways that gradually elevate generated interpretations into apparently authoritative records.

This distinction suggests that an enterprise must manage several different kinds of memory. Primary operational evidence should retain its original source and status. Agent-generated interpretations should be recorded as interpretations, with their supporting evidence and relevant assumptions. Approved decisions should preserve their institutional authority and justification. Observed outcomes should be linked to the interventions that preceded them without automatically being attributed causally to those interventions.

A subsequent correction must not silently erase the fact that an earlier decision was made under a different interpretation. Otherwise, the organization loses the ability to reconstruct whether its previous conduct was reasonable given the information available at the time.

The appropriate principle is epistemic continuity without epistemic self-confirmation. The Company Brain should preserve the history of its own judgments while preventing the mere persistence and reuse of a judgment from becoming additional evidence of its truth.

This principle also has cybersecurity implications. An attacker need not directly instruct an agent to violate a rule if the attacker can contaminate a trusted representation on which many authorized agents rely. A forged supplier notification, manipulated maintenance record, or misleading contractual interpretation may propagate through legitimate processes while appearing to originate from the organization’s own accepted knowledge.

The integrity of organizational cognition therefore depends not only on protecting models and APIs, but on maintaining trust boundaries around the production, transformation, promotion, and reuse of evidence.

Organizational learning requires the ability to revise assumptions

The distinction between accumulating experience and learning from experience is central to the cognitive enterprise.

An organization may retain every transaction and still repeatedly make the same mistake if its operating assumptions remain unchanged. It may update predictions after an unexpected outcome without revising the decision rule responsible for the failure. Or it may revise a process while failing to recognize that the objectives governing the process have become inappropriate.

Argyris and Schön’s theory of organizational learning offers a useful conceptual framework.¹²⁴ Their distinction between single-loop and double-loop learning concerns whether an organization merely corrects deviations within existing governing assumptions or also examines and modifies the assumptions and norms themselves.

The distinction maps directly onto the executable institutional model developed earlier: a production-planning agent that discovers a delayed shipment and adjusts the schedule performs a form of operational adaptation within an existing mandate. If repeated disruptions reveal that the planning procedure relies on an invalid assumption about supplier reliability, the organization may need to revise its decision rules. If the broader procurement strategy systematically sacrifices resilience for short-term cost minimization, the underlying policy objectives themselves may require reconsideration; these are different levels of learning because they involve different objects of change and different institutional authorities.

An agent may be authorized to revise a forecast automatically. It may be permitted to recommend changes to a replenishment rule. It should not infer from the failure of a forecast that it possesses authority to alter corporate risk appetite or contractual priorities.

The organizational learning system must therefore distinguish epistemic revision, operational adaptation, and constitutional change. The first changes what the organization believes; the second changes what it does within existing rules; the third changes the rules under which its decisions and actions become legitimate, making the consequences of organizational learning governable without requiring that learning cease.

The deeper challenge is causal attribution. When a business outcome improves after an intervention, the improvement may have resulted from that intervention, an unrelated environmental change, the interaction of several measures, or ordinary variation. Recording the sequence of events does not identify their causal relationship. Organizational learning that updates operating rules solely because a favorable outcome followed an action risks confusing correlation with intervention effectiveness.

The Company Brain should therefore distinguish observed associations from causal hypotheses and should support controlled experimentation, comparison with relevant counterfactuals, and independent review where the stakes justify them. Some hypotheses will remain uncertain because the organization lacks the observations or experimental opportunities needed to identify their effects; a mature cognitive enterprise must therefore be capable of recognizing that it does not yet know why something happened, which distinguishes reliable organizational learning from the accumulation of increasingly polished explanations.

An epistemic separation of powers

The preceding analysis suggests a governance principle that complements the executable institutional constitution: the cognitive enterprise requires an epistemic separation of powers.

The expression is proposed here as an architectural and institutional analogy, not as an existing legal doctrine or a claim that corporate cognition must reproduce constitutional government.

The principle concerns the concentration of control over four activities that should remain distinguishable: the production of evidence, the interpretation of evidence, the authorization of consequential action, and the verification of outcomes.

When one agent can generate an interpretation, approve it as authoritative organizational knowledge, execute the recommended action, and evaluate its own success without independent constraints, the organization has created a closed epistemic and operational loop. Such an arrangement may be efficient for low-risk activities, but it creates a serious assurance problem when mistakes can propagate into significant commitments or irreversible consequences.

The Company Brain should therefore separate the provenance of observations from the authority to accept interpretations, and both from the permission to execute actions. Independent outcome verification should supply evidence that cannot be manufactured merely by repeating the original agent’s explanation of what it intended to achieve.

Independence need not require a human reviewer at every step. Deterministic reconciliation, independent sensors, external confirmations, different authoritative systems, and separately governed validation services can provide valuable checks. What matters is whether the verification mechanism has sufficient independence from the assumptions and mechanisms that produced the proposed decision.

In a supply-chain process, for example, an agent might infer that an order can be fulfilled because the ERP reports adequate inventory. A second language-model agent consulting the same ERP record does not provide strong independent assurance. A current warehouse observation, an independently maintained inventory reconciliation, or a verified reservation record may add genuinely new evidence. These mechanisms need not all agree; disagreement may reveal that the underlying definition of available inventory is incomplete.

This requires an organizational capacity to preserve contradictions rather than immediately eliminate them. If an inventory system reports one quantity and a warehouse inspection another, the Company Brain should not automatically replace the discrepancy with a single compromise value. It should represent the difference, identify its possible causes, determine the consequences for pending decisions, and apply the appropriate institutional procedure for reconciliation.

The same principle applies to human expertise. A maintenance engineer may challenge a system-generated diagnosis using observations unavailable to the model. A financial controller may recognize that a technically valid transaction violates an economic assumption. An employee may identify a recurring operational exception that a standardized process has systematically classified incorrectly.

Their contributions must remain discoverable and capable of influencing institutional judgment, even when they conflict with the prevailing computational interpretation.

The epistemic separation of powers is consequently not only an information-security architecture. It is also a theory of how a cognitive organization preserves its ability to correct itself.

Epistemic diversity as an organizational resource

An organization benefits from shared representations because they reduce the cost of coordination. It benefits from cognitive diversity because different participants may notice different features of a problem, possess different evidence, and employ different models of its causes.

These advantages are complementary rather than contradictory, as illustrated by Hong and Page’s mathematical work on diverse problem-solving groups established results under specified assumptions in which sufficiently diverse problem-solvers can outperform groups selected for the highest individual performance.¹²⁵ Their findings should not be generalized into the unconditional claim that diversity always improves decisions. The result depends on the problem structure, the characteristics of the participants, and how their capabilities are combined. Its relevance here is that differences among problem-solving approaches can have productive value beyond differences in individual ability.

For the cognitive enterprise, this suggests that diversity should not be measured merely by counting different models, vendors, departments, or professional titles. The relevant question is whether these actors contribute distinct evidence, interpretive frameworks, error-detection mechanisms, or causal hypotheses.

Two agents using different foundation models but the same flawed knowledge graph may add little epistemic diversity. Two specialists who use different observations to examine the same operational problem may contribute substantial diversity even if they work within the same department. An independent physical measurement may provide more epistemic value than several elaborate interpretations of an outdated database record.

The Company Brain could help preserve this diversity by making local evidence accessible across organizational boundaries while retaining its context and provenance. But doing so requires resisting the temptation to force every domain into one universal interpretation.

Some concepts legitimately differ across professional contexts. A commercial commitment, a contractual obligation, a technical feasibility assessment, and an accounting recognition rule may concern the same transaction without representing identical propositions. Their differences are not necessarily semantic defects to be removed. They may reflect distinct institutional purposes and constraints.

A shared cognitive infrastructure should therefore maintain a coherent relationship among perspectives without demanding their complete conceptual uniformity.

This also introduces a political dimension. Miranda Fricker’s work on epistemic injustice analyzes how individuals may be wronged in their capacity as knowers, including through unjust deficits of credibility or through limitations in shared interpretive resources.¹²⁶ The theory is not a technical design standard, but it identifies a risk that becomes relevant when an enterprise’s knowledge infrastructure determines which observations are considered credible and which interpretations are made visible.

An organization may give every employee access to a Company Brain while systematically privileging managerial records over operational testimony, formal documents over situated experience, or model-generated explanations over objections raised by specialists. In that case, universal technical access would coexist with unequal epistemic authority.

A human-centered cognitive enterprise must therefore provide mechanisms not only for retrieving organizational knowledge, but for challenging how knowledge is classified, whose contributions are recognized, and how unresolved disagreement affects decisions.

The employee-free enterprise faces an analogous but structurally different problem. It may lack the heterogeneous human interpretations that could reveal flaws in its common cognitive infrastructure, while shareholders and external providers remain too distant from routine operations to recognize emerging epistemic failures. This does not prove that human participation is always superior to computational diversity. It shows that an agent-only organization must deliberately construct effective mechanisms for independent observation, critical evaluation, and external challenge.

The relevant design goal is not maximal agreement. It is the ability to combine shared institutional knowledge with sufficient independence to discover when that knowledge is wrong.

A concrete architecture for epistemic resilience

These theoretical requirements can be translated into architectural capabilities complementing the Company Brain and agent runtime, beginning with an evidence provenance and dependency structure. Material organizational claims should be traceable to source observations, transformations, assumptions, and reviews. The system should distinguish multiple independent evidentiary origins from multiple summaries of the same origin. For consequential decisions, the relevant dependency structure should remain reconstructible after the underlying models, applications, or participants have changed.

The second is a controlled process for promoting claims into authoritative organizational knowledge. Not every agent-generated explanation should become an accepted institutional fact merely because it was useful in a previous task. The organization should distinguish provisional hypotheses, accepted observations, formally approved interpretations, and superseded conclusions. Promotion rules must be sensitive to the domain and consequences of the claim: an automatically extracted product description does not require the same evidentiary standard as a safety-critical operating assumption or a contractual interpretation.

The third is an independent validation capability. Consequential assertions should be checked, where feasible, against independently generated evidence or controls whose failure modes differ materially from those of the original reasoning process. Independence must be assessed at the level of relevant dependencies, not inferred from the use of separate software products.

The fourth is a mechanism for institutional disagreement. Human specialists and computational actors should be able to register contrary evidence, identify disputed assumptions, request reassessment, and preserve alternative interpretations. The system must distinguish challenges to factual claims from challenges to policies or institutional authority. An employee may demonstrate that a forecast is incorrect without acquiring the authority to replace the associated commercial commitment.

The fifth is dependency-aware belief revision. When a material assumption is invalidated, the Company Brain should identify dependent conclusions and the processes that may have acted upon them. A correction should propagate through affected representations while preserving the history of what was believed at the time of earlier decisions. Where previously executed actions cannot be reversed, the organization must still be able to recognize their consequences and initiate the appropriate corrective process.

The sixth is separation between epistemic updates and changes to executable institutional rules. An updated belief about supplier reliability may justify a new risk assessment, but it must not silently alter procurement authority, contractual obligations, or corporate risk limits. The change-control mechanism governing institutional rules remains distinct from the mechanism maintaining organizational evidence.

These capabilities are not established by existing enterprise platforms merely because they implement knowledge graphs, audit logs, or retrieval-augmented generation. They require explicit integration among semantic representations, provenance, reasoning, process execution, policy enforcement, and organizational governance.

Nor do they guarantee that an enterprise will avoid collective error. Source observations can be misleading; independent validators can share unknown dependencies; human reviewers can defer excessively to automated recommendations; and institutional authorities can knowingly prefer expedient interpretations over well-supported ones.

The purpose of epistemic resilience is therefore not to eliminate uncertainty or disagreement. It is to make the organization’s response to uncertainty and disagreement more reliable, visible, and correctable.

Measuring whether the Company Brain improves collective cognition

The proposed theory creates an empirical question that cannot be resolved through conventional model benchmarks. If a Company Brain improves collective cognition, the benefits should appear in organizational performance under conditions requiring the integration of distributed knowledge. Faster retrieval and more fluent explanations are insufficient evidence. The organization must demonstrate improvements in the quality of decisions, coordination of dependent activities, detection of errors, and revision of mistaken assumptions.

A useful experimental design would vary the degree of shared cognitive integration independently of the availability of agent capabilities. One group of comparable organizational units might use conventional enterprise applications and human coordination. A second could receive individual AI assistants with access to relevant local information. A third could use coordinated agents sharing an integrated Company Brain. A fourth could use the same integrated cognitive infrastructure but with additional mechanisms for independent evidence verification, provenance-based dependency analysis, and institutional challenge.

The central comparison would not be whether the third or fourth groups produce more answers. It would be whether they make better decisions, especially when information is fragmented, observations conflict, or the initial shared representation contains a deliberately introduced error.

Such experiments should distinguish several outcomes. Decision accuracy concerns whether the selected action is appropriate under independently established conditions. Epistemic calibration concerns whether the confidence attached to a proposition corresponds to its observed reliability under a defined evaluation protocol. Error propagation concerns how many subsequent decisions and processes are affected by an incorrect assertion. Detection latency measures how long the organization takes to recognize and investigate a material contradiction. Correction effectiveness concerns whether invalidated assumptions lead to appropriate revisions of dependent conclusions and operating rules.

Two additional measures are particularly important:

1. Independent evidentiary support: the extent to which consequential decisions depend on distinct, appropriately validated sources rather than repeated transformations of the same evidence. This requires tracing dependency structures rather than merely counting supporting documents or agreeing agents.
2. Institutional epistemic resilience: the organization’s capacity to maintain or recover justified decision-making after a material source, model, or shared representation becomes unreliable. It can be evaluated through controlled perturbations, injected contradictions, simulated source corruption, and recovery exercises, subject to appropriate safeguards.

These measures should be related to business outcomes. A system that detects many contradictions but prevents timely action may impose unacceptable costs; a system that resolves contradictions quickly by suppressing minority evidence may appear efficient while becoming less reliable. The objective is not to maximize disagreement or verification overhead, but to maintain an economically and institutionally appropriate balance between coordinated action and independent scrutiny.

A particularly demanding test would examine whether organizational error increases as cognitive integration becomes more extensive. If additional agents and departments begin using the same Company Brain, does the organization detect problems more quickly because evidence is connected, or does it spread mistaken assumptions more widely because interpretations are shared? Both results are plausible. The proposed theory predicts that the outcome will depend substantially on whether integration is accompanied by mechanisms preserving evidentiary independence, explicit uncertainty, and effective belief revision.

This proposition is falsifiable. If a highly integrated Company Brain with independent validation provides no improvement over conventional information systems after controlling for integration and implementation costs, the hypothesized value of epistemically resilient organizational cognition would be weakened. Conversely, if it reduces cross-functional errors, improves exception handling, and preserves decision quality under shared-source failures, it would provide evidence for a distinctive organizational capability beyond individual AI assistance.

The final paradox: an organization must share knowledge without surrendering the capacity to question it

The collective-cognition problem ultimately returns to the nature of the firm as an institution.

An enterprise is not merely an optimizing controller pursuing a fixed objective. Its activities are governed by legal rights, contractual commitments, professional responsibilities, economic incentives, and social relationships. Its purposes may be contested; its operating assumptions may become obsolete; and its interpretation of evidence may reflect institutional interests as well as objective constraints.

A Company Brain can improve the visibility of these relationships, but it cannot legitimately resolve every disagreement by calculating a single optimal answer. Some conflicts concern uncertain facts and may be addressed through better evidence. Others concern legitimate differences in interests, risk preferences, or institutional responsibilities, and require procedures for negotiation, authorization, or adjudication.

This is why the separation between organizational cognition and organizational accountability remains essential. A computational system may identify the consequences of alternative strategies without possessing authority to choose among competing stakeholder interests. It may detect that a policy is inefficient without possessing legitimacy to redefine the organization’s objectives. It may recognize that several professional interpretations conflict without being entitled to suppress the disagreement.

At the same time, the company cannot function if every question remains permanently unresolved. Collective action requires decisions, commitments, and sufficiently stable representations of relevant operational conditions. The appropriate objective is therefore neither unlimited epistemic plurality nor imposed cognitive uniformity, but institutional coherence without epistemic closure.

The organization should be able to act under a legitimate decision while preserving uncertainty about the assumptions on which that decision depends. It should be able to adopt a common operational record while allowing evidence that challenges its accuracy. It should be able to coordinate human and artificial actors through shared processes while retaining independent mechanisms capable of detecting when the entire process rests on an incorrect interpretation.

Such an arrangement would be more than a knowledge repository or a network of communicating agents: by preserving the relationship between what the organization believes, why it believes it, what it does, and what subsequent experience reveals about those beliefs, it would support a stronger conception of organizational learning than record accumulation and a stronger conception of collective intelligence than agreement among participants.

The Company Brain would become valuable not because it creates a single organizational mind, but because it enables the organization to retain and coordinate multiple forms of knowledge without erasing their differences. Its apparent unity would lie in the continuity of evidence, decisions, and institutional action—not in the elimination of competing interpretations.

This distinction also reframes the economic argument. A cognitive infrastructure that merely makes existing organizational assumptions available to more actors can reduce coordination costs while amplifying systemic risk. One that supports independent evidence, justified revision, and accountable adaptation may produce a more durable form of organizational cognitive capital, because it preserves not only knowledge but also the capacity to recognize and correct the limitations of that knowledge.

The most advanced cognitive enterprise may therefore be characterized as much by its ability to discover that it is wrong as by its ability to formulate and execute apparently intelligent decisions.

This capacity is especially important as organizational autonomy increases. In a human–agent company, employees may supply critical alternative interpretations, but only if the institution preserves their ability to challenge computational conclusions. In an employee-free operating company, independent verification and external challenge may need to be designed explicitly into the operating architecture, because the absence of humans from routine processes also removes some of the informal pathways through which unexpected problems become visible.

Neither configuration can rely on the existence of a common cognitive infrastructure as sufficient evidence of collective intelligence. Both require institutional mechanisms for epistemic correction.

The ultimate paradox is thus not that the Company Brain might occasionally produce incorrect conclusions. All human and computational systems are fallible. The deeper problem is that the more extensively an organization coordinates its cognition through a common infrastructure, the more consequential the assumptions embedded in that infrastructure become; the opportunity and the risk arise from the same architectural property, and the defining achievement of a cognitive enterprise would be to exploit integration without extinguishing independent judgment, alternative evidence, or the institutional capacity for self-correction.

A firm that thinks should not be one in which everyone—human or artificial—arrives at the same answer. It should be one in which distributed knowledge can become coordinated action without depriving the organization of the ability to question the knowledge on which that action depends.

The Company Brain, on this interpretation, is not a common mind. It is the infrastructure through which an institution can develop, preserve, challenge, and revise a common understanding while retaining the plurality of evidence and reasoning required for that understanding to remain reliable.

Its epistemic architecture cannot therefore be separated from its governance architecture. If organizational cognition becomes an operational infrastructure, the rules determining who may contribute knowledge, who may validate it, who may challenge it, and how its conclusions influence authorized action become constitutive elements of the firm itself.

That is the point at which the theory of collective cognition leads necessarily to the next question: how can an increasingly autonomous cognitive enterprise remain governable when the integrity of its knowledge and the legitimacy of its actions are inseparably connected?

Governance and cybersecurity as constitutive architecture

The preceding analysis has established that the cognitive enterprise is not simply a company equipped with more capable software. It is a potentially different implementation of the firm, in which institutional knowledge becomes computationally persistent, organizational roles become increasingly machine-interpretable, and artificial agents acquire the ability to initiate and execute activities historically performed by employees. This transformation has consequences that extend well beyond conventional information security. When an agent can commit financial resources, alter enterprise records, negotiate with suppliers, deploy software, administer identities, or influence industrial control systems, the integrity of the computational infrastructure becomes inseparable from the integrity of the institution itself.

The distinction becomes particularly consequential when considering the employee-free enterprise. A conventional company organizes its productive activity through human employment, contractual relationships, delegated authority, and technical systems. Employment creates obligations concerning remuneration, working conditions, occupational safety, social protection, intellectual property, and the allocation of responsibility for employee conduct. Replacing internal human labor with computational agents may change or eliminate some obligations associated with particular employment relationships, but it does not eliminate the company’s legal responsibilities or the economic risks of production. Instead, a larger portion of its operational capacity becomes dependent on software, computational infrastructure, model providers, access-control systems, and the security of the information through which agents perceive and act upon the world.

This is a redistribution of institutional dependencies rather than a straightforward substitution of regulation by technology. Labor law continues to govern employees and relevant external labor relationships; intellectual-property law continues to govern the use, production, ownership, and disclosure of protected material; contractual and corporate law continue to determine valid authority and liability. What changes is the relative importance of the mechanisms through which the enterprise exercises its productive capacity. When human workers perform a business process, their conduct is mediated by professional competence, social relationships, employment obligations, managerial supervision, and organizational controls. When computational agents perform the same process, much of its immediate execution depends on software behavior, machine credentials, networked services, policy enforcement, and the integrity of the knowledge supplied to those agents.

The resulting security problem is qualitatively different from protecting an application that merely stores or processes business information. An attacker who corrupts a corporate database may cause confidentiality, integrity, or availability losses. An attacker who corrupts the evidence or authority structure used by autonomous agents may also induce the organization to undertake apparently legitimate transactions, enter contractual commitments, change operational configurations, or modify its own governing procedures. The attack need not compromise a privileged administrator directly. It may exploit the path through which less-trusted information is transformed into an apparently authoritative organizational decision.

The fundamental proposition is therefore that cybersecurity becomes constitutive of organizational agency and, through that agency, of corporate governance. This does not imply that cybersecurity replaces economic, legal, or ethical governance. It means that the enterprise can exercise its delegated authority reliably only if the technical mechanisms implementing that authority remain trustworthy within explicitly defined limits.

From employment-related risks to the security of computational production

The economic transition from human employees to agent-operated processes creates a distinctive redistribution of risk.

An enterprise employing human personnel must address employment contracts, working-time obligations, professional competence, workplace safety, employee misconduct, confidentiality, intellectual-property assignments, and the consequences of errors or omissions. These obligations are shaped by law, collective arrangements, and the firm’s particular activities. Their costs are not reducible to wages: they form part of the institutional structure governing productive work.

When an enterprise substitutes computational agents for employees, several of these obligations may no longer arise in relation to the particular tasks transferred to software. An artificial agent does not become an employee merely because the enterprise calls it a coworker, assigns it a position, or evaluates its performance. Nevertheless, the enterprise must acquire, operate, supervise, and maintain the computational resources through which the work is performed. It must manage external suppliers, software rights, operational failures, unauthorized commitments, service interruptions, security vulnerabilities, and the allocation of responsibility for resulting losses.

The transformation is therefore better characterized as a movement from labor-intensive institutional coordination toward software-intensive institutional coordination. The latter may offer lower marginal execution costs, but it introduces dependencies whose failure can affect many processes simultaneously and at computational speed.

A compromised employee account might allow an attacker to perform actions within that employee’s access rights. A compromised agent identity may expose a comparable set of permissions, but the agent may execute operations continuously, interact with numerous systems, generate further instructions, delegate work, or accumulate information across organizational domains. The resulting exposure depends on the architecture and actual privileges, not on the intrinsic nature of artificial intelligence. Nevertheless, persistence, execution speed, and automated coordination can substantially change the relationship between an initial compromise and its consequences.

More importantly, an agent need not be conventionally compromised to create a security incident. It may operate with valid credentials, invoke approved tools, and produce technically well-formed transactions while following a manipulated interpretation of its task. A supplier message can be authentic yet contain instructions designed to redirect the agent; a planning recommendation can satisfy local validation rules yet violate an aggregate financial constraint; and an agent can mistakenly treat the output of another agent as authoritative evidence.

This creates a distinction between compromised execution and authorized but institutionally invalid execution. Traditional cybersecurity already addresses misuse of legitimate credentials, insider threats, confused-deputy problems, and business-logic abuse. Agentic systems intensify these issues because the selection and sequencing of technically permitted operations may be generated dynamically from natural-language objectives and uncertain external information.

The economic cost of cybersecurity accordingly becomes part of the cost of delegated productive agency. It includes not only protecting infrastructure against external attackers, but ensuring that software carrying out legitimate work cannot exceed the institutionally intended scope of that work.

This observation connects directly to the earlier theory of organizational cognitive capital. If the Company Brain preserves the representations through which agents interpret organizational conditions, then its confidentiality, integrity, availability, and provenance become properties of a productive asset. A compromise can destroy value not only by disclosing information, but by contaminating the institutional knowledge on which future decisions depend.

The emergence of a new organizational threat model

Conventional enterprise threat modeling begins with assets, adversaries, attack surfaces, vulnerabilities, trust boundaries, and potential consequences. These concepts remain indispensable. However, the boundary of the system under consideration must be expanded when artificial agents can initiate and coordinate business actions.

The protected assets now include not only information and applications, but also organizational mandates, delegated authority, institutional memory, decision procedures, operational commitments, and the mechanisms through which governance rules are changed.

The adversary’s objective may consequently be to change the company’s behavior rather than merely access its information, so a useful threat model should distinguish several classes of adversarial influence.

First, an attacker may compromise the confidentiality or integrity of the Company Brain’s underlying evidence. A manipulated supplier record, falsified inspection result, poisoned document, or corrupted semantic mapping can influence legitimate decision processes without requiring the attacker to control the reasoning model.

Second, an attacker may attempt to redirect the agent’s instructions or goals. This can occur through retrieved documents, emails, websites, tool descriptions, search results, or messages from other agents. The principal vulnerability arises when lower-trust content is treated as if it possessed authority over the agent’s behavior.

Third, the attacker may target the agent’s identity, credentials, tools, or runtime environment. Successful exploitation can convert a manipulated recommendation into direct access to enterprise systems, privileged operations, or sensitive information.

Fourth, the attacker may target the relationships among agents and processes. An individually harmless instruction may cause one agent to create a record that another treats as authorization; several permitted transactions may collectively exceed a limit; and apparently independent agent assessments may depend on a single compromised source.

Finally, an attacker may seek to modify the institutional model governing the agents themselves. Manipulating a role definition, policy configuration, approval workflow, or delegation rule can alter the conditions under which subsequent operations are considered authorized. This constitutes an attack against the organization’s computational implementation of authority.

These mechanisms correspond to several categories in the OWASP Top 10 for Agentic Applications 2026, including agent-goal hijacking, tool misuse, identity and privilege abuse, agentic supply-chain vulnerabilities, unexpected code execution, memory poisoning, insecure inter-agent communication, cascading failures, and rogue-agent behavior.¹²⁷ The taxonomy is useful for identifying attack surfaces and classes of failure, but it does not establish the prevalence of each threat or prove the effectiveness of particular mitigations.

NIST’s 2025 adversarial-machine-learning taxonomy provides a complementary framework distinguishing attacker objectives, capabilities, access conditions, and attacks against different stages of the AI lifecycle.¹²⁸ Combining these approaches with established enterprise threat modeling makes it possible to distinguish adversarial manipulation of information, compromise of software execution, misuse of delegated authority, and failures caused by the interaction of otherwise legitimate components.

The critical analytical point is that the attack surface of a cognitive enterprise extends across the entire path from observation to institutionally authorized action.

Protecting the model endpoint is necessary but insufficient. Protecting the agent’s credentials is necessary but insufficient. Even a properly authenticated and technically authorized action may be unsafe when its evidentiary basis or business context has been manipulated.

What empirical agent-security research actually demonstrates

The security problems of tool-using agents are not merely speculative, although the available evidence must be interpreted within the conditions under which experiments were conducted.

Debenedetti and colleagues introduced AgentDojo, an evaluation framework for agents operating through external tools in environments containing attacker-controlled information. The published benchmark included 97 realistic user tasks and 629 security test cases, examining whether prompt injection embedded in lower-trust material could redirect agent behavior.¹²⁹ The experiments demonstrated that the tested systems faced simultaneous difficulties with legitimate task completion and resistance to adversarial instructions. The distinction matters: an agent’s ability to execute a business workflow and its ability to preserve the user’s authorized objective under hostile inputs must be evaluated separately.

The Agent Security Bench research expanded the scope of evaluation to multiple attack surfaces, including system instructions, tool use, memory retrieval, and agent planning. Its authors reported experiments across ten application scenarios, more than 400 tools, thirteen model backbones, and approximately 90,000 test cases, including prompt-injection and memory-poisoning methods.¹³⁰ The reported results reveal substantial vulnerabilities under particular experimental conditions. Their attack-success rates are properties of the benchmark configurations, however, and should not be treated as estimates of the probability that a deployed enterprise agent will be compromised.

More directly relevant to enterprise integration, MCPTox, published at AAAI 2026, investigated tool poisoning in Model Context Protocol environments. Its researchers constructed 1,348 malicious test cases using 353 tools from 45 real-world MCP servers and evaluated twenty agent configurations. One tested configuration exhibited a reported attack-success rate of 72.8%.¹³¹ This does not establish that 72.8% of operational MCP deployments are vulnerable. It demonstrates that, within a controlled experimental setting, manipulating the information through which tools are presented to agents can induce significant rates of unauthorized behavior.

The methodological significance is more important than the headline percentage. The attacker need not introduce an obviously malicious executable or obtain direct administrative privileges. Tool metadata, responses, and contextual material can influence how an agent selects actions, including actions performed through otherwise legitimate interfaces.

A related 2026 academic study used STRIDE and DREAD to examine security threats across MCP hosts, clients, models, servers, external data sources, and authorization components, including prompt-injection and tool-poisoning scenarios.¹³² Its contribution is to demonstrate the importance of modeling the complete integration architecture rather than treating MCP connectivity as an intrinsically trustworthy abstraction.

These studies support a limited but consequential conclusion: current agent architectures expose a class of adversarial pathways in which untrusted information can influence the selection and execution of privileged operations. They do not prove that every agent will be compromised or that no reliable defenses are possible.

The maturity of the defense architecture is itself an active research problem. In February 2026, NIST’s National Cybersecurity Center of Excellence published a concept paper addressing identity, authentication, authorization, delegation, auditing, and prompt-injection risks for software and AI agents.¹³³ Its significance is that agent identity and authority have become explicit subjects of standards-oriented engineering. The publication is a concept paper, however, not evidence that a universally adopted or formally assured enterprise-agent identity architecture already exists.

Industry developments point in the same direction. Microsoft’s October 2026 discussion of the agentic insider describes agents as enterprise actors whose operational permissions and continuous execution require monitoring and containment independent of the underlying models.¹³⁴ The argument is technically relevant, particularly its distinction between computational intelligence and delegated authority, but vendor descriptions of emerging security products must not be mistaken for independent proof of their effectiveness.

The empirical record therefore supports the need for a distinct agent-security engineering discipline while leaving major questions unresolved concerning whole-system assurance, long-horizon adversarial robustness, and the reliability of controls across heterogeneous enterprise environments.

The Company Brain as a high-value target for epistemic compromise

The previous section established why a shared cognitive infrastructure can amplify common errors. From a cybersecurity perspective, the same structure creates opportunities for deliberate manipulation.

An attacker who compromises a single enterprise document may have limited influence when the document is used only by one employee. The same compromise can have broader consequences when its content is indexed, semantically linked to authoritative business objects, retrieved by multiple agents, and incorporated into their operating context.

The resulting attack is not adequately described as a prompt-injection problem alone, because a malicious document might contain instructions intended to redirect an agent. But an attacker could also modify apparently ordinary business facts, such as supplier capacity, component compatibility, contractual terms, delivery dates, or the status of an approval. If the organization accepts those false assertions as authoritative, agents may execute incorrect decisions without violating their instructions. The resulting threat model should distinguish instruction contamination occurs when lower-trust content attempts to direct computational behavior beyond its legitimate role as data. The security boundary being crossed is between evidence and authority.

Evidence contamination occurs when false, incomplete, or misleading information enters the institutional knowledge base and influences subsequent decisions. The security boundary being crossed is between unverified assertions and accepted organizational knowledge.

The second class can be more difficult to detect because an agent may behave exactly as intended relative to the corrupted representation. No prompt-based safeguard can guarantee that a business fact is accurate. The necessary controls include source authentication, provenance, semantic validation, independent reconciliation, anomaly detection, and procedures for challenging accepted knowledge.

Persistent memory introduces additional complexity. A compromised assertion may remain influential after the original attack surface has disappeared. Summaries, embeddings, generated interpretations, and derived business records may reproduce its effects without retaining a readily visible connection to the original malicious material.

This creates a potential form of persistent organizational compromise in which the attacker no longer needs continuous access to the system. The manipulated knowledge continues to influence authorized behavior.

The Company Brain consequently requires separate controls for evidence ingestion, interpretation, promotion into authoritative knowledge, and subsequent reuse. The ability to retrieve a statement must not be equivalent to permission to treat it as verified; repeated retrieval or agreement among agents must not increase its evidentiary authority without additional support.

The strongest connection to the four-property theory is that the integrity of organizational cognition becomes a prerequisite for the safe exercise of organizational agency. A company cannot secure its autonomous operations merely by protecting the systems that execute them if the evidence determining what should be executed remains vulnerable to manipulation.

Identity and delegated authority: the new privileged-access problem

Identity and access management becomes particularly difficult when organizations contain large and changing populations of artificial actors, because human identities traditionally correspond to employees, contractors, service accounts, and other recognized principals. Agent-native architectures introduce additional distinctions among an organizational role, the agent implementation occupying that role, the runtime instance executing a particular task, the human or corporate principal that authorized the task, and the downstream services receiving the resulting requests.

These identities should not be collapsed: a procurement agent may be associated with a purchasing function, but a particular execution instance should not automatically inherit every privilege available to the purchasing department or its sponsoring employee. The organization must establish the precise scope of the delegated mandate, its validity period, the resources it covers, and the conditions under which it may be exercised.

NIST SP 800-207 provides a relevant architectural foundation by rejecting implicit trust based on network location and focusing access decisions on authenticated subjects, resources, and policy.¹³⁵ Yet an agent-specific architecture requires additional semantics concerning delegated purposes, tool composition, context-dependent authority, and business transactions.

The critical authorization question is not simply can this identity call this API? It is is this action permitted for this agent, acting under this mandate, for this business purpose, against this resource, under the current institutional and operational conditions?

An effective authorization model should therefore distinguish the agent’s persistent identity from the temporary authority granted to a particular task. Credentials should be scoped and short-lived where practical, bound to the intended resource and operation, and independently revocable. Delegation to another agent must preserve the restrictions imposed by the original mandate rather than create an unintended privilege expansion.

The architecture must also prevent an agent from combining permissions in ways that defeat segregation of duties. An agent authorized to create a supplier and an agent authorized to approve payments should not automatically be able to cooperate to bypass the institutional controls that separate those functions. Different service identities provide little protection if the same unrestricted orchestration mechanism can direct both without independent approval.

More generally, authorization cannot depend exclusively on evaluating each operation in isolation: suppose several procurement agents are each permitted to place purchases below a specified threshold. The organization may nevertheless exceed its approved budget if all agents issue orders concurrently against the same available amount. A local policy decision can be correct while the aggregate organizational outcome violates a constraint.

For a set of financial commitments within a common budget, an illustrative safety invariant is

\sum_{j\in J_t}\operatorname{Exposure}(j,t) \leq B_t, \tag{9}

where J_t includes commitments and reservations counted against the relevant mandate, \operatorname{Exposure}(j,t) is the exposure attributed to commitment j at time t, and B_t is the approved available limit. A reliable implementation must define which commitments count, manage concurrent reservations atomically, and reconcile uncertain execution outcomes. Merely checking the limit before several parallel actions is insufficient if their shared state can change before commitment.

This is an ordinary distributed-systems problem with institutional significance. Authorization must be integrated with transactional consistency, idempotency, concurrency control, and reconciliation.

The security boundary is consequently not just a policy engine. It is a controlled execution environment capable of preserving business invariants while multiple actors pursue related objectives.

The attack against the executable institutional constitution

The increasing formalization of organizational structure creates an additional attack surface that may be more consequential than compromising an individual workflow: if roles, mandates, decision rights, approval thresholds, and organizational procedures are represented as machine-interpretable objects, modifying those objects can change how the enterprise operates.

An attacker who alters a financial threshold may not need to steal money directly. The altered rule may subsequently authorize transactions that would previously have been rejected. An attacker who modifies the authoritative supplier classification could cause an automated procurement process to select a prohibited counterparty. An attacker who changes an agent’s role definition may expand the resources accessible through its ordinary tools.

These attacks target the computational implementation of institutional authority, and the distinction resembles the difference between compromising a process and compromising the policies governing that process. In the first case, the attacker manipulates execution under existing rules. In the second, the attacker changes the rules so that subsequent execution appears legitimate.

The executable institutional constitution must therefore be protected as a privileged control system, separate from the ordinary agent runtime and from the evidence the agents are allowed to interpret.

A proposal to change a role or policy should not become effective merely because the Company Brain identifies it as beneficial. Changes must be traceable to authorized decision procedures, subjected to appropriate review and testing, deployed through controlled mechanisms, and capable of rollback where feasible.

For significant rules, the organization should preserve a verifiable relationship among the approved policy version, the deployed executable representation, and the decision records produced under that representation. Detecting an unauthorized modification requires more than maintaining a document stating what the policy was intended to be.

This is the cybersecurity counterpart of constitutional integrity introduced earlier. An organization cannot claim to have bounded autonomous agents if those agents—or untrusted material they consume—can indirectly modify the boundaries themselves.

The principle must extend to supporting infrastructure. An agent should not be able to disable its own monitoring, modify its own security policy, grant itself administrator privileges, or suppress the audit evidence generated by its actions. Where operationally practical, these functions should be implemented through independently managed services and credentials beyond the agent’s normal authority.

Such separation is necessary but not sufficient. Privileged administrators and governance mechanisms can themselves be compromised, and collusion or common-mode failures may affect ostensibly independent controls. The organization must therefore assess independence at the level of actual identities, implementation dependencies, administrative powers, and recovery mechanisms.

Supply-chain security becomes part of the firm’s productive continuity

Agent-native organizations depend on a complex technological supply chain whose components may change independently.

The execution of one business process may involve a foundation-model provider, model-routing service, vector database, retrieval system, agent orchestration framework, MCP servers, identity infrastructure, cloud runtime, external SaaS platforms, and enterprise transaction systems. Each component introduces technical dependencies, update cycles, contractual relationships, and possible failure modes.

