Enterprise Intelligence - Research for Understanding Organised Human Systems

Position paper, version 1.0

Abstract

Artificial intelligence research has organised itself around two frontiers of understanding: the symbolic world of language and information, and the physical world of objects, causality and embodiment. This paper argues for a third. Organised human systems, the firms and institutions through which almost all consequential human activity is conducted, constitute a distinct and unexamined form of intelligence. They perceive, remember, learn, judge, coordinate and adapt, yet no discipline studies them as intelligent systems in their own right. Management science studies their outcomes retrospectively from outside, cognitive science studies the individual mind, and artificial intelligence studies the machine.

The result is that we possess a more precise account of what makes one language model better than enterprise-intelligence-research-for-understanding-organised-human-systemsanother than of what makes one organisation's judgement better than another's, despite the second question governing the allocation of most of the world's capital and the working lives of most of its people.

This paper sets out the case for enterprise intelligence as a research field, identifies the principal obstacle as observational rather than theoretical, specifies what an adequate instrument would need to capture, proposes an initial classification of organisational distortions analogous to the classification of individual cognitive biases, and defines what would count as progress. The central empirical claim advanced throughout is that organisations typically contain substantially more intelligence than they are able to express, and that the difference is produced by structural conditions that are identifiable, measurable and correctable.

1. Introduction

Yann LeCun has observed that a four-year-old child has processed, through vision alone, a volume of data far exceeding the text corpus used to train the largest language models, and argues from this that language is the narrow channel: whatever produces general intelligence is occurring in the sensory one, where machines have no equivalent. Closing that gap, on his account, is the substantive problem facing the field (LeCun, 2022). 

I accept the argument and want to extend its logic. If the limitation of present systems is that they model text rather than worlds, then it is worth asking which worlds are worth modelling. The physical world is one. There is another, of comparable complexity and greater immediate consequence, which has attracted almost no attention as an object of study: the world of organised human systems.

The claim of this paper is that enterprises constitute a form of intelligence, that this intelligence is poorly understood, and that understanding it is a legitimate and tractable scientific programme rather than a business application of techniques developed elsewhere.

Three frontiers can be distinguished. The first is information, comprising language, code, image and the recorded symbolic output of human activity. Progress here has been extraordinary. The second is the physical world, comprising causality, embodiment and environmental prediction. Progress here is active and contested. The third is organised human systems. Progress here is negligible, not because the problem is intractable but because it has not been posed.

2. What is an enterprise?

The definitional question is prior to everything else, and existing answers are unsatisfying for present purposes.

Economics answers through transaction costs: the firm exists because coordinating certain activities through internal authority is cheaper than coordinating them through markets (Coase, 1937; Williamson, 1975). This explains the boundary of the firm without describing what happens inside it. Law answers through the legal person, which is a fiction of convenience. Organisational theory answers through structure, culture and process, which are descriptions of apparatus rather than of function.

None of these treats the enterprise as a cognitive system, though the intellectual materials for doing so have existed for some time. Simon's account of administrative behaviour treated organisations explicitly as structures for making decisions under bounded rationality, and remains the closest thing the field has to a foundational text (Simon, 1947; 1955). Galbraith framed organisational design as an information-processing problem, in which structure exists to manage uncertainty about the work to be performed (Galbraith, 1974). Beer's cybernetic models treated the firm as a viable system with sensory, regulatory and adaptive subsystems (Beer, 1972). Ashby's law of requisite variety, that a regulator must possess at least as much variety as the system it regulates, has direct and largely unexploited implications for the limits of centralised authority in complex organisations (Ashby, 1956).

This lineage stalled. The cybernetic tradition lacked the data and computational capacity to test its models; the decision-theoretic tradition was absorbed into economics and lost its organisational specificity; and the field that inherited the subject matter, management studies, moved toward narrative case work and survey instruments that do not accumulate.

For the purposes of this programme I propose a working definition. An enterprise is a distributed cognitive system that perceives an environment through partial and mediated channels, maintains a model of itself and its situation, allocates authority over classes of decision, produces judgement through the interaction of many agents under conditions of incomplete information and misaligned incentive, and adapts, or fails to adapt, on the basis of feedback it may or may not be able to receive.

The definition is offered as an instrument rather than a truth claim. Its value is that it makes specific research questions askable.