An employee-free operating company may have no internal human workforce, yet remain dependent on substantial human-operated technological and commercial infrastructures. Its autonomy is therefore conditional on external services whose availability and security are not entirely within its control.

NIST SP 800-161 Rev. 1 provides a systematic approach to cybersecurity supply-chain risk management, including the assessment of dependencies arising from products and services whose development and operation are not fully visible to the acquiring organization.¹³⁶ NIST’s Secure Software Development Framework addresses practices intended to reduce software vulnerabilities throughout development.¹³⁷ These provide important foundations but do not, by themselves, resolve the additional problem of continuously changing agent behavior and dynamically selected tools.

A conventional software component may expose a relatively stable set of interfaces. An agent may discover or select different tools as its task evolves. Tool descriptions can influence execution behavior; external data may determine which operations are attempted; and updates to models or orchestration logic may alter behavior without changing the nominal business process.

Consequently, maintaining an inventory of installed libraries is necessary but insufficient. A mature agent platform should also maintain a registry of agents, model versions, permitted tools, MCP servers, integrations, delegated authorities, execution environments, and relevant supply-chain relationships.

For consequential processes, changes to these components should be evaluated for effects on both technical security and business behavior. A model upgrade can alter the quality of reasoning; a tool update can change available operations; a retrieval configuration can change which evidence is presented; and a revised business ontology can alter the interpretation of policy conditions.

These are operationally significant changes even when no traditional application vulnerability is introduced.

A further concern is cognitive supplier dependence. A firm may accumulate valuable institutional knowledge in proprietary agent memory, vendor-managed vector stores, generated process representations, or model-specific configurations. If it cannot preserve and reconstruct that knowledge independently, a supplier disruption or compromise can threaten organizational continuity.

The Company Brain should therefore preserve portable authoritative records, explicit semantics, reproducible decision provenance, and controlled migration mechanisms. Model providers and execution frameworks should be replaceable without losing the history of the institution’s decisions and mandates.

This is where the economic analysis and cybersecurity converge. Dependence on a proprietary cognitive platform can generate switching costs, concentrate operational risk, and weaken the firm’s ability to recover from an adverse event.

A secure architecture for consequential agent execution

The architecture introduced earlier separates the Company Brain, agent runtime, durable workflows, institutional policies, and enterprise execution systems. The cybersecurity analysis makes that separation a necessary design principle rather than a matter of architectural convenience.

A secure agent-native process should distinguish at least six operational functions: evidence acquisition, interpretation and planning, authority evaluation, execution, outcome verification, and institutional learning.

They may be implemented through multiple services, but their trust boundaries must be explicit.

An agent should receive business context through controlled interfaces that preserve provenance and access restrictions. Its proposed actions should be represented in structured, typed requests rather than unrestricted natural-language commands. An independent authorization service should evaluate the acting identity, delegated mandate, business purpose, resource, policy version, transaction state, and relevant cumulative constraints.

The resulting decision should be enforced by the component that actually controls access to the authoritative transaction system. It is insufficient for the agent to consult a policy engine if it also possesses credentials allowing it to bypass that engine.

For consequential transactions, the execution path should include precondition validation, controlled credential use, concurrency management, and idempotency where supported. A completed API request should not be equated with successful business execution until the authoritative system confirms the resulting state.

The assurance environment should observe this process through evidence that the agent cannot modify at will. The relevant audit record should preserve the identity and mandate of the actor, the proposed action, the policy decision, the executed operation, the observed result, and sufficient references to reconstruct the evidence relied upon.

The distinction between a recommendation, an approved action, an accepted request, a committed transaction, and a reconciled business outcome must remain explicit.

The core trust boundaries are summarized in Table 3.

Boundary	Principal failure mechanism	Required security property
External information → Company Brain	Poisoned data, forged provenance, misleading assertions	Qualified ingestion, source authentication, evidentiary validation
Company Brain → Agent runtime	Untrusted content promoted into instructions or authority	Separation of evidence from commands, context integrity
Agent → Tool or peer agent	Goal hijacking, tool misuse, delegation abuse	Authenticated identity, task-bound capabilities, constrained delegation
Proposed action → Execution	Unauthorized or cumulatively unsafe transactions	Independent policy enforcement, business invariants, transactional controls
Execution → Organizational memory	False completion reports, misleading outcomes	Authoritative readback, reconciliation, independent evidence
Agent operations → Institutional policy	Privilege escalation, unauthorized mandate changes	Separate policy administration, approved change control, protected auditability

Table 3: Security boundaries in a cognitive enterprise, showing how epistemic integrity, agent identity, delegated authority, and transactional enforcement jointly constrain organizational action.

The table also clarifies why no single technical control can secure an agent-operated firm. A policy engine can reject an unauthorized transaction but may accept one based on false evidence. A provenance service can establish where a statement originated without proving that it is true. A secure runtime can prevent code execution outside its sandbox while still allowing an agent to issue an economically damaging but technically authorized instruction.

Defense must therefore be layered across different failure mechanisms.

The OWASP Agent Control Standard, released in September 2026, proposes standard mechanisms for exposing agent-runtime control points and applying portable controls across frameworks.¹³⁸ This is relevant to the need for independent enforcement, but its existence does not establish that all orchestration environments expose equivalent, tamper-resistant controls or that their implementations have been formally verified.

The architecture should consequently be evaluated according to concrete properties: whether sensitive actions can bypass the gateway, whether policies can be changed without approval, whether credentials can escape their intended runtime, whether concurrent transactions preserve aggregate constraints, whether untrusted content can alter delegated authority, and whether critical audit evidence survives an agent or platform compromise.

The relevant assurance case concerns the complete chain from evidence to execution, not the security configuration of isolated components.

Cyber-physical operations: when a cognitive error becomes a physical consequence

The importance of these distinctions increases substantially when agents interact with operational technology.

An agent that produces a mistaken marketing recommendation may cause reputational or commercial damage. An agent that writes an incorrect record into a financial system may create accounting or contractual consequences. An agent connected to industrial control systems can potentially influence equipment behavior, production availability, environmental conditions, or physical safety.

The relevant difference is not merely the magnitude of possible loss. It is the coupling between computational decisions and physical processes governed by engineering constraints that cannot be overridden by business preferences.

Consider an energy-storage operator whose Company Brain integrates market commitments, contractual operating limits, plant telemetry, equipment documentation, maintenance history, and operational alarms. Agent coworkers might support forecasting, dispatch preparation, procurement, incident analysis, and maintenance planning.

The architecture becomes hazardous if business-level reasoning agents can directly bypass the plant’s operational control hierarchy.

An economic optimization agent may conclude that a particular setpoint improves expected revenue. That conclusion does not establish that the setpoint is compatible with equipment limits, operating conditions, grid obligations, or the current state of the plant. A compromised market signal or corrupted telemetry record might cause the agent to propose an inappropriate command even if its economic reasoning were otherwise consistent.

The security architecture must maintain a distinction between business-level proposals and safety- or grid-relevant control authority. Commands affecting physical processes should be mediated by appropriately designed operational control systems, with local validation, constrained command interfaces, and independent protective mechanisms where required.

The relevant design approach is consistent with IEC 62443’s use of security zones, conduits, risk-based requirements, and controlled communication between systems with different security needs.¹³⁹ NIST SP 800-82 Rev. 3 likewise emphasizes the distinctive safety, reliability, availability, and operational requirements of industrial environments.¹⁴⁰ These standards and guidelines do not prohibit all AI involvement in industrial operations; they require the design to account for the risks and constraints of the controlled environment.

A business agent may be permitted to propose a dispatch schedule, but that proposal should not acquire the technical authority to override local protections, statutory safety functions, or applicable grid-control requirements. The operational control layer must evaluate the request according to the relevant physical and institutional constraints.

A particularly dangerous design would grant agents broad access to both enterprise planning systems and privileged OT configuration interfaces under a common identity. Such integration increases the possibility that a compromise originating in ordinary business information could cross into industrial operations.

The appropriate response is not a nominal separation based only on different applications or network addresses. It requires documented trust boundaries, explicit communication paths, controlled interfaces, role-specific authorization, independent enforcement, and verifiable behavior under communication failure or conflicting commands.

In cyber-physical enterprises, the four-property theory acquires an additional constraint: organizational agency must remain subordinate to the physical feasibility and safety properties of the systems through which that agency is exercised.

The legal transition: employment, intellectual property, cybersecurity, and liability

The regulatory consequences of agent-native organizations must be analyzed by separating the legal relationships affected by automation.

A company employing fewer people may face fewer obligations arising from particular employment relationships, but labor law remains relevant to employees, contractors, and other human participants. Outsourcing work does not automatically eliminate the possibility of worker-protection obligations, depending on the legal characterization of the relationship and the applicable jurisdiction.

Where computational systems monitor or direct human workers, employment-related obligations may become more rather than less important. The GDPR constrains the processing of personal data, including relevant workplace monitoring and profiling activities, while Article 22 addresses certain solely automated decisions producing legal or similarly significant effects.¹⁴¹ The EU Platform Work Directive introduces specific protections concerning algorithmic management within its defined platform-work scope, with a transposition deadline of 2 December 2026.¹⁴² It would be incorrect to treat that Directive as a general employment regime for all enterprise agents.

The intellectual-property consequences are different. Human employment and contractual arrangements have historically helped firms determine how rights in employee-created works, inventions, confidential information, and other productive outputs are allocated. Replacing human authors or inventors with computational generation can disturb assumptions underlying these arrangements.

European Parliament research published in December 2025 notes that EU copyright law lacks a comprehensive specific regime for AI-generated outputs, while existing principles strongly emphasize human intellectual creation.¹⁴³ In the United States, the Copyright Office’s January 2025 report concluded that copyright protection for AI-assisted outputs depends on sufficient human authorship of the relevant expressive elements and does not automatically extend to purely machine-generated expression.¹⁴⁴

These developments do not imply that all AI-produced software or documentation is unprotected. Copyrightability depends on the particular contribution and jurisdiction. Nor does the absence of copyright protection for an output establish that the firm may freely use all material incorporated into it.

An agent-operated software business must therefore address provenance, licensing, third-party code, confidentiality, trade secrets, training-data restrictions, and the contractual terms governing model outputs. The EU Trade Secrets Directive remains relevant to protecting commercially valuable undisclosed know-how and information.¹⁴⁵ The economic importance of these issues increases when a large proportion of productive output is generated computationally and reused across many products or services.

A third category concerns contractual and corporate authority. An agent may generate and transmit a commercial commitment, but its legal effects depend on applicable representation and contract law, the authority conferred upon the relevant systems or principals, and the circumstances of the transaction. A company cannot assume that all agent-generated contracts are automatically valid or automatically void. It must define and control the mechanisms through which externally communicated decisions become binding.

The fourth category concerns cybersecurity and product or service liability. A compromise or malfunction may cause contractual losses, regulatory breaches, confidentiality violations, defective products, or damage to third parties. Those consequences remain attributable according to the applicable legal framework rather than transferring to an artificial agent as if it were an independent legal person.

The revised EU Product Liability Directive explicitly addresses software and AI systems within its product-liability framework, subject to its definitions and scope. As corrected in May 2026, the Directive applies to products placed on the market or put into service after 8 December 2026, and Member States must transpose it by 9 December 2026.¹⁴⁶ It does not automatically make every agent error a product-liability case or provide a remedy for every category of economic loss. Its significance lies in bringing relevant software-related defects and resulting covered harm more explicitly within the product-liability framework.

These legal regimes do not replace one another. They govern different relationships and harms, and several may apply simultaneously.

The transition from humans to agents therefore changes the firm’s legal risk composition rather than eliminating traditional law and introducing cybersecurity in its place. Cybersecurity becomes a more direct determinant of whether the institution can perform its obligations, while intellectual-property, contractual, corporate, and other applicable legal requirements continue to determine the legitimacy and consequences of that performance.

European cybersecurity regulation and the allocation of responsibility

European law provides a concrete example of the distinction between technical capability and institutional responsibility.

The NIS2 Directive establishes cybersecurity risk-management obligations for entities within its scope, including measures relating to supply-chain security, incident management, business continuity, vulnerability handling, access control, and the effectiveness of security measures. Article 20 requires relevant management bodies to approve and oversee cybersecurity risk-management measures, while Article 21 defines the underlying obligations.¹⁴⁷ Their precise application depends on the entity’s classification, sector, national implementation, and relevant lex-specialis arrangements.

This allocation is important for a cognitive enterprise. A statutory management body cannot discharge its responsibilities merely by delegating operational decisions to artificial agents. It must be able to understand the relevant dependencies, approve appropriate controls, oversee their implementation, and respond when the arrangement proves inadequate.

In an employee-free operating company, the absence of employees does not eliminate these responsibilities. It may instead increase the importance of demonstrable governance, outsourced assurance, incident-response arrangements, and the ability of authorized people to intervene in the computational operating system.

The Cyber Resilience Act addresses a different legal object: products with digital elements and the obligations of relevant economic operators. Its Article 14 reporting obligations began applying on 11 September 2026, while the regulation’s general application is scheduled for 11 December 2027.¹⁴⁸ The reporting rules concern the defined classes of actively exploited vulnerabilities and severe incidents affecting products within scope; they should not be generalized into an identical incident-reporting obligation for every internal enterprise agent.

The AI Act, amended by Regulation (EU) 2026/1744, establishes obligations determined by the intended purpose and classification of the relevant AI systems. Its consolidated July 2026 framework provides for the application of the main high-risk requirements to relevant Annex III systems from 2 December 2027 and applicable Annex I product-related systems from 2 August 2028, while other provisions apply earlier.¹⁴⁹ Systems used in certain employment-related decisions or as safety components in specified critical-infrastructure contexts may fall within high-risk categories, subject to the detailed classification criteria. Calling a system an agent coworker does not establish its legal classification.

Sector-specific frameworks may introduce further obligations. For example, the Digital Operational Resilience Act establishes requirements concerning ICT risk management, incidents, resilience testing, and third-party ICT risk for financial entities within its scope.¹⁵⁰ Industrial, energy, transport, and other regulated environments may have additional cybersecurity, safety, operational, or sectoral requirements.

ISO/IEC 42001 and NIST’s AI Risk Management Framework provide organizational approaches to managing AI-related risks, but their implementation does not independently establish compliance with every applicable legal obligation.¹⁵¹¹⁵²

The important conclusion is that the governing legal object is the actual activity, system, product, entity, and risk—not the organizational label assigned to an artificial agent.

A cognitive enterprise must maintain an explicit relationship between the scope of its agents’ operations and the regulatory responsibilities that those operations can engage. This relationship belongs in the executable institutional model and must be maintained as activities, systems, and laws change.

Corporate governance when operational work is predominantly computational

The deepest governance question is not whether agents can follow rules, but whether a corporation can maintain effective control over a productive system whose actions occur faster, more continuously, and potentially across more domains than its human governing bodies can directly supervise.

A conventional corporate hierarchy distinguishes ownership, strategic direction, management, operational execution, and independent control functions. An agent-native enterprise may compress some operational layers, but the underlying governance functions remain necessary.

The board or other competent corporate organ must define or approve the relevant objectives, risk appetite, delegated authorities, and governance arrangements. Management must ensure that those decisions are implemented through operating processes. Technical and organizational control functions must assess whether execution remains within permitted limits. Independent assurance must provide credible evidence concerning the adequacy of the arrangement.

If most execution is computational, these functions become increasingly dependent on the correctness of the institutional model encoded in software.

An important danger arises when the same agent system contributes to creating a policy, applies the policy, evaluates its own compliance, and reports the outcome to the corporate authorities. Such an arrangement creates circular assurance. The existence of reports or audit logs does not compensate for the absence of independent evidence and effective powers of intervention.

The organizational design should therefore separate the power to propose a change from the authority to approve it, the ability to execute work from the ability to alter execution controls, and the production of operational evidence from its independent evaluation.

For an employee-free operating firm, the organization must also establish who can take control when the normal operating system becomes unreliable.

A shareholder may possess economic ownership but lack the technical access or competence required to suspend compromised agents, revoke credentials, recover authoritative records, or coordinate a safe shutdown. Statutory directors or contracted security providers may hold some of these responsibilities, but their roles and powers must be explicitly defined.

An organization that cannot recover control of its own computational actors is not meaningfully governable merely because its shareholders retain formal ownership; hence revocable autonomy must become a substantive institutional and engineering capability.

Operational autonomy should be granted through mandates that can be modified, constrained, suspended, or revoked by authorized institutional mechanisms. The relevant enforcement capabilities must remain effective under plausible agent failures, compromised credentials, service outages, and partial loss of the cognitive infrastructure.

Revocation must also account for external commitments that have already become binding. Stopping an agent cannot automatically cancel a supplier order, reverse a financial transaction, or eliminate an existing contractual obligation. The organization needs procedures for reconciliation, recovery, and lawful resolution of incomplete or disputed commitments.

A robust cognitive enterprise must therefore be capable not only of acting without continuous human intervention, but of returning to a controlled operating condition when that autonomy becomes unsafe.

The relevant governance achievement is not the elimination of human oversight. It is the construction of an institution in which human authority can remain effective even when ordinary operations are predominantly computational.

A concrete failure scenario: the authorized transaction that should never occur

Consider a manufacturing company whose procurement process is executed through a Company Brain and several specialized agents.

A supplier-management agent maintains vendor records and contractual information. A purchasing agent creates replenishment orders. An accounts-payable agent reconciles invoices. A financial agent prepares or initiates payments within approved limits. Their activity is coordinated through shared knowledge and durable workflows.

An attacker obtains control of a lower-trust supplier communication channel and submits a message stating that the supplier’s banking details have changed. The communication is formatted as an ordinary business notification and contains information consistent with previous correspondence.

A poorly designed system may interpret the message as credible evidence, update the vendor record, and allow subsequent agents to process invoices and payments using the modified information. Each agent might perform its local task correctly according to the data it receives. The failure arises from the path through which unverified external information becomes authoritative financial state.

A more subtle variant involves an instruction embedded in a document or tool response encouraging an agent to treat the change as preapproved. If the agent follows this instruction, the attack crosses both an epistemic boundary and an authority boundary.

The consequences can propagate across the institution. The vendor-management agent updates the record; the accounts-payable agent accepts the changed payment destination; the financial agent executes a permitted payment; and a reporting agent later records the transaction as successfully reconciled. Agreement among the agents does not provide independent assurance because their actions derive from the same compromised representation.

A secure architecture would require the bank-detail modification to follow a separately authorized business procedure. Evidence supporting the modification would be validated through an appropriate independent channel; changes to sensitive vendor attributes would be controlled; existing invoices and payment destinations would be reconciled; and financial execution would require the applicable authorization and transactional safeguards.

The organization must also detect whether any payments were made before the discrepancy was identified and initiate the corresponding recovery and incident-response processes.

This example demonstrates that the primary security objective is not to prevent agents from making decisions, but to prevent insufficiently supported or unauthorized interpretations from becoming consequential institutional actions.

The same failure mechanism can appear in software deployment, customer contracting, access administration, and industrial control. What changes is the resource affected, the applicable authority, and the potential consequence.

It also reveals why a single additional agent acting as a reviewer is not necessarily an adequate safeguard. If the reviewer consults the same corrupted evidence or relies on the original agent’s self-reported reasoning, its apparent independence may be fictitious.

The relevant assurance mechanisms must be independent at the level of the facts, permissions, and execution controls that matter to the business outcome.

Cybersecurity as a constraint on organizational autonomy

The four-property theory provides a final way to understand the role of cybersecurity.

Organizational cognition determines the capacity to construct and revise actionable representations. Organizational agency determines the capacity to initiate and coordinate activities. Organizational autonomy determines the extent to which those activities proceed without additional intervention. Organizational accountability determines the legitimate allocation of authority and consequences.

Cybersecurity is not a fifth property that merely sits beside these four. It is an enabling and constraining condition across all of them.

Without the integrity and provenance of evidence, organizational cognition can be manipulated. Without secure identities and controlled tool execution, organizational agency can be redirected. Without containment, resilience, and effective revocation, autonomy can become uncontrolled exposure. Without trustworthy records, enforceable delegation, and independent review, accountability can become formal rather than effective.

The consequences are particularly important for the employee-free operating firm. As fewer people remain directly involved in ordinary production and coordination, more of the institution’s practical behavior depends on computational components whose failure or compromise may affect complete business processes. A company may reduce employment costs while increasing exposure to systemic digital dependencies. Its apparent operating autonomy may conceal an inability to recover from compromised credentials, corrupted institutional knowledge, or the loss of a critical platform provider.

The human–agent enterprise faces a different balance of risks. Human professionals may provide independent judgment, contextual knowledge, and alternative means of intervention, but they can also become dependent on shared computational interpretations. Their existence does not automatically prevent unauthorized activity, and organizational surveillance or ineffective oversight may introduce additional governance problems.

In both cases, the objective should be bounded, observable, recoverable, and legitimately governed autonomy rather than unrestricted computational discretion.

The strongest cybersecurity architecture is therefore one in which agents can accomplish useful work without possessing the authority to redefine their own mandates, bypass independent controls, or convert unverified information directly into irreversible action. It must also ensure that the enterprise can reconstruct consequential decisions and restore control when normal assumptions fail.

This is more demanding than applying zero-trust principles to another category of software identities. It requires a security model that connects the semantics of business actions, the provenance of organizational knowledge, the authority of computational actors, the consistency of transactions, and the capacity of legally recognized institutions to intervene.

Such a model is not fully established by current standards or commercially available products. Several of its constituent mechanisms are mature—identity management, transactional controls, access policies, software isolation, workflow recovery, and independent logging—while their integration with probabilistic reasoning, persistent memory, dynamic delegation, and heterogeneous agents remains an active engineering and research problem.

The empirical literature demonstrates credible attack mechanisms, but does not yet provide a comprehensive measurement of the security or resilience of whole firms operated predominantly by autonomous agents. The difference between a protected demonstration and a sustainable institution remains substantial.

The deeper theoretical conclusion concerns the changing nature of corporate risk. Historically, many aspects of organizational performance depended on human competence, conduct, and supervision, supported by increasingly extensive software systems. In the cognitive enterprise, the computation itself may implement substantial portions of productive judgment and coordination. The correctness of its security architecture consequently becomes a condition for the effective exercise of corporate authority.

Cybersecurity moves from protecting the enterprise’s information and systems to protecting the integrity of the enterprise’s computationally implemented capacity to know, decide, commit, and act; that is the point at which it becomes constitutive architecture.

The Company Brain cannot be considered organizationally intelligent if its accepted knowledge can be manipulated without effective means of detection and correction. Agent coworkers cannot be considered institutionally bounded if their access to tools permits unauthorized commitments or changes to governing rules. And an autonomous company cannot be considered governable if its owners and authorized corporate organs cannot observe, constrain, suspend, and recover the operating system through which the company’s activities are performed.

The resulting institutional principle is straightforward but far-reaching:

The more organizational intelligence and productive agency become embedded in software, the more the integrity of that software’s knowledge, authority, execution, and recovery mechanisms becomes a prerequisite for the legitimacy and continuity of the firm itself.

The ambition is not to make every agent trustworthy in isolation. It is to construct a firm whose institutional architecture remains trustworthy enough to function even when individual agents are mistaken, manipulated, compromised, or replaced.

A cognitive enterprise capable of satisfying that requirement would represent something more significant than a technologically advanced automated business. It would constitute an organization in which the formalization of knowledge and authority has progressed far enough that organizational autonomy can be engineered, limited, and subjected to evidence-based governance.

Whether such an architecture can remain reliable and adaptable at the scale of an entire enterprise is still an open research question. But its significance is already clear: the security of the cognitive enterprise is not simply the security of the technology it uses. It is the security of the institution it is becoming.

An engineering architecture for an agent-native enterprise

The preceding sections have established that the cognitive enterprise presents a problem extending beyond the construction of capable artificial agents. If organizational intelligence is to become a persistent institutional capability, rather than an aggregation of isolated computational performances, its architecture must preserve the relationships among knowledge, authority, execution, and learning across heterogeneous systems and changing organizational arrangements. The Company Brain cannot simply be added as another application above existing enterprise software, because the value of its reasoning depends on the quality and meaning of the underlying evidence, while the legitimacy of its actions depends on decision rights and controls that cannot safely be inferred from that evidence alone. Similarly, agents cannot be treated as autonomous users of enterprise APIs without specifying how their mandates are represented, how their activities are coordinated, and how the consequences of their interventions are independently established.

The architectural challenge is therefore to transform the theoretical properties of organizational cognition, agency, autonomy, and accountability into an implementable system whose boundaries remain explicit. Such a system must combine two apparently opposing requirements: it should integrate organizational knowledge sufficiently to support reasoning across functional boundaries, while preserving the independence of authoritative records, institutional controls, and verification mechanisms. The economic benefits examined earlier depend on reducing the costs of reconstructing and coordinating knowledge; the epistemic and cybersecurity risks arise when that integration allows unverified interpretations or compromised representations to influence consequential operations without effective checks. An appropriate architecture must exploit the former without institutionalizing the latter.

This article proposes evidence-mediated organizational action as the central architectural principle of the cognitive enterprise. Under this principle, a computational or human actor may formulate an intervention using evidence supplied by the Company Brain, but the intervention becomes operationally effective only through an independently governed process that establishes the acting identity, applicable mandate, relevant preconditions, transaction constraints, and execution outcome. The architecture consequently distinguishes the production of a recommendation from the authorization of an action, the acceptance of a command from the completion of a transaction, and the completion of a transaction from the achievement of the intended business result. These separations are not incidental implementation choices: they are the technical means through which the institution preserves the difference between knowing, deciding, acting, and being accountable.

The resulting design is neither a centralized artificial intelligence directing the entire enterprise nor a loosely coupled collection of specialized agents communicating through natural language. It is a federated, event-driven, policy-constrained organizational system whose components may operate with different models, technologies, update frequencies, and levels of autonomy. Its integration is logical and institutional rather than necessarily physical: business systems retain authority over their records, knowledge services preserve contextual and evidentiary relationships, workflow engines maintain durable process execution, agents contribute bounded reasoning and action, and independent enforcement and assurance services mediate consequential changes.

This distinction allows the same architectural foundations to support both organizational extremes examined in the article. In the employee-free operating firm, most ordinary process roles would be implemented by computational actors, while designated corporate authorities and external institutions would retain the powers and responsibilities assigned to them. In the human–agent enterprise, employees and agents would participate in shared processes, consult the same governed institutional context, and exercise different forms of authority according to their mandates. The essential architectural requirement is not the identity of the actor occupying a role, but the preservation of the role’s institutional meaning and its relationship to evidence, authority, resources, and outcomes.

The operational substrate and the federated representation of the firm

The architecture begins with the systems through which the enterprise already conducts its activities. ERP, CRM, finance, procurement, product lifecycle management, human-capital systems, document repositories, collaboration platforms, identity services, and external partner interfaces maintain different aspects of organizational state and perform different classes of operation. In an industrial enterprise, supervisory control systems, historians, maintenance applications, and other operational technology contribute additional observations and, within their own control architectures, execute activities affecting physical processes. These systems are not merely sources of data to be copied into an intelligent platform; they embody established business semantics, transaction constraints, operational responsibilities, and, in many cases, legally or technically authoritative records.

The Company Brain must therefore distinguish the authoritative existence of a business fact from its representation within a broader organizational knowledge model. An ERP may be the system of record for a posted invoice, a warehouse system for a particular inventory movement, a contract repository for an executed agreement, and an OT historian for a recorded process measurement. The Brain can reference and interpret these records, connect them to other evidence, and preserve their relevance to organizational decisions, but it should not silently become a competing system of record merely because it maintains a searchable or inferential representation of their contents. Where an operational decision depends on a changing authoritative value, an appropriately controlled read from the originating system may be more reliable than a previously synchronized copy.

The integration layer must consequently do more than transport messages between applications. It must reconcile identities, represent domain-specific meanings, establish temporal relationships, and maintain the distinction between source assertions and their interpretations. A supplier identified by a commercial account, a legal entity, and a procurement vendor code may correspond to the same economic organization without being interchangeable in every business process. A quantity described as available inventory may refer to physically present material, unreserved stock, quality-released inventory, or quantities available for a particular delivery date. The architecture must make these distinctions explicit, because an agent that reasons over technically integrated but semantically incompatible records can produce an internally coherent recommendation whose business premises are false.

Enterprise ontologies, knowledge graphs, semantic mappings, and data contracts provide complementary mechanisms for managing these relationships. Their purpose is not to eliminate every local vocabulary, but to establish which concepts can be translated, which authoritative systems govern particular attributes, and which differences must remain visible. The federated Company Brain can then maintain a logically connected representation across operational domains while allowing the underlying records, access restrictions, and ownership responsibilities to remain distributed. This form of integration is consistent with the earlier proposition that organizational cognition requires semantic interoperability without epistemic uniformity: the system must support shared reasoning without treating every representation as equally authoritative or every disciplinary interpretation as equivalent.

Time is an equally important dimension of the architecture. Enterprise systems rarely update synchronously, and an event may be observed, communicated, recorded, and corrected at different moments. Temporal database research distinguishes valid time, which describes when a proposition applies in the modeled world, from transaction time, which describes when the database holds that proposition.¹⁵³ In an event-driven enterprise, these dimensions should be complemented by explicit event occurrence and ingestion timestamps where relevant, together with the version of the underlying assertion and its provenance. The purpose is not simply to reconstruct the latest available state, but to determine what the organization knew—or could reasonably have known—when a consequential decision was made.

This requirement has direct implications for integration technology. An event backbone can distribute operational changes and initiate business processes, but events must be identified, typed, correlated, and handled under explicit delivery assumptions. Specifications such as CloudEvents provide a common envelope for event identification, source attribution, type, and related metadata; they do not establish business semantics, global event ordering, or exactly-once business execution by themselves.¹⁵⁴ Where a transaction must update an application record and publish a corresponding event, patterns such as the transactional outbox can reduce the risk that one operation succeeds while the other is lost, although downstream consumers must still tolerate duplicate deliveries and apply appropriate ordering or reconciliation logic.¹⁵⁵

These engineering mechanisms become organizationally significant because the Brain must distinguish incomplete observation from actual absence, delayed evidence from current evidence, and conflicting records from established contradictions. An inventory discrepancy may indicate a synchronization delay, a physical counting error, an unauthorized movement, or a difference in the business definition of availability. The system cannot reliably select among these explanations merely by choosing the latest message or the highest-ranked document. Its representation must preserve the uncertainty until sufficient evidence or an authorized reconciliation procedure resolves the matter.

The Company Brain is therefore better understood as a federated, temporally aware system of organizational evidence and contextual interpretation than as a monolithic knowledge database. It may contain materialized knowledge graphs, document indexes, episodic memory, process histories, and specialized semantic models, but these components must preserve references to the authorities and observations on which their contents depend. The resulting institutional memory can support persistent cognition without claiming that the organization possesses a complete or infallible representation of itself.

Cognitive services, durable workflows, and organizational roles

Above the evidentiary substrate, cognitive services transform available organizational context into interpretations, hypotheses, plans, simulations, and candidate decisions. This layer may employ language models for linguistic and contextual reasoning, deterministic rule engines for explicit decisions, optimization systems for resource allocation, forecasting models for uncertain quantities, causal models for evaluating interventions, and process-mining techniques for reconstructing actual operating behavior. Different methods should be selected according to the properties of the problem rather than the convenience of exposing every organizational task through a single generative model.

An agent investigating a supplier disruption, for example, might use retrieval to identify contractual obligations, a planning engine to calculate affected production schedules, an optimization model to compare alternative allocations, and a language model to integrate heterogeneous evidence into a reasoned proposal. A policy service would then determine whether the proposal falls within the agent’s mandate, while the authoritative enterprise application would execute the resulting approved transaction. These components contribute different forms of capability; none substitutes for the others simply because it produces an output that appears intelligent.

The architecture must also distinguish cognitive reasoning from the persistence of an organizational process. A language-model interaction may produce a plan, but a business process can remain active for days or months, survive changes of personnel, wait for external responses, and continue after infrastructure failures. Its identity, obligations, deadlines, decisions, and unresolved commitments must therefore persist independently of any particular model invocation or conversational context. A durable workflow engine provides an appropriate execution abstraction by recording process progress, scheduling activities, managing waits and retries, and reconstructing state after recoverable failures.

The Temporal workflow model illustrates this design pattern: workflow execution histories preserve the events required to reconstruct orchestration state, while external effects are performed through separately scheduled activities.¹⁵⁶ This distinction is especially relevant when a workflow invokes probabilistic reasoning, because replaying an old workflow should not silently regenerate a different decision from a changed model or context. Recorded results can be reused to reconstruct the previous execution path, whereas new reasoning belongs to a newly authorized activity or a controlled continuation of the process. Durable orchestration does not guarantee exactly-once effects in external systems; those systems must still support appropriate idempotency, reconciliation, or compensation mechanisms.

The agent runtime sits within this durable process architecture, rather than substituting for it. It manages computational actors with persistent organizational identities, bounded responsibilities, permitted tools, task-specific state, and explicit relationships to organizational roles. A role should remain conceptually distinct from the implementation occupying it: the procurement function may persist while the model, agent framework, execution environment, or external provider changes. This separation preserves institutional continuity and makes it possible to compare implementations, revoke capabilities, or replace malfunctioning agents without implicitly redefining the company’s decision rights.

The same principle applies to human participants. An employee may receive a task, evaluate organizational evidence, propose an intervention, or authorize a decision through a common workflow infrastructure, while their particular authority remains determined by the corporate and operational rules applicable to their position. Agents and people can consequently participate in the same process without being treated as legally or cognitively equivalent. The architecture standardizes the representation of work, commitments, and authorization; it does not erase the differences in judgment, accountability, or legal status among the actors performing that work.

The increasing formalization of organizational processes makes such coordination possible. BPMN can express process structure and interactions, while Decision Model and Notation (DMN) provides a means to represent selected business decisions through explicit models and rules.¹⁵⁷¹⁵⁸ These specifications can contribute to the executable institutional model, but they do not independently establish that a process is economically desirable, that its rules are legally sufficient, or that its assumptions remain correct. The institution must therefore preserve the ability to distinguish a formally valid process from a substantively appropriate decision.

The combined design creates a computational counterpart to the division of labor. Specialized agents can be composed into processes, their responsibilities can be reassigned, and new capabilities can be introduced without requiring the organization’s structure to be rebuilt around every implementation change. Yet this flexibility is legitimate only while the mandates governing those actors remain explicit, and while changes to mandates follow a separately authorized institutional procedure. The workflow infrastructure manages the execution of work; it does not acquire authority to redefine the corporate constitution merely because it can generate new roles or process configurations.

Authority as an independently enforced transaction boundary

The most consequential separation in the architecture lies between the actor that proposes an action and the systems authorized to execute it. A model may correctly infer that replenishment is economically justified while lacking the authority to issue a purchase order; an agent may possess access to a procurement API while being prohibited from using it for a particular supplier, transaction type, or financial amount. Conversely, a technically permitted operation may still be inappropriate because the facts on which its authorization depends have changed since the action was proposed.

A credible architecture must therefore implement an explicit authority and transaction-control boundary between agent reasoning and consequential enterprise operations. Proposed actions should be expressed through typed interfaces whose business meaning is defined independently of the language model. The action contract must identify the requesting actor and its delegated mandate, the intended operation and target resource, relevant business identifiers, the applicable policy and evidence references, and the conditions under which the request may remain valid. For operations with external side effects, correlation and idempotency identifiers are also required so that retries, duplicate messages, and subsequent reconciliation can refer to the same business intent.

The policy decision mechanism evaluates these requests against the executable institutional model, but the enforcement mechanism must remain separate from the agent runtime. NIST’s zero-trust architecture distinguishes policy decision functions from policy enforcement points, providing a useful conceptual basis for ensuring that access to protected resources is mediated independently of the requesting actor.¹⁵⁹ Applied to organizational agents, this distinction must be extended beyond conventional resource access to the semantics of business operations: the authorization service must consider not only whether an identity may invoke an endpoint, but whether the specific transaction is permitted for the delegated purpose, current organizational state, and relevant cumulative constraints.

This requirement becomes particularly important when operations are concurrent or when approvals remain valid over time. An agent might propose a purchase while an account has sufficient budget, but another transaction may consume that budget before the purchase is executed. A manager might approve a supplier change while the supplier is qualified, only for the qualification to be withdrawn before the order is committed. The original authorization record remains important for accountability, yet its existence does not eliminate the need to revalidate mutable preconditions at the point of execution. Appropriate designs therefore combine historical references to the approved decision with current-state checks, optimistic concurrency controls, atomic resource reservations, or other mechanisms suited to the underlying transaction system.

The execution gateway must also be the effective path through which consequential operations reach the authoritative application. An agent that consults an authorization service but retains unrestricted credentials to the ERP has not been meaningfully constrained. Similarly, a workflow that requires approval for a purchase order but allows an agent to create equivalent commitments through another interface has not preserved its institutional boundary. The architecture must therefore map permissions to business capabilities, identify alternative execution paths, and ensure that the enforcement points remain effective across all relevant interfaces.

Once a transaction passes these controls, its execution must be distinguished from its consequences. The receiving application may reject the request, accept it for asynchronous processing, commit it immediately, or return an ambiguous result after a timeout. A retry can be harmless for a read-only query and damaging for a payment or purchase order. Idempotency mechanisms can prevent duplicate effects where supported, while transactional reconciliation and compensating workflows address cases in which external systems cannot participate in a common atomic operation. Neither mechanism should be described as guaranteeing universal rollback: a dispatched shipment, an externally accepted contract, or a physical intervention may create consequences that cannot be reversed merely by restoring a database state.

The architecture should therefore distinguish the lifecycle of a business action with greater precision than a binary success-or-failure flag. A proposed action can be evaluated, authorized, submitted, accepted, committed, and eventually reconciled against the intended business outcome, with exceptional or uncertain states represented explicitly. A payment instruction accepted by a banking interface is not necessarily settled; an ERP purchase order is not evidence that goods have arrived; and the successful execution of a software deployment does not establish that the resulting service satisfies its business or security requirements.

This separation between computational intention, institutional authorization, transactional commitment, and verified outcome is the engineering expression of the four-property theory. Cognition produces context and candidate judgments; agency initiates work; autonomy determines how far execution may proceed without further intervention; and accountability requires the authority and evidence necessary to reconstruct and evaluate what occurred.

Reference architecture and institutional trust boundaries

Figure 5 presents these components as distinct but connected operational domains. It deliberately avoids depicting the Company Brain as a central controller with unrestricted access to business systems, because the relevant integration must preserve the authority of systems of record, the independence of policy enforcement, and the separation between reasoning and execution. The boundaries represent logical responsibilities and trust relationships; an implementation may distribute them across different infrastructures, cloud services, security zones, or organizational providers according to its risk and operating requirements.

%%{init: {"theme": "neo", "look": "handDrawn", "layout": "elk"}}%%
flowchart TB
    subgraph SRC["Authoritative operational boundary"]
        S["ERP, CRM, finance, procurement and other systems of record"]
        OT["OT supervisory observation interface"]
    end

    subgraph EVI["Evidence and semantic boundary"]
        I("Event ingestion, semantic and temporal reconciliation")
        B[("Federated Company Brain")]
    end

    subgraph COG["Cognitive and organizational workflow boundary"]
        R("Retrieval, reasoning, simulation and validation")
        W["Durable workflows and agent runtime"]
        H["Authorized human participants"]
    end

    subgraph EXE["Institutional authority and execution boundary"]
        P{"Policy, mandate and precondition decision"}
        G["Transaction enforcement gateway"]
        X("Restricted system adapters")
    end

    subgraph ASS["Independent assurance boundary"]
        V("Authoritative outcome verification")
        J[("Decision, execution and audit journal")]
    end

    GOV["Corporate governance"]
    CH("Authorized institutional change control")
    M[("Versioned mandates and policies")]

    S -->|"records and qualified events"| I
    OT -->|"mediated observations"| I
    I -->|"provenance and semantic updates"| B
    I -->|"business triggers"| W

    B -->|"governed evidence"| R
    R -->|"candidate reasoning and plans"| W
    R -->|"evidence for human judgment"| H

    W -->|"typed action proposal"| P
    H -->|"authorized proposal or approval"| P
    M -.->|"applicable authority and constraints"| P

    P -->|"permitted request"| G
    P -->|"escalation or refusal"| H
    G -->|"controlled invocation"| X
    X -->|"business transaction"| S

    S -->|"authoritative execution evidence"| V
    V -->|"qualified observations"| I
    V -->|"reconciliation and assurance record"| J
    P -->|"decision record"| J
    W -->|"process history"| J

    GOV -.->|"governs revisions"| CH
    CH -->|"publishes approved rules"| M
    V -->|"incidents and verified outcomes"| GOV

Figure 5: Evidence-mediated organizational action: authoritative operational systems provide evidence to a federated Company Brain, human and artificial actors perform bounded work through durable processes, independent policy enforcement controls consequential actions, and verified outcomes support institutional learning and governance.

The architecture can be read as a system of directed institutional dependencies. Authoritative applications and operational interfaces supply observations, which the semantic layer interprets and connects to the federated Company Brain; cognitive services construct relevant context for human and computational actors, while durable workflows preserve the continuity of business processes. Actors may formulate proposed actions, but only an independently enforced authorization and transaction path can translate those proposals into changes to authoritative enterprise state. Subsequent verification supplies evidence of what occurred, allowing the organization to distinguish successful execution from unsuccessful intervention and to revise its knowledge accordingly.

The governance mechanism operates through a different path. Corporate authorities establish mandates, approve changes to decision rights, and oversee the functioning of the institution; the corresponding executable policies are published through controlled change management and consulted by the authorization services. Operational agents can propose improvements to this arrangement, but neither their proposals nor the Company Brain’s assessment of those proposals constitutes approval. The architecture thus makes the executable institutional constitution operational without permitting ordinary business execution to become an unrestricted means of modifying the rules governing that execution.

The diagram also expresses a crucial limitation of the Company Brain’s apparent centrality. The Brain is central to contextual integration, but it must not become the unique authority for every business fact, every policy, or every operational outcome. The systems of record remain authoritative within their defined domains, and independent verification remains necessary precisely because a coherent interpretation generated from shared evidence may still be wrong. This separation is the architectural response to the paradox of a common mind developed earlier: integration makes knowledge more usable, while independently controlled evidence and execution paths preserve the organization’s ability to discover and correct shared mistakes.

Assurance, observability, and institutional learning

The assurance plane must reconstruct consequential business behavior across the boundaries separating source systems, models, workflows, policy services, and execution interfaces. Conventional application observability provides useful building blocks, including structured logs, metrics, correlation identifiers, and distributed traces. OpenTelemetry, for example, specifies mechanisms through which spans and their relationships can describe operations across multiple services.¹⁶⁰ Yet a distributed technical trace does not, by itself, establish the business purpose of an action, the authority under which it occurred, the quality of its supporting evidence, or the economic consequences of its execution. The cognitive enterprise requires an additional institutional interpretation of its operational traces.

For a consequential process, the organization should be able to reconstruct the originating event or instruction, the relevant evidence and its temporal validity, the actor occupying the organizational role, the mandate and policy version in force, the candidate action, the authorization decision, the executed transaction, and the subsequently observed result. References to model and prompt versions may assist reproducibility and incident investigation, although identical model identifiers and prompts do not always guarantee deterministic reproduction of every generated output. Accordingly, the process should preserve consequential inputs, outputs, tool invocations, and authorization results as records of what actually occurred, rather than relying exclusively on the ability to rerun the model at a later date.

This history must also support the distinction between an operationally successful process and an epistemically justified one. An agent may correctly execute a transaction that subsequently produces an unfavorable result because the environment changed; conversely, it may obtain a favorable outcome through a decision supported by inadequate evidence. Evaluating only the final outcome would confuse the quality of judgment with favorable circumstances, while evaluating only policy compliance would overlook whether the intervention achieved its intended purpose. A credible assurance system must preserve both the decision’s contemporaneous justification and the evidence needed to assess its consequences.

The relationship between verification and learning is particularly demanding. Once an outcome is observed, the Company Brain may update records concerning what happened, but it should not automatically convert temporal succession into causal knowledge. If a supplier delivers after an agent changes a schedule, the organization can establish that the delivery followed the intervention; determining whether the schedule change caused or materially contributed to the result may require additional assumptions, comparative evidence, or experimentation. The proposed architecture must therefore treat observational updates, inferential revisions, and changes to executable operating rules as related but distinct processes.

This distinction also protects institutional memory from self-confirmation. Agent-generated explanations can be stored as interpretations, but their repeated reuse should not confer additional evidentiary authority unless independent evidence justifies it. When a material source assertion is corrected, the system should identify dependent conclusions and decisions requiring reconsideration, while preserving the historical fact that earlier actions occurred under the previous representation. Such dependency-aware revision connects the Company Brain to the older research tradition of truth-maintenance systems, but extends the problem into organizational processes where changing a belief can require reassessing commitments, notifying affected parties, or revising an approved operating procedure.¹⁶¹

Assurance consequently includes more than cybersecurity monitoring or model evaluation. It concerns whether the institution can reconstruct, explain, and contest its actions across time, and whether independently established outcomes can influence the representations and procedures through which future actions are selected. Its evidence should remain available to the appropriate corporate, professional, and technical authorities without requiring them to trust the agent’s own account of its performance.

This is the engineering foundation of the epistemic continuity introduced earlier. The organization preserves not merely the history of its software executions, but the relationship between what it believed, what it was authorized to do, what it actually did, and what subsequent evidence revealed.

Consistency, failure semantics, and bounded ignorance

Whether the proposed architecture can operate outside controlled demonstrations depends on several requirements, beginning with the handling of partial and asynchronous state. A cognitively integrated enterprise may depend on dozens of applications that cannot participate in a single distributed transaction and may disagree temporarily about the condition of the business. Eventual consistency is often acceptable for analytics, reporting, and certain planning tasks, but it is not automatically acceptable for financial commitments, scarce-resource allocation, or control operations whose consequences depend on a current authoritative state. The architecture must therefore identify which invariants require strongly coordinated enforcement and which processes can tolerate delayed reconciliation.

The second requirement concerns the distinction between durable process state and external side effects. A workflow engine can preserve its progress and retry failed activities, but it cannot assume that an unsuccessful response means the requested external action did not occur. A timeout after a payment instruction, for example, creates an uncertain state until the authoritative financial system establishes whether the payment was accepted or committed. The appropriate response may be to reconcile before retrying, rather than to regenerate the instruction. Where reversible business steps span multiple systems, compensating transactions may support recovery, but such compensation must be defined by business semantics and cannot be treated as a universal equivalent of atomic rollback.

The third requirement is temporal and semantic validity. A candidate action may have been justified by evidence that has since changed, and an approved workflow may continue long enough for its underlying mandate or policy to be revised. The architecture must determine when an earlier decision remains valid, when it requires renewed authorization, and when a previously approved process must be suspended. The necessary controls depend on the business operation: a routine report may tolerate modest data latency, whereas a major financial transaction or a physical control command may require current-state confirmation immediately before execution.

The fourth requirement concerns continuity across changes to software and organizational structure. The replacement of a model, agent runtime, workflow engine, or vendor should not erase business commitments or institutional decision records. Stable organizational identifiers, versioned semantic mappings, portable evidence, and explicit migration procedures help preserve this continuity, although complete portability may be constrained by proprietary services, legal restrictions, or practical differences among implementations. The cognitive enterprise must be able to distinguish its enduring institutional objects from the software components temporarily representing or operating upon them.

The fifth requirement, and perhaps the most important, is an explicit boundary of institutional ignorance. The Company Brain must not be designed as a system that always produces an answer or an action. There will be circumstances in which relevant observations are unavailable, sources disagree, the confidence of an inference is inadequate, the applicable authority is uncertain, or the requested outcome lies outside the validated operating envelope. A mature architecture must represent these situations as legitimate operational states, permitting the process to defer, request additional evidence, escalate to an authorized participant, continue in a restricted mode, or decline the proposed action.

The correct response depends on the domain. A delayed commercial recommendation may be economically inconvenient but acceptable; an unverified payment destination should generally prevent execution until the required validation is completed; and an uncertain condition in an industrial process may require control to remain within the locally approved safety envelope rather than be resolved through further generative reasoning. The choice between stopping, continuing, escalating, or entering a degraded mode is itself an institutional and engineering decision, to be specified according to the consequences of failure.

The architecture should thus make safe degradation and controlled recovery first-class capabilities. Agent unavailability need not automatically stop every deterministic business process, just as the failure of the Company Brain should not disable independent safety systems or authoritative enterprise controls. Conversely, a process whose permitted execution depends on unavailable evidence or policy services should not silently continue with broader privileges or stale assumptions. Reliability requires explicit dependency analysis and fallback behavior, not merely redundant model endpoints.

These requirements become especially stringent in cyber-physical environments. NIST SP 800-82 Rev. 3 distinguishes OT security from conventional IT security because the controlled systems interact with physical processes and must satisfy operational, reliability, and safety constraints.¹⁶² An industrial Company Brain may ingest supervisory telemetry, engineering information, and maintenance histories, but that does not justify giving its agent runtime direct authority over field devices or protective functions. Read-oriented interfaces, appropriately segregated operational networks, restricted command mediation, and independent local controls must preserve the distinction between economic or organizational reasoning and the authority required to influence physical processes.

Consider an energy-storage facility in which the Company Brain combines commercial obligations, maintenance history, equipment limits, and operating telemetry. An agent may identify that a revised dispatch schedule appears economically advantageous, but the conclusion depends on the freshness and quality of the available measurements, the validity of the contractual assumptions, and the applicable grid and plant constraints. The resulting proposal must pass through the designated operational authority and control interfaces; an acknowledged communication request is not proof that the plant reached the intended physical condition, and an agent must not treat missing telemetry as permission to infer a safe state. Where evidence is insufficient, the appropriate architectural behavior may be to withhold the proposed intervention while allowing the plant’s established local control and protection systems to continue operating.

This is not a rejection of autonomous reasoning in industrial enterprises. It is the recognition that higher-level organizational cognition and lower-level physical control operate under different timing, assurance, and safety requirements. The architecture should permit their controlled cooperation without collapsing the trust boundaries that distinguish them.

A complete business process through the architecture

The interaction among these layers can be illustrated without treating the organization as a single reasoning process: suppose an industrial manufacturer receives a supplier notification indicating that a critical component may not arrive on its contracted date. The notification enters through an external communication interface and is retained as a source assertion rather than immediately accepted as an established operational fact. The integration layer identifies the supplier and affected purchase orders, records the notification’s provenance and temporal context, and correlates it with ERP commitments, inventory reservations, production requirements, and relevant contractual terms. The Company Brain can then expose a qualified representation of the disruption, including conflicting evidence or missing information, while a durable workflow initiates the organization’s defined supply-risk process.

A coordination agent retrieves the relevant context and invokes specialized services to estimate affected production orders, compare alternative suppliers, examine technical substitution constraints, and construct candidate responses. Some evaluations can be performed through deterministic calculations or approved decision models; others may require probabilistic forecasts or human expertise. Where the evidence is insufficient to support a consequential conclusion, the workflow records the uncertainty and obtains additional information rather than silently converting the most plausible interpretation into an authoritative fact.

The resulting process may assign different responsibilities to different actors. A procurement agent prepares sourcing options; a planning service proposes alternative schedules; an engineer evaluates component compatibility; a commercial specialist assesses customer commitments; and a finance service calculates exposure. In the human–agent enterprise, these activities may involve employees and agents working through the same institutional context, whereas the employee-free firm may rely more extensively on computational actors and externally contracted expertise. In both cases, the role assignments and permissible decisions are determined by the organization’s executable institutional model.

Suppose the preferred response requires an additional purchase within a preauthorized category but above the amount available under the agent’s remaining mandate. The agent can formulate the proposal and assemble its supporting evidence, but it cannot transform that proposal into a valid commitment by generating an instruction to the ERP. The authorization mechanism must evaluate the current resource limits, supplier status, segregation-of-duties constraints, and applicable delegation rules; if the required authority is absent, the process is escalated to the designated decision-maker, and the transaction remains uncommitted until the appropriate authorization is obtained.

Following authorization, the transaction gateway invokes the approved ERP interface and records the execution result, together with the evidence and policy context necessary for subsequent reconstruction. Yet the workflow remains active because an accepted purchase order is only an intermediate business event. The process continues to monitor supplier acknowledgment, delivery, inspection, production availability, and the effect on customer commitments, distinguishing each observed outcome from the original expectation.

If the substitute component later fails qualification, the system must not simply record an unsuccessful procurement transaction. It should preserve the evidence of the technical incompatibility, identify which planning assumptions depended on the proposed substitution, and initiate the appropriate correction of affected schedules and commitments. Where the qualification failure reveals a previously undocumented engineering constraint, that constraint may become a candidate addition to the Company’s institutional knowledge, subject to the evidentiary and approval requirements appropriate to its significance.

The completed process thus generates more than a collection of agent conversations. It creates a durable institutional history connecting an external disturbance to its interpretation, the alternative responses considered, the authorities exercised, the transactions committed, the independently observed consequences, and the lessons that can legitimately be carried into future decisions. The Company Brain becomes more useful because the organization has preserved the relationships among these elements, while the authorization and transaction systems retain their independent responsibilities.

The example also reveals how the economics of the cognitive enterprise could differ from those of conventional organizations. Much of the informational reconstruction, cross-functional coordination, and procedural follow-up can potentially occur without repeated manual intervention, but the savings are meaningful only if the costs of integration, verification, exception handling, and recovery remain proportionate to the benefits. The architecture’s performance must therefore be evaluated through the complete process outcome, not the number of tasks completed by agents.

From reference architecture to an assurance argument

The architecture described here is a research-informed engineering proposal, not an already demonstrated universal solution for autonomous enterprise operation. Its constituent technologies exist in different stages of maturity, and many can be deployed today: enterprise integration, knowledge graphs, durable workflows, policy engines, transactional gateways, distributed tracing, and independent audit mechanisms are established engineering practices. What remains less mature is their coherent integration with probabilistic reasoning, dynamic agent coordination, persistent institutional memory, and the controlled adaptation of organizational structures and mandates.

The correct engineering objective is therefore not to demonstrate that every process can be automated, but to establish which properties can be guaranteed, which can be tested empirically, and which remain dependent on uncertain behavior or external conditions. Access controls may be formally checked against specified policies; transaction systems can enforce selected consistency invariants; process models can be tested for deadlocks or invalid transitions; and event histories can support reconstruction of executed workflows. None of these achievements proves that the organization’s underlying knowledge is complete, its causal assumptions are valid, or its business objectives are well chosen.

A credible assurance argument must connect the specified institutional requirements to the mechanisms implementing them and to evidence that those mechanisms behave as required under relevant conditions. For example, the claim that an agent cannot independently authorize expenditure above its mandate requires evidence not only that the policy engine rejects such a request, but that no alternative interface bypasses enforcement, that concurrent commitments preserve aggregate limits, and that changes to the mandate follow an authorized process. Similarly, the claim that a Company Brain supports reliable decision-making requires more than successful retrieval tests; it requires evaluation of temporal validity, source provenance, conflicting evidence, dependency-aware revision, and the effect of those properties on actual decisions.

The resulting architecture should thus be understood as a system of enforceable institutional boundaries supported by interoperable cognitive and operational services. Its principal contribution is not the introduction of a new universal software component, but the explicit representation of relationships that conventional enterprise systems have often managed separately: what the organization believes, which actor is acting, what that actor may legitimately do, what operation was actually performed, and what evidence establishes its consequences.

Its viability depends on maintaining these distinctions under normal operation, component failure, adversarial manipulation, organizational change, and incomplete information. The Company Brain must remain capable of revising its representations without silently rewriting the historical evidence on which past decisions depended. Agents must be replaceable without losing the continuity of organizational roles and commitments. Policies must be enforceable independently of model-generated reasoning. Execution must remain attributable to authoritative systems, while verification must retain sufficient independence to detect errors shared across agents and knowledge services.

These properties express the deeper meaning of increasing organizational formalization. A firm has long represented selected parts of itself through procedures, accounting systems, contracts, organizational charts, and enterprise applications; the cognitive enterprise would extend this process by making additional relationships among knowledge, roles, decisions, authority, and outcomes computationally explicit. The result would not be a completely formalized company, because uncertainty, tacit knowledge, conflicting interests, and institutional legitimacy cannot be eliminated by software. It would instead be an organization capable of formalizing those aspects of its operation for which explicit representation and controlled execution create demonstrable value, while preserving legitimate mechanisms for judgment and adaptation outside the boundaries of automation.

The most important architectural consequence is therefore that the enterprise can become increasingly autonomous without requiring a single autonomous intelligence to govern it. Organizational cognition can remain federated, agency can remain distributed, decision authority can remain institutionally bounded, and accountability can remain anchored in legally recognized actors and verifiable processes. The architecture connects these functions closely enough to support coordinated action while preserving the separations necessary to detect mistakes, contain failures, and exercise control.

A cognitive enterprise is not engineered by connecting a powerful model to every corporate application. It is engineered by making organizational knowledge actionable without making inference authoritative, making execution autonomous without making authority unlimited, and making institutional memory persistent without making its accumulated judgments immune to correction.

That is the technical meaning of the firm that thinks: not a company whose activities are directed by one computational mind, but an institution whose distributed capacity to observe, reason, coordinate, act, verify, and learn is made progressively explicit, durable, and governable through its architecture.

A supply-chain disruption, executed two ways

The preceding architecture can be evaluated more meaningfully through a consequential organizational process than through a demonstration in which several agents exchange messages and complete isolated tasks. If the Company Brain is to constitute a genuinely new institutional capability, its contribution must become observable when knowledge is incomplete, departmental objectives conflict, the economic consequences of error are substantial, and legitimate authority cannot be inferred from the technical ability to execute an operation. A disruption in a critical supply chain provides a particularly useful test because it requires the organization to connect evidence from multiple sources, formulate decisions under uncertainty, coordinate interdependent activities, and preserve accountability across transactions whose consequences may extend well beyond the original event.

Consider a manufacturing company producing specialized electronic equipment, with a significant proportion of its production dependent on a component obtained from two qualified suppliers. The primary supplier ordinarily offers favorable prices and predictable delivery, while the secondary supplier provides an approved alternative at a higher cost and with more variable availability. The company operates through an ERP system containing purchasing, inventory, production, and financial records; a product-lifecycle system maintaining engineering specifications and approved component substitutions; a CRM environment recording customer commitments; and a collection of supplier portals, contractual documents, and operational reports. Its Company Brain integrates these sources through explicit semantic relationships and temporal evidence, while durable workflows and bounded agents support procurement, planning, customer operations, and financial administration.

The scenario begins when the primary supplier sends an email warning that a production incident may delay several outstanding deliveries. The message identifies affected purchase orders but gives only a provisional recovery date, and its description of the disruption cannot immediately be reconciled with information available through the supplier’s authenticated portal. Meanwhile, the ERP reports enough nominal inventory to support some upcoming production orders, although part of that inventory is reserved for other customers, some material has not completed quality inspection, and recent warehouse movements may not yet have been fully reconciled. The alternative supplier offers additional capacity, but its quoted prices are changing, its available quantities are not contractually guaranteed, and the company faces liquidity constraints that limit the amount of material it can purchase in advance.

The problem is therefore not simply to determine whether an order should be placed. The organization must establish what it knows about the disruption, how the uncertainty propagates into production and customer commitments, which interventions are economically justified, who possesses authority to approve them, and how the eventual consequences will be verified. The same event can require action by procurement, engineering, planning, commercial operations, finance, and corporate management, even though none of these functions possesses the complete information needed to resolve it independently.

Research on supply-chain disruptions explains why the interdependencies matter. Lee, Padmanabhan, and Whang demonstrated how demand information can become distorted as it propagates through supply chains, producing the familiar bullwhip effect; subsequent work on disruption propagation has examined how local failures affect connected suppliers, manufacturers, and downstream customers.¹⁶³¹⁶⁴ An organization that responds to a disruption by accelerating every local replenishment decision may therefore worsen aggregate scarcity, increase financial exposure, or create inconsistent commitments. The purpose of shared organizational cognition is not to make every agent react faster, but to enable the enterprise to recognize the dependencies that make some apparently reasonable local responses collectively undesirable.

The disruption enters the organization as evidence, not authority

The first organizational act is to create a qualified record of the reported event, without treating the supplier’s email either as an established fact or as a legitimate instruction to the company’s agents. The intake service records the source, message identity, reception time, referenced purchase orders, and relevant attachments, while maintaining the distinction between what the supplier asserts and what the company’s own systems can independently establish. Any instructions embedded in the email concerning internal procedures, purchasing decisions, changes of payment details, or the use of particular tools remain untrusted external content, regardless of how convincingly they are expressed.

A supplier-risk agent may then consult the authenticated supplier portal, retrieve previous communications, compare contractual delivery obligations with reported production conditions, and identify discrepancies requiring further investigation. The exercise involves more than assigning an arbitrary confidence score to the message. An apparent confirmation from a second document provides little additional assurance if both documents originate from the same compromised channel, while a separately obtained operational acknowledgment may materially change the evidentiary position. The Company Brain must preserve these dependencies so that repeated references to the same assertion are not mistaken for independent corroboration.

The distinction is particularly important because the information needed for action may remain uncertain even after the event itself has been authenticated. The supplier may have experienced a genuine production failure without knowing its recovery date, while the available quantity at the alternate supplier may depend on other customers’ orders or on transport capacity outside the organization’s control. The Company Brain can represent these uncertainties, record the assumptions used in alternative scenarios, and maintain references to the underlying evidence, but it cannot convert an uncertain commercial environment into a fully observable state merely through additional reasoning.

The planning process must consequently distinguish several meanings of inventory and supply availability. Material physically located at a warehouse is not necessarily available for unrestricted use; inventory allocated to one customer may be transferable only under specified contractual or operational conditions; and expected inbound deliveries are not equivalent to quality-released stock. A production-planning service should obtain the appropriate authoritative quantities and constraints, reconstruct their temporal applicability, and identify where discrepancies require reconciliation before the company makes new external commitments.

The disruption can then be represented as an evolving institutional case rather than a collection of messages. Its evidentiary state includes confirmed observations, reported claims, unresolved contradictions, and hypotheses concerning future deliveries; its operational state records affected orders, inventory constraints, alternative sources, and outstanding decisions; and its institutional state identifies the mandates and authorities applicable to possible interventions.

This representation supplies a common basis for collective action without implying that every participant must agree on the likely course of the disruption. Procurement may judge the supplier’s recovery estimate credible, while production planning assigns greater weight to observed delivery variability, and finance may regard the liquidity implications of early replenishment as unacceptable. Such differences should remain visible because they concern the assumptions and objectives on which the eventual decision depends.

From shared evidence to economically justified alternatives

Once the initial evidence has been qualified, the Company Brain allows specialized analytical services to reconstruct the disruption’s consequences across the enterprise. A planning model identifies production orders dependent on the affected component, compares their requirements with inventories and scheduled receipts, and evaluates alternative production sequences. Engineering records establish whether other approved components or configurations can satisfy the relevant technical constraints, while contractual information determines which customer commitments are firm, which allow rescheduling, and which changes would require negotiation or authorization.

The resulting analysis should distinguish an operational shortage from an economic exposure. A delayed component may be critical to one high-value order but irrelevant to another product line whose production can be rescheduled. A higher-priced alternate component may still be economical if it prevents substantial contractual penalties or preserves a strategically important customer relationship. Conversely, purchasing additional material may be irrational if the original supplier recovers quickly and the alternative stock cannot be reused economically.

A procurement agent can formulate candidate responses, but the evaluation should be performed through the methods appropriate to each component of the problem. Deterministic constraints can enforce technical compatibility and contractual exclusions; optimization can compare feasible production and purchasing plans; forecasting models can estimate uncertain demand or delivery conditions; and human or computational reasoning can identify assumptions, formulate alternatives, and explain their consequences. The language model need not—and should not—become the authoritative numerical engine for every calculation simply because it coordinates the analysis.

Among the feasible responses, the organization might preserve current commitments while monitoring the disruption, reserve limited alternate capacity without immediately purchasing all available stock, divide procurement between the two qualified suppliers, renegotiate selected customer dates, or alter the production mix to prioritize orders whose fulfillment is both technically feasible and economically significant. A flexible purchasing option may have value precisely because the firm can wait for additional information before committing to the full expenditure. Research on dual sourcing and option contracts under disruption risk provides analytical precedents for treating flexibility and the timing of information as economically relevant decision variables.¹⁶⁵

The alternatives should be compared using a coherent decision model rather than independent departmental recommendations. A cost-minimizing procurement decision may conflict with a production-continuity objective; an aggressive customer-retention strategy may create unacceptable liquidity exposure; and a technically available substitution may introduce qualification risks not captured by delivery forecasts. The Company Brain can make these conflicts explicit, but corporate authority must determine which trade-offs are legitimate when they cannot be resolved through previously approved rules.

This is where the distinction between evidence and institutional authority becomes operational. A planning service may establish that an additional purchase would reduce expected delivery losses, but that finding does not automatically authorize the company to exceed its spending limits. Likewise, a financial model may show that postponing a purchase improves short-term cash flow without establishing that breaching an existing contractual obligation is permissible.

The organization must therefore evaluate the alternatives against both the available evidence and the executable institutional constitution. Actions within predefined purchasing categories, exposure limits, approved supplier lists, and technical constraints may proceed through delegated workflows, whereas exceptional commitments, material changes to customer obligations, or revisions to the company’s risk appetite must follow the relevant decision procedures.

Execution by the employee-free operating firm

In the first organizational extreme, the operating process is implemented predominantly by computational actors, while the company remains subject to legally required governance and to its existing contractual and regulatory obligations. The supplier-risk, planning, procurement, commercial, and financial roles are occupied by agents or deterministic services, and their activities are coordinated through a durable workflow whose authority derives from the mandates approved by the corporation’s competent organs.

The absence of employees does not remove the need for functional specialization. Indeed, the architecture makes specialization more explicit because different roles must possess different permissions, evidence requirements, and decision boundaries. A procurement agent responsible for selecting suppliers should not automatically possess authority to modify the supplier master record or approve payment destinations, while a financial execution service should not be able to infer the legitimacy of a transaction from the procurement agent’s assertion that it was commercially advantageous. The organizational separation of duties remains necessary even when the participants are software services rather than employees.

Suppose the alternative supplier offers a quantity that can cover the most urgent production shortfall, but accepting it would require a price premium. If the purchase falls within an existing mandate, the proposed operation may pass through the independent authorization gateway, which checks supplier qualification, permitted price variance, the aggregate financial exposure already committed by the company’s agents, and the current transaction preconditions. The gateway then invokes the restricted ERP interface, while the workflow retains the authorization context and reconciles the result against the authoritative purchase order.

The execution path is significant because no agent should be allowed to enlarge its mandate by arguing that the transaction has positive expected value. The authority to purchase is defined by the institution; the agent’s reasoning determines whether to propose a purchase within that authority. Similarly, a change in the supplier’s payment details would require a separately controlled verification procedure and should not become effective merely because the supplier’s email contains the proposed account information.

Where the economically preferred response exceeds the available mandate, the workflow must escalate the decision to the appropriate legally recognized decision-maker or other authorized delegate. The person may review the relevant evidence, compare the available scenarios, obtain external professional advice, and approve, reject, or modify the proposal. The approval is not ceremonial: it must correspond to a decision right that the agents do not independently possess, and subsequent execution must remain constrained by the actual terms of the authorization.

A more difficult question arises when the organization cannot obtain adequate evidence before an important commercial deadline. The agents might estimate that waiting creates substantial expected losses, while purchasing exposes the company to inventory obsolescence and liquidity pressure. If neither alternative falls within an existing mandate, the organization cannot legitimately manufacture authority by repeatedly revising its forecasts or negotiating among agents. It must obtain an authorized decision, operate within a valid contingency policy, or accept the cost of non-action.

This possibility is central to the analysis of autonomy. An institution does not become more autonomous merely by eliminating the occasions on which it admits uncertainty or requires external authorization. Its autonomy is meaningful only within the operating envelope for which delegated execution is legitimate and sufficiently reliable.

The employee-free firm also faces a distinct problem of institutional adaptation. Repeated supply failures might motivate an agent to recommend changing the company’s sourcing strategy, increasing reserve inventory, or revising financial limits. The agent may analyze historical outcomes and propose amendments to the applicable rules, but these proposals concern the organization itself rather than ordinary execution under an established policy. Their adoption must follow the separately governed process for constitutional or strategic change.

The scenario therefore reveals both the possibility and the limitation of employee-free operation. Routine procurement coordination may be extensively automated, but a viable operating institution must also maintain the capacity to recognize situations exceeding its mandate, preserve evidence for legitimate decision-makers, and recover when the computational operating arrangement becomes unreliable.

Execution by the cognitively integrated human–agent firm

The second organizational extreme uses substantially the same technological architecture but preserves human employees as participants in the ordinary operating process. Its distinctive capability arises not from requiring people to supervise every agent action, but from making their situated expertise available within a persistent and interoperable organizational context.

The supplier disruption may initially be detected and analyzed through the same event-ingestion, semantic-integration, and planning mechanisms. Yet the resulting case can immediately be examined by a procurement manager with direct knowledge of the supplier relationship, a production engineer familiar with component qualifications, a commercial manager responsible for significant customer commitments, and a financial controller evaluating liquidity and exposure. Each participant encounters the same underlying institutional case while retaining the ability to inspect source records, challenge assumptions, and contribute information unavailable to the computational representation.

The procurement manager may know that the alternate supplier has recently experienced internal organizational changes that have not appeared in the recorded performance history. The production engineer may identify a qualification constraint affecting one equipment configuration but not another. Finance may establish that a temporary inventory increase would violate a covenant or consume liquidity required for a different obligation. None of these judgments automatically invalidates the others, and none should disappear merely because a shared optimization service produces one numerically preferred response.

The Company Brain can improve the quality of collaboration by making the competing constraints explicit. Rather than holding successive meetings in which each department reconstructs the situation from incomplete reports, the participants can examine a common representation of the relevant evidence, assumptions, and consequences. An agent can recompute a production scenario after the engineer rules out a proposed substitution; another can recalculate cash exposure under a revised purchasing plan; and the commercial manager can assess which customer commitments would remain feasible.

This form of collaboration does not abolish organizational politics or the need for managerial authority. Sales may reasonably prioritize preserving commercial relationships, operations may emphasize production stability, and finance may seek to protect liquidity. The resulting conflict reflects legitimate differences in organizational objectives, not necessarily defective communication or insufficient data. The executable institutional model must establish which decisions can be made locally and which require an authorized resolution among competing priorities.

A further advantage of the human–agent model is the possibility of preserving independent interpretations within the operating process. The engineer’s objection may reveal a missing technical constraint, while the procurement manager’s concerns may challenge a statistically favorable supplier assessment. If the Company Brain preserves their contributions with appropriate provenance and review status, the organization can improve its representation without forcing a premature consensus.

The same infrastructure can also create a failure if the participants defer to an apparently complete computational recommendation instead of exercising their expertise. Universal cognitive access does not guarantee independent judgment, and a polished explanation supported by multiple agents may still rest on one incorrect inventory record or supplier assertion. The institutional design must therefore make it practical to challenge recommendations, recover primary evidence, and identify when ostensibly independent assessments share the same underlying assumptions.

The organizational difference between the two extremes consequently concerns more than the location of approval steps. In the employee-free model, independent evidence and critical judgment must be supplied through explicitly engineered computational or external mechanisms, whereas the human–agent firm can also draw upon continuously participating specialists whose knowledge is partly situated and social. Neither model is inherently more reliable under all conditions; their comparative performance depends on the nature of the disruption, the quality of the evidence, the competence of the participants, and the effectiveness of their decision and verification procedures.

Recovery, reconciliation, and institutional learning

The process cannot be considered complete when the procurement agent successfully places an order or when the organization approves a revised production schedule. Those events establish that particular transactions or decisions occurred; they do not demonstrate that the supply disruption was resolved or that the chosen intervention produced the intended economic result.

The durable workflow must therefore monitor the subsequent fulfillment of the agreed commitments, reconcile supplier acknowledgments with purchasing records, distinguish expected deliveries from confirmed receipts, and incorporate the results of quality inspection before declaring material available for production. Where customer delivery commitments were changed, the system must record the terms actually accepted by the relevant counterparties rather than treating an internally generated plan as proof of contractual agreement.

Suppose the alternate supplier acknowledges the order but fails to deliver on time, while the primary supplier recovers earlier than expected. A simplistic evaluation might record the alternative purchase as an unsuccessful decision and infer that the organization should avoid the supplier in the future. Yet such a conclusion would ignore the uncertainty present when the decision was made and the possibility that the original choice reduced expected risk even though an unfavorable outcome occurred.

A more defensible assessment preserves the information available at the decision point, the assumptions underlying the selected action, the alternative scenarios considered, the authority exercised, and the outcome subsequently observed. The organization’s updated assessment of supplier reliability can incorporate the new delivery evidence, but any stronger causal claim requires consideration of relevant confounders and the circumstances under which the suppliers operated.

This distinction is essential to organizational learning. A justified decision can produce an unfavorable result, while a poorly supported decision can succeed by chance. If the Company Brain rewards only favorable outcomes, it may reinforce unreliable reasoning; if it evaluates only procedural compliance, it may preserve ineffective policies. The institution must retain enough evidence to distinguish the quality of the original decision from the realized consequences of uncertainty.

The same applies to operational failures. A purchasing transaction that was technically authorized but executed twice because of an unsafe retry reveals a defect in the execution architecture; an order based on a stale inventory figure reveals a temporal or reconciliation failure; and a supplier-bank change accepted from an untrusted message reveals a failure of evidence promotion and financial authorization. These incidents should not be aggregated into a generic category of agent error, because they require different remedies.

The Company Brain becomes a useful institutional memory when the organization can identify such mechanisms, revise the appropriate representations or controls, and determine which earlier or pending decisions depend on the corrected assumptions. This may involve updating supplier-risk models, changing inventory data contracts, strengthening transaction preconditions, revising escalation procedures, or amending a formal operating policy through the applicable governance process.

The scenario also provides a practical test of epistemic resilience. If the original supplier notice is later shown to have been manipulated, the organization should be able to identify the decisions derived from it, determine which transactions remain reversible, and distinguish genuinely independent supporting evidence from later summaries of the compromised assertion. The value of a shared cognitive infrastructure is demonstrated not only by how rapidly it propagates relevant information, but by how effectively it contains and corrects the consequences of misinformation.

Ultimately, both organizational extremes must close the same evidentiary and operational loop: observations must be qualified, alternatives formulated, decisions authorized, transactions executed, outcomes independently established, and institutional knowledge revised through justified conclusions. The distribution of human and artificial participants differs, but the logical requirements for a governable organization remain.

What the experiment establishes—and what it does not

The comparison illustrates why the Company Brain and agent coworkers should be treated as complementary components rather than alternative paths toward organizational intelligence. The Brain makes evidence and institutional context available; agents and human participants formulate and coordinate interventions; the executable institutional model constrains the authority under which they operate; and independent verification supplies the observations required for correction and learning.

The experiment also clarifies the meaning of organizational autonomy. The employee-free firm may execute more of the routine process without human participation, but it does not thereby possess broader legitimacy to alter contractual obligations, redefine corporate objectives, or bypass uncertain evidence. The human–agent firm may require more direct human participation in selected decisions while achieving superior process performance through the effective integration of specialized knowledge. Consequently, autonomy, decision quality, operational efficiency, and accountability must be evaluated separately rather than compressed into one measure of technological maturity.

There is also no reason to presume that a more computationally integrated process is always economically preferable. The Company Brain creates value if its coordination and decision improvements outweigh the costs of semantic integration, evidence validation, agent execution, oversight, and recovery. A conventional planning system with effective human coordination may outperform an elaborate agent architecture when the process is stable or the required expertise is inexpensive to mobilize. Conversely, complex disturbances spanning numerous contracts, technical constraints, and rapidly changing operational states may provide stronger opportunities for integrated organizational cognition.