3. The case for treating enterprises as intelligent systems

Six faculties are ordinarily taken as constitutive of intelligence. Each has an organisational analogue that is real, observable and imperfectly understood.

Perception. Organisations take in the world through customers, employees, financial signal, suppliers, regulators, competitors and instrumentation. Perception is necessarily partial, and it is distorted in patterned rather than random ways: channels with formal reporting structures are over-weighted, channels carrying threatening information are attenuated, and information that contradicts an established position tends to be received later and with a higher evidentiary burden than information that confirms it. Organisational blind spots are structural features rather than accidents, and to my knowledge no systematic taxonomy of them exists.

Memory. Organisational memory is distributed across individuals, artefacts, routines and culture (Walsh and Ungson, 1991). A substantial portion is tacit and resides only in people (Polanyi, 1966; Nonaka, 1994). This memory decays, and the decay is invisible from inside: an organisation that has lost knowledge does not know it once possessed it, which makes the loss unmeasurable by ordinary means and its consequences unattributable.

Learning. Organisational learning has been studied extensively (Levitt and March, 1988; Argyris, 1977), and the core distinction, between adjusting action within existing assumptions and revising the assumptions themselves, remains the most useful in the literature. What is missing is any account of learning rate, or of the structural conditions that govern it. Two organisations receiving identical signal will update differently, and we cannot currently predict which will update or by how much.

Judgement. Collective judgement is not the aggregate of individual judgements. Woolley and colleagues demonstrated a measurable general factor of collective intelligence in small groups, and found it correlated with conversational turn-taking distribution and social sensitivity rather than with the average or maximum individual intelligence of members (Woolley et al., 2010). This is among the most important empirical findings available to this field and it has not been extended to organisational scale.

Coordination. The capacity to direct many agents toward an outcome none could produce alone. This is the faculty organisations are conventionally understood to possess, and the only one with a mature literature.

Metacognition. Whether an organisation can observe its own reasoning, recognise flawed patterns in it, and correct them. This is the least studied faculty and, I would argue, the most consequential. An organisation with accurate self-knowledge can correct almost anything else. Without it, every other capability is deployed on a mistaken picture of the situation.

The crucial property across all six is emergence. Organisational intelligence is a property of the system rather than of its members, which is why an organisation composed of exceptional individuals can be reliably unintelligent. Enterprises belong to the class of systems, alongside immune systems, colonies and nervous systems, whose behaviour cannot be derived from the properties of components examined in isolation. This has an immediate methodological consequence: studying organisational intelligence by studying the people inside organisations is a category error, and much of the existing literature commits it.

4. The observation problem

The principal obstacle to this programme is not theoretical. It is that organisational cognition leaves almost no usable trace.

Every major advance in machine learning has followed the assembly of a corpus. There is no corpus of organisational cognition, and the documentary residue that enterprises generate is not one.

Financial records capture consequences, aggregated past the point of attribution and separated by many steps from the judgements that produced them. Calendars capture the allocation of time, which is not the allocation of attention and reveals nothing about resolution. Correspondence captures communication, a filtered projection of belief shaped as much by what is safe to write as by what is held to be true. Transcripts capture speech. A decision is not speech, and the relationship between the two is unreliable in both directions.

What constitutes an act of organisational cognition is a structured object that no ordinary enterprise system records: the class of decision, the authority formally holding it, the agent who in fact made it, the basis and the confidence declared at the time, the alternatives that were live and those foreclosed before articulation, the latency and its location, the dissent registered and its consequences, and the eventual outcome linked back to the originating judgement rather than absorbed into general performance.

Two features of this make the observation problem harder than it first appears.

The first is the absent portion. A significant fraction of organisational cognition occurs in the space between what is believed and what may be said, and this fraction is absent from every record by construction. Organisational silence has been studied as a phenomenon (Morrison and Milliken, 2000), and psychological safety has been shown to predict learning behaviour in teams (Edmondson, 1999), but neither line of work produced an observational instrument. Silence leaves no trace. A model trained on organisational records learns the organisation's comfortable self-presentation, and the most consequential unexamined assumption in any enterprise is precisely the one appearing in no document because nobody has ever been positioned to raise it.

The second is the observer effect, which in this domain is severe enough to be a primary design constraint. Any instrument inside a social system alters the system, and instruments understood to produce evidence about individuals corrupt within a single cycle as participants optimise the measure rather than the underlying condition. The failure mode is well documented in the general form given by Goodhart, and it is acute here because the quantities of interest, confidence, dissent, uncertainty and suppression, are exactly those a rational participant has an interest in misreporting when observed.