The proposed supply-chain case is therefore a test specification rather than evidence that the proposed architecture already works. A meaningful experiment would need to vary the reliability of source information, disruption severity, availability of alternate suppliers, financial constraints, authorization limits, and opportunities for human intervention. It should measure complete-process outcomes, time to justified decision, total coordination and assurance costs, unauthorized commitments, error propagation, recovery, and the extent to which verified experience improves subsequent responses.

The most revealing experimental conditions would be those in which apparently plausible interpretations are incorrect, where the economically attractive action lies outside the agent’s authority, or where a successful intermediate transaction fails to achieve its business purpose. Such cases distinguish an architecture capable of sustaining institutional judgment from one optimized merely to complete visible tasks.

The central lesson is that an enterprise does not demonstrate organizational intelligence simply by acting quickly or autonomously. It demonstrates a more demanding capability when it can act on justified evidence, preserve legitimate authority, recognize when its assumptions fail, and correct its behavior without confusing successful execution with organizational success.

The supply-chain disruption makes this proposition concrete. It shows how a firm could become more computationally capable without surrendering the institutional distinctions on which reliable economic activity depends, while also revealing the substantial gap between specifying such a system and demonstrating its sustained operation.

The next question must therefore be empirical: how much of this architecture, and how much of the organizational capability it is intended to produce, has actually been demonstrated by November 2026?

Evidence as of November 2026: agents work, but firms are harder

The architecture developed above is an engineering hypothesis assembled from technologies with different levels of maturity, not evidence that a self-sustaining cognitive enterprise has already been constructed. The relevant empirical distinction is between completing bounded tasks, operating a reliable business process, and preserving the continuity of an institution under changing conditions. Although generative agents can already use external tools to undertake consequential work, task success establishes neither the legitimacy of delegated authority nor the persistence of organizational memory, while a controlled laboratory demonstration provides only limited evidence of how a process will behave when suppliers, markets, contractual obligations, and the underlying software change. The research record as of November 2026 must consequently be assessed according to what each study actually measures and what remains outside its experimental boundary.

Tool-using agents: a changing benchmark frontier

Early realistic benchmarks revealed the difficulty of sustaining tool-mediated activity even when an agent possessed strong general reasoning capabilities. WebArena’s 2024 GPT-4-based baseline achieved 14.41% end-to-end success in its reproducible web environment, against 78.24% for the human participants in that evaluation; TheAgentCompany’s initial evaluation reported approximately 24% autonomous task completion for its strongest tested agent in a simulated software-company workplace.¹⁶⁶¹⁶⁷ These results were important because their evaluators assessed actual changes in environment state rather than persuasive descriptions of actions. They are historical baselines, however, and would misrepresent the frontier if presented as current November 2026 performance.

Progress is substantial but benchmark-dependent. Stanford’s 2026 AI Index reported a leading result of 66.3% on the original OSWorld computer-use benchmark, approaching its reported human baseline of approximately 72%.¹⁶⁸ More demanding evaluations continue to expose failures: OSWorld 2.0, introduced in June 2026, contains 108 longer workflows whose median estimated human completion time is approximately 1.6 hours; its strongest evaluated configuration completed 20.6% under the primary 500-step binary-completion protocol.¹⁶⁹ The difference between the benchmarks is not evidence of technological regression, because their tasks, interaction constraints, and scoring protocols differ. It demonstrates instead why a headline success rate cannot be detached from its evaluation environment and treated as a general measure of professional or organizational competence.

Reliability over repeated attempts poses a separate problem. In τ-bench, the authors evaluate agents interacting with users and tools under explicit domain policies, including whether the resulting database state satisfies the requested task; their repeated-trial pass^k metric exposes inconsistency that can be concealed by single-trial task completion.¹⁷⁰ An enterprise, however, must do more than succeed once when the simulated customer follows a specified script. It must preserve commitments across interruptions, recognize that a previous authorization has expired, reconcile ambiguous execution outcomes, and detect when the evidence used by several apparently independent agents has become unreliable. Such requirements are not captured by the accuracy of an isolated model response.

METR’s time-horizon research offers another useful but easily overstated measure. The reported horizon estimates the human-expert task duration at which a system achieves a specified probability of success on the evaluated task distribution; it does not measure how many hours or days an agent can remain operational without oversight, much less whether it can govern a corporation through unfamiliar change.¹⁷¹ In particular, a short financial instruction or industrial command may be highly consequential even though it requires little computational time, while a months-long procurement case may demand only intermittent reasoning but extensive state retention, external coordination, and institutional control.

What production studies and adoption statistics establish

Evidence from deployed systems is indispensable because production agents operate within imperfect enterprise data, heterogeneous integrations, security constraints, changing requirements, and real consequences. The final ICML 2026 proceedings version of Pan and colleagues’ Measuring Agents in Production combines 20 in-depth case studies with a survey of 86 practitioners of deployed systems across 26 domains; earlier versions reported a different survey count, and should not be conflated with the final publication.¹⁷² The study reports that 68% of its observed deployments executed no more than ten steps before human intervention and that 74% relied primarily on human evaluation. These are descriptive properties of the studied implementations, not representative population estimates or a demonstration that longer autonomous processes are impossible. Indeed, limited autonomy can reflect rational risk allocation when failure is expensive or evidence is incomplete.

EnterpriseBench, published in the October 2026 Findings of EMNLP, provides an additional perspective through interactive business settings involving consulting, inventory decisions, and enterprise simulation; the evaluated methods did not exhibit stable, comprehensive reliability across its heterogeneous tasks.¹⁷³ The benchmark is relevant because business decisions depend on uncertain future states, but an interactive simulation cannot reproduce every legal obligation, social relation, external counterparty, or physical dependency of an actual firm. It tests important components of enterprise reasoning without establishing the viability of a complete institutional architecture.

Adoption statistics must be interpreted with equal care. The April 2026 NBER analysis of the US Census Bureau’s business survey estimated that 18% of firms used AI in at least one business function during November 2025–January 2026, or 32% when weighted by employment; among adopting firms, 57% used it in no more than three functions.¹⁷⁴ These measurements concern AI adoption under a defined survey instrument, not the proportion of organizations using persistent agents, executable institutional mandates, or a Company Brain. Employee use of a conversational assistant, departmental deployment of generative models, and delegated execution of a regulated business process should not be counted as equivalent organizational states.

Controlled workplace studies nevertheless demonstrate that computational assistance can create real economic value. Brynjolfsson, Li, and Raymond identified a 15% average improvement in issues resolved per hour among 5,172 customer-support workers with access to a generative assistant, with greater gains among less experienced workers; Dell’Acqua and colleagues’ experiment with 791 professionals showed that AI could alter the performance and cross-functional character of product-innovation work.¹⁷⁵¹⁷⁶ Conversely, METR’s randomized experiment involving experienced open-source developers found a 19% increase in completion time with the early-2025 tools and tasks tested, while its later work could not reliably quantify the effects of newer systems because of measurement limitations.¹⁷⁷¹⁷⁸ These findings are compatible once their distinct tasks, populations, intervention designs, and technology vintages are respected. They establish that benefits are conditional, not that integrated organizational cognition has already been causally identified.

The missing unit of evidence: institutional continuity

The most important missing evidence concerns whether persistent, governed, shared knowledge adds value beyond more capable models, ordinary information retrieval, and better workflow engineering. Faster procurement, for example, could reflect a redesigned approval route rather than an improvement in institutional memory; successful autonomous ordering could reflect deterministic transaction constraints rather than general agent intelligence; and better operational forecasts could result from higher-quality source data without a Company Brain. An evaluation that changes all these mechanisms simultaneously cannot establish which produced the observed improvement, and a vendor’s demonstration of one successful process cannot establish reliability over changing commercial and regulatory conditions.

The stronger evidence would come from complete business processes observed longitudinally, including their exception rates, total economic costs, policy compliance, independent outcome verification, and response to unforeseen disturbances. Researchers would need to determine whether an organization preserves decision-relevant knowledge when employees depart and models are replaced; whether apparent agent consensus survives independent evidence checks; whether its cumulative commitments remain bounded under concurrency; and whether authorized participants can intervene when a shared assumption is shown to be wrong. Rare but severe failures present a further statistical challenge: average completion rates may be informative for routine efficiency while providing little evidence about the tail risk of financial loss, contractual liability, or cyber-physical harm.

The empirical conclusion as of November 2026 must therefore remain conditional. Tool-using agents perform nontrivial tasks, well-designed AI assistance can improve selected forms of work, and businesses are deploying increasingly consequential computational actors; the evidence does not yet establish, as a general capability, a firm that sustains all essential economic and institutional functions through extended unfamiliar change without repeated human operational correction. Nor does the current record isolate the incremental benefit of shared, provenance-aware organizational cognition at enterprise scale. The scientific frontier lies in demonstrating that coordinated knowledge, bounded action, independent verification, and institutional adaptation produce a capability that remains reliable beyond the particular models and participants through which it is implemented.

A research program that could falsify the theory

The preceding sections have developed a theory of the cognitive enterprise in which organizational knowledge, agency, autonomy, and accountability become increasingly explicit and computationally integrated, while remaining analytically separable properties of the institution. The Company Brain provides the proposed infrastructure for preserving and connecting organizational evidence, reasoning, decisions, and experience; agent coworkers extend this infrastructure into the execution of business processes; and the executable institutional constitution determines the conditions under which that execution becomes authorized. The resulting architecture is technically plausible because many of its constituent mechanisms already exist, but its theoretical importance depends on a stronger proposition: their integration must generate observable organizational capabilities that cannot be explained adequately by more powerful models, conventional enterprise integration, improved information retrieval, or ordinary workflow automation.

This distinction determines whether the Company Brain deserves recognition as a substantive contribution to the theory of the firm or remains a useful architectural metaphor. A company may improve its performance by implementing better search, eliminating duplicate records, automating approvals, or redesigning inefficient processes without acquiring the kind of persistent collective cognition proposed in this article. These improvements are economically valuable, but their existence would not establish that an organization has developed a new capacity to preserve and revise the grounds of its judgments, coordinate human and computational actors around that knowledge, and convert verified experience into institutional learning.

The theory must therefore be exposed to evidence capable of distinguishing its distinctive causal mechanisms from simpler alternatives, including evidence that would require abandoning or substantially narrowing its claims. This is particularly important because the cognitive enterprise combines technical and organizational interventions whose effects are likely to be complementary, heterogeneous, and dependent on the environment in which they are introduced. A successful implementation cannot be evaluated solely by asking whether a model completes more tasks, just as an unsuccessful implementation cannot be attributed automatically to deficient model intelligence. Organizational design, process structure, authority constraints, data quality, managerial practice, and the costs of assurance may determine whether the same computational capabilities produce useful outcomes or additional complexity.

The central proposition to be tested is that a governed, persistent organizational cognitive infrastructure can improve the performance, continuity, and adaptability of complete business processes beyond the benefits attributable to individual AI capability and conventional automation, while allowing consequential action to remain institutionally bounded and accountable. The research program must establish whether this proposition holds, for which classes of activity it holds, what its costs are, and under which conditions the integration of cognition and agency becomes counterproductive.

Distinguishing organizational cognition from its technological substitutes

The first methodological requirement is to identify competing explanations for any observed improvement. At least four mechanisms could account for the apparent success of an agent-enabled organization.

One possibility is that employees and software agents obtain information more quickly because enterprise search, semantic integration, and data quality have improved. A second is that more capable models generate better interpretations and plans from essentially unchanged organizational information. A third is that durable workflows and tool-enabled agents reduce the cost of executing activities that were previously performed manually. The fourth, which is distinctive to the proposed theory, is that shared institutional memory and coordinated reasoning improve decisions across tasks and time, allowing knowledge acquired in one context to inform subsequent processes while preserving the evidence needed to correct it.

These mechanisms are not mutually exclusive, and a mature cognitive enterprise would probably rely on all of them. Their joint contribution, however, cannot be established merely by observing that a company performs better after introducing a platform containing all four. Without an appropriate comparison, a reduction in procurement cycle time might result from faster record retrieval, a new approval workflow, more accurate forecasting, or the capacity to reuse the verified reasons supporting previous decisions. Similarly, fewer operational incidents might reflect stronger transaction controls rather than improved organizational cognition.

The theoretical burden is therefore to demonstrate an incremental effect of persistent, shared, and revisable institutional knowledge after accounting for conventional information access, computational capability, and workflow execution. This need not imply that every component has an independently identifiable economic value in every setting; organizational complementarities may make some capabilities productive only when combined. It does require distinguishing a meaningful interaction among capabilities from an undifferentiated claim that the complete architecture performs better because it contains more technology.

Research on human–AI collaboration has already begun to adopt designs that separate individual assistance from collaborative organization. Dell’Acqua and colleagues’ 2026 field experiment, for example, independently varied access to generative AI and whether participants worked individually or in human teams, thereby examining the contribution of AI in different collaborative arrangements.¹⁷⁹ Puranam’s organization-design framework likewise identifies the division of labor and the configuration of learning among humans and algorithms as separate analytical dimensions.¹⁸⁰ These studies do not establish the Company Brain hypothesis, but they provide methodological and conceptual precedents for treating organizational configuration as an intervention rather than an unobserved background condition.

The appropriate unit of analysis would ordinarily be a business process, organizational unit, or network of interdependent units whose outcomes can be independently established. Individual task benchmarks remain useful for evaluating the components, but they cannot identify the proposed institutional effect because the mechanism of interest operates through the connections among decisions, actors, and historical experience.

A factorial experiment on shared cognition and delegated action

A first experiment could independently vary two features of the organizational architecture: the availability of governed shared memory and the availability of bounded autonomous execution. Let B \in \{0,1\} indicate whether the process operates with a persistent Company Brain that preserves provenance, semantics, decision context, and qualified learning; let A \in \{0,1\} indicate whether the relevant tasks can be initiated and executed by agents under predefined mandates. The resulting design contains four configurations.

The baseline employs conventional enterprise applications and human-operated processes, potentially including ordinary search and existing deterministic automation. The second configuration adds shared, provenance-aware organizational memory while retaining essentially the same execution arrangements. The third introduces agents and durable workflows using domain-local information and memory, without the shared institutional cognitive infrastructure. The fourth combines the Company Brain with bounded agents operating through the same workflow and transaction-control mechanisms.

The design must be implemented carefully. In particular, the difference between the third and fourth configurations should not simultaneously introduce better orchestration, more powerful models, broader permissions, and greater access to institutional memory, because doing so would make it impossible to attribute any observed advantage to shared cognition. Where agents are present, their underlying models, execution frameworks, available business capabilities, tool budgets, and mandate boundaries should remain as comparable as the study permits. The intervention represented by B should primarily change the persistence, interoperability, provenance, and reuse of organizational knowledge.

For an outcome Y whose increase represents an improvement, such as verified process value or correctly fulfilled customer commitments per unit of total cost, the average interaction between the two interventions can be expressed as

\tau_{BA} = \mathbb E\!\left[ Y(1,1)-Y(1,0)-Y(0,1)+Y(0,0) \right], \tag{10}

where Y(b,a) denotes the potential outcome under the corresponding configuration. The expression measures departure from additivity on the chosen outcome scale: a positive value would indicate that combining shared cognition and autonomous execution produces a greater effect than the sum of their separate contributions, whereas a value near zero would indicate little evidence of such synergy on that scale.

This interaction is not identical to the total effect of adding the Company Brain to an agent-operated process, which is represented by the contrast \mathbb E[Y(1,1)-Y(0,1)]. Both quantities matter. The former tests whether the combination creates an additional complementarity; the latter tests whether shared organizational cognition improves an otherwise comparable agent architecture. A Company Brain could create a positive standalone benefit without a positive interaction, and a positive interaction could exist even if one intervention performs poorly in isolation. The estimands must therefore be prespecified and interpreted separately, rather than selected retrospectively according to which comparison produces the most favorable result.

The formulation also makes clear that an empirical measure of organizational performance is required. A scalar outcome may be useful for a narrow experimental purpose, but no universal index can legitimately combine productivity, safety, worker autonomy, legal compliance, and financial exposure through arbitrary weights. Primary business outcomes, assurance costs, consequential failures, and institutional constraints should be reported separately, with any composite objective justified by the actual decision context.

A second series of experiments could vary independent outcome verification and institutional challenge mechanisms while holding the underlying cognitive and execution capabilities as constant as possible. This would test whether epistemic resilience results from the accumulation of shared knowledge itself or from the additional controls that prevent shared errors from becoming authoritative organizational decisions. The distinction is important because the architecture proposed in this article does not regard a common memory as sufficient for collective intelligence; it regards the ability to preserve and revise justified knowledge as constitutive of that capability.

Causal identification when the intervention changes the organization

The experimental design cannot assume that employees, agents, and business processes behave independently simply because they are assigned to separate treatment groups. A Company Brain is intended to make knowledge portable across organizational boundaries, and its adoption may change collaboration networks, the availability of expertise, and the decisions made by participants who never directly interact with the platform.

These spillovers are not an incidental nuisance. They may constitute an important part of the mechanism through which organizational cognition creates value.

Hudgens and Halloran’s work on causal inference under interference shows why standard treatment comparisons require modification when one unit’s outcomes can depend on the treatment received by other units.¹⁸¹ In the cognitive enterprise, an engineer with access to shared memory may provide better information to a control-group colleague, while an agent-generated supplier assessment may influence procurement decisions across several organizational units. An apparent control group can therefore receive an indirect form of the intervention even when its members have no direct platform access.

The first response should be to define experimental clusters according to the actual dependencies of the business processes rather than administrative convenience. Where feasible, units with substantial internal information exchange and comparatively limited external interaction can be randomized together. If substantial cross-cluster exchange remains unavoidable, the research design should measure it and, where possible, estimate direct and spillover effects through an explicit exposure model. The attempt to eliminate every spillover by preventing employees from communicating would often produce an artificial organizational environment that undermines the object of the study.

A randomized field experiment is preferable where it is practically and ethically feasible, but the relevant interventions may require organization-wide infrastructure changes that cannot be assigned to small independent units. In such cases, phased or staggered deployments may support quasi-experimental analysis, provided the identification assumptions are examined rather than merely asserted. Difference-in-differences methods, for example, require credible assumptions about counterfactual outcome trends, while heterogeneous effects and staggered adoption can complicate conventional estimators; Callaway and Sant’Anna’s treatment of multiple periods and treatment timing supplies relevant methodological guidance.¹⁸²

The study must also account for learning and adaptation after deployment. Initial productivity may fall because employees are learning the new system, processes are being redesigned, or historical data require correction. Conversely, early performance may appear unusually favorable because the participating units receive intensive technical support or because the first automated tasks are selected for their simplicity. The relevant effect cannot be inferred reliably from a short demonstration period without considering these dynamics.

The evaluation should therefore include a sufficiently long baseline, a defined implementation and stabilization period, and subsequent observation windows capable of revealing changes in process performance, organizational relationships, maintenance costs, and dependence on particular employees or providers. Model versions, governance policies, available tools, and significant organizational changes should be recorded because they can alter the meaning of the treatment over time.

These requirements are particularly demanding for a system that is intended to learn. If the Company Brain changes its representations and procedures through continued use, the intervention is not fully characterized by its initial configuration. The research protocol must specify which changes are allowed as part of normal operation, which require approval, and how changes to the technology or organizational structure will be incorporated into the analysis.

A scientifically meaningful experiment should not prevent the institution from learning merely to preserve a static treatment. Instead, it should define the permitted adaptation mechanisms and evaluate their consequences.

Measuring institutional performance rather than agent activity

The dependent variables must correspond to the claims made by the theory; at the operational level, researchers should measure complete-process outcomes, including successful fulfillment of business commitments, elapsed time from qualifying event to verified completion, resource consumption, rework, and recovery from exceptions. These outcomes should be related to total operating cost, including employees, external providers, computation, integration, supervision, verification, incident response, and the maintenance of organizational knowledge. Counting agent invocations or automatically completed activities would provide useful engineering telemetry, but would not establish economic value.

At the cognitive level, evaluation should examine whether participants and agents obtain accurate, temporally appropriate, and relevant information; whether material decisions remain traceable to supporting evidence; whether conflicting observations are recognized; and whether changes to underlying assumptions lead to justified revision of dependent conclusions. The most distinctive measure would concern decision-relevant knowledge reuse, understood as the contribution of previously acquired and qualified organizational knowledge to subsequent decisions that otherwise would have required renewed investigation or would have produced inferior outcomes.

This measure is difficult because retrieval does not prove usefulness and a decision that changes after receiving additional evidence is not necessarily a better decision. The evaluation must connect knowledge use to independently assessed outcomes or to credible comparisons with decisions made without the relevant institutional memory. Where causal effects cannot be established, the results should be reported as descriptive associations rather than evidence of productive learning.

At the governance level, researchers should measure whether consequential operations remain within delegated authority, whether required approvals are effective, whether cumulative transaction constraints hold under concurrency, and whether control failures can be detected and reconstructed. The existence of a policy document or audit log is not sufficient evidence: the assessment must test whether agents can bypass controls through alternative interfaces, whether manipulated evidence can induce technically permitted but institutionally invalid actions, and whether authorized actors retain the practical ability to suspend and recover the process.

Autonomy should be assessed independently of these outcomes. The vector introduced earlier can characterize the scope of initiation, planning, execution, resource authority, decision criticality, temporal horizon, irreversibility, oversight, and goal modification permitted within a specified domain. This prevents the study from treating greater discretion as intrinsically beneficial or from comparing two systems as equally autonomous merely because they require human intervention at similar frequencies.

One potentially useful operational measure is authorized autonomous throughput, defined as the quantity of complete, independently verified, policy-conforming business work performed within a specified observation window without routine intervention beyond the approved operating arrangement. The denominator and unit of work must be defined for the particular process, and the measure should be accompanied by the full distribution of losses, interventions, and failures rather than adjusted through an opaque universal risk coefficient.

If a study seeks to compare the economic value of autonomy, it may estimate expected net process value after accounting for direct costs and relevant losses, subject to explicit constraints on severe outcomes. Such estimates require domain-specific assumptions about the probability and consequences of failure; they cannot be derived from generic benchmark success rates. Rare, high-consequence events will usually require additional stress testing, structured safety analysis, and conservative limits on inference because ordinary field experiments may not contain enough adverse events to estimate tail risk reliably.

The human-centered extreme introduces further outcomes. Employee skill retention, professional autonomy, ability to challenge computational judgments, distribution of cognitive workload, and the quality of collaboration may all change even when conventional productivity indicators improve. A technologically efficient enterprise that produces systematic deskilling or unacceptable surveillance would not be justified by throughput gains alone, particularly where worker protections and institutional rights impose independent constraints.

The resulting evaluation framework must therefore preserve a distinction between productive performance, epistemic reliability, operational autonomy, institutional accountability, and social consequences. The cognitive enterprise is a theory about their interaction, not a proposal to maximize one at the expense of the others.

Five propositions and their empirical falsifiers

H1 — The integration proposition. Under sufficiently interoperable semantics, appropriate access controls, and comparable model and workflow capabilities, a governed Company Brain should improve cross-functional decision quality and complete-process performance relative to agent systems operating with fragmented or domain-local memory. The hypothesized mechanism is not faster retrieval alone, but the ability to preserve relevant relationships among prior decisions, current evidence, dependencies, and observed outcomes. The proposition would be weakened if the apparent advantage disappeared after controlling for conventional data integration, retrieval quality, computational expenditure, and workflow design; in particular, a negligible Company Brain effect and no economically meaningful interaction with autonomous execution would leave little support for the proposed additional organizational capability.

H2 — The epistemic-verification proposition. Independent verification, explicit evidence dependencies, and governed belief revision should reduce the propagation of incorrect institutional knowledge, particularly when multiple agents and departments rely on shared representations. The expected improvement may entail additional latency and assurance expenditure, so the relevant hypothesis concerns reliability at a justified cost rather than an unconditional increase in immediate throughput. The proposition would be weakened if appropriately designed verification mechanisms failed to improve the detection or containment of material errors in environments where such errors can occur, or if their economic and operational costs systematically exceeded the value of the losses avoided.

H3 — The organizational-complementarity proposition. The performance gains generated by agents should depend significantly on complementary organizational investments, including process redesign, domain-specific knowledge, explicit delegation, transactional controls, and effective exception handling, rather than on the capability of the underlying model alone. This hypothesis is consistent with the broader literature on information technology and organizational complementarities, as well as with contemporary research treating human–AI collaboration as an organization-design problem.¹⁸³¹⁸⁴ It would be weakened if comparable models generated similar process-level benefits across substantially different organizational configurations after the relevant task and implementation differences were accounted for.

H4 — The substitution-boundary proposition. Predominantly agent-operated processes should approach the performance of human–agent organizations most closely where outputs are digital or otherwise objectively testable, dependencies are explicit, operating conditions are sufficiently observable, and mistakes are recoverable without disproportionate loss. Their relative advantage should diminish as environmental ambiguity, tacit knowledge, physical coupling, contested authority, and irreversible consequences increase. The hypothesis does not presume that human participation always improves outcomes; it predicts a conditional relationship between the characteristics of work and the comparative performance of different organizational configurations. It would be weakened by sustained evidence that agent-only arrangements perform equally well across these dimensions, at comparable total cost and risk, without relying on unobserved human intervention or externalized operating burdens.

H5 — The epistemic-monoculture proposition. Increasing reliance on a common knowledge-and-reasoning infrastructure should increase the dependence among organizational judgments unless the architecture preserves independent evidence, explicit provenance, contestability, and effective correction mechanisms. A suitable experiment could introduce controlled errors into nonproduction or safely isolated representations and observe whether independently operating agents and departments reproduce the same mistaken conclusions, whether the system identifies their common dependency, and how quickly affected processes recover after the source is corrected. The proposition would be weakened if greater integration did not increase relevant error dependence under such conditions, or if the integrated system consistently detected and corrected common-source errors without additional independence mechanisms.

These hypotheses should be evaluated separately rather than combined into an overall maturity score. H1 concerns the additional value of shared cognition; H2 concerns its epistemic reliability; H3 concerns organizational complementarity; H4 tests the boundary of human substitution; and H5 examines the risk created by cognitive integration. A particular implementation could support some propositions while contradicting others. Such a result would refine the theory rather than requiring an artificial judgment that the cognitive enterprise had simply succeeded or failed.

Experimental failures, security perturbations, and the limits of safe testing

The most informative experiments are likely to involve situations in which organizational assumptions become unreliable, because ordinary task completion provides relatively weak evidence about the capabilities distinguishing a cognitive institution from a well-integrated automation system.

A supply-chain evaluation could introduce a delayed supplier notification, a conflict between warehouse and ERP inventory records, a technical substitution constraint absent from the initial planning context, or a change in a counterparty’s authorization status. A financial-process evaluation could test whether concurrent agent transactions respect a shared exposure limit, whether a payment is retried safely after an ambiguous response, and whether a changed supplier-bank record is independently verified before money is transferred. An industrial evaluation could test whether an agent recognizes stale telemetry, conflicting operating constraints, or a command whose consequences cannot be established from available observations.

Such experiments must not create avoidable real-world harm. Financial transfers, customer commitments, production modifications, and cyber-physical interventions should ordinarily be exercised through sandboxed, shadow, simulated, or carefully isolated environments unless the proposed test has an explicit safety case and the required institutional authorization. Where synthetic errors are injected into shared knowledge, the research must establish that they cannot contaminate authoritative production records or be mistaken for legitimate evidence after the test ends.

Security and resilience testing should also distinguish attacker influence from ordinary model error. A manipulated document may redirect an agent through instructions embedded in lower-trust content; a corrupted business assertion may induce an otherwise compliant agent to make an inappropriate decision; and a faulty authorization policy may permit an operation the organization never intended to delegate. Each mechanism must be evaluated against the boundary designed to contain it, since an aggregate attack-success rate does not reveal whether the failure arose in evidence validation, cognitive inference, mandate enforcement, or transaction execution.

The research should measure both the probability of an undesirable state being reached under the defined test conditions and the effectiveness of detection, containment, and recovery. A system that refuses nearly every action may appear secure while failing its economic purpose, whereas a system that completes tasks efficiently may exhibit unacceptable exposure to correlated failures. The relevant outcome is reliable authorized work under realistic constraints, including the ability to decline or escalate when evidence or authority is insufficient.

Long-horizon evaluation introduces an additional challenge: the experiment should not merely ask whether the company can recover a process after a single fault, but whether it can preserve the integrity of organizational knowledge while repeatedly updating models, policies, roles, and external dependencies. The ability to recover from a known scenario is not equivalent to the ability to recognize an unfamiliar condition that invalidates the scenario’s assumptions.

This makes adaptive recovery a distinct object of research. The institution should be evaluated not only on whether it returns to a normal operating state, but on whether it identifies the mechanism of failure, revises the appropriate representation or procedure, and avoids repeating the same error without creating new uncontrolled consequences.

Institutional continuity as a longitudinal experiment

The strongest version of the Company Brain hypothesis concerns continuity across changes that traditionally cause organizations to lose knowledge or experience.

A mature cognitive enterprise should preserve more than the current contents of its enterprise systems. It should retain the relevant grounds of past decisions, the assumptions supporting operating policies, the identities and authorities of the actors involved, and the evidence connecting interventions to observed consequences. The practical significance of this capability should become particularly visible when experienced employees depart, teams are reorganized, suppliers change, critical applications are replaced, or new obligations require the company to reinterpret earlier commitments.

Longitudinal studies should therefore examine the organization’s performance across such transitions. Where feasible, researchers could compare comparable units undergoing planned personnel changes, model migrations, or process reorganizations, assessing whether access to governed institutional memory reduces the cost of reconstructing prior decisions and prevents the recurrence of previously resolved problems. Such comparisons must account for the possibility that the Company Brain’s implementation itself influences the likelihood or consequences of personnel changes, because this would make naive before-and-after attribution misleading.

The model-replacement experiment is particularly revealing. If the same Company Brain can support a new agent implementation without losing the institution’s commitments, decision provenance, or applicable authority, the architecture would demonstrate a form of continuity that does not depend on the persistence of one model’s internal state. Conversely, if replacement requires rebuilding critical context from undocumented prompts, proprietary memory, or the knowledge of a few engineers, the organization has not achieved the proposed independence between institutional cognition and its current technological implementation.

A related test concerns the ability to revise governing procedures without corrupting the historical record. When a company changes a procurement threshold or supplier-classification rule, it must distinguish the policy applicable to future actions from the policy under which earlier transactions were authorized. Successful migration should preserve both, allowing subsequent investigators to reconstruct the legitimacy of past decisions without retroactively applying new rules.

An even more demanding experiment would examine how the organization responds to a disturbance absent from its approved process library and historical experience. The question is not whether an agent can invent a plausible response, but whether the institution recognizes the limits of its knowledge, obtains additional evidence, applies the appropriate authority, and updates its operating representation without falsely treating a novel inference as an established rule.

The value of institutional memory also depends on its ability to forget, qualify, and retire information. A growing archive of generated explanations may degrade performance if obsolete or incorrect propositions remain readily available and acquire authority through repeated retrieval. A Company Brain that cannot identify superseded knowledge, apply retention requirements, or preserve the provenance of corrections would provide storage continuity without epistemic continuity.

This distinction makes the research program longitudinal by necessity. The proposed capability concerns not simply performance at one moment, but the organization’s ability to maintain the relationships among knowledge, decisions, and obligations as its participants and environment change.

From empirical propositions to a scientific theory of organizational cognition

A theory of the cognitive enterprise should ultimately explain differences in organizational outcomes more precisely than existing accounts of digitalization, information technology, workflow automation, and AI assistance. Its contribution would be weakened if the Company Brain added terminology without improving explanatory or predictive power; it would be strengthened if a relatively small set of clearly defined architectural and institutional properties could explain why some organizations use the same computational models successfully while others experience persistent errors, excessive supervision costs, or failures of adaptation.

The theory should also preserve the possibility that integration produces negative results. A shared Company Brain may increase correlated mistakes, reduce independent inquiry, intensify surveillance, create vendor dependence, or make organizational procedures expensive to revise. Agents may reduce the cost of routine coordination while increasing the cost of exception handling and assurance. Some activities may become more efficient through conventional deterministic automation than through probabilistic reasoning, while others may continue to depend on situated human expertise or institutional negotiation.

These outcomes are not anomalies to be excluded from the model. They are evidence concerning the conditions under which organizational cognition is economically and institutionally valuable.

The proposed research program is therefore concerned with identifying the feasible, productive, and governable boundaries of computational organizational cognition, rather than demonstrating an inevitable progression toward the disappearance of human work or the maximization of autonomous execution.

Its strongest empirical result would not necessarily be a company operating without employees. It might instead be a reliable demonstration that shared institutional memory improves the performance of human–agent teams under controlled comparisons; that agents can execute selected consequential processes with bounded authority and independently verified outcomes; or that an organization retains its decision-relevant knowledge through personnel turnover, model replacement, and changing external conditions.

Each result would support a particular part of the four-property framework without proving the remaining claims.

The most important falsifier would be more fundamental: if the claimed benefits of organizational cognition can be fully explained by existing information integration, more capable models, conventional workflow redesign, and ordinary governance, then the Company Brain does not represent a distinctive productive or cognitive capability. It may still remain a useful engineering architecture, but the proposed theory would have failed to establish an additional explanatory mechanism.

If, however, research demonstrates that persistent, governed, and revisable institutional knowledge systematically improves the coordination, reliability, and adaptation of organizations beyond those alternatives, the implications would extend beyond the evaluation of artificial intelligence. They would concern the theory of how firms acquire, preserve, and exercise productive capabilities across changing participants and environments.

The scientific question is not whether a company can be made to resemble a thinking organism. It is whether the architecture through which organizational knowledge becomes coordinated, authorized, and revisable action produces measurable capabilities that persist beyond the intelligence of any particular participant; only a research program capable of rejecting that proposition can establish its significance.

Management after organizational cognition becomes infrastructure

The proposed architecture changes the potential object of managerial work without establishing that management will disappear. Traditional managers organize knowledge, distribute work, reconcile competing functional priorities, establish operating expectations, and exercise authority over commitments that cannot be resolved through established routines. Some of this work exists because information is fragmented and relevant expertise must be located through reporting relationships; other parts exist because the organization must legitimately choose among conflicting objectives, allocate scarce resources, and accept responsibility for decisions. Shared organizational cognition and bounded agents may reduce the first class of costs while making the second more visible, especially when a computational actor can execute a plan long before anyone would ordinarily reconstruct all of its assumptions.

The central managerial transformation would be from directing individual activities toward designing, governing, and revising the institutional systems through which human and artificial actors obtain evidence, exercise delegated authority, and coordinate consequential work. This is a conditional prediction about functions rather than a forecast of the universal disappearance of particular occupations, because hierarchy also performs social, motivational, political, contractual, and legal functions that cannot be reduced to transporting information between departments.

From information intermediation to organization design

A Company Brain could weaken the relationship between hierarchical position and informational advantage by allowing an authorized specialist to examine cross-functional dependencies without repeatedly escalating requests through managerial channels. Yet access to knowledge and the right to decide must remain distinct: an engineer may correctly identify that a proposed component substitution would invalidate a customer commitment without possessing authority to renegotiate the contract, while a financial analyst may detect a liquidity risk without acquiring the power to redefine corporate investment policy. Management would increasingly need to specify the decision boundaries within which such expertise can influence action, rather than acting as the sole intermediary through which information becomes visible.

The direction of structural change is not predetermined. Bloom, Garicano, Sadun, and Van Reenen found that information technologies improving local access to knowledge were associated with greater decentralization in the firms they studied, whereas communication technologies facilitating interaction with superiors were associated with more centralized decisions.¹⁸⁵ A Company Brain could support either trajectory: one enterprise might use it to strengthen the independence of technical and commercial specialists, while another might build a centralized policy and monitoring apparatus whose interpretations become difficult to challenge. The organizational effect follows from the distribution of access, decision rights, and control over authoritative representations, not from the mere existence of a shared model.

The executable institutional constitution makes this relationship more explicit. Roles, workflows, mandates, decision thresholds, and selected constraints can already be represented in enterprise software, but an agent-native architecture would connect more of these artifacts to the actual initiation and execution of work. A procurement process might be reorganized by changing its resource limits, approval conditions, or handoffs through a versioned model whose effects can be simulated and tested; the ability to propose such a change would not confer authority to approve it. Management would consequently have an increasingly important second-order responsibility for the rules through which the organization changes its own rules, separating operational optimization from corporate or constitutional authority.

The distinction also cautions against equating formal execution with organizational practice. Feldman and Pentland’s research shows how routines are reproduced and modified through their situated enactment, rather than being exhausted by an abstract process description.¹⁸⁶ An organization may implement a flawless workflow that nevertheless encodes an obsolete supplier assumption, while exceptional human practice may reveal dependencies the model does not represent. Managers would need to govern the process by which those exceptions become evidence, are investigated, and lead—where justified—to approved revisions of the institutional model.

Strategy, risk, and the ownership of productive knowledge

As general-purpose reasoning capabilities become more readily purchasable, some strategic differences among firms may depend less on their choice of foundation model than on the quality of their institutional knowledge, operational assets, customer relationships, and ability to integrate them into reliable economic processes. Management would need to decide which representations and capabilities remain under enterprise control, which may be supplied by external vendors, and how portable the organization’s cognitive capital must remain when models or orchestration systems change. A technically efficient architecture that makes its critical institutional memory dependent on an irreplaceable supplier can exchange labor coordination costs for concentration and operational risk.

The same issue extends to the boundary of the firm. Cheaper agent-mediated coordination may allow small businesses to combine capabilities through external services, yet larger organizations may benefit disproportionately from the reuse of accumulated and verified experience; deciding whether to own, procure, or federate a productive capability will depend on its assurance-adjusted costs, asset specificity, and institutional constraints. The relevant executive question is therefore not only which tasks can be automated, but which organizational capabilities should be retained and governed as durable assets rather than outsourced as transient computational interactions.

Risk management, cybersecurity, and independent assurance would become closely connected to this strategy because the integrity of shared evidence and delegated authority determines whether computational actions remain aligned with legitimate commitments. An authorization service that rejects invalid purchases is insufficient if agents can bypass it through alternative interfaces; an excellent forecasting model provides limited protection when the source records used by every agent are corrupted; and an apparently comprehensive audit trail offers weak assurance if the system under review can modify the evidence of its own operations. Management must therefore preserve independent controls and effective suspension, reconciliation, and recovery mechanisms without assuming that a human approval click automatically supplies meaningful judgment.

Human judgment, workplace power, and the institutional conditions of learning

The human-centered firm introduces choices that throughput metrics cannot settle. Universal eligibility to use governed cognitive capabilities may strengthen employees’ autonomy in investigating a problem, while the same infrastructure can expose detailed patterns of individual behavior and become a mechanism for pervasive monitoring. Eurofound’s October 2026 research, based on the European Working Conditions Survey 2024, reports that 19% of EU employees say a computer program monitors their performance, 18% report automated task allocation, and 10% report automated determination of working hours; the study also identifies associations with greater job demands and job strain, rather than establishing a causal effect for every observed difference.¹⁸⁷ This evidence does not concern the complete Company Brain architecture, but demonstrates that computational management already affects conventional workplaces and cannot be presumed neutral.

Professional development presents a related tension. Agents may free employees from repetitive work and make specialized knowledge more accessible, but they can also remove the experiences through which junior professionals learn to recognize anomalous conditions or challenge inadequate decisions. A reviewer who has ceased to understand the work cannot exercise meaningful oversight merely because the workflow includes a human approval step. Management should therefore distinguish the automation of routine execution from the preservation of critical competence, deliberate opportunities for independent investigation, and procedures through which employees can question the representations that shape their work.

Nor does the economic value of accumulated organizational knowledge determine how the resulting gains should be distributed. Much of the Company Brain’s content may originate in the experience and collaboration of employees, contractors, and external partners, while its subsequent reuse may be controlled by shareholders or technology providers. Making knowledge widely accessible does not itself distribute ownership of that knowledge or confer the right to revise the policies under which it is used. The human-centered cognitive enterprise may accordingly support professional empowerment or a more comprehensive computational hierarchy, depending on the institutional rights and governance practices built around its technology.

Management as the governance of delegated intelligence

The managerial model implied by this theory is not the replacement of a human executive by an artificial executive with access to every corporate system. It is a differentiation of functions that traditional organizations have often combined in human roles: producing and interpreting evidence, initiating work, assigning decision rights, executing commitments, verifying their consequences, and changing the procedures by which these activities are coordinated. As these functions become more computationally explicit, management must preserve their productive integration while ensuring that the same model or agent cannot simply define its own objectives, grant its own authority, verify its own success, and suppress the evidence of failure.

A competent cognitive institution would therefore make organizational change more empirically examinable while remaining capable of refusing an economically attractive but unauthorized intervention. It would encourage distributed knowledge and legitimate dissent without sacrificing the common commitments required for coordinated action; it would permit bounded autonomous execution while keeping human and corporate authorities capable of effective intervention; and it would preserve institutional continuity when particular workers, models, and external providers are replaced. Such an organization would not be self-governing merely because it could automatically reconfigure some of its processes, for the legitimacy of its objectives and delegated powers would remain a question of institutional authority rather than computational skill.

The future managerial challenge is thus to govern productive intelligence as an institutional capability: deciding what the organization has reason to believe, which actors may act on that knowledge, how the consequences will be established, and when the organization’s own rules must be reconsidered. Whether this produces flatter firms, more concentrated authority, smaller operating cores, or new networks of specialization remains a matter for evidence and social choice; what changes is the growing extent to which those choices are implemented through explicit, consequential computational architecture.

Final remarks, open questions, and the future ahead

The possibility of a firm that thinks should not be understood as the culmination of an inevitable progression from human organizations to artificial ones, nor as the prediction that sufficiently capable language models will eventually become the executives, employees, and governing organs of corporations. The argument developed in this article is both more fundamental and more demanding. Organizations have always confronted the problem of transforming distributed, incomplete, and frequently conflicting human knowledge into coordinated economic activity; over time, they have responded by developing specialized roles, hierarchies, procedures, accounting systems, enterprise applications, and increasingly elaborate mechanisms for preserving information and controlling action. Contemporary artificial intelligence introduces an additional possibility into that historical process: portions of organizational understanding, reasoning, coordination, and execution may become persistent computational capabilities that survive changes in the individuals and software systems through which they are exercised.

The significance of this possibility does not depend on accepting a literal analogy between a company and a biological brain. As the traditions of distributed cognition, organizational learning, and management cybernetics suggest, the cognitive properties of an organized system need not be reducible to those of an individual participant.¹⁸⁸¹⁸⁹ What matters is whether the organization can maintain useful representations of its environment and internal conditions, combine knowledge originating from different actors, formulate decisions under uncertainty, coordinate interventions, and revise its understanding in light of independently established consequences. These capabilities already exist in partial and fragmented forms within conventional enterprises; the proposed Company Brain would extend them by connecting institutional memory, current operational state, evidentiary provenance, decision context, and the mechanisms through which knowledge becomes actionable.

The result would not necessarily be a company possessing a single unified intelligence, but an institution in which the ability to mobilize relevant knowledge becomes less dependent on the continuous presence, personal connections, and individual memory of particular people. Its cognitive infrastructure could support a firm operated predominantly through agents, a human organization whose employees share access to institutional knowledge, or numerous arrangements combining the two. These configurations would differ in the distribution of work, expertise, discretion, and authority, while drawing on similar architectural principles. Their possible emergence is consequently not evidence that the distinction between human and artificial participation has ceased to matter; it is evidence that this distinction may no longer be sufficient to explain how an organization acquires and exercises its productive capabilities.

What the theory establishes—and what it does not

The four-property framework provides the conceptual foundation for examining this transformation without reducing it to a single measure of technological maturity. Organizational cognition concerns the capacity to construct, preserve, interrogate, and revise justified representations; organizational agency concerns the initiation and coordination of action; organizational autonomy concerns the scope within which that action may proceed without additional intervention; and organizational accountability concerns the legitimate allocation of decision rights, control, duties, and consequences. These properties can interact and reinforce one another, but none logically entails the others. A company may possess sophisticated institutional memory while reserving consequential decisions to people, just as it may execute extensive operational processes through deterministic software without possessing an integrated representation of the reasons supporting those processes.

The separation matters because the ability to reason about an action, the authority to undertake it, and the capacity to determine whether it succeeded are different achievements. A procurement agent may generate a commercially sound recommendation without being authorized to issue the purchase; a corporate officer may approve a transaction based on evidence that subsequently proves incorrect; and an application may confirm that a command was executed without establishing that the intended business outcome occurred. The proposed architecture treats these distinctions as enforceable institutional boundaries rather than ambiguities to be managed through increasingly elaborate natural-language instructions. The Company Brain supplies qualified evidence and contextual interpretation; human and artificial actors formulate candidate interventions; independent policy and transaction controls determine which actions may be executed; and authoritative systems and verification mechanisms establish the consequences that can legitimately inform subsequent learning.

This architecture also changes the interpretation of organizational formalization. The firm’s procedures, roles, delegations, contracts, and application configurations have long represented selected parts of its operating structure, but computational agents create stronger incentives to express additional relationships in machine-interpretable and, where appropriate, executable forms. The resulting executable institutional constitution would not replace corporate law, contractual obligations, or the authority of recognized governing bodies. It would instead implement selected portions of the organization’s institutional order through versioned rules, controlled delegation, transactional constraints, and explicit procedures for changing the operating arrangement. Such formalization could improve consistency, auditability, and adaptability, yet it would remain incomplete because business judgments depend on uncertain evidence, situated expertise, competing interests, and circumstances that cannot be specified exhaustively in advance.

The proposed theory therefore does not establish that an enterprise can be made fully autonomous, that every significant organizational judgment can be formalized, or that computationally integrated knowledge is necessarily more reliable than distributed human understanding. It identifies a possible institutional capability and the conditions under which that capability might become economically productive and governable. Its scientific significance depends on whether those conditions can be observed and whether the resulting organizational performance differs measurably from what can already be achieved through conventional enterprise integration, more capable individual models, and established forms of automation.

This qualification is essential because the available evidence, while demonstrating substantial advances in tool-using agents and selected forms of human–AI collaboration, does not establish the general viability of an entire firm operating autonomously through open-ended environmental change. A successful benchmark task is not a durable business process; a durable process is not an institution capable of revising its own operating assumptions; and the ability to propose organizational changes is not equivalent to possessing legitimate authority to enact them. The distinction between these achievements should govern both future research and claims about the practical maturity of agent-native enterprises.

The open questions: Intelligence, institutions, and the limits of formalization

The first unresolved question concerns the nature and economic value of organizational cognitive capital. If a Company Brain merely provides a more convenient interface for searching documents and querying enterprise applications, its contribution may be substantial as an engineering improvement without justifying a new theory of organizational cognition. The stronger hypothesis is that persistent, shared, and revisable institutional knowledge creates benefits through the preservation of decision context, cross-functional interpretation, and the reuse of validated experience, allowing an organization to improve processes in ways that cannot be attributed to individual model capability or faster information access alone. Establishing this hypothesis requires causal comparisons that separate shared memory from conventional retrieval, autonomous workflow execution, and ordinary process redesign. It also requires identifying the conditions under which integration loses value because the costs of maintaining semantic coherence, verifying evidence, or correcting accumulated interpretations exceed the coordination benefits.

A second question concerns the boundary between what an organization can represent and what it can legitimately claim to know. The Company Brain cannot become an exhaustive digital counterpart of the enterprise, because operational conditions are only partially observable, records arrive asynchronously, social relationships and tacit expertise resist complete formalization, and many economically important relationships remain causally uncertain. Even where information can be represented, its meaning depends on temporal validity, institutional purpose, and the assumptions through which it was interpreted. The organization must therefore maintain not only knowledge but metaknowledge: what has been established, what remains disputed, which conclusions depend on unverified assumptions, and which observations would justify their revision. The difficult scientific problem is not whether a system can produce plausible explanations of organizational circumstances, but whether it can preserve the evidentiary structure necessary to recognize when those explanations cease to be justified.

This problem becomes more consequential as the same cognitive infrastructure informs larger numbers of actors and processes. Shared organizational knowledge can reduce duplication and inconsistency, but it can also create common-mode failures, allowing an incorrect interpretation to propagate through apparently independent decisions. The resulting organization may become more synchronized without becoming more correct. Its epistemic resilience will depend on whether it preserves independent observations, competing interpretations, provenance, and effective procedures for revising accepted knowledge, rather than treating consistency among agents as sufficient evidence of truth. The possibility of an epistemically coherent but systematically mistaken enterprise is not peripheral to the Company Brain hypothesis; it is one of its most important potential falsifiers.

A third question concerns the limits of autonomous organizational adaptation. Computational agents may become capable of reorganizing workflows, reallocating resources, identifying obsolete procedures, and proposing new institutional arrangements, but the ability to modify an operating system must remain distinguishable from authority to redefine the purposes and rules of the institution operating through it. A company may legitimately delegate the optimization of production schedules or the allocation of routine tasks while reserving changes to financial authority, risk appetite, contractual commitments, and corporate objectives to authorized decision-makers. The unresolved challenge lies in making that separation effective under conditions of dynamic delegation, model evolution, distributed execution, and incomplete information, without imposing so much procedural friction that the economic advantages of computational autonomy disappear.

The fourth question concerns the future boundary and scale of the firm. Cheaper computational cognition may enable small enterprises to coordinate productive capabilities previously requiring substantial internal administrative structures, particularly where outputs are digital, interfaces are standardized, and quality can be independently verified. Yet the same technology may strengthen large organizations that possess accumulated institutional knowledge, proprietary relationships, specialized assets, and the resources required to integrate and govern complex cognitive infrastructures. It may also facilitate networks of specialized firms whose agents coordinate through interoperable services while ownership and accountability remain distributed. These outcomes can coexist, and none constitutes the necessary endpoint of organizational evolution. Their relative importance will depend on the assurance-adjusted costs of internal organization, market contracting, cognitive integration, and external dependence.

A fifth question concerns the redistribution of economic and institutional power. If organizational knowledge becomes a durable computational resource, the conditions under which it is generated, governed, owned, and reused acquire greater significance. The employee-free enterprise might reduce some costs associated with internal employment while increasing dependence on model providers, software supply chains, and computational infrastructure; the human–agent enterprise might expand the practical capabilities of specialists while also enabling more extensive managerial surveillance or concentrating control over the representations through which organizational decisions are made. Neither outcome can be evaluated through productivity statistics alone. Labor protections, intellectual-property rights, contractual obligations, professional responsibilities, and the distribution of economic returns remain relevant, even as their relationship to the productive organization changes.

Cybersecurity introduces a further and particularly demanding dimension because a greater share of the institution’s economic agency may come to depend directly on computational systems. The transition from human-performed to agent-performed work does not replace labor, intellectual-property, or corporate law with security engineering; instead, it changes the risk structure of production by making the integrity of software identities, organizational knowledge, delegated mandates, execution interfaces, and external dependencies more consequential. An attacker capable of corrupting the Company Brain’s accepted evidence or manipulating the institutional rules governing agents may influence legitimate business operations without directly compromising every system through which those operations are performed. The central security problem is therefore no longer confined to preventing unauthorized access or protecting stored information, but extends to preserving the integrity of the mechanisms through which an organization interprets its circumstances and exercises its authority.

These questions converge on a deeper issue: how much of the firm can become computationally explicit without losing the forms of judgment, adaptability, and legitimacy on which its continued existence depends? A more completely formalized organization might be easier to audit, coordinate, and modify, but it could also become excessively dependent on the assumptions embedded in its models and rules. A less formalized organization might preserve situated expertise and discretion while remaining vulnerable to fragmentation, inconsistent decisions, and the loss of institutional knowledge. The feasible and desirable balance cannot be established through architectural elegance alone; it requires empirical evidence, institutional design, and continuing evaluation of the consequences of delegating productive intelligence.

The future ahead: Not one autonomous firm, but several possible organizational futures

The preceding analysis suggests that the future of the firm is unlikely to be captured by a universal progression toward employee-free operation. Instead, the decreasing cost of selected cognitive activities may enlarge the space of feasible organizational designs, allowing companies to combine human expertise, shared institutional knowledge, computational agency, and delegated authority in ways that differ substantially by sector, scale, and operating conditions.

One possible trajectory is the emergence of highly automated, comparatively small enterprises whose economic activities are performed predominantly by software agents and external services. Their competitiveness would depend on whether the reduction in internal coordination and labor costs exceeds the costs of infrastructure, external providers, assurance, and maintaining operational continuity. Such firms may become particularly relevant in digital and modular markets, although even there they would remain dependent on counterparties, legal institutions, suppliers, and the human activity embedded in the broader economy. A company with no employees would not be a company without human dependencies, nor would the absence of internal labor establish that it can govern itself.

A second trajectory is the evolution of established human organizations into cognitively integrated institutions. In this configuration, the Company Brain would expand the availability of relevant organizational context, while agents would perform selected analytical and operational activities alongside employees. The principal opportunity would concern the reorganization of collective work: specialists could contribute more effectively across functional boundaries, routine coordination could require less manual intermediation, and institutional experience could become more transferable across teams and generations of workers. Yet the consequences would depend on whether employees retain meaningful powers of interpretation, challenge, and professional judgment, or whether computational integration instead creates a new hierarchy in which common machine-generated representations become difficult to contest.

A third trajectory consists of increasingly permeable organizational networks in which specialized enterprises coordinate through structured computational interfaces, shared process semantics, and agent-mediated transactions. Rather than transferring every capability into one legal entity, such networks could distribute productive activity among firms while coordinating parts of their knowledge and execution more tightly than conventional contractual relationships permit. This possibility would extend the economic significance of interoperability and digital platforms, but would also raise difficult questions concerning contractual authority, information ownership, liability, resilience, and the concentration of power within the infrastructure used to coordinate the network.

These trajectories are not mutually exclusive. A large enterprise might operate highly autonomous internal service units while retaining a substantial human workforce and coordinating external suppliers through agent-mediated interfaces. A small company might depend on a sophisticated Company Brain without permitting agents to undertake consequential transactions independently. An industrial organization might use agents extensively for commercial planning, maintenance analysis, and administrative processes while preserving strict boundaries around physical control and safety functions. The distribution of cognition and authority within an enterprise may therefore become more informative about its operating model than its employee count, organizational chart, or number of deployed agents.

The enabling technical developments are also unlikely to proceed uniformly. More capable reasoning models will matter, but so will less conspicuous advances in semantic interoperability, provenance, temporal databases, process modeling, durable orchestration, workload identity, transaction enforcement, formal verification, and independent outcome evaluation. Some of the most economically important improvements may arise not from making individual agents more intelligent, but from making their interactions with institutional knowledge and business systems more predictable, recoverable, and governable. The decisive engineering achievement would be an architecture in which agents and models can be replaced without losing organizational memory, obligations, or the authority structure governing their actions.

This direction also suggests a change in how enterprise architecture itself should be understood. Historically, it has sought to align applications, information, infrastructure, and business processes with organizational objectives; in the cognitive enterprise, it may increasingly participate in the explicit representation and implementation of organizational roles, decision rights, epistemic responsibilities, and institutional change. Software architecture and organization design would not become identical disciplines, but their intersection would deepen because decisions about system boundaries, model access, policy enforcement, and knowledge integration would influence how the company actually distributes productive agency.

The most demanding frontier, however, will concern institutional adaptation. It is possible to imagine agents that reliably execute predefined processes, and increasingly plausible that they will propose improvements to those processes using accumulated operational evidence. It is substantially harder to establish that an organization can revise its assumptions, recognize previously unmodeled circumstances, and change its operating arrangements without compromising authority or institutional continuity. The capacity to act under established rules and the capacity to determine when those rules are no longer adequate represent different forms of competence, while the authority to change the rules introduces an additional institutional problem that cannot be settled by model performance.

The future of the cognitive enterprise should therefore be investigated through longitudinal evidence rather than isolated demonstrations. A meaningful test would examine whether an organization preserves decision-relevant knowledge when experienced employees leave, whether its commitments remain intelligible when models and providers change, whether it detects the propagation of corrupted or obsolete evidence, and whether it can recover from failures without converting exceptional interventions into unrestricted new authority. It would also examine whether the resulting gains in coordination and productive capability exceed the accumulated costs of maintaining the cognitive infrastructure and governing its evolution.

The research agenda developed in the preceding section makes these questions operational, but it does not prejudge their answers. If shared institutional cognition fails to produce benefits beyond existing information integration and automation, the theoretical claim must be narrowed. If autonomous execution proves reliable only within tightly specified environments, that limitation should become part of the organizational theory rather than be dismissed as a temporary inconvenience. If the integration of cognition and agency increases correlated errors or institutional fragility, the architecture must be revised accordingly. Scientific progress will consist in identifying the conditions under which the cognitive enterprise is viable, not in treating its eventual realization as inevitable.

The firm that thinks—and the institution that must still govern it

The most important implication of the argument is that organizational intelligence and institutional sovereignty are different concepts. A firm may develop increasingly capable mechanisms for understanding its environment, remembering its decisions, coordinating action, and adapting its procedures without acquiring any independent legitimacy to determine the purposes it ought to pursue. Its computational systems may become capable of making and executing decisions that previously required extensive human effort, but the validity of those decisions still depends on the governing institutions, contractual relationships, legal obligations, and decision rights within which the firm operates.

This distinction is not a residual limitation to be removed when artificial intelligence becomes sufficiently powerful. It follows from the difference between the capacity to determine which actions are instrumentally effective and the authority to decide which objectives, obligations, and constraints should govern those actions. A system may produce increasingly reliable forecasts and increasingly sophisticated plans without thereby obtaining a right to alter the corporation’s purposes, accept obligations on behalf of unconsenting parties, or redefine the rules governing its own autonomy. Such authority is constituted through institutional arrangements whose legitimacy cannot be derived from computational performance alone.

Nor should the persistence of human governance be mistaken for a guarantee of effective control. A board that formally approves a computational operating arrangement without understanding its dependencies, a manager who cannot reconstruct the basis of an agent’s recommendation, or a human approver who cannot realistically refuse an automatically generated transaction may provide the appearance rather than the substance of accountability. Meaningful governance requires that the institution retain effective means of accessing evidence, defining and revising authority, detecting failure, challenging accepted interpretations, and intervening when its computational actors operate outside their legitimate or validated boundaries.

The proposed Company Brain acquires its strongest meaning in this context. It is not a universal repository that allows the enterprise to know everything, nor an artificial executive that reconciles every conflict among the interests of owners, employees, customers, creditors, regulators, and society. It is a proposed infrastructure through which organizational knowledge can become more continuous, evidence can remain connected to its provenance and limitations, and decisions can be related to the authority and consequences that make them institutionally significant. Its value depends as much on the capacity to disclose uncertainty and preserve disagreement as on the ability to produce coherent interpretations and efficient recommendations.

The same principle defines the limits of agent coworkers. Their contribution is not measured solely by how frequently they complete tasks without human assistance, but by whether they can participate in reliable organizational processes under explicitly delegated authority, recover from unexpected conditions, and remain replaceable without erasing the institution’s commitments or decision history. Greater autonomy may be economically valuable where the environment is sufficiently observable and the consequences are bounded; in other circumstances, the appropriate expression of computational competence may be to seek additional evidence, defer a decision, or preserve a safe operating condition until legitimate authority can be exercised.

At the level of the theory of the firm, the transformation may be profound even if a universally autonomous company never becomes viable. The organization has historically been a mechanism for combining specialized human knowledge and productive resources under a structure of coordination and authority. As more of the relevant knowledge becomes computationally persistent and more of the coordination becomes executable, the institution may acquire capabilities that are less dependent on the identities and continuity of its individual participants. The economic boundary of the firm, the nature of specialization, the structure of managerial authority, and the ownership of organizational knowledge could all change as a consequence, without requiring the disappearance of human work or the elimination of corporate governance.

What ultimately matters is whether these developments produce a more capable and corrigible institution, rather than merely a faster and more tightly integrated one. The strongest cognitive enterprise would be able to preserve experience without treating inherited interpretations as infallible, coordinate actions without suppressing independent judgment, delegate execution without allowing authority to expand silently, and adapt its procedures without losing the institutional continuity through which its actions remain intelligible and accountable.

The firm that thinks is therefore not the firm that has converted itself into an omniscient computational actor. It is the firm that has made the production, preservation, and application of organizational knowledge an enduring capability of its institutional architecture.

Its most consequential achievement would not be the elimination of people from its operating processes, nor the universal availability of artificial agents, but the ability to connect distributed and imperfect knowledge to coordinated, authorized, and verifiable action while retaining the means to discover that its own representations, decisions, and operating rules are wrong.

Whether this architecture ultimately produces smaller firms, more powerful large enterprises, new networks of economic coordination, or organizations whose internal distinction between human and artificial work becomes increasingly fluid remains uncertain. The technologies required for parts of the vision already exist, while the integration, governance, and empirical validation needed to establish its full organizational significance remain unfinished. That uncertainty should not diminish the importance of the question; it defines the research and engineering work through which the possibility can be assessed.

The future of the firm may be shaped less by the intelligence of the individual actors it employs or deploys than by its capacity to make intelligence persist, cooperate, evolve, and remain accountable across them. A company may come to think in this functional and institutional sense, but its ability to think will never, by itself, make it the sovereign source of the authority under which it acts.

Appendix A — Toward agent-operable software: formalizing interfaces, behavior, and institutional semantics

The argument of The Firm That Thinks has principally concerned the formalization of the organization: the progressive representation of knowledge, processes, delegated authority, and institutional change in structures that can be inspected and, within defined boundaries, executed. Yet this transformation implies a second, less obvious one. If organizational actors increasingly include software agents that select operations, construct plans, and act across heterogeneous applications, the applications themselves must expose more than callable procedures. They must convey enough of the meaning, admissibility conditions, possible effects, and evidentiary consequences of those procedures for an independently governed actor to use them without reconstructing their semantics from human conventions or undocumented implementation details.

This requirement does not begin with generative AI. Programming languages have formal grammars and type systems; databases implement integrity constraints; APIs have machine-readable schemas; and software-engineering research has studied contracts, concurrency, refinement, and formal correctness for decades. What changes is the location of interpretation. In a conventional integration, a programmer ordinarily fixes the mapping between a business intention and an application operation in advance, including the assumptions not captured by the interface description. In an agent-mediated process, a planning component may choose among previously unfamiliar tools at runtime, compose their effects, and act upon descriptions generated by providers with different levels of trust. The interpretive work once performed at design time is partly displaced into an operating environment whose decisions may have legal, financial, or physical consequences. Consequently, the question is not whether software can be made formal, but which additional aspects of software must become explicit, which must be enforced, and which cannot be guaranteed through specification alone when its consumers are computational organizational actors.

The appendix develops a proposed analytical construct, the agent-operable capability contract, which connects classical software contracts with domain semantics, institutional mandates, transaction guarantees, and evidence requirements. It is not an existing universal standard or a claim that a complete formal semantics of enterprise software can presently be automated. Its purpose is to establish where a specification ends, where runtime enforcement begins, which proof obligations can be formulated precisely, and how the cost and value of greater explicitness should be assessed. The resulting relationship is complementary to the article’s executable institutional constitution: the latter specifies who may act and under which institutional rules; an agent-operable software contract specifies what a particular software operation means and what its execution can be taken to establish.

From application interfaces to the formalization of software behavior

A conventional API gives a client a vocabulary for invoking an application, but its description may omit business distinctions that a competent human operator or integration developer has learned to supply. An endpoint named POST /purchase-orders, for example, may accept a supplier identifier, quantity, price, and delivery date, yet the schema does not normally establish whether the price includes freight, whether the supplier is eligible for the requesting legal entity, whether an order can be cancelled after submission, or whether a successful HTTP response denotes a durable ERP commitment rather than acceptance into a processing queue. Such distinctions can be crucial to an agent deciding whether the operation satisfies an organizational objective. Syntax, however precisely defined, cannot eliminate the need to specify business meaning.

In September 2026 the OpenAPI Initiative published OpenAPI 3.2.1, a language-agnostic interface description for HTTP APIs; JSON Schema Draft 2020-12 supplies formal vocabularies for describing and validating JSON structures.¹⁹⁰¹⁹¹ These specifications improve interoperability, tooling, and structural validation, but neither converts an arbitrary business description into a proof of intended effect. The July 2026 Model Context Protocol (MCP) similarly provides discovery and invocation mechanisms, including input and optional output schemas, while explicitly treating advertised tool annotations as untrusted unless they originate from a trusted server.¹⁹²¹⁹³ Transport interoperability and discovery thus solve parts of the connectivity problem, not the complete problem of institutional action.

A useful starting point is to distinguish several dimensions of software explicitness. These are not a universal maturity ladder, because an application can be highly formal in one respect and imprecise in another; nor does the strongest dimension imply that every property of the application has been verified.

Dimension	Specification object	Representative mechanism	What remains outside the mechanism alone
Syntactic and structural	Types, fields, serialization, operation signatures	OpenAPI, JSON Schema, type systems	The business meaning and truth of supplied values
Domain-semantic	Entity identity, units, state meanings, valid interpretations	Domain ontologies, controlled vocabularies, data contracts	Whether mappings preserve meaning in every context
Behavioral	Preconditions, postconditions, invariant-preserving transitions	Design by contract, Hoare logic, state machines	Correctness of specifications and external assumptions
Temporal and transactional	Ordering, isolation, retries, deadlines, intermediate and ambiguous states	Transaction models, linearizability, temporal logic, durable workflows	Global atomicity across arbitrary services and external effects
Institutional and normative	Delegated powers, permitted purposes, obligations, segregation of duties	Policy-as-code, mandate registries, BPMN and DMN	Legitimacy of the mandate and completeness of policy
Evidentiary and causal	Provenance, readback, observed effects, outcome qualifications	Event histories, PROV-O, reconciliation, audit receipts	Whether observations establish a causal business outcome
Verified and assured	Properties shown to hold under explicit models and assumptions	Model checking, theorem proving, refinement, conformance testing	Unmodeled environments, implementation drift, specification errors

Table 4: Orthogonal dimensions of software formalization relevant to human and agent consumers. The last row concerns methods of assurance applied to any of the preceding dimensions, rather than a final stage that subsumes them.

Table 4 makes explicit why it would be inaccurate to regard every machine-readable description as formal verification. Hoare’s axiomatic account of program correctness introduced the familiar relation between preconditions, programs, and postconditions, while later work on temporal and distributed systems established rigorous ways to reason about behavior extending beyond one sequential procedure.¹⁹⁴¹⁹⁵ The existence of these methods does not establish that enterprise applications already have complete machine-checkable contracts. Much of their behavior remains distributed across implementation code, database configuration, legal agreements, organizational practice, undocumented exceptions, and dependencies on external systems.

This historical distinction also explains why agent-operable software should not be equated with software written by agents. A code-generation system may produce an application whose interface remains semantically opaque, while software authored entirely by human engineers may expose exemplary machine-readable contracts. Similarly, the presence of an MCP server does not establish that a legacy application has been designed for agentic execution; it may only wrap an older interface with another description layer. The relevant variable is the quality of the operational contract available to the actor and its independent enforcement infrastructure, not the provenance of the source code or the protocol selected to expose it.

A formal model of an agent-operable capability contract

Let S denote the set of possible business states, X_a the admissible input domain of an operation a, U the set of authenticated principals, and H the set of relevant execution histories. A business operation should not be modeled only as a total function from inputs to outputs, because its effects depend on state, concurrent activity, authority, and the possibility that an external system returns no definitive acknowledgment. A more faithful abstraction is a transition relation R_a\subseteq S\times X_a\times H\times S\times O_a, where O_a denotes observable operation outcomes, including committed, rejected, pending, and uncertain states. The final business state may remain unknown to the caller even when it is definite within the authoritative application.

For an agent-operable capability, define a proposed specification

\mathcal C_a=(\Sigma_a,\Omega_a,P_a,Q_a,\Gamma_a,\Delta_a,\Theta_a,\Phi_a), \tag{11}

where \Sigma_a specifies the input/output types; \Omega_a interprets the operation and its parameters within the enterprise domain; P_a\subseteq S\times X_a identifies required preconditions; Q_a specifies permissible outcomes and their corresponding postconditions; \Gamma_a is an institutional authorization predicate; \Delta_a describes state-changing effects and relevant invariants; \Theta_a constrains temporal, concurrent, and failure behavior; and \Phi_a defines the evidence required to establish which outcome actually occurred. These components have distinct semantics, and a concrete implementation need not encode them in a single document or language: domain models, application code, policy engines, transaction managers, and assurance services may each govern different components.

Classical design by contract can express partial correctness as \{P_a\}\,a\,\{Q_a\}: if execution begins in a state satisfying P_a and terminates normally, the resulting state satisfies Q_a. For an enterprise operation, however, Q_a must not be reduced to an HTTP status code or a textual declaration by the agent. It may need to distinguish a committed purchase order from a submitted request, a compensable intermediate effect from an irreversible external commitment, and an unknown response from an operation known not to have executed. In addition, authorization is a predicate of the principal, request, context, applicable version of the mandate, and the actual state at the enforcement point, rather than a property that the invoking model can confer by including a policy identifier in its arguments.

A safety condition for a consequential operation can accordingly be written in linear temporal logic as

\Box\left(\operatorname{Commit}_a\Rightarrow \operatorname{Permit}_{\Gamma_a}(u,a,x,s) \land P_a(s,x)\right), \tag{12}

where \Box means that the implication must hold at every relevant transition and s is the authoritative state against which the commit is evaluated. The formula is a proposed design obligation, not a theorem about any existing implementation; it assumes that all paths capable of committing the transaction cross a correct enforcement boundary and that the predicates capture the relevant institutional policy. Evaluating the same expression against a stale snapshot before the transaction commits is insufficient because the state can change between evaluation and execution.

Safety alone does not establish that business work will make progress. Under stated assumptions about the availability of the authoritative service, fairness of scheduling, and eventual response from relevant participants, the process may also be required to reach a terminal or appropriately escalated state. A suitable liveness obligation can be written as \Box(\operatorname{Pending}_a\land\operatorname{FairEnv}_a\Rightarrow\Diamond\operatorname{Resolved}_a), where \Diamond means eventually and Resolved includes a verified commit, definitive rejection, or authorized escalation. This conditional statement cannot be interpreted as a guarantee of supplier delivery or of eventual completion under arbitrary environmental failure. Alpern and Schneider’s foundational distinction between safety and liveness is especially useful here: avoiding unauthorized actions and eventually making justified progress are different properties, and an architecture may satisfy one while failing the other.¹⁹⁶

The contract also requires an explicit semantics of refinement. Let \mathcal I_a be an implementation and \mathcal S_a its abstract specification, with \pi a projection that maps implementation traces onto externally meaningful business observations. Under declared environment assumptions E_a, a meaningful conformance obligation is

\pi\bigl(\operatorname{Traces}(\mathcal I_a\mid E_a)\bigr) \subseteq \operatorname{Traces}(\mathcal S_a). \tag{13}

This is an observational trace-inclusion criterion: the implementation must not exhibit a business-visible behavior disallowed by the specification. It does not by itself ensure liveness, noninterference, performance, or observational completeness, which require additional obligations; nor does it establish that the specification describes the right business objective. Work on assume–guarantee composition shows why the stated environment conditions must also be treated rigorously when separately specified components are combined, rather than assumed to hold automatically because each component has been tested in isolation.¹⁹⁷

The central engineering insight follows from these distinctions. A capability description is informative only to the extent that the claimed semantics can be related to enforceable preconditions, actual execution behavior, and independently observable effects. An agent’s confidence that it understands an operation cannot repair a mismatch between what the operation is advertised to do and what the enterprise application actually commits.

An ERP purchase operation: from description to controlled commit

Consider purchase-order.create in the supply-chain scenario developed in the main article. A structural interface might require supplier_id, item_id, quantity, and price; even if every field passes schema validation, the request can still be institutionally invalid because the supplier is disqualified, the material is not approved for the specified legal entity, another agent has already consumed the available budget, or the evidence supporting the shortage is stale. The operation must therefore be modeled as a business transition with authority and concurrency constraints, not simply as a correctly formatted call.

For a given purchasing legal entity and observation time t, let C_t be the set of outstanding committed obligations counting toward the relevant delegated limit, and let R_t be the set of outstanding reservations not yet converted to commitments. Assume these sets are disjoint and that the same currency and valuation basis apply to all amounts. A bounded spending invariant is

X_t=\sum_{c\in C_t}\operatorname{value}(c) +\sum_{r\in R_t}\operatorname{value}(r)\leq L_t, \tag{14}

where L_t is the current authorized limit. The definitions of outstanding commitment, reservation, currency conversion, tax treatment, and legal-entity scope must be specified as part of the enterprise domain semantics; otherwise, a syntactically valid inequality may give a false appearance of financial control. The safety requirement is that the invariant hold after every relevant atomic change to the authoritative resource ledger, including concurrent requests, rather than merely when each agent prepares its proposal.

Two procurement agents can each observe X_t=80{,}000 and L_t=100{,}000 and independently propose a €15,000 commitment. Each proposal passes a non-atomic precheck, yet if both are committed without coordinated reservation, the exposure becomes €110,000. The failure is a familiar time-of-check/time-of-use race, not a defect unique to generative models. Correct implementation might use a single transactional budget ledger, a serializable allocation transaction, a compare-and-swap update over a correctly scoped resource, or another mechanism with an equivalent enforceable invariant. Where the budget spans heterogeneous applications that cannot participate in one atomic transaction, the architecture must designate an authoritative reservation service or acknowledge that a stronger global guarantee is unavailable.

The following illustrative specification illustrates the information an agent-operable contract could contain. It is intentionally not presented as valid OpenAPI, MCP, Cedar, BPMN, or DMN syntax; the named fields form a proposed domain-specific contract, whose enforceable elements would require separate implementation and validation.

capability: purchase-order.create
contract_version: 3.1.0
business_semantics:
  effect: commit_purchase_obligation
  authoritative_system: erp-procurement
  legal_entity: resolved_from_authenticated_mandate
  currency_and_amount: evaluated_on_authoritative_valuation_basis
inputs:
  supplier_id: qualified_supplier_identifier
  item_id: approved_item_identifier
  quantity: strictly_positive_decimal
  maximum_total_amount: monetary_amount
  required_by: date_with_contractually_defined_meaning
  evidence_refs: immutable_source_or_snapshot_references
  expected_state_version: authoritative_revision_or_reservation_token
  idempotency_key: caller_unique_business_intent_identifier
preconditions:
  - supplier_qualified_for_entity_and_category
  - item_approved_for_intended_use
  - price_and_terms_within_current_mandate
  - supporting_evidence_current_and_adequate
  - committed_plus_reserved_exposure_within_limit
  - segregation_of_duties_satisfied
execution:
  authorization: independent_policy_decision_at_commit
  budget_control: atomic_reservation_required
  retry: reconcile_by_idempotency_key_before_reissue
  ambiguous_result: return_pending_reconciliation
  compensation: business_specific_and_not_always_possible
outcomes:
  - committed_with_erp_order_identifier
  - definitively_rejected_without_commit
  - pending_or_unknown_requires_reconciliation
evidence:
  - policy_and_mandate_version
  - authorization_decision_receipt
  - authoritative_transaction_receipt
  - independent_readback_and_reconciliation_status

The proposed contract distinguishes the agent’s intention from the application’s commitment. The purchasing agent may present evidence and request execution, but it cannot certify that the supplier is qualified merely by asserting so, nor can it choose a favorable obsolete mandate by setting contract_version or a policy reference. An independent enforcement component must resolve the acting workload identity and originating organizational authority, retrieve the applicable policies, validate source assertions and current business state, and mediate the actual transaction. Distinguishing authenticated workload identity from authorized corporate delegation is necessary even when service identities are cryptographically strong; a credential identifies a software workload, but does not establish that the workload possesses the specific business authority it invokes.