An adequate instrument must therefore satisfy a demanding specification. It must capture decision provenance, latency and its location, the variance between designed and exercised authority, the topology of suppression, and outcome linkage. It must record conditions rather than evaluate persons, and this must hold structurally rather than by assurance, since participants infer an instrument's real function from its consequences within days of encountering it. It must apply identically across organisations, since bespoke observation yields case studies and management literature has accumulated tens of thousands of those without converging on a science. And it must operate in real time, since retrospective reconstruction of reasoning is unreliable and systematically flattering.

THE LEAD ICON has built such an instrument and is deploying it as a thirty-day observational cycle within operating enterprises. Its design follows the specification above, and its purpose within this research programme is the assembly of the first comparable corpus of organisational cognition. The claims in the following section are testable against that corpus and are offered as hypotheses rather than findings.

5. Toward a classification of organisational distortion

The study of individual judgement was transformed less by theory than by naming. Anchoring, availability, loss aversion and the rest gave researchers and practitioners a vocabulary of specific, testable distortions, and the names outlived several of the explanations attached to them (Kahneman and Tversky, 1974).

Organisational judgement has no equivalent catalogue. There are isolated named phenomena, groupthink being the best known (Janis, 1972), and there is descriptive work on sensemaking failure (Weick, 1993; 1995) and defensive routine (Argyris, 1977). What does not exist is a systematic classification of the distortions arising specifically because cognition is distributed across people, authority, incentive and time rather than held within a single mind.

I propose eight candidates as an initial classification. Each is stated in a form intended to be falsifiable and measurable against the instrumentation described above. None should be regarded as established.

Authority–competence divergence. Decision rights drift over time toward positions holding organisational status rather than positions holding the relevant information. The mechanism is that authority is allocated by hierarchy, which is stable, while information distribution is determined by proximity to the work, which changes. Hayek's argument about the epistemic advantage of dispersed knowledge in economies applies with equal force inside firms, where it is rarely acknowledged (Hayek, 1945). Predicted signature: measurable distance between the locus of decision authority and the locus of relevant information, increasing with organisational tenure of the decision class.

Escalation gravity. Decisions migrate upward under uncertainty regardless of where competence sits, because upward transfer reduces individual exposure at no individual cost. The distortion is self-reinforcing: each escalation lowers the threshold for the next, and the accumulated load at the apex then slows every decision routed through it, which increases uncertainty further down. Predicted signature: escalation rate rising over time within decision classes, independent of decision complexity.

Silence concentration. The topics least able to be raised are systematically those of greatest consequence, because consequence and threat correlate. This produces an inverse relationship between a subject's importance and its discussability, meaning the organisation's perceptual blind spot is not randomly distributed but concentrated precisely where clear sight matters most. Predicted signature: subjects identified as unraisable clustering around the organisation's largest capital commitments, most senior individuals and most load-bearing assumptions.

Assumption fossilisation. Beliefs that were tested at the point of formation become load-bearing without being retested, and the probability of re-examination declines with age, while the model's dependence on them increases. The result is that the assumptions most critical to the operating model are the least likely to have been examined recently. Predicted signature: inverse correlation between an assumption's structural load and the recency of its last formal test.

Confidence inflation under observation. Declared confidence rises with the seniority of the audience, decoupling stated certainty from actual certainty and rendering the organisation's aggregate self-assessment unreliable in a consistent direction. The phenomenon is illustrated by the well-documented episode at Ford in 2006, in which every executive of a company losing billions reported operational status as green until the reporting instrument was demonstrated to be safe. Predicted signature: systematic divergence between confidence declared upward and outcome realised, with divergence increasing by hierarchical distance.

Consensus foreclosure. Agreement reached early suppresses the variance that would have improved the decision. This is distinct from groupthink, where the mechanism is cohesion; here the mechanism is timing, and it operates in groups with no unusual social pressure. Once a position is established in a group, the cost of introducing an alternative rises sharply, so the quality of a decision is partly a function of how quickly the first plausible option was articulated. Predicted signature: inverse relationship between time-to-first-consensus and subsequent decision quality, controlling for complexity.

Attention capture by the reversible. Leadership attention flows toward urgent reversible decisions and away from slow irreversible ones, because the former produce visible resolution