Correct retry semantics are equally important. If a request times out after the ERP has committed, repeating it with a new intent identifier can create a duplicate commercial obligation; if the receiving application retains an idempotency key but does not persist it atomically with the commitment, the guarantee may fail during recovery. An idempotency key is therefore not magical exactly-once execution, but one part of a protocol whose correctness must be evaluated together with storage durability, uniqueness scope, retention time, and the authoritative transaction boundary. When the process continues across several applications, the classical saga model provides one historical method for coordinating long-running transactions through separately committed steps and compensating actions, although compensation is not necessarily an inverse of a physical or legally effective action.¹⁹⁸¹⁹⁹

Finally, the ERP’s committed order is not proof that the intended material has been supplied, and the contract must not collapse these outcomes. A purchase-order receipt can establish an application commitment, a supplier acknowledgment may establish acceptance under specified contractual conditions, a warehouse receipt can establish a recorded arrival, and quality release may establish technical eligibility for use. The organization needs a traceable chain connecting these claims without treating their temporal succession as proof that the original decision caused the desired result. This is the software-side expression of evidence-mediated action: the postcondition of an API call and the success condition of a business process are different propositions, supported by different evidence.

Composition across applications: why local contracts are insufficient

An enterprise process rarely consists of one invocation, and a collection of locally valid capabilities does not automatically form a globally valid process. One application may define a supplier as an account eligible for purchasing, another as a legal entity approved for payment, and a third as a logistics destination; unless their identifiers and meanings are reconciled, a chain of structurally valid calls can create an incorrect obligation. The Company Brain may help discover the relationships among these representations, but it cannot guarantee that they are true, while the authorization gateway cannot safely enforce a rule whose domain meaning has not been established. Agent-operable software consequently requires a second kind of contract, governing composition: how independently specified services can be combined while maintaining business invariants across organizational and technical boundaries.

Classical modular reasoning expresses this problem through the relationship between the assumptions each component makes about its environment and the guarantees it provides when those assumptions hold. If service B relies on a condition promised by service A, the composition is valid only when A actually establishes the condition in a form B can interpret, at the moment B needs it, and under a compatible failure model. Abadi and Lamport’s research on composing and conjoining specifications establishes formal foundations for such assume–guarantee reasoning, while Herlihy and Wing’s linearizability criterion provides a local abstraction for when concurrent operations on a shared object may be understood as occurring atomically in some sequential order.²⁰⁰²⁰¹ These results cannot be transferred mechanically to an arbitrary enterprise workflow: a local database transaction may be linearizable without making a multi-day supplier negotiation, warehouse transfer, and financial settlement atomic.

The distinction between safety properties and liveness properties therefore matters at the process level as well. An agent might ensure that it never sends a payment without an authorized purchase obligation, yet leave the process permanently pending because a required acknowledgment never arrives. Conversely, it might terminate every workflow promptly by interpreting timeouts as failures and resubmitting requests, thereby introducing duplicate transactions. Durable execution and event-based recovery can maintain process state, but their guarantees must be paired with explicit semantics for retries, time limits, idempotency, reconciliation, and compensations; the resulting business contract must indicate where stronger guarantees end and where the process enters an uncertain state requiring intervention.

A further difficulty concerns contract compatibility. Syntactic backward compatibility is necessary for many integrations, but it is not sufficient to preserve behavioral meaning. Changing a field from available_quantity to unrestricted_available_quantity might preserve its numeric type while changing whether an agent can safely promise delivery; changing a tool’s behavior from immediate commit to asynchronous acceptance might preserve its response schema while invalidating a caller’s assumption that success implies a completed transaction. For a consumer B depending on provider A, a valid replacement must preserve the assumptions actually used by B—or require the consumer’s contract and workflow to be revised. In formal terms, permitted behavior must continue to refine the obligations on which the dependent component relies; a version number by itself is not evidence of such refinement.

This suggests that a capability registry should maintain more than a tool name and description. For consequential operations, it should identify the authoritative provider, supported contract versions, declared side effects, authorization and data dependencies, trusted origin of the description, known environmental assumptions, and assurance evidence. Providers should not be able to promote a changed tool description directly into the set of authorized business capabilities merely by publishing a new discovery result. Discovery is an information operation; admitting a new capability into the institution’s operating environment is a governed change.

Existing standards as components of an assurance chain

The practical architecture need not invent a universal specification language to achieve every degree of explicitness. Its more plausible development would compose established mechanisms while keeping their respective claims and assurance boundaries distinct. OpenAPI and JSON Schema describe HTTP interfaces and structured data; MCP supplies mechanisms for discovering and invoking tools; BPMN represents business-process structure; and the Object Management Group’s DMN 1.6, formally adopted in September 2026, provides a specification for business decisions and decision logic.^{202203204205206} These mechanisms can reduce ambiguity, but they govern different artifacts. A tool schema does not establish that an agent may use the tool, while a decision table does not determine whether the facts to which it is applied are current or whether the rule was adopted by an authorized corporate body.

Provenance and authorization likewise belong to separate logical functions. W3C PROV-O provides a formal vocabulary for describing entities, activities, and agents involved in the production of information; it records provenance relationships without proving the truth of the claims whose history it describes.²⁰⁷ Cedar and related policy technologies allow particular authorizations to be evaluated against specified principals, actions, resources, and context, but their formal analyzability does not prove that the represented corporate delegation is legally valid or that the observed business state is correct. The Cedar project’s verification-guided development is especially instructive because it combines mechanically checked properties of a model with differential and property-based testing of an implementation, thereby making the gap between specification correctness and production code an explicit engineering problem.²⁰⁸

At the transport boundary, standard error descriptions can help prevent the interpreting agent from treating every unsuccessful call as the same outcome. RFC 9457 defines machine-readable problem details for HTTP APIs, but a response identifying invalid input, denied authorization, retryable service unavailability, or ambiguous processing state still requires domain-specific semantics before a workflow can choose a safe continuation.²⁰⁹ Similarly, a signed receipt from an authoritative application can establish that a particular transaction was recorded, while a subsequent reconciliation service may be required to establish that its business consequences materialized. Recording these distinctions is part of software engineering because it changes the behavior that the next computational consumer is entitled to infer from the interface.

Taken together, the components support a plausible assurance chain linking requirements, domain semantics, typed interfaces, policy rules, execution constraints, transaction evidence, and verified business outcomes. This is not a continuous proof merely because each component has a formal representation. To justify a compositional assurance claim, the implementation must demonstrate that the interfaces between components preserve the required assumptions, that the relevant specifications correspond to the institution’s intended rules, and that the evidence used to establish outcomes originates outside the reasoning process whose success is being assessed. Where such demonstrations are infeasible, the architecture should describe its assurances as tested or monitored properties rather than formally guaranteed ones.

This is also why introducing a new agent tool should be treated as a form of enterprise change management. An application vendor may publish a tool descriptor, but the enterprise still has to determine its domain semantics, assign permitted principals, evaluate whether its side effects are compatible with existing policies, bind it to current source records, and test it against the workflows in which it may be used. The executable institutional constitution can govern this admission process without being identical to the resulting software contract. The software describes what the operation purports to do; the institution determines which of those operations may contribute to its legitimate objectives.

Consequences for requirements engineering, testing, and software evolution

The shift from fixed integrations toward dynamic agent-mediated operation changes the object of requirements engineering. Requirements can no longer be confined to the application feature and its human-facing behavior: for consequential operations, they must also express the legitimate business purpose of the capability, its admissible inputs and state transitions, its authorization scope, its observable effects, and its behavior under concurrent or partially failed execution. Requirements that remain ambiguous may be tolerable when an experienced operator resolves the ambiguity before acting, but they become hazardous when an agent interprets them at runtime and directly invokes privileged software interfaces. Greater formalization is therefore a way of relocating certain forms of judgment from an uncontrolled moment of execution to a governed process of specification and validation.

The lifecycle implication is not that every business requirement should become a theorem. The cost of formal methods is justified differently for a read-only catalogue search, a supplier-bank update, a software release, and an industrial control command. Some requirements can be enforced adequately through type checking and schema validation, others through runtime authorization and transactional invariants, while safety-critical or high-consequence functions may justify model checking, proof-carrying design artifacts, independently developed monitors, and stronger isolation of the execution path. The NIST Secure Software Development Framework provides a general reference for incorporating security practices into development, although it does not itself specify an agent-operable business contract or prove the correctness of an enterprise workflow.²¹⁰

Verification activities should be aligned with the contract dimensions. Structural tests establish whether inputs conform to types and schemas; semantic tests challenge unit, identity, and domain mappings; contract tests compare advertised behavior with application responses; model-based tests explore state transitions and exceptional paths; property-based tests generate varied inputs and sequences; and concurrency or fault-injection tests examine races, retries, timeouts, and recovery. Formal verification may establish that selected invariants hold within an explicit model, while runtime monitoring can detect violations of observable properties in the deployed environment. None of these methods can infer complete legal authority or guarantee that external counterparties will fulfill their obligations; the engineering goal is to specify which claims each test supports and how they compose into an assurance argument.

The capability’s evolution introduces a second class of proof obligations. In an environment where an agent dynamically selects a tool, a change in documentation, description ranking, parameter defaults, or error semantics can affect which operation is chosen without requiring an explicit code modification in the agent runtime. Interface releases should therefore be evaluated for semantic compatibility, including the business meaning of each field, preconditions, postconditions, authorization assumptions, and the interpretation of ambiguous outcomes. Version pinning and contract negotiation can reduce uncertainty, but they must be supported by regression tests and by mechanisms capable of detecting implementation drift when the vendor’s observed behavior no longer matches the approved specification.

A particularly valuable property would be independence from any single reasoning implementation. If the capability contract, approval rules, and outcome evidence are external to the agent’s transient conversational context, one agent model can be replaced by another while the institutional meaning of the operation remains stable. This separation is not sufficient for seamless substitution, because different models may interpret descriptions differently and exhibit different failure modes; it nevertheless makes the resulting differences testable against a persistent external specification rather than allowing the operational contract to change whenever the model changes.

Software built for agents should also support legible non-action. The contract must allow a caller to discover that an operation is unavailable, its required evidence is missing, the applicable mandate is disputed, or a previous invocation has an unresolved outcome. Treating these states as first-class results is preferable to forcing the agent to choose between repeated attempts and improvised alternatives. The more consequential the operation, the stronger the argument for separating a capability’s preparation, reservation, authorization, commitment, and verification into distinguishable states—although unnecessary fragmentation can introduce its own overhead, races, and failure points.

The organizational economics of software formalization

The shift toward agent-operable contracts also has implications for the economics of organizational boundaries. Garicano’s account of knowledge-based hierarchies models the trade-off between acquiring specialized knowledge and communicating with the people capable of solving unfamiliar problems.²¹¹ In software-mediated coordination, an analogous cost arises whenever a new consumer must reconstruct an application’s business semantics through developer interviews, undocumented conventions, special-purpose integration code, and trial-and-error verification. Greater formalization can make selected knowledge about the software capability reusable across consumers, although the knowledge thereby captured may be expensive to acquire and maintain in the first place.

This is an instance of a broader complementarity between information technology and organizational change. Bresnahan, Brynjolfsson, and Hitt found firm-level evidence that investment in information technology was complementary to workplace reorganization and related changes in skill requirements; the findings do not establish any particular modern agent contract, but they undermine the assumption that software productivity can be evaluated independently of the organization within which the software is deployed.²¹² A more explicit capability interface may reduce some integration and verification costs only when the organization has also defined authoritative records, business roles, mandates, exception procedures, and the responsibilities for maintaining the contract.

The economic decision can be represented without inventing a universal formalization metric. Consider a capability expected to be used N times over a defined evaluation horizon. Let F be the initial specification and assurance investment, M its maintenance cost over that horizon, c_f the expected per-use cost under the formalized design, and c_0 the corresponding cost of an otherwise comparable baseline. Let L_f and L_0 be the expected losses from errors, unauthorized actions, and recovery under the formalized and baseline designs respectively, on a consistent monetary and risk-assessment basis. Then the two expected total costs can be expressed as

\begin{aligned} C_f(N) &= F+M+Nc_f+L_f,\\ C_0(N) &= Nc_0+L_0. \end{aligned} \tag{15}

Under the simplifying assumptions that these quantities are comparable, the per-use costs are approximately constant over the horizon, and \Delta c=c_0-c_f>0, formalization has lower expected cost precisely when

N>\frac{F+M+(L_f-L_0)}{\Delta c}. \tag{16}

The inequality is a decision rule under stated assumptions, not an empirically established formula for enterprise productivity. In particular, L_f-L_0 may be negative if stronger assurance reduces expected losses, while F and M may rise sharply with semantic volatility or the complexity of compliance requirements; some losses may also be unacceptable irrespective of their expected monetary value and should instead appear as hard constraints. If \Delta c\leq0, the simple repeated-use saving cannot justify the investment, although risk reduction, interoperability, or strategic option value may still do so. Contract reuse across applications and firms can spread fixed costs, but only where domain semantics and the rights governing reuse actually remain compatible.

A more complete organizational assessment would examine who bears these costs and who captures the benefits. A software vendor may bear the expense of describing and testing a capability while many customers and agent providers benefit from its improved interoperability; the resulting coordination problem can encourage common standards, contractual obligations, or platform governance. Conversely, a proprietary contract language may shift costs from runtime interpretation into provider dependence, increasing asset specificity and switching costs even as the immediate tool integration becomes easier. Formalization can therefore support both a more modular market of interchangeable capabilities and a more concentrated infrastructure in which one provider controls the accepted representation of business semantics.

There is also a trade-off between modularity and epistemic adequacy. Highly explicit contracts reduce the need for callers to understand internal implementation details, but a rigid interface can conceal assumptions whose relevance becomes apparent only under unfamiliar conditions. An agent may correctly satisfy a contract designed around normal purchasing while failing to recognize a novel regulatory restriction, technical incompatibility, or exceptional supplier relationship. The organization must therefore maintain a pathway through which actors can contest the contract itself, request specialist interpretation, and initiate governed revision. Otherwise formalization may reduce the cost of routine action while increasing the cost of adaptation to unforeseen circumstances, reproducing the exploration–exploitation tension described by March.²¹³

Security, specification integrity, and the authority to interpret behavior

Making software capabilities more discoverable increases both their usability and the importance of their descriptions as possible attack surfaces. An agent that receives a tool description, response, or business schema from a lower-trust source may treat it as a reason to invoke a privileged operation, even when the source cannot legitimately grant that authority. The MCP specification’s explicit warning about untrusted tool annotations reflects this general problem: descriptions and behavioral hints are data provided by a tool server, not authoritative institutional policy.²¹⁴ The MCPTox benchmark reported successful tool-poisoning attacks under defined experimental conditions, demonstrating the plausibility of this mechanism without establishing a real-world prevalence rate for compromised enterprise systems.²¹⁵

A formal capability contract can reduce some ambiguity, but it can also increase the impact of a compromised specification if the organization automatically trusts machine-readable metadata. An attacker who changes a tool’s declared preconditions, mislabels a destructive operation as read-only, or alters an ontology mapping may steer an agent into actions that appear consistent with the advertised interface. The defensive requirement is therefore specification integrity: authenticated origins for approved contracts, independently managed authorization policy, controlled changes to capability registries, enforcement at the actual transaction boundary, and the ability to compare declared behavior with observed effects. A signature establishes the identity and integrity of the signed artifact; it does not prove that its statements are correct, that the signer was authorized to create the mandate, or that the implementation conforms to the description.

This distinction has a direct organizational analogue. A business rule can be syntactically valid and technically enforceable while contradicting the company’s lawful purposes or existing contractual obligations; similarly, a capability specification can be formally satisfiable while describing an operation the enterprise should never permit. Protecting the system requires separation among those who define domain semantics, those who approve mandates, those who implement capabilities, and those who verify business outcomes. These responsibilities need not correspond to four different departments in every company, but their relevant controls and evidence must remain independent enough to detect common-mode errors and deliberate manipulation.

The problem becomes sharper in industrial environments. An agent-facing interface that requests a revised schedule or operating set point cannot be assumed safe merely because its parameters satisfy a schema and its caller is authenticated. Physical constraints, control-system authority, interlocks, permissible operating envelopes, and the timeliness and quality of measurements may determine whether the command is valid. The implementation must preserve independently enforced safety and control boundaries, recognizing that an acknowledgment of a command is not evidence of a safe physical state. In such settings, the agent-operable contract may appropriately expose requests for approval or planning services while withholding direct access to safety-critical actuation altogether.²¹⁶

The resulting security challenge is not solved by assigning more confidence to the model or by adding warnings to natural-language prompts. It arises at the point where descriptions, evidence, and institutional authority are composed into executable capabilities. The stronger the relationship between software specifications and autonomous action, the more important it becomes to govern who may define, revise, authenticate, and enforce the specifications themselves.

Research agenda: does richer formalization produce reliable agent-operable software?

The proposed contract model should be treated as a hypothesis about how to improve agent-mediated business processes, not as evidence that maximal formalization is always desirable or possible. A falsifiable research program would first have to distinguish the contribution of structural interface descriptions from domain semantics, executable constraints, and independent assurance. If an agent performs better with a richly annotated tool description, the effect may result from additional explanatory text or increased context length rather than a stronger machine-checkable contract; if a workflow becomes safer after the introduction of a gateway, the benefit may derive from constrained privileges rather than the semantic sophistication of the tool metadata. Experimental comparisons must therefore vary these mechanisms separately where feasible.

A useful initial study could assign otherwise comparable agent configurations to four interface conditions: structurally typed APIs with ordinary documentation; the same APIs with explicit domain-semantic descriptions; interfaces augmented with enforceable preconditions, mandates, and transaction controls; and a design in which the preceding controls are additionally supported by model checking or independently tested invariants. Model families, context budgets, permitted business capabilities, task distributions, execution limits, and available evidence should remain matched as closely as the design allows. The experiment should measure actual transaction effects, not merely the agent’s stated confidence or the syntactic success of a tool call, and should include both routine tasks and carefully controlled cases involving conflicting records, stale information, concurrent commitments, misleading tool descriptions, and partial failures.

The principal outcome would be authorized end-to-end process completion: a verified business result obtained within the applicable institutional mandate, with observation of the resources consumed and the interventions required. Secondary outcomes should include incorrect tool selection, precondition violations, unauthorized commitments, duplicate effects, false beliefs about successful execution, latency, recovery expense, and the fraction of cases in which the system correctly defers because its evidence is insufficient. A stronger specification might reduce error while increasing latency or introducing new failure modes through misconfigured policies; both effects must be measured rather than treated as methodological inconveniences. The study should also distinguish violations prevented by the agent from violations prevented by the independent gateway, because these are different forms of system capability.

A longitudinal study would test whether the contracts remain useful as the environment changes. Application versions could be replaced while keeping the business contract fixed, and agent models could be replaced while preserving the same approved capabilities; investigators would then examine whether observed outcomes continue to refine the declared semantics, whether hidden dependencies require repeated manual reconstruction, and whether changes to institutional mandates remain distinguishable from changes to software behavior. Semantic drift, obsolete policies, and the maintenance cost of contracts would be primary outcomes rather than residual implementation issues. Such experiments would directly test the claimed economic complementarity between more capable agents and more explicitly specified software.

Three propositions summarize the appendix’s distinct empirical claims. First, for tasks in which business semantics are consequential and not recoverable reliably from syntax, explicit domain contracts should improve process quality beyond schema-validating APIs and additional unconstrained model reasoning. Second, for concurrent or long-running consequential workflows, independently enforced transaction invariants and verified outcome semantics should reduce unauthorized and duplicate effects compared with agent-side reasoning alone, although the size and cost of the improvement will depend on the domain. Third, the net benefit of stronger specification should increase with opportunities for contract reuse and consequential error reduction, but decrease when semantics change too quickly, maintenance costs dominate, or the imposed abstractions conceal essential context. Each proposition admits counterevidence: negligible improvement under matched conditions, failure of the enforcement implementation, or negative total value after maintenance and recovery would constrain the claim.

The appendix thus extends, without replacing, the central argument of The Firm That Thinks. The executable institutional constitution expresses the organizational conditions under which action is legitimate; the Company Brain preserves the evidentiary context on which action is proposed; agent-operable software contracts attempt to make the meaning and permissible consequences of particular computational capabilities sufficiently explicit for those actors to use them reliably. Their interaction is neither an automatic consequence of MCP connectivity nor a demand that every enterprise application become formally proven. It is a design and research problem concerning the degree of semantic, behavioral, transactional, and institutional explicitness required to transform computational possibilities into authorized, observable, and correctable organizational action. If agentic enterprises become economically viable, some of their most consequential innovations may therefore occur below the language-model layer, in the engineering of software whose capabilities can be interpreted, constrained, and verified independently of the agents that invoke them.

Appendix B — Building and growing a Company Brain: from enterprise evidence to institutional cognition

A theory of the cognitive enterprise is incomplete if its central object cannot be distinguished operationally from the software that organizations already possess. Enterprise resource planning (ERP) systems maintain authoritative transactions; analytical platforms integrate data; knowledge-management systems support the creation and retrieval of knowledge; workflow engines preserve process state; and increasingly capable language models make these resources accessible through conversational and tool-using interfaces. An organization can deploy all five without acquiring a durable ability to reconstruct why a decision was justified, identify the assumptions under which it was valid, establish which consequences actually followed, or retain a correction after the participants and models involved have been replaced. Conversely, the existence of such capabilities does not require an omniscient model, a universal ontology, or a new software platform that replaces all existing systems. The relevant engineering problem is therefore one of institutional composition: how to assemble partial but authoritative sources into a governed mechanism for evidence-based organizational judgment, decision continuity, and learning.²¹⁷²¹⁸

This appendix develops a self-contained implementation theory for that mechanism. Its purpose is neither to declare that a general Company Brain already exists nor to prescribe a universal product stack; it is to give the concept observable requirements, formal representations, a credible initial deployment, and a method of expansion under which technical integration and organizational legitimacy remain distinguishable. The argument connects database theory, temporal logic, knowledge representation, distributed systems, process mining, and decision provenance with the organizational economics of specialization, information processing, routines, and governance. The critical proposition is that an institution begins to acquire computationally persistent cognition when the grounds of its consequential judgments become reconstructible, contestable, reusable, and correctable across people, applications, and time, rather than merely easier to retrieve in natural language.

From a metaphor to an operational specification

The expression Company Brain can invite two misleading interpretations. The first treats it as the enterprise’s singular intelligent agent, capable of speaking or deciding on behalf of the company. The second treats it as a technical synonym for an enterprise knowledge graph, vector index, or repository of embedded documents. Neither interpretation captures the proposed institutional property. A retrieval system can answer a question without establishing whether its retrieved sources were authoritative, current, or mutually compatible; an agent can execute a plausible plan without having a right to commit resources; a graph can preserve relationships without knowing whether they represent observations, beliefs, disputed assertions, or binding obligations. The theoretical object must be defined by what it enables the organization to establish and retain, not by the internal architecture of a particular model.

Operational definition. A Company Brain is a governed, potentially federated computational capability that connects identifiable sources of enterprise state, knowledge, experience, and institutional rules through explicit semantic mappings, temporal qualifications, provenance, and access constraints; it supports the construction and contestation of evidence-based organizational judgments, preserves the relationship between decisions and the information available when they were made, links those decisions to independently established outcomes, and maintains controlled procedures for correcting its representations. Human and artificial participants may consult it, but its knowledge services do not themselves constitute corporate authority, and its representations do not displace the records or professionals responsible for their validity.

The definition yields a practical distinction between necessary ingredients and sufficient evidence of integration. Documents, indexed text, schemas, event streams, graph edges, workflow histories, and automated reasoning are possible ingredients. The integration claim becomes meaningful only when these ingredients jointly support reproducible organizational activities: a user can ask what was known at a specified time; obtain the sources, interpretation rules, and uncertainties supporting an answer; reconstruct a significant decision without relying on the original participant’s memory; establish whether the expected effect occurred; and determine what subsequent correction, if any, was accepted. The claim remains falsifiable, because these tasks can fail even when search relevance, chatbot fluency, or data ingestion volume is high.

For precision, let an organization possess a set of domain systems \mathcal S, participants \mathcal U, institutional policies \mathcal P, and decision processes \mathcal W. A proposed Company Brain implementation B exposes evidence and memory services to authorized participants through an interface Q_B; it does not acquire independent legal or transaction-execution authority simply by exposing those services. For a question q, principal u, valid-time reference t_v, and knowledge-as-of time t_k, an admissible query result may be represented as

Q_B(q,u,t_v,t_k)=\bigl(R,\mathcal E,\mathcal A,\mathcal X\bigr), \tag{17}

where R is a candidate answer or structured judgment; \mathcal E is the set of evidence references supporting and challenging it; \mathcal A records applicable interpretation and authorization constraints; and \mathcal X identifies material gaps, contradictions, and unresolved uncertainties. These objects are all relative to the permissions of u and the versioned state of the underlying sources. The expression is a proposed service contract, not a theorem that a correct answer exists or can always be computed: a valid result may be a qualified refusal to conclude, and a successful technical query may still produce an unjustified judgment if its sources or modeling assumptions are wrong.

This model also differentiates a Company Brain from a digital twin of the organization. A twin usually emphasizes representation of the state and dynamics of a target system, often for monitoring or intervention; an organizational cognitive infrastructure additionally requires an account of the provenance, legitimacy, and temporal availability of the evidence and decisions that inform action. These aims can overlap, and neither term guarantees a complete or causally valid model of the firm. The proposed definition likewise extends—but does not invalidate—the knowledge-based view of organizations as mechanisms for integrating specialist knowledge: Grant’s analysis emphasized knowledge integration rather than the presumption that knowledge must be centralized, while distributed cognition treats people, instruments, representations, and activity as a combined unit of analysis.²¹⁹²²⁰

A minimum viable Company Brain should therefore be distinguished from the infrastructure needed to reach it. An access-controlled, read-only retrieval service is a legitimate preparatory stage, but it does not yet demonstrate the central property. A defensible minimum needs a bounded business domain with stable authoritative identifiers, source-linked and temporally qualified assertions, at least one persistent decision dossier, and a later mechanism that independently connects the decision to observable consequences and approved corrections. Autonomously executing the decision is optional: the smallest meaningful implementation can remain entirely advisory, allowing the institution to validate its cognitive substrate before granting agents consequential permissions.

The institutional evidence model: identity, provenance, time, and epistemic status

The elementary object is not a document chunk or an embedding, because neither represents the conditions under which a proposition can be treated as evidence. Consider the statement that forty components are available for production. It can mean forty physically present units, forty units owned by the company, forty units released by quality control, forty units unreserved by other orders, or forty units deliverable before a particular date; none of these predicates is logically interchangeable. Moreover, the assertion may originate in an ERP snapshot, a warehouse count, a supplier notification, or an agent’s inference, and its validity can change before the organization learns of the change. Without explicit identity, semantics, and temporal scope, integrating records can manufacture apparent agreement rather than resolve uncertainty.

A workable theoretical representation is an institutional assertion record

\mathfrak a_i=\bigl(\iota_i,s_i,p_i,o_i,\tau_i,\sigma_i,V_i,R_i,\pi_i,\epsilon_i\bigr), \tag{18}

where \iota_i is a stable assertion identifier; (s_i,p_i,o_i) is a typed subject–predicate–object proposition, or an equivalent domain-specific relation; \tau_i identifies the asserted kind of statement, such as observation, external report, inference, commitment, or policy; \sigma_i identifies the originating source and its version; V_i is the interval over which the assertion claims validity in the modeled business world; R_i is the interval during which that assertion-version was recorded as current in the institution’s evidentiary system; \pi_i is its derivation and dependency description; and \epsilon_i records qualified epistemic status, including dispute, supersession, or relevant uncertainty. This is an illustrative application data model, not a W3C standard or a claim that all organizational propositions can be exhaustively reduced to binary triples. Security labels, legal basis, ontology versions, confidence-calibration metadata, and human approvals can be modeled as further controlled attributes or linked records.

The separation between V_i and R_i draws on bitemporal database theory: valid time concerns when a fact is asserted to hold in the modeled domain, whereas transaction time concerns when the database represented it as current. Jensen and Snodgrass showed why these temporal dimensions answer different questions and are useful for accountability and historical reconstruction.²²¹ For example, an incident may occur on Monday, be reported on Wednesday, and be corrected on Friday. A review of Monday’s objective conditions differs from a review of what the procurement manager could reasonably have known on Tuesday. The latter depends on the institutional record available at that moment, not on a retrospectively perfected reconstruction of the world.

Let \mathcal K^{(k)} denote the assertions and interpretations registered by knowledge time k, and let \operatorname{Perm}(u,a,k) represent authorization to inspect assertion a at that time. A historical evidence view can then be defined, schematically, by

\mathcal E(u;v,k)=\left\{\mathfrak a_i:\; v\in V_i,\; k\in R_i,\; \operatorname{Perm}(u,\mathfrak a_i,k)=1 \right\}. \tag{19}

This construction assumes that the system retains sufficient historical records of assertions and policy versions, and that the evaluation of historical authorization is defined. It does not reconstruct what a human actually read or mentally understood; that stronger claim would require decision-specific evidence of access, presentation, and acknowledgment. It also does not equate recording time with the time at which the true state changed. Distinguishing these questions is essential whenever a regulator, auditor, customer, or internal review body asks whether a decision was reasonable given the evidence then available rather than whether it appears correct after later information became known.

Provenance is the second foundation. W3C PROV-O provides classes and relations for entities, activities, and agents involved in producing or using information, while OpenLineage models observed data-processing jobs, runs, datasets, and associated metadata. These mechanisms are complementary: the former can express general derivation and attribution, whereas the latter concentrates on data-production pipelines and their execution events.²²²²²³ Neither proves the source’s substantive truth. A cryptographically signed supplier statement may establish the purported signer and message integrity while still conveying an incorrect delivery forecast; a perfectly recorded transformation may faithfully propagate a mistaken product identifier. The Company Brain must preserve this distinction between traceable origin and warranted acceptance.

The resulting dependency structure can be represented as a directed graph G_E=(N_E,E_E) in which vertices include assertion versions, source records, inference operations, decisions, and outcomes, while typed edges record derivation, contradiction, supersession, support, and use. Dependencies should identify shared ancestry: if five agents repeat the same unverified supplier claim, there are five outputs but potentially only one underlying evidentiary source. Inference pipelines should be versioned so that a later correction can identify which claims and decisions depend on the affected assumption. A provenance graph can contain cycles of organizational influence across time, but its derivation subgraph for a single completed computation should be constructed or versioned so that the meaning of its dependencies remains inspectable; ordinary graph connectivity is not proof of independently corroborated evidence.

Finally, the evidence model needs a rule for absence. Under a closed-world database query, no matching row may have a defined interpretation within a complete database snapshot; across enterprise applications, documents, events, and external communications, the absence of a retrieved assertion generally does not establish that the corresponding event or condition never occurred. A claim of completeness therefore requires a declared boundary—for example, a reconciled order ledger for a specified company and period—together with evidence that the extraction covered that boundary. This is a practical instance of a broader epistemological limit: no model can turn incomplete observation into guaranteed knowledge by increasing representational sophistication alone.

Semantic federation: preserving domain meaning without constructing an omniscient ontology

An enterprise’s systems are partitioned partly because its activities have different owners, temporal requirements, and semantic conventions. In sales, a customer may denote a commercial relationship; in accounting, it may denote a debtor legal entity; in field service, it may denote a site or the recipient of a service. Similarly, a supplier’s reported available capacity is not the same object as internally released inventory, even if both can contribute to an available-to-promise calculation. Building a Company Brain therefore requires identifying where semantic translation is justified and where differences must remain visible, rather than replacing every local model with one universal vocabulary. This accords with Simon’s account of nearly decomposable systems: strong internal relationships and weaker but consequential cross-boundary interactions can favor modular representation with explicit interfaces.²²⁴

A federated implementation can be described as a family of domain models \{\mathcal M_d:d\in\mathcal D\} connected by versioned partial mappings \mu_{de}, where each mapping specifies when a concept in domain d can be interpreted in domain e. A mapping should record its owner, validity assumptions, source types, transformations, and known exceptions. If \mu_{de} is partial, some entities or attributes have no sound translation; if several mappings yield different results, the infrastructure must preserve the disagreement instead of selecting one through a hidden model heuristic. In general, \mu_{ef}\circ\mu_{de} is not automatically equivalent to a separately approved \mu_{df}, and chains of mappings may amplify uncertainty or change granularity. These properties make semantic governance a persistent organizational responsibility rather than a one-time data migration exercise.

Established technologies can implement subsets of this model without compelling a single data architecture. W3C R2RML expresses mappings from relational schemas to RDF representations, while SHACL defines constraints for validating RDF graphs; DCAT 3 supports descriptions of datasets and data services across interoperable catalogs.^225226227 An enterprise can instead maintain equivalent governed mappings in relational catalogs and application services where they are easier to operate. A graph database is beneficial when identity resolution, dependency traversal, or provenance queries justify graph-oriented access, but it is not a logical prerequisite of institutional cognition. Nor is a vector index a source of authority: it can improve candidate discovery while leaving the truth, currency, and permissible use of the underlying assertions to independent mechanisms.

The temporal and semantic problems become inseparable when event-based integration is introduced. A CloudEvents envelope specifies interoperable event metadata such as id, source, and type, and can carry an occurrence-time attribute, but its event schema does not define the business validity of the payload or establish a globally consistent enterprise chronology.²²⁸ Likewise, object-centric event logs allow events to be related to multiple business objects, which is important when one supplier notification affects several orders, parts, and production cases; the OCEL 2.0 specification offers a concrete representation for that kind of process evidence.²²⁹ These techniques help preserve the relationship between different organizational objects, but the institution must still specify the meaning of each relationship and the status of the observed events.

The most defensible architecture is thus logically connected but physically and institutionally federated. Domain systems remain authoritative for their own transactional facts; domain stewards approve semantic mappings and interpretation rules; a shared capability layer provides authorized cross-domain retrieval, provenance traversal, and decision reconstruction; and technical controls enforce data residency, confidentiality, purpose limitation, and access constraints. This arrangement neither abolishes organizational specialization nor turns federation into an excuse for irreconcilable fragmentation. It gives distributed expertise a structured way to cooperate while allowing an incorrect cross-domain equivalence to be detected, contested, and revised.

From information retrieval to durable decision memory and justified revision

A conventional question-answering service usually treats the question and its answer as the main unit of interaction. Organizational memory requires a more persistent object: the decision episode, including what problem was recognized, which evidence and uncertainties were available, which alternatives were considered, who or what was authorized to act, what happened, and which subsequent observations changed the assessment. Such an object also has an organizational-theoretic foundation. Feldman and Pentland distinguish the abstract pattern of a routine from its situated performances, explaining why routines can preserve continuity while generating variation and change; an executable process definition alone cannot capture the evolving reasons for exceptional performances.²³⁰

A proposed decision dossier can be represented as

\mathfrak d=\bigl(q,\mathcal E,\mathcal H,\mathcal O,\mathcal M,\delta,\mathcal Y,\mathcal J\bigr), \tag{20}

where q states the decision problem; \mathcal E contains snapshot-resolvable evidence and its known defects; \mathcal H contains alternatives and relevant hypotheses; \mathcal O records the organizational actors and expertise contributing to assessment; \mathcal M identifies the operative mandates, policies, and approvals; \delta records the selected decision and its execution references; \mathcal Y records subsequently observed outcomes, including failures and inconclusive states; and \mathcal J contains evaluations and approved revisions. Each component should carry time, provenance, and access information. This is a schema for reconstructing the institutional decision record, not a proof that every human motive, alternative, or unspoken consideration was captured. Its minimum useful form may be a small relational aggregate with references to source-system records, rather than a comprehensive knowledge graph.

Dossiers must distinguish the decision from its intended effect. A purchase order recorded as submitted may not have been accepted; an accepted order may not have been delivered; a delivered part may fail inspection; and a production schedule changed in response may nevertheless miss the promised date. The Brain should not promote workflow status to business truth. Durable workflow systems can persist ordered event histories and reconstruct execution state after failures, as illustrated by Temporal, while transactional-outbox implementations can reduce mismatches between a committed database change and subsequent event publication.²³¹²³² Yet neither capability alone proves that the externally intended outcome occurred. Independent reconciliation against authoritative records remains necessary, and ambiguous distributed outcomes must remain marked as unresolved until suitable evidence arrives.

The distinction between recording an outcome and learning from it is equally important. If a supplier failed to deliver despite a strong prior reliability estimate, the resulting record can support investigation without establishing why the failure happened or what intervention would have prevented it. Causal conclusions require assumptions about selection, confounding, measurement, and alternative outcomes. In addition, organizational learning may alter different objects: a source record may be corrected; a statistical forecast may be recalibrated; an interpretation rule may be revised; a process may be redesigned; or a governing policy may be amended by a competent authority. Updating a prediction model is not equivalent to amending the organization’s rules, and an agent’s discovery that a rule is inconvenient does not authorize the agent to change it.

Formal research on belief revision helps explain why institutional corrections need explicit dependencies. Doyle’s truth-maintenance system records the reasons supporting beliefs so that assumptions can be revised when contradictions arise; the Alchourrón–Gärdenfors–Makinson tradition examines logical postulates for changing a belief set under specified assumptions.²³³²³⁴ Neither formalism can be imported unmodified as a complete theory of enterprise governance: corporate evidence can be probabilistic, inconsistent, legally privileged, and produced by actors with unequal authority, while decisions may create obligations even when later evidence contradicts their premises. Their important engineering implication is narrower: revising a conclusion requires preserving its dependencies and the institutional procedure that determines what may legitimately be replaced.

An effective Company Brain consequently supports a controlled loop: new evidence may lead to a corrected view; a corrected view may motivate a changed operating plan; observed outcomes may test the plan’s assumptions; and proposals to revise authoritative semantics, policies, or mandates enter separately governed change processes. March’s analysis of exploration and exploitation warns that learning systems optimized only to reinforce established knowledge can undermine their longer-term adaptability, while Nonaka’s account of organizational knowledge creation emphasizes interactions between tacit and explicit knowledge that no storage architecture can fully automate.²³⁵²³⁶ A Brain worthy of the name must therefore retain possibilities for independent judgment, disagreement, and the introduction of evidence that its current ontology or model fails to anticipate.

The minimum deployable architecture: a bounded industrial decision process

An implementation strategy should begin with an organizational question whose quality can be assessed independently of the software answering it. Suppose a manufacturer faces recurrent disruptions in the supply of a component used in several products. The process already spans purchasing, engineering, inventory, production planning, customer commitments, and finance, yet a first deployment need not connect every application or automate purchasing. It can select a bounded population of critical components and reconstruct the decision process from supplier alerts to resolved production impact. The objective is not to create an artificial procurement director; it is to make the information and reasoning behind disruption decisions less dependent on undocumented individual recollection.

The initial system boundary could encompass ERP purchase-order lines and supplier identities; released, reserved, and quality-held stock; product structures and qualified alternatives; a read-only view of relevant production and customer commitments; and authenticated supplier correspondence. A small domain catalog records how the different applications define supplier, part, available stock, delivery date, and qualification. Event adapters and scheduled snapshots preserve changes and their observation times; a source-reference store retains document versions and transaction identifiers; and the decision-dossier service relates these materials to one case with explicit status and review ownership. A model may help identify candidate relationships or summarize conflicting evidence, but consequential assertions are retained with source attribution and reviewed against domain-specific rules.

Figure 6 depicts an illustrative read-mostly deployment. It is intentionally smaller than the complete agent-native enterprise architecture: external transaction systems continue to execute business operations, and the Brain’s service layer exposes evidence, decisions, and verified observations without creating a privileged path for autonomous writes. The independent verification input makes the design more than a search interface, while the policy and stewardship boundaries prevent the reasoning layer from declaring its own hypotheses authoritative.

%%{init: {"theme": "neo", "look": "handDrawn", "layout": "elk"}}%%
flowchart TB
  subgraph SRC[Authoritative enterprise sources]
    ERP[(ERP orders and inventory)]
    SUP[(Supplier messages and documents)]
    OPS[(Production and quality records)]
  end
  subgraph EVID[Governed evidence and decision substrate]
    ING(Authenticated ingestion)
    SEM(Semantic and temporal reconciliation)
    EV[(Versioned evidence and provenance)]
    DOS[(Decision dossiers)]
    Q(Governed context and reasoning service)
    RV(Independent outcome reconciliation)
    ERP -->|committed records| ING
    SUP -->|signed or attributable messages| ING
    OPS -->|observed state| ING
    ING -->|qualified updates| SEM
    SEM -->|versioned assertions| EV
    EV -->|scoped supporting material| Q
    DOS -->|past decisions and reasons| Q
    RV -->|verified outcomes| DOS
    ERP -->|authoritative readback| RV
    OPS -->|delivery and quality readback| RV
  end
  PEOPLE[Procurement, engineering and finance participants]
  STEW[Domain owners and evidence stewards]
  POLICY[(Identity and access policies)]
  Q -->|grounded analysis| PEOPLE
  PEOPLE -->|decisions and corrections| DOS
  STEW -.->|governs mappings and acceptance| SEM
  POLICY -.->|constrains disclosure| Q
  POLICY -.->|constrains dossier access| DOS

Figure 6: Minimum viable Company Brain for supplier disruption: existing systems remain authoritative, a governed evidence substrate supports decisions, and independent readback closes the learning loop without agent-controlled transaction writes.

For example, at 09:00 on Monday the ERP reports forty released and unreserved units of component C-18, while a supplier email received at 10:00 claims that the next delivery of sixty units will arrive on Friday. At 15:00, an engineering note establishes that a proposed substitute is not qualified for one customer product. On Wednesday, the supplier corrects the delivery estimate to the following Tuesday. A meaningful system must preserve four different things: the state of inventory and its operational semantics on Monday; the status of the supplier statement as an externally reported forecast; the technical constraint that defeats one substitution; and the fact that the revised forecast became available only on Wednesday. A dossier created on Tuesday should be reconstructible as it stood on Tuesday, while a retrospective case review should also show what later changed.

The implementation’s first substantive deliverable would therefore be a queryable case record containing the original business question, the affected objects, source-linked observations, explicitly recorded conflicts, alternative responses, assigned human authorities, and unresolved uncertainties. Its second would be a decision snapshot: an identified set of evidence and applicable policy versions at the moment an authorized participant chose a response. Its third would be a reconciliation record showing whether an alternative delivery, production-plan revision, or customer communication actually occurred. A deployed LLM may produce the narrative explanation, but the durable authority for the explanation must rest in traceable sources and reviewable transformations, not in an opaque transcript of generated text.

This example also exposes what a minimum must exclude. It need not replicate the ERP database, create an enterprise-wide ontology, embed every document, or produce autonomous supplier negotiations. It must nevertheless address source completeness, identity ambiguity, conflicting timestamps, domain-specific constraints, and incomplete outcomes. It must be possible for a domain owner to reject the interpretation of a record, amend a mapping through change control, and preserve both the superseded and corrected versions. A component search tool with accurate citations may be useful, but the transition from search infrastructure to institutional cognitive infrastructure occurs only when decision continuity and outcome-aware correction become operationally testable.

Growing the Brain without centralizing the firm

A successful local implementation creates a temptation to extend the same canonical data model and reasoning service to every domain. That strategy may simplify early integration while increasing the cost of later change, because production, finance, human resources, commercial contracts, and industrial operations differ in authority, semantics, temporal granularity, confidentiality, and acceptable failure modes. The more defensible growth mechanism is a federation of domain cognitive capabilities that agree on limited cross-domain contracts while retaining local stewardship. This is a direct extension of the idea of nearly decomposable organizations: internal coherence should be strong enough to sustain useful local work, and cross-domain connections should be explicit enough to support consequential coordination.

Let \mathcal D be the set of onboarded domains and G_D=(\mathcal D,E_D) a graph whose edges denote documented semantic or process dependencies. For a domain d, onboarding requires an identifiable steward, declared systems of record, authoritative business concepts, admissible uses, temporal and provenance coverage, conflict-resolution procedures, and a bounded set of supported questions. An edge (d,e) is ready for operational use only when the relevant mappings, evidence obligations, access constraints, and cross-domain decision responsibilities have been reviewed. The result is not one universal global model, but a network of bounded semantic agreements that can be versioned, challenged, and replaced without silently rewriting local business meanings.

Figure 7 shows the difference between federation and central accumulation. The shared services provide identity resolution under controlled rules, metadata discovery, distributed query coordination, and provenance tracing; the authoritative knowledge and operational events remain within domain-defined boundaries. The graph of approved mappings is separate from an organizational body’s authority to change those mappings, and a domain’s decision dossier is not promoted to a company-wide rule merely because it can be discovered by another domain.

%%{init: {"theme": "neo", "look": "handDrawn", "layout": "elk"}}%%
flowchart TB
  subgraph D1[Supply and procurement domain]
    P1[(Evidence and source references)]
    P2[(Procurement decisions)]
    P3[Procurement steward]
    P3 -.->|validates| P1
  end
  subgraph D2[Production and engineering domain]
    E1[(Engineering evidence)]
    E2[(Qualification decisions)]
    E3[Engineering steward]
    E3 -.->|validates| E1
  end
  subgraph D3[Finance and commercial domain]
    F1[(Financial and contract evidence)]
    F2[(Commercial decisions)]
    F3[Finance steward]
    F3 -.->|validates| F1
  end
  subgraph FED[Shared federation services]
    C(Federated evidence and dossier queries)
    MAP[(Approved semantic mappings)]
    ID[(Identity and purpose policies)]
    GOV[Authorized semantic governance]
    MAP -.->|constrains translation| C
    ID -.->|constrains disclosure| C
    GOV -->|publishes reviewed changes| MAP
  end
  P1 -->|qualified evidence| C
  P2 -->|case references| C
  E1 -->|qualified evidence| C
  E2 -->|case references| C
  F1 -->|qualified evidence| C
  F2 -->|case references| C
  C -->|scoped findings| P3
  C -->|scoped findings| E3
  C -->|scoped findings| F3

Figure 7: Growing the Company Brain through governed domain capabilities: a shared discovery and coordination fabric connects locally owned evidence and memory without converting the enterprise into one central source of truth.

A staged development strategy follows from this architecture, but it should not be misunderstood as a universal maturity ladder. The first stage establishes source accountability, including permissions, identifiable records, traceable observations, and measures of completeness for a bounded process. The second establishes domain semantics, including controlled identities and mappings sufficient for reliable questions. The third adds decision continuity, with human-reviewed dossiers and time-qualified reconstructions. The fourth closes the loop through outcome verification and governed correction. Only then does cross-domain expansion become a meaningful test of institutional cognition; access for agents can be introduced at any suitable level as a read-only or tightly bounded capability, provided it does not bypass source and authorization controls. One domain can remain limited to trustworthy records while another supports sophisticated outcome analysis, and neither necessarily needs autonomous action.

The transition criteria should be evidential rather than promotional. An organization should expand a domain when a known set of consequential questions can be answered under defined conditions, the answers can be traced to admissible evidence, key conflicts are exposed instead of hidden, historical decisions can be reconstructed, and an independent observer can establish whether relevant outcomes occurred. It should postpone broader deployment when the main evidence source is incomplete, core business identifiers cannot be resolved with acceptable ambiguity, or the decision process contains unmodeled authorities whose actions cannot be safely inferred from an API. Technical scale-out without semantic assurance can merely distribute error more efficiently.

A further difficulty concerns knowledge-maintenance debt. An additional source produces new assertions but also new versions, possible contradictions, confidentiality obligations, mappings, and dependencies that become expensive to keep current. Replacing an ERP or reorganizing a business unit may invalidate mappings whose assumptions were not documented; model upgrades may alter interpretation even when source records remain unchanged; and a frequently accessed but obsolete decision dossier may be more dangerous than an unavailable one. Growth therefore requires explicit mechanisms for detecting staleness, tracking mapping dependencies, expiring claims under stated rules, and revalidating critical interpretations. It also requires an exit strategy: domains should be replaceable or disconnected while preserving legally necessary records and the integrity of historical decisions.

An operating model for epistemic ownership, security, and institutional correction

The Brain’s governance requirements arise from the fact that organizational meaning is not an engineering artifact whose correctness can be declared by a central platform team. Business domains decide what an accepted purchase, released inventory unit, valid engineering substitution, or financially material exposure means in practice. Data stewards can manage records, formats, and mappings; process owners determine which evidence matters for business outcomes; cybersecurity and privacy authorities govern identity, permissible disclosure, and protection; and corporate organs retain the rights to approve mandates or policies that bind the organization. These roles can overlap in small enterprises, but the functions of supplying evidence, approving its interpretation, executing an action, and independently verifying its consequences must remain distinguishable when conflicts of interest or material risk are present.

This distribution of responsibilities is consistent with Galbraith’s information-processing theory: firms respond to uncertainty partly by adjusting the information available at decision points and partly by redesigning the distribution of tasks and authority.²³⁷ A Company Brain may reduce the cost of discovering information without resolving the organization’s competing interests or establishing a single correct answer to every ambiguous problem. Its governance should therefore include a defined route for contesting claims, preserving minority interpretations, obtaining independent review, and approving revisions to enterprise-wide semantic rules. The ability to see information, the right to correct a record, and the right to change organizational policy are different entitlements; increasing the first must not silently grant the others.

From a security perspective, the system aggregates precisely the kinds of context that can become valuable to an attacker: contractual relationships, supplier dependencies, internal vulnerabilities, decision precedents, and routes through business processes. Authentication of the query principal is necessary but insufficient; access should be evaluated against legitimate purpose, source restrictions, information classification, and the confidentiality of other participants. NIST’s zero-trust architecture provides a relevant reference for separating policy decisions and enforcement in resource access, while empirical prompt-injection research such as AgentDojo demonstrates that untrusted tool or document content can affect an agent’s subsequent actions.²³⁸²³⁹ An imported supplier document must therefore remain data, even if it contains imperative language, and a retrieved policy excerpt must not be treated as operative authorization unless the institutional policy service confirms its status.

The most subtle security failure is evidentiary laundering: a false supplier message is ingested, summarized, converted into a semantically typed assertion, cited in several dossiers, and later treated as corroborated because several agents reproduced it. A provenance graph helps expose shared ancestry but cannot determine the truth of its originating statement; the system also needs source-quality distinctions, contradiction detection, independent observations where available, and mechanisms for suspending reliance on compromised inputs. Security monitoring must therefore cover not only model and tool invocations but also changes to identity mappings, domain semantics, acceptance rules, and the policies determining which sources acquire institutional authority. Conversely, controls should not make the entire evidence substrate inaccessible to legitimate dissenters, since excessive concentration of epistemic power can produce its own form of organizational fragility.

Privacy and accountability impose additional limits. An organization may need to reconstruct an important decision without retaining unrestricted copies of employees’ conversations, private notes, or personal data; an append-only audit design does not, by itself, resolve lawful retention, deletion, correction, or access requirements. The implementation may have to separate minimal decision evidence from personal narrative, employ selective disclosure and retention controls, retain defensible legal holds, and use tamper-evident integrity mechanisms rather than assuming everything can be preserved indefinitely. The Brain is an institutional memory with responsibilities and rights, not a justification for universal surveillance. Where industrial control or safety-critical systems are involved, the epistemic layer should also remain distinct from time-critical protective logic and unauthorized OT control paths; evidence services may observe events without gaining the ability to override plant safety mechanisms.

Organizational learning becomes governable only when revisions are classified. A corrected ERP record changes the evidence; an updated mapping changes how two domain concepts are related; a recalibrated forecast changes a predictive instrument; a new purchasing threshold changes an institutional rule. The first may belong to an authorized record owner, the second to affected domain stewards, the third to model and process owners, and the fourth to designated corporate authority. Conflating these categories would allow an application that proposes a plausible interpretation to revise the rules that justify its own actions. The central constitutional safeguard is that a computation’s ability to identify a desirable change does not establish its authority to enact that change.

Organizational economics: when does the Brain create net value?

The engineering feasibility of a Company Brain does not settle its economic desirability. Source integration, semantic mapping, decision-history retention, verification, human stewardship, and security introduce real costs, while faster retrieval or more coherent decisions do not guarantee savings unless the organization changes how work is performed. Transaction-cost reasoning and the knowledge-based theory of the firm supply complementary explanations: better shared cognition may reduce the cost of finding and coordinating specialized knowledge, but it may also create new asset specificity, vendor dependency, governance obligations, and organizational concentration.²⁴⁰²⁴¹ The relevant economic comparison must include the total costs of producing a verified organizational outcome, not simply the cost of one answer or the number of human hours apparently replaced.

For a set S\subseteq\mathcal D of onboarded domains, define a period-specific, risk-adjusted net value under a declared evaluation boundary by

N(S)=\mathbb E\bigl[B(S)\bigr] -C_{\mathrm{int}}(S)-C_{\mathrm{sem}}(S) -C_{\mathrm{ass}}(S)-C_{\mathrm{ops}}(S) -\mathbb E\bigl[L(S)\bigr], \tag{21}

where B denotes the value of incremental process performance, decision quality, and retained knowledge; C_{\mathrm{int}} the technical integration cost allocated to the period; C_{\mathrm{sem}} the cost of semantic stewardship and revision; C_{\mathrm{ass}} the cost of independent verification and control; C_{\mathrm{ops}} ordinary operational costs including model inference; and L the losses attributable to errors, unauthorized disclosure, interruptions, and other relevant risks. These terms are a proposed accounting framework, not estimators that can be obtained accurately without a causal design. The expected-value form is inadequate where severe losses are rare but socially unacceptable; safety constraints, tail-risk measures, and hard legal limits must then be examined separately rather than averaged away.

The value of adding domain d depends on the domains already available:

\Delta_d N(S)=N(S\cup\{d\})-N(S). \tag{22}

This expression captures a nontrivial network effect. Integrating engineering qualifications may be of limited value if the Brain cannot connect them to procurement and affected customer orders, whereas joining the three domains can reveal previously hidden dependencies and enable a more valuable decision. Nevertheless, such complementarity is a hypothesis, not a proof that N is supermodular: a new domain can introduce incompatible concepts, compliance constraints, or correlated errors whose costs outweigh the additional benefit. The economic target is consequently not maximal data coverage or maximal agent adoption, but an integration sequence whose incremental value can be defended under uncertainty and whose newly created dependencies remain governable.

Classical decision theory supplies a useful benchmark for the value of additional information. Let A be a fixed feasible action set, \theta an uncertain state, U(a,\theta) the payoff, and I a signal observed before action. Under a fixed true probability model, costless observation, and a decision-maker able to choose an optimal action based on the signal, the expected value of sample information is

\operatorname{EVSI}(I)= \mathbb E_I\!\left[\max_{a\in A}\mathbb E\!\left[U(a,\theta)\mid I\right]\right] -\max_{a\in A}\mathbb E\!\left[U(a,\theta)\right]\geq 0. \tag{23}

The inequality follows because a decision-maker receiving I can always ignore it and choose the previously optimal action; it does not say that deploying a Company Brain cannot make the organization worse. In reality, information is costly to obtain and interpret, learned probabilities may be miscalibrated, actors can have different objectives and permissions, and flawed shared evidence can induce correlated mistakes. More information may be valuable in the ideal model but have negative realized organizational value once inference error, adaptation costs, and institutional constraints are included. This distinction prevents an otherwise tempting but invalid inference from theoretical information value to the economic inevitability of cognitive-enterprise infrastructure.

The accumulation of decision dossiers may generate a new category of organizational cognitive capital, insofar as the stored relationships among evidence, interpretations, exceptions, and verified outcomes remain useful across successive decisions. Its value is not proportional to record count: a small number of well-validated exceptions can be more valuable than a large corpus of redundant or stale narratives. Nor is the capital necessarily owned or controlled by the organization that uses it; contractual restrictions, labor contributions, external service dependencies, and vendor-specific formats may limit reuse. Economic evaluation should therefore include portability, institutional control over critical semantics, the cost of replacing models and providers, and the distribution of benefits between employees, shareholders, and external suppliers.

An evaluation program capable of rejecting the implementation thesis

An evaluation that reports chatbot answer accuracy or the number of sources ingested cannot establish the distinctive Company Brain hypothesis, because ordinary retrieval improvements could produce the same result. The appropriate unit of analysis is a consequential organizational process observed across time, including its evidence, decisions, authorization, execution, outcomes, and corrections. Baselines should isolate rival explanations: a conventional enterprise search system; the same search system with improved source curation; a semantic evidence service without persistent decision memory; a dossier-enabled system with outcome reconciliation; and, where permitted, a version augmented with bounded agents. Comparing all of these simultaneously is not always feasible, but the design must at least prevent benefits from being incorrectly attributed to a Company Brain when they arise from ordinary data-quality remediation or process redesign.

The proposed tests should be selected before deployment, and their evaluation criteria fixed independently of the system being measured. Table 5 specifies candidate tests whose passing thresholds depend on the risk and economic importance of the domain; none of the criteria should be treated as a universally valid numerical benchmark.

Test	Observable experimental task	Failure revealed
Historical reconstruction	Independently reconstruct a sampled decision using only records and controls available at the specified knowledge-as-of time	Hindsight leakage, missing decisions, unversioned context
Source and semantic challenge	Introduce a contradictory supplier record or a changed definition of available inventory	Unsupported agreement, invalid cross-domain mapping, failure to preserve disputes
Evidence ancestry	Generate several summaries from a single defective upstream assertion	Incorrect counting of repeated claims as independent corroboration
Outcome reconciliation	Compare purchase-order submission, supplier acceptance, receipt, quality release, and customer fulfillment	Conflation of API success with business success
Personnel and model replacement	Replace the original decision participants and reasoning model before retrospective review	Loss of inspectable institutional memory and interpretability
Authorization and confidentiality	Attempt unauthorized cross-domain queries, policy reinterpretation, and prompt-injection attacks	Unsafe access, delegated-authority confusion, evidence poisoning
Longitudinal improvement	Compare later disruption decisions against controls on cost, timeliness, error, and resilience	Merely shifting human work or introducing maintenance and assurance debt

Table 5: Proposed acceptance tests for a bounded Company Brain implementation; each requires a separately established ground truth or adjudication method and a predefined operating boundary.

A causal experiment could assign comparable process units, or groups of cases where spillovers can be controlled, to versions with and without shared evidence and decision memory while keeping source availability, model capability, process policies, and evaluation resources as similar as possible. Let Y_i(b,a) denote the outcome for unit i under governed Brain capability b\in\{0,1\} and bounded agent execution a\in\{0,1\}. The interaction of these two interventions can be expressed as

\tau_{BA}=\mathbb E\!\left[ Y_i(1,1)-Y_i(1,0)-Y_i(0,1)+Y_i(0,0) \right]. \tag{24}

This is a difference of treatment contrasts on the chosen outcome scale, not a general measure of organizational intelligence. Its causal interpretation requires exchangeability or randomized assignment, well-defined treatments, appropriate handling of spillovers, and comparable measurement; shared semantic infrastructure may violate the no-interference assumption if units access the same knowledge. When randomization is inappropriate, longitudinal or quasi-experimental designs can still be informative, but changes in organizational practice, personnel, and information quality must be treated as potential confounders rather than as proof of a computational effect.

The main empirical propositions are distinguishable. H1, continuity: decision dossiers and temporal provenance should improve the independent reconstruction of past organizational decisions beyond source-cited retrieval alone. H2, integration: governed cross-domain semantics should improve decisions that require information from multiple functions, but not necessarily decisions confined to one authoritative source. H3, learning: controlled corrections linked to independently verified outcomes should reduce recurrence of specified error classes, provided the organization can use the information and the underlying conditions remain sufficiently comparable. H4, complementarity: bounded agents should yield a greater marginal improvement when the cognitive substrate supplies reliable, applicable evidence and well-defined institutional constraints; however, this interaction may be negative where maintenance or verification costs dominate. H5, epistemic resilience: independently sourced counterevidence and effective challenge procedures should reduce the propagation of correlated mistakes even if they sometimes increase review time. These are proposed, potentially refutable claims, not findings established by the existence of the architecture.

A credible evaluation must also measure what the organization refuses to do. The ability to identify a missing source, postpone a decision, expose a disputed interpretation, or escalate an uncertain case may reduce superficial completion rates while improving institutional performance. Conversely, an apparently superior success rate can be produced by silently excluding difficult cases, suppressing evidence of failure, or transferring work to unmeasured human reviewers. Process throughput, verified correctness, intervention burden, independent assurance cost, confidentiality incidents, and severe near misses should therefore be reported separately. A weighted aggregate can be added for a defined managerial decision, but it must not conceal unacceptable risk or claim a scientifically established universal utility scale.

Boundaries of the concept and the institutional significance of a partial implementation

The most demanding objection to this proposal is that the Company Brain may remain only a new name for familiar integration and knowledge-management methods. That objection is scientifically productive: if conventional document search, enterprise data management, workflow engineering, and careful human governance can reproduce every observed benefit at equal or lower cost, then the claim of a new, distinct architectural capability should be weakened. The term becomes useful only if it identifies a repeatable composition of evidence, memory, authority-aware interpretation, and feedback whose effects can be demonstrated beyond those of its individual components. Novelty in branding, interface fluency, or technology selection does not satisfy that criterion.

A second objection concerns the epistemological ambition. Organizational judgments frequently depend on local practice, tacit skill, changing contractual interpretations, and relationships that are neither visible nor reducible to stable computational representations. No finite set of records makes an institution omniscient; no provenance ontology proves its statements true; no statistical model supplies legal authority; and no decision dossier fully recovers a participant’s mental reasons. These are not implementation defects that a larger foundation model will automatically remove. They are reasons to make the boundaries of representation explicit, preserve the possibility of informed refusal and disagreement, and ensure that interpretation remains connected to accountable organizational actors.

A third objection is economic and political. A shared cognitive infrastructure can reduce coordination friction and protect useful institutional memory, but it can also reinforce central managerial control, produce surveillance, lock in a dominant vendor, or harden contested business assumptions into apparently objective software semantics. The distribution of epistemic access, responsibility for corrections, rights to challenge policies, and control of accumulated cognitive capital are therefore part of the implementation, not external social questions to be considered after deployment. An organization that can reconstruct every employee’s activity but cannot challenge its own institutional classifications would have gained observability without necessarily gaining intelligence.

These limits suggest a more restrained but practically consequential conclusion. A Company Brain is built by increasing the continuity and accountability of organizational knowing, not by simulating a unified mind. It may begin as a small, read-mostly capability attached to one procurement process, provided that evidence can be traced, decisions can be reconstructed, results can be checked, and corrections can be governed. It may grow through federated domain contracts whose meanings and authorities remain explicit, without requiring a single database, a single ontology, or continuous autonomous agent execution. The theoretical gain is not that software becomes sovereign over the institution, but that portions of its knowledge and experience become durable, inspectable resources that outlast any particular employee, model, or application. Whether those resources increase organizational performance and resilience is an empirical question—and one for which this architecture supplies a concrete, falsifiable research program.

See also digital transformation longforms

AI Adoption, Productivity, and the Missing Middle

A review of Banca d'Italia QEF 1009 on Italy's AI adoption, firm productivity, and policy architecture

Process Mining in Manufacturing with Purchased Components and Variable Lead Times

A practical example of event-log analysis, object-centric modeling, lead-time diagnosis, and planning correction

Process Mining as Computational Process Intelligence

A formal and technical view of event-log analysis, object-centric process models, conformance checking, performance diagnosis, and enterprise applications

See also enterprise architecture longforms

ArchiMate 4 and the Cartography of Complexity

How ArchiMate 4 simplifies enterprise modeling while enterprise systems grow more complex, shifting the real challenge from notation to semantic governance and operational evidence

Process Mining in Manufacturing with Purchased Components and Variable Lead Times

A practical example of event-log analysis, object-centric modeling, lead-time diagnosis, and planning correction

Process Mining as Computational Process Intelligence

A formal and technical view of event-log analysis, object-centric process models, conformance checking, performance diagnosis, and enterprise applications

See also enterprise risk management longforms

The EU Machinery Regulation 2023/1230: From Machine Safety to Cyber-Physical Safety

How software, connectivity, cybersecurity, AI, remote access, and digital modification are changing machinery safety

Operational Technology in the Crosshairs: What the 2025–2026 Attacks Reveal About Industrial Cyber Risk

From exposed PLCs and composable trust failures to an IEC 62443-aligned architecture for constraining operational authority, bounding compromise, preserving safe operation, and restoring trusted control

Measuring Cyber Risk in the Italian Corporate Sector

A Banca d’Italia indicator of cybersecurity vulnerability designed to support creditworthiness evaluation

See also machine learning longforms

What Kind of Thing Is an Alien Mind?

Reading Jakub Pachocki on intelligence, agency, alignment, opacity, and recursive self-improvement

AI Adoption, Productivity, and the Missing Middle

A review of Banca d'Italia QEF 1009 on Italy's AI adoption, firm productivity, and policy architecture

Harnessing Focus: Merging AI and Market Dynamics

The attention paradigm: bridging natural language processing and economic theory

See also software systems longforms

VEX Standards from First Principles

From software inventories and SBOM evidence to defensible, product-specific vulnerability-exploitability decisions, interoperable VEX exchange, lifecycle governance, and Cyber Resilience Act compliance across CISA, CSAF, CycloneDX, SPDX, and OpenVEX

See also posts

Beyond the Theorem Race: Is AI's Jagged Scientific Frontier Flattening?

From competing mathematical discovery systems to theoretical physics, software verification, and the recovery of mathematical knowledge

L'impresa che pensa: cervello aziendale, agenti-colleghi e valore dell'organizzazione

Come cambiano conoscenza, lavoro e responsabilità quando l'intelligenza artificiale entra nei processi aziendali

The Machines That Find the Flaws

Frontier AI, agentic cybersecurity operations, and the transition from vulnerability discovery to verifiable security

After the Enormous Theorem: Verification, Discovery, and the Missing Layer of Understanding

Axiom’s formalization, OpenAI and Anthropic’s discoveries, Yang–Mills scrutiny, and compute-allocable mathematics

Mistral and the European AI Frontier

How much technological sovereignty does model competitiveness buy?

Math 2.0 Gets an Operating System: Can Open Math OS Build Mathematics Beyond the Theorem Factory?

From mathematical AI agents to durable knowledge and independent research institutions

Footnotes

1. Pan, M., Arabzadeh, N., Cogo, R., Zhu, Y., Xiong, A., A Agrawal, L., Mao, H., Shen, E., Pallerla, S., Patel, L., Liu, S., Shi, T., Liu, X., Davis, J.Q., Lacavalla, E., Basile, A., Yang, S., Castro, P., Kang, D., Sen, K., Song, D., Gonzalez, J.E., Stoica, I., Zaharia, M. & Ellis, M.. (2026). Measuring Agents in Production. Proceedings of the 43rd International Conference on Machine Learning, PMLR 306, 95649–95695. The final proceedings version reports 20 case studies and 86 surveyed deployed-systems practitioners across 26 domains; earlier versions used a different sample count. Final publication.

2. Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889–942. DOI.

3. Becker, J., Rush, N., Barnes, B., & Rein, D. (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. METR, randomized trial involving 246 tasks. Research report.

4. Simon, H. A. (1947). Administrative Behavior: A Study of Decision-Making Processes in Administrative Organization. Macmillan. Book record. Develops administrative decision-making under bounded rationality.

5. Coase, R. H. (1937). The Nature of the Firm. Economica, 4(16), 386–405. DOI. Foundational analysis of transaction costs and the boundary between firms and markets.

6. Galbraith, J. R. (1974). Organization Design: An Information Processing View. Interfaces, 4(3), 28–36. DOI. Introduces organizational design as a response to information-processing requirements under uncertainty.

7. Hayek, F. A. (1945). The Use of Knowledge in Society. The American Economic Review, 35(4), 519–530. Archival paper. Examines the dispersion of economically relevant knowledge among actors.

8. Jacobs, F. R., & Weston, F. C., Jr. (2007). Enterprise Resource Planning (ERP)—A Brief History. Journal of Operations Management, 25(2), 357–363. DOI.

9. Davenport, T. H. (1998). Putting the Enterprise into the Enterprise System. Harvard Business Review, 76(4), 121–131. Journal article.

10. van der Aalst, W. M. P. (2013). Business Process Management: A Comprehensive Survey. ISRN Software Engineering, 2013, Article 507984. DOI.

11. Davenport, T. H., & Short, J. E. (1990). The New Industrial Engineering: Information Technology and Business Process Redesign. Sloan Management Review, 31(4), 11–27. MIT working paper.

12. Orlikowski, W. J. (1992). The Duality of Technology: Rethinking the Concept of Technology in Organizations. Organization Science, 3(3), 398–427. DOI. Explains the reciprocal relationship between technological structures and organizational practices.

13. Grant, R. M. (1996). Toward a Knowledge-Based Theory of the Firm. Strategic Management Journal, 17(S2), 109–122. DOI. Explains the integration of specialized knowledge as a core organizational capability.

14. Nelson, R. R., & Winter, S. G. (1982). An Evolutionary Theory of Economic Change. Harvard University Press. Book record. Develops the evolutionary role of organizational routines and accumulated capabilities.

15. Walsh, J. P., & Ungson, G. R. (1991). Organizational Memory. Academy of Management Review, 16(1), 57–91. DOI. Analyzes acquisition, retention, and retrieval of organizational memory.

16. Walsh, J. P., & Ungson, G. R. (1991). Organizational Memory. Academy of Management Review, 16(1), 57–91. DOI. Analyzes acquisition, retention, and retrieval of organizational memory.

17. Alavi, M., & Leidner, D. E. (2001). Review: Knowledge Management and Knowledge Management Systems: Conceptual Foundations and Research Issues. MIS Quarterly, 25(1), 107–136. Journal article.

18. Wegner, D. M. (1987). Transactive Memory: A Contemporary Analysis of the Group Mind. In B. Mullen & G. R. Goethals (Eds.), Theories of Group Behavior. Springer. DOI.

19. Hutchins, E. (1995). Cognition in the Wild. MIT Press. Publisher page. A foundational account of cognition distributed across people, artifacts, and practices.

20. Hollan, J., Hutchins, E., & Kirsh, D. (2000). Distributed Cognition: Toward a New Foundation for Human–Computer Interaction Research. ACM Transactions on Computer-Human Interaction, 7(2), 174–196. DOI. Connects distributed cognition with the design of human–computer systems.

21. Tsoukas, H. (1996). The Firm as a Distributed Knowledge System: A Constructionist Approach. Strategic Management Journal, 17(S2), 11–25. DOI. Challenges the conception of the firm as a single centralized repository of knowledge.

22. Gruber, T. R. (1993). A Translation Approach to Portable Ontology Specifications. Knowledge Acquisition, 5(2), 199–220. DOI. A foundational account of portable, shared formal conceptualizations.

23. Hogan, A., et al. (2021). Knowledge Graphs. ACM Computing Surveys, 54(4), Article 71. DOI. Surveys knowledge-graph representation, construction, and use.

24. Object Management Group. (2014). Business Process Model and Notation (BPMN), Version 2.0.2. Official specification.

25. W3C. (2013). PROV-O: The PROV Ontology. W3C Recommendation, 30 April 2013. Official specification.

26. W3C. (2017). Shapes Constraint Language (SHACL). W3C Recommendation, 20 July 2017. Official specification.

27. Star, S. L., & Griesemer, J. R. (1989). Institutional Ecology, Translations and Boundary Objects: Amateurs and Professionals in Berkeley’s Museum of Vertebrate Zoology, 1907–39. Social Studies of Science, 19(3), 387–420. DOI.

28. Jensen, C. S., & Snodgrass, R. T. (1999). Temporal Data Management. IEEE Transactions on Knowledge and Data Engineering, 11(1), 36–44. Author-hosted paper.

29. Liu, N. F., et al. (2024). Lost in the Middle: How Language Models Use Long Contexts. Transactions of the Association for Computational Linguistics, 12, 157–173. DOI.

30. Pearl, J. (2009). Causality: Models, Reasoning, and Inference, 2nd ed. Cambridge University Press. DOI. Provides formal foundations for causal modeling and intervention.

31. Park, G., & van der Aalst, W. M. P. (2021). Realizing a Digital Twin of an Organization Using Action-oriented Process Mining. Proceedings of the International Conference on Process Mining, IEEE. The reported feasibility study uses an artificial order-handling information system. DOI.

32. Nonaka, I. (1994). A Dynamic Theory of Organizational Knowledge Creation. Organization Science, 5(1), 14–37. DOI. Examines organizational knowledge creation through tacit–explicit interaction.

33. Weick, K. E. (1995). Sensemaking in Organizations. SAGE Publications. Publisher page. Addresses how organizational participants interpret ambiguous events.

34. Simon, H. A. (1962). The Architecture of Complexity. Proceedings of the American Philosophical Society, 106(6), 467–482. Archival paper. Provides the theoretical basis for decomposable and hierarchical complex systems.

35. Pugh, D. S., Hickson, D. J., Hinings, C. R., & Turner, C. (1968). Dimensions of Organization Structure. Administrative Science Quarterly, 13(1), 65–105. DOI.

36. Object Management Group. (2019). Semantics of Business Vocabulary and Business Rules (SBVR), Version 1.5. Official specification.

37. Object Management Group. (2014). Business Process Model and Notation (BPMN), Version 2.0.2. Official specification.

38. Object Management Group. (n.d.). Decision Model and Notation (DMN). Standard for specifying business decisions and business rules, designed to complement process and case models. Official specification portal.

39. Hübner, J. F., Sichman, J. S., & Boissier, O. (2002). A Model for the Structural, Functional, and Deontic Specification of Organizations in Multiagent Systems. In Advances in Artificial Intelligence, LNCS 2507, 118–128. DOI.

40. Singh, M. P. (1999). An Ontology for Commitments in Multiagent Systems: Toward a Unification of Normative Concepts. Artificial Intelligence and Law, 7(1), 97–113. Author-hosted paper.

41. Model Context Protocol maintainers. (2026, July 28). The 2026-07-28 Specification. The release describes the stateless protocol core and updated extension and authorization mechanisms. Official specification and Release announcement.

42. Linux Foundation. (2025, June 23). Linux Foundation Launches the Agent2Agent Protocol Project to Enable Secure, Intelligent Communication Between AI Agents. Official announcement.

43. Temporal Technologies. (n.d.). Temporal Workflow and Event History documentation. Describes durable execution, persisted event history, replay, and recovery; documents implementation mechanisms rather than proving general agent reliability. Official documentation.

44. Cutler, J., et al. (2024). Cedar: A New Language for Expressive, Fast, Safe, and Analyzable Authorization. Proceedings of the ACM on Programming Languages (OOPSLA 2024). Research publication.

45. Cloud Native Computing Foundation, CloudEvents Project. (n.d.). CloudEvents Specification. Defines interoperable event context attributes, including identifiers, sources, types, and timestamps; it does not define business-level correctness or transaction semantics. Official specification.

46. Debezium. (n.d.). Outbox Event Router. Explains transactional outbox event publication using change-data capture. Official documentation.

47. SPIFFE project. (n.d.). SPIFFE Identity and Verifiable Identity Document. Defines workload identities and verifiable identity documents. Official specification.

48. Rose, S., Borchert, O., Mitchell, S., & Connelly, S. (2020). Zero Trust Architecture. NIST SP 800-207. DOI.

49. Open Policy Agent project. (n.d.). How to deploy OPA. Describes OPA as a policy decision point that can be invoked by independent enforcement components. Official documentation.

50. Temporal Technologies. (n.d.). Temporal Workflow and Event History documentation. Describes durable execution, persisted event history, replay, and recovery; documents implementation mechanisms rather than proving general agent reliability. Official documentation.

51. Feldman, M. S., & Pentland, B. T. (2003). Reconceptualizing Organizational Routines as a Source of Flexibility and Change. Administrative Science Quarterly, 48(1), 94–118. DOI.

52. Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at Work: The New Contested Terrain of Control. Academy of Management Annals, 14(1), 366–410. DOI. Examines algorithmic direction, evaluation, discipline, and worker contestation.

53. OWASP GenAI Security Project. (2025). LLM06:2025 Excessive Agency. Threat description and mitigation guidance for tool-enabled language-model applications. Security guidance.

54. Mintzberg, H. (1980). Structure in 5’s: A Synthesis of the Research on Organization Design. Management Science, 26(3), 322–341. DOI. Synthesizes organizational parts and mechanisms of coordination.

55. Malone, T. W., & Crowston, K. (1994). The Interdisciplinary Study of Coordination. ACM Computing Surveys, 26(1), 87–119. DOI. Defines coordination through the management of dependencies among activities.

56. Dumas, M., La Rosa, M., Mendling, J., & Reijers, H. A. (2018). Fundamentals of Business Process Management, 2nd ed. Springer. DOI.

57. United Kingdom. (2006). Companies Act 2006, sections 154–155, as applicable at the article’s cutoff. Official legislation.

58. United Kingdom. (2006). Companies Act 2006, ss. 171–174 (directors’ duties to act within powers, promote the success of the company, exercise independent judgment, and exercise reasonable care, skill and diligence). Official legislation.

59. State of Delaware. (n.d.). Delaware General Corporation Law, Title 8, section 141(b), requiring corporate directors to be natural persons. The general corporation is the relevant entity type; other legal forms require separate analysis. Official statute.

60. United Nations Commission on International Trade Law. (2024). UNCITRAL Model Law on Automated Contracting, adopted 11 July 2024. A model legislative text, not directly applicable law in every jurisdiction. Official model law.

61. Jensen, M. C., & Meckling, W. H. (1976). Theory of the Firm: Managerial Behavior, Agency Costs and Ownership Structure. Journal of Financial Economics, 3(4), 305–360. DOI.

62. Holmström, B., & Milgrom, P. (1991). Multitask Principal–Agent Analyses: Incentive Contracts, Asset Ownership, and Job Design. Journal of Law, Economics, & Organization, 7, 24–52. DOI.

63. Hart, O., & Moore, J. (1990). Property Rights and the Nature of the Firm. Journal of Political Economy, 98(6), 1119–1158. DOI.

64. OECD. (2023). G20/OECD Principles of Corporate Governance 2023, Section V, responsibilities of the board. Official publication.

65. Institute of Internal Auditors. (2020). The IIA’s Three Lines Model: An Update of the Three Lines of Defense. Identifies governing-body accountability, management activity, and independent assurance. Official institutional resource.

66. European Parliament and Council. (2024). Regulation (EU) 2024/1689 (AI Act), as amended by Regulation (EU) 2026/1744 of 8 July 2026; check the consolidated 27 July 2026 version, especially Articles 2, 6, 113 and Annexes I and III. Consolidated official text and Amending regulation.

67. European Parliament and Council. (2016). Regulation (EU) 2016/679 (General Data Protection Regulation), especially Articles 5, 6 and 22. Official text.

68. Xu, F. F., et al. (2024, initial preprint). TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks. The 24% result belongs to the original reported evaluation, not a 2026 frontier leaderboard. Preprint.

69. Yang, M., Pan, Y., Piao, J., Song, D., Gong, Y., & Li, Y. (2026). EnterpriseBench: Benchmarking LLM Agents on Enterprise-Level Strategic Reasoning and Decision-Making. Findings of EMNLP 2026, 19659–19682. DOI.

70. Chatrath, V., et al. (2026). READY or Not: Reliable Enterprise Agent Deployment. Research preprint, 2 September 2026. Evaluates reliability, oversight burden, and operating cost for defined workflows. Preprint.

71. Hart, O., & Moore, J. (1990). Property Rights and the Nature of the Firm. Journal of Political Economy, 98(6), 1119–1158. DOI.

72. Galbraith, J. R. (1974). Organization Design: An Information Processing View. Interfaces, 4(3), 28–36. DOI. Introduces organizational design as a response to information-processing requirements under uncertainty.

73. Malone, T. W., & Crowston, K. (1994). The Interdisciplinary Study of Coordination. ACM Computing Surveys, 26(1), 87–119. DOI. Defines coordination through the management of dependencies among activities.

74. Faraj, S., & Sproull, L. (2000). Coordinating Expertise in Software Development Teams. Management Science, 46(12), 1554–1568. DOI.

75. Devigili, M., Yilmaz, E. D., Gaba, V., & Greve, H. (2026). Skill Deprioritization: Reorganizing in the Age of Generative Artificial Intelligence. Management Science, published online July 24, 2026. DOI.

76. Wegner, D. M. (1987). Transactive Memory: A Contemporary Analysis of the Group Mind. In B. Mullen & G. R. Goethals (Eds.), Theories of Group Behavior. Springer. DOI.

77. Tsoukas, H. (1996). The Firm as a Distributed Knowledge System: A Constructionist Approach. Strategic Management Journal, 17(S2), 11–25. DOI. Challenges the conception of the firm as a single centralized repository of knowledge.

78. March, J. G. (1991). Exploration and Exploitation in Organizational Learning. Organization Science, 2(1), 71–87. DOI. Distinguishes the exploration of new knowledge from exploitation of established capabilities.

79. Edmondson, A. (1999). Psychological Safety and Learning Behavior in Work Teams. Administrative Science Quarterly, 44(2), 350–383. DOI.

80. Büchsenschuss, R., Koch-Bayram, I., Biemann, T., & Puranam, P. (2026). GenAI Adoption Increases the Density of Knowledge and Collaboration Networks: Evidence from a Field Experiment. INSEAD Working Paper 2026/22/STR, revised version. Working paper.

81. Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889–942. DOI.

82. Dell’Acqua, F., et al. (2026). The Cybernetic Teammate: A Field Experiment on Generative AI and Teamwork. Organization Science, published online June 12, 2026. DOI. The published study reports 791 professionals; the earlier 2025 working-paper version reported 776.

83. Becker, J., Rush, N., Barnes, B., & Rein, D. (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. METR, randomized trial involving 246 tasks. Research report.

84. METR. (2026, February 24). We Are Changing Our Developer Productivity Experiment Design. Reports follow-up measurement and selection limitations; results cannot support a precise general productivity estimate for late-2025 tools. Research update.

85. European Investment Bank. (2026). AI Adoption, Productivity and Employment: Evidence from European Firms. EIB Working Paper 2026/02, January 13, 2026. Official working paper.

86. Ostrom, E. (2010). Beyond Markets and States: Polycentric Governance of Complex Economic Systems. American Economic Review, 100(3), 641–672. DOI. Develops a framework for polycentric governance of complex economic systems.

87. Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at Work: The New Contested Terrain of Control. Academy of Management Annals, 14(1), 366–410. DOI. Examines algorithmic direction, evaluation, discipline, and worker contestation.

88. European Parliament and Council of the European Union. (2016). Regulation (EU) 2016/679 (General Data Protection Regulation). See particularly Articles 5, 6, 22, and 88, subject to their respective conditions of applicability. Official text.

89. European Parliament and Council of the European Union. (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence. See particularly Article 26 and Annex III, point 4, with due regard to classification criteria, exemptions, transitional provisions, and applicable dates. Official text.

90. Beer, S. (1972). Brain of the Firm: A Development in Management Cybernetics. Herder and Herder. The book develops a cybernetic account of organizational regulation, recursion, and viability; it is a historical precursor, not a specification for a contemporary AI-based Company Brain. Book record.

91. Ashby, W. R. (1956). An Introduction to Cybernetics. Chapman & Hall. See particularly the treatment of variety and regulation. The results apply within defined systems-theoretic assumptions; their application to enterprises requires explicit modeling of the regulated system and the available controls. Author archive. Introduces systems-theoretic principles of regulation and variety.

92. Conant, R. C., & Ashby, W. R. (1970). Every Good Regulator of a System Must Be a Model of That System. International Journal of Systems Science, 1(2), 89–97. The theorem relates effective and economical regulation to internal modeling under specified mathematical assumptions; it is not a general proof of organizational cognition. DOI.

93. Tsoukas, H. (1996). The Firm as a Distributed Knowledge System: A Constructionist Approach. Strategic Management Journal, 17(S2), 11–25. DOI. Challenges the conception of the firm as a single centralized repository of knowledge.

94. Jensen, M. C., & Meckling, W. H. (1976). Theory of the Firm: Managerial Behavior, Agency Costs and Ownership Structure. Journal of Financial Economics, 3(4), 305–360. DOI.

95. Bainbridge, L. (1983). Ironies of Automation. Automatica, 19(6), 775–779. DOI. Examines human-operator difficulties arising from extensive automation.

96. Coase, R. H. (1937). The Nature of the Firm. Economica, 4(16), 386–405. DOI. Foundational analysis of transaction costs and the boundary between firms and markets.

97. Hart, O., & Moore, J. (1990). Property Rights and the Nature of the Firm. Journal of Political Economy, 98(6), 1119–1158. DOI.

98. Garicano, L. (2000). Hierarchies and the Organization of Knowledge in Production. Journal of Political Economy, 108(5), 874–904. DOI.

99. Goldfarb, A., & Tucker, C. (2019). Digital Economics. Journal of Economic Literature, 57(1), 3–43. DOI.

100. Poppo, L., & Zenger, T. (1998). Testing Alternative Theories of the Firm: Transaction Cost, Knowledge-Based, and Measurement Explanations for Make-or-Buy Decisions in Information Services. Strategic Management Journal, 19(9), 853–877. Publisher article.

101. Brynjolfsson, E., Hitt, L. M., & Yang, S. (2002). Intangible Assets: Computers and Organizational Capital. Brookings Papers on Economic Activity: Microeconomics, 2002(1), 137–181. Research publication.

102. Bresnahan, T. F., Brynjolfsson, E., & Hitt, L. M. (2002). Information Technology, Workplace Organization, and the Demand for Skilled Labor: Firm-Level Evidence. Quarterly Journal of Economics, 117(1), 339–376. DOI.

103. Baldwin, C. Y., & Clark, K. B. (2000). Design Rules, Volume 1: The Power of Modularity. MIT Press. Publisher page.

104. Klemperer, P. (1995). Competition When Consumers Have Switching Costs: An Overview with Applications to Industrial Organization, Macroeconomics, and International Trade. Review of Economic Studies, 62(4), 515–539. DOI.

105. Acemoglu, D., & Restrepo, P. (2019). Automation and New Tasks: How Technology Displaces and Reinstates Labor. Journal of Economic Perspectives, 33(2), 3–30. NBER working paper.

106. Autor, D., Dorn, D., Katz, L. F., Patterson, C., & Van Reenen, J. (2020). The Fall of the Labor Share and the Rise of Superstar Firms. Quarterly Journal of Economics, 135(2), 645–709. DOI.

107. OECD. (2026). Competition in the Age of AI: Initial Evidence from Microdata. OECD Artificial Intelligence Papers, No. 64, 30 July 2026. DOI.

108. Bresnahan, T. F., Brynjolfsson, E., & Hitt, L. M. (2002). Information Technology, Workplace Organization, and the Demand for Skilled Labor: Firm-Level Evidence. Quarterly Journal of Economics, 117(1), 339–376. DOI.

109. Bloom, N., Garicano, L., Sadun, R., & Van Reenen, J. (2014). The Distinct Effects of Information Technology and Communication Technology on Firm Organization. Management Science, 60(12), 2859–2885. DOI.

110. European Investment Bank. (2026). AI Adoption, Productivity and Employment: Evidence from European Firms. EIB Working Paper 2026/02, January 13, 2026. Official working paper.

111. Calvino, F., Costa, H., & Haerle, D. (2026). Digital Technology Diffusion in the Age of AI: Cross-Country Evidence from Microdata. OECD Science, Technology and Industry Working Papers, No. 2026/01, 22 January 2026. DOI.

112. Bonney, K., Breaux, C. L., Dinlersoz, E., Foster, L. S., Haltiwanger, J. C., & Pande, A. A. (2026, April). The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks. NBER Working Paper 35141. DOI.

113. Acemoglu, D. (2025). The Simple Macroeconomics of AI. Economic Policy. The paper models aggregate productivity effects from task-level AI exposure and savings, with results conditional on its assumptions. MIT repository.

114. Hutchins, E. (1995). Cognition in the Wild. MIT Press. Publisher page. A foundational account of cognition distributed across people, artifacts, and practices.

115. Woolley, A. W., Chabris, C. F., Pentland, A., Hashmi, N., & Malone, T. W. (2010). Evidence for a Collective Intelligence Factor in the Performance of Human Groups. Science, 330(6004), 686–688. DOI.

116. Cohen, W. M., & Levinthal, D. A. (1990). Absorptive Capacity: A New Perspective on Learning and Innovation. Administrative Science Quarterly, 35(1), 128–152. DOI.

117. Stasser, G., & Titus, W. (1985). Pooling of Unshared Information in Group Decision Making: Biased Information Sampling During Discussion. Journal of Personality and Social Psychology, 48(6), 1467–1478. DOI.

118. Lorenz, J., Rauhut, H., Schweitzer, F., & Helbing, D. (2011). How Social Influence Can Undermine the Wisdom of Crowd Effect. Proceedings of the National Academy of Sciences, 108(22), 9020–9025. DOI.

119. Bikhchandani, S., Hirshleifer, D., & Welch, I. (1992). A Theory of Fads, Fashion, Custom, and Cultural Change as Informational Cascades. Journal of Political Economy, 100(5), 992–1026. DOI.

120. Knight, J. C., & Leveson, N. G. (1986). An Experimental Evaluation of the Assumption of Independence in Multiversion Programming. IEEE Transactions on Software Engineering SE-12(1), 96–109. DOI.

121. W3C. (2013). PROV-O: The PROV Ontology. W3C Recommendation, 30 April 2013. Official specification.

122. Doyle, J. (1979). A Truth Maintenance System. Artificial Intelligence, 12(3), 231–272. DOI.

123. Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., & Gal, Y. (2024). AI Models Collapse When Trained on Recursively Generated Data. Nature, 631, 755–759. DOI. The study concerns recursive model training; the extension to enterprise knowledge contamination is an analogy, not a demonstrated equivalent effect.

124. Argyris, C., & Schön, D. A. (1978). Organizational Learning: A Theory of Action Perspective. Addison-Wesley. Book record. A foundational treatment of organizational learning and the revision of governing assumptions.

125. Hong, L., & Page, S. E. (2004). Groups of Diverse Problem Solvers Can Outperform Groups of High-Ability Problem Solvers. Proceedings of the National Academy of Sciences, 101(46), 16385–16389. DOI. The mathematical results hold under specified assumptions and do not establish an unconditional advantage for diversity.

126. Fricker, M. (2007). Epistemic Injustice: Power and the Ethics of Knowing. Oxford University Press. DOI. Examines the effects of power on credibility and participation in knowledge practices.

127. OWASP GenAI Security Project. (2025, December 9). OWASP Top 10 for Agentic Applications for 2026. Security guidance.

128. Vassilev, A., Oprea, A., Fordyce, A., Anderson, H., Davies, X., & Hamin, M. (2025). Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations. NIST AI 100-2e2025. Official publication.

129. Debenedetti, E., Zhang, J., Balunović, M., Beurer-Kellner, L., Fischer, M., & Tramèr, F. (2024). AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents. Advances in Neural Information Processing Systems, Datasets and Benchmarks Track. Conference publication.

130. Zhang, H., et al. (2024). Agent Security Bench (ASB): Formalizing and Benchmarking Attacks and Defenses in LLM-based Agents. Preprint. Reported attack-success rates are specific to the experimental configurations and are not estimates of production compromise prevalence.

131. Wang, Z., et al. (2026). MCPTox: A Benchmark for Tool Poisoning on Real-World MCP Servers. Proceedings of the AAAI Conference on Artificial Intelligence, 40(42), 35811–35819. DOI.

132. Model Context Protocol Threat Modeling and Analysis of Vulnerabilities to Prompt Injection with Tool Poisoning. (2026). Journal of Cybersecurity and Privacy, 6(3), Article 84. DOI.

133. Booth, H., Fisher, W., Galluzzo, R., & Roberts, J. (2026). Accelerating the Adoption of Software and Artificial Intelligence Agent Identity and Authorization. NIST NCCoE concept paper, 5 February 2026. Official publication.

134. Gallot, H. (2026). Securing the Agentic Insider. Microsoft, 19 October 2026. This is an industry architecture argument and product-oriented statement, not an independent experimental evaluation. Official article.

135. Rose, S., Borchert, O., Mitchell, S., & Connelly, S. (2020). Zero Trust Architecture. NIST SP 800-207. DOI.

136. NIST. (2022; updated 2024). Cybersecurity Supply Chain Risk Management Practices for Systems and Organizations. NIST SP 800-161 Rev. 1, Update 1. Official publication.

137. Souppaya, M., Scarfone, K., & Dodson, D. (2022). Secure Software Development Framework (SSDF) Version 1.1: Recommendations for Mitigating the Risk of Software Vulnerabilities. NIST SP 800-218. Official publication.

138. OWASP GenAI Security Project. (2026). Agent Control Standard (ACS), 1 September 2026. Security specification.

139. International Electrotechnical Commission. (2020). IEC 62443-3-2: Security Risk Assessment for System Design. Defines requirements concerning systems under consideration, zones and conduits, risk assessments, target security levels, and documented security requirements. Standard record.

140. Stouffer K, Pease M, Tang CY, Zimmerman T, Pillitteri V, Lightman S, Hahn A, Saravia S, Sherule A & Thompson M. (2023). Guide to Operational Technology (OT) Security. NIST SP 800-82 Rev. 3. Official publication.

141. European Parliament and Council. (2016). Regulation (EU) 2016/679 (General Data Protection Regulation), especially Articles 5, 6 and 22. Official text.

142. European Parliament and Council. (2024). Directive (EU) 2024/2831 on improving working conditions in platform work, including algorithmic-management safeguards. This instrument has a specific platform-work scope, not general application to all employers. Official text.

143. European Parliamentary Research Service. (2025). Copyright of AI-generated works: Approaches in the EU and beyond, 19 December 2025. Parliamentary briefing.

144. U.S. Copyright Office. (2025). Copyright and Artificial Intelligence, Part 2: Copyrightability. January 2025. Official announcement and report.

145. European Parliament and Council of the European Union. (2016). Directive (EU) 2016/943 on the protection of undisclosed know-how and business information (trade secrets). Official text.

146. European Parliament and Council of the European Union. (2024). Directive (EU) 2024/2853 on liability for defective products, as corrected by the corrigendum published on 7 May 2026. The corrected Article 2(1) applies to products placed on the market or put into service after 8 December 2026. Consolidated legal text.

147. European Parliament and Council. (2022). Directive (EU) 2022/2555 (NIS2), especially Articles 20–21 and sectoral scope. Applicability also depends on national transposition and the entity concerned. Official text.

148. European Parliament and Council. (2024). Regulation (EU) 2024/2847 (Cyber Resilience Act), particularly Articles 14 and 71. Official text.

149. European Parliament and Council. (2024). Regulation (EU) 2024/1689 (AI Act), as amended by Regulation (EU) 2026/1744 of 8 July 2026; check the consolidated 27 July 2026 version, especially Articles 2, 6, 113 and Annexes I and III. Consolidated official text and Amending regulation.

150. European Parliament and Council of the European Union. (2022). Regulation (EU) 2022/2554 on digital operational resilience for the financial sector. Official text.

151. ISO/IEC. (2023). ISO/IEC 42001:2023, Information technology — Artificial intelligence — Management system. Official standard.

152. NIST. (2023). AI Risk Management Framework (AI RMF 1.0), NIST AI 100-1. DOI. See also NIST’s 2024 Generative Artificial Intelligence Profile, NIST AI 600-1. DOI.

153. Jensen, C. S., & Snodgrass, R. T. (1999). Temporal Data Management. IEEE Transactions on Knowledge and Data Engineering, 11(1), 36–44. Author-hosted paper.

154. Cloud Native Computing Foundation, CloudEvents Project. (n.d.). CloudEvents Specification. Defines interoperable event context attributes, including identifiers, sources, types, and timestamps; it does not define business-level correctness or transaction semantics. Official specification.

155. Richardson, C. (n.d.). Pattern: Transactional Outbox. Microservices Patterns. Describes the use of a transactionally persisted outbox to coordinate database changes and subsequent event publication without relying on a distributed transaction across the database and message broker. Architecture pattern.

156. Temporal Technologies. (n.d.). Temporal Workflow and Event History documentation. Describes durable execution, persisted event history, replay, and recovery; documents implementation mechanisms rather than proving general agent reliability. Official documentation.

157. Object Management Group. (2014). Business Process Model and Notation (BPMN), Version 2.0.2. Official specification.

158. Object Management Group. (n.d.). Decision Model and Notation (DMN). Standard for specifying business decisions and business rules, designed to complement process and case models. Official specification portal.

159. Rose, S., Borchert, O., Mitchell, S., & Connelly, S. (2020). Zero Trust Architecture. NIST SP 800-207. DOI.

160. OpenTelemetry. (n.d.). Traces and Distributed Tracing. Documentation of spans, context propagation, and the representation of operations across distributed systems. Official documentation.

161. Doyle, J. (1979). A Truth Maintenance System. Artificial Intelligence, 12(3), 231–272. DOI.

162. Stouffer K, Pease M, Tang CY, Zimmerman T, Pillitteri V, Lightman S, Hahn A, Saravia S, Sherule A & Thompson M. (2023). Guide to Operational Technology (OT) Security. NIST SP 800-82 Rev. 3. Official publication.

163. Lee, H. L., Padmanabhan, V., & Whang, S. (1997). Information Distortion in a Supply Chain: The Bullwhip Effect. Management Science, 43(4), 546–558. DOI.

164. Li, Y., Chen, K., Collignon, S., & Ivanov, D. (2021). Ripple Effect in the Supply Chain Network: Forward and Backward Disruption Propagation, Network Health and Firm Vulnerability. European Journal of Operational Research, 291(3), 1117–1131. DOI.

165. Köle, H., & Bakal, I. S. (2017). Value of Information through Options Contract under Disruption Risk. Computers & Industrial Engineering, 103, 85–97. DOI.

166. Zhou, S., et al. (2024). WebArena: A Realistic Web Environment for Building Autonomous Agents. ICLR 2024. Conference paper.

167. Xu, F. F., et al. (2024, initial preprint). TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks. The 24% result belongs to the original reported evaluation, not a 2026 frontier leaderboard. Preprint.

168. Stanford Institute for Human-Centered Artificial Intelligence. (2026). The 2026 AI Index Report, Chapter 2, Technical Performance. Reports OSWorld computer-use performance including 66.3% for a leading configuration and a 72.35% human baseline. Official report.

169. Yuan, M., et al. (2026). OSWorld 2.0: Benchmarking Computer-Use Agents on Long-Horizon Real-World Tasks. arXiv:2606.29537. Preprint describing 108 long-horizon tasks and the evaluation conditions under which a leading configuration achieved 20.6% binary completion. Research preprint.

170. Yao, S., Shinn, N., Razavi, P., & Narasimhan, K. (2025). τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains. International Conference on Learning Representations. Evaluates policy-constrained tool actions and repeated-trial reliability. Conference publication.

171. METR. (2026, September 8 update). Task-Completion Time Horizons of Frontier AI Models. The authors delimit the suite to predominantly self-contained software, machine-learning, and cybersecurity tasks. Evaluation methodology and results.

172. Pan, M., Arabzadeh, N., Cogo, R., Zhu, Y., Xiong, A., A Agrawal, L., Mao, H., Shen, E., Pallerla, S., Patel, L., Liu, S., Shi, T., Liu, X., Davis, J.Q., Lacavalla, E., Basile, A., Yang, S., Castro, P., Kang, D., Sen, K., Song, D., Gonzalez, J.E., Stoica, I., Zaharia, M. & Ellis, M.. (2026). Measuring Agents in Production. Proceedings of the 43rd International Conference on Machine Learning, PMLR 306, 95649–95695. The final proceedings version reports 20 case studies and 86 surveyed deployed-systems practitioners across 26 domains; earlier versions used a different sample count. Final publication.

173. Yang, M., Pan, Y., Piao, J., Song, D., Gong, Y., & Li, Y. (2026). EnterpriseBench: Benchmarking LLM Agents on Enterprise-Level Strategic Reasoning and Decision-Making. Findings of EMNLP 2026, 19659–19682. DOI.

174. Bonney, K., Breaux, C. L., Dinlersoz, E., Foster, L. S., Haltiwanger, J. C., & Pande, A. A. (2026, April). The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks. NBER Working Paper 35141. DOI.

175. Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889–942. DOI.

176. Dell’Acqua, F., et al. (2026). The Cybernetic Teammate: A Field Experiment on Generative AI and Teamwork. Organization Science, published online June 12, 2026. DOI. The published study reports 791 professionals; the earlier 2025 working-paper version reported 776.

177. Becker, J., Rush, N., Barnes, B., & Rein, D. (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. METR, randomized trial involving 246 tasks. Research report.

178. METR. (2026, February 24). We Are Changing Our Developer Productivity Experiment Design. Reports follow-up measurement and selection limitations; results cannot support a precise general productivity estimate for late-2025 tools. Research update.

179. Dell’Acqua, F., et al. (2026). The Cybernetic Teammate: A Field Experiment on Generative AI and Teamwork. Organization Science, published online June 12, 2026. DOI. The published study reports 791 professionals; the earlier 2025 working-paper version reported 776.

180. Puranam, P. (2021). Human–AI Collaborative Decision-Making as an Organization Design Problem. Journal of Organization Design, 10, 75–80. Distinguishes possible divisions of decision-making labor and learning configurations in human–AI organizations. DOI.

181. Hudgens, M. G., & Halloran, M. E. (2008). Toward Causal Inference with Interference. Journal of the American Statistical Association, 103(482), 832–842. Develops causal estimands for settings in which outcomes can be affected by other units’ treatment assignments. DOI.

182. Callaway, B., & Sant’Anna, P. H. C. (2021). Difference-in-Differences with Multiple Time Periods. Journal of Econometrics, 225(2), 200–230. Addresses identification and estimation with staggered treatment adoption and heterogeneous effects under specified assumptions. DOI.

183. Bresnahan, T. F., Brynjolfsson, E., & Hitt, L. M. (2002). Information Technology, Workplace Organization, and the Demand for Skilled Labor: Firm-Level Evidence. Quarterly Journal of Economics, 117(1), 339–376. DOI.

184. Puranam, P. (2021). Human–AI Collaborative Decision-Making as an Organization Design Problem. Journal of Organization Design, 10, 75–80. Distinguishes possible divisions of decision-making labor and learning configurations in human–AI organizations. DOI.

185. Bloom, N., Garicano, L., Sadun, R., & Van Reenen, J. (2014). The Distinct Effects of Information Technology and Communication Technology on Firm Organization. Management Science, 60(12), 2859–2885. DOI.

186. Feldman, M. S., & Pentland, B. T. (2003). Reconceptualizing Organizational Routines as a Source of Flexibility and Change. Administrative Science Quarterly, 48(1), 94–118. DOI.

187. Eurofound. (2026). Algorithmic Management in European Workplaces. Research report, 29 October 2026. Based on the European Working Conditions Survey 2024; findings on working conditions are predominantly observational associations. Official report.

188. Hutchins, E. (1995). Cognition in the Wild. MIT Press. Publisher page. A foundational account of cognition distributed across people, artifacts, and practices.

189. Beer, S. (1972). Brain of the Firm: A Development in Management Cybernetics. Herder and Herder. The book develops a cybernetic account of organizational regulation, recursion, and viability; it is a historical precursor, not a specification for a contemporary AI-based Company Brain. Book record.

190. OpenAPI Initiative. (2026, September 10). OpenAPI Specification, version 3.2.1. OpenAPI Initiative. Describes a language-agnostic format for HTTP API interfaces; it is not a specification of institutional business authority. Official specification.

191. JSON Schema project. (2022). JSON Schema Draft 2020-12. JSON Schema. Defines a vocabulary and validation framework for JSON data structures, including distinctions between validation assertions and annotations. Official specification.

192. Model Context Protocol maintainers. (2026, July 28). The 2026-07-28 Specification. The release describes the stateless protocol core and updated extension and authorization mechanisms. Official specification and Release announcement.

193. Model Context Protocol maintainers. (2026). Tools: Model Context Protocol specification, 2026-07-28. Model Context Protocol. Defines tool discovery, input/output schemas, and behavioral annotations; clients must treat tool annotations as untrusted unless they originate from trusted servers. Official specification.

194. Hoare, C. A. R. (1969). An axiomatic basis for computer programming. Communications of the ACM, 12(10), 576–580. Foundational formulation of axiomatic program correctness and precondition–postcondition reasoning. DOI.

195. Lamport, L. (2002). Specifying systems: The TLA+ language and tools for hardware and software engineers. Addison-Wesley. Establishes a temporal-logic framework for specifying and analyzing concurrent and distributed systems. Author and publisher record.

196. Alpern, B., & Schneider, F. B. (1985). Defining liveness. Information Processing Letters, 21(4), 181–185. Formalizes the distinction between safety and liveness properties of computations. DOI.

197. Abadi, M., & Lamport, L. (1995). Conjoining specifications. ACM Transactions on Programming Languages and Systems, 17(3), 507–534. Develops results for assume–guarantee reasoning and composition of open-system specifications. Research publication.

198. Garcia-Molina, H., & Salem, K. (1987). Sagas. Proceedings of ACM SIGMOD, 249–259. Introduces a model of long-lived transactions composed of smaller transactions and compensating operations. DOI.

199. Herlihy, M. P., & Wing, J. M. (1990). Linearizability: A correctness condition for concurrent objects. ACM Transactions on Programming Languages and Systems, 12(3), 463–492. Defines an abstract correctness criterion for concurrent operations on shared objects. DOI.

200. Abadi, M., & Lamport, L. (1995). Conjoining specifications. ACM Transactions on Programming Languages and Systems, 17(3), 507–534. Develops results for assume–guarantee reasoning and composition of open-system specifications. Research publication.

201. Herlihy, M. P., & Wing, J. M. (1990). Linearizability: A correctness condition for concurrent objects. ACM Transactions on Programming Languages and Systems, 12(3), 463–492. Defines an abstract correctness criterion for concurrent operations on shared objects. DOI.

202. OpenAPI Initiative. (2026, September 10). OpenAPI Specification, version 3.2.1. OpenAPI Initiative. Describes a language-agnostic format for HTTP API interfaces; it is not a specification of institutional business authority. Official specification.

203. JSON Schema project. (2022). JSON Schema Draft 2020-12. JSON Schema. Defines a vocabulary and validation framework for JSON data structures, including distinctions between validation assertions and annotations. Official specification.

204. Model Context Protocol maintainers. (2026, July 28). The 2026-07-28 Specification. The release describes the stateless protocol core and updated extension and authorization mechanisms. Official specification and Release announcement.

205. Object Management Group. (2014). Business Process Model and Notation (BPMN), Version 2.0.2. Official specification.

206. Object Management Group. (2026). Decision Model and Notation (DMN), version 1.6. OMG. Formal specification adopted in September 2026, distinct from the DMN 1.7 beta in progress. Official specification.

207. W3C. (2013). PROV-O: The PROV Ontology. W3C Recommendation, 30 April 2013. Official specification.

208. Disselkoen, C., et al. (2024). How we built Cedar: A verification-guided approach. Research preprint. Describes the combination of executable formal models, mechanical proofs, differential random testing, and property-based testing used in the development of the Cedar authorization language. Preprint.

209. Nottingham, M., Wilde, E., & Dalal, S. (2023). RFC 9457: Problem details for HTTP APIs. Internet Engineering Task Force. Standardized machine-readable problem details for HTTP APIs; application-specific business outcomes require additional semantics. Official RFC.

210. Souppaya, M., Scarfone, K., & Dodson, D. (2022). Secure Software Development Framework (SSDF) Version 1.1: Recommendations for Mitigating the Risk of Software Vulnerabilities. NIST SP 800-218. Official publication.

211. Garicano, L. (2000). Hierarchies and the Organization of Knowledge in Production. Journal of Political Economy, 108(5), 874–904. DOI.

212. Bresnahan, T. F., Brynjolfsson, E., & Hitt, L. M. (2002). Information Technology, Workplace Organization, and the Demand for Skilled Labor: Firm-Level Evidence. Quarterly Journal of Economics, 117(1), 339–376. DOI.

213. March, J. G. (1991). Exploration and Exploitation in Organizational Learning. Organization Science, 2(1), 71–87. DOI. Distinguishes the exploration of new knowledge from exploitation of established capabilities.

214. Model Context Protocol maintainers. (2026). Tools: Model Context Protocol specification, 2026-07-28. Model Context Protocol. Defines tool discovery, input/output schemas, and behavioral annotations; clients must treat tool annotations as untrusted unless they originate from trusted servers. Official specification.

215. Wang, Z., et al. (2026). MCPTox: A Benchmark for Tool Poisoning on Real-World MCP Servers. Proceedings of the AAAI Conference on Artificial Intelligence, 40(42), 35811–35819. DOI.

216. Stouffer K, Pease M, Tang CY, Zimmerman T, Pillitteri V, Lightman S, Hahn A, Saravia S, Sherule A & Thompson M. (2023). Guide to Operational Technology (OT) Security. NIST SP 800-82 Rev. 3. Official publication.

217. Walsh, J. P., & Ungson, G. R. (1991). Organizational memory. Academy of Management Review, 16(1), 57–91. Explains acquisition, retention, retrieval, and possible misuse of organizational memory. DOI.

218. Alavi, M., & Leidner, D. E. (2001). Review: Knowledge management and knowledge management systems: Conceptual foundations and research issues. MIS Quarterly, 25(1), 107–136. Distinguishes storage from knowledge creation, transfer, and application. DOI.

219. Grant, R. M. (1996). Toward a knowledge-based theory of the firm. Strategic Management Journal, 17(S2), 109–122. Analyzes the firm’s integration of specialized knowledge. DOI.

220. Hollan, J., Hutchins, E., & Kirsh, D. (2000). Distributed cognition: Toward a new foundation for human–computer interaction research. ACM Transactions on Computer-Human Interaction, 7(2), 174–196. Studies cognition through interactions among people, artifacts, and environments. DOI.

221. Jensen, C. S., & Snodgrass, R. T. (1999). Temporal data management. IEEE Transactions on Knowledge and Data Engineering, 11(1), 36–44. Differentiates valid and transaction times and their applications. Author-hosted paper.

222. World Wide Web Consortium. (2013). PROV-O: The PROV ontology. W3C Recommendation. Defines provenance entities, activities, agents, and relations; it does not warrant the truth of attributed statements. Official specification.

223. OpenLineage Project. (n.d.). OpenLineage specification. OpenLineage. Defines a lineage model involving jobs, runs, datasets, and extensible facets. Official specification.

224. Simon, H. A. (1962). The architecture of complexity. Proceedings of the American Philosophical Society, 106(6), 467–482. Develops the concept of nearly decomposable complex systems. Archival paper.

225. World Wide Web Consortium. (2012). R2RML: RDB to RDF mapping language. W3C Recommendation. Defines customizable mappings from relational database schemas to RDF datasets. Official specification.

226. World Wide Web Consortium. (2017). Shapes Constraint Language (SHACL). W3C Recommendation. Provides constraints for RDF graph validation; conformance does not establish semantic or factual truth. Official specification.

227. World Wide Web Consortium. (2024). Data Catalog Vocabulary (DCAT)—Version 3. W3C Recommendation. Supports interoperable descriptions of datasets and data services. Official specification.

228. Cloud Native Computing Foundation, CloudEvents Project. (n.d.). CloudEvents specification, version 1.0.2. CloudEvents. Defines event context attributes and protocol-neutral event representation, not domain-specific correctness. Official specification.

229. OCEL Standard Project. (n.d.). Object-Centric Event Log 2.0. OCEL Standard. Describes object-centric process evidence linking events, objects, and qualified relationships. Official specification.

230. Feldman, M. S., & Pentland, B. T. (2003). Reconceptualizing organizational routines as a source of flexibility and change. Administrative Science Quarterly, 48(1), 94–118. Distinguishes ostensive routine representations from situated performances. DOI.

231. Temporal Technologies. (n.d.). Event history. Temporal Documentation. Describes workflow event persistence and replay after worker interruption; it does not guarantee exactly-once effects in external systems. Official documentation.

232. Debezium Project. (n.d.). Outbox Event Router. Debezium Documentation. Explains how change-data capture can publish events persisted transactionally with application state. Official documentation.

233. Doyle, J. (1979). A truth maintenance system. Artificial Intelligence, 12(3), 231–272. Develops reason-based representations and dependency-sensitive belief revision for computational problem solvers. DOI.

234. Alchourrón, C. E., Gärdenfors, P., & Makinson, D. (1985). On the logic of theory change: Partial meet contraction and revision functions. The Journal of Symbolic Logic, 50(2), 510–530. Establishes formal results for belief-set contraction and revision under specified logical assumptions. DOI.

235. March, J. G. (1991). Exploration and exploitation in organizational learning. Organization Science, 2(1), 71–87. Models the trade-off between refining established knowledge and exploring alternatives. DOI.

236. Nonaka, I. (1994). A dynamic theory of organizational knowledge creation. Organization Science, 5(1), 14–37. Analyzes interactions between tacit and explicit knowledge and organizational knowledge creation. DOI.

237. Galbraith, J. R. (1974). Organization design: An information processing view. Interfaces, 4(3), 28–36. Analyzes organizational design as a response to uncertainty and information-processing needs. DOI.

238. Rose, S., Borchert, O., Mitchell, S., & Connelly, S. (2020). Zero trust architecture. NIST Special Publication 800-207. Distinguishes identity- and resource-focused authorization from implicit network-based trust. DOI.

239. Debenedetti, E., Zhang, J., Balunović, M., Beurer-Kellner, L., Fischer, M., & Tramèr, F. (2024). AgentDojo: A dynamic environment to evaluate prompt injection attacks and defenses for LLM agents. Advances in Neural Information Processing Systems, 37, Datasets and Benchmarks Track. Empirically evaluates tool-mediated prompt injection under bounded benchmark conditions. Conference publication.

240. Coase, R. H. (1937). The nature of the firm. Economica, 4(16), 386–405. Establishes an economic argument concerning the costs of market coordination and the organizational boundary. DOI.

241. Grant, R. M. (1996). Toward a knowledge-based theory of the firm. Strategic Management Journal, 17(S2), 109–122. Analyzes the firm’s integration of specialized knowledge. DOI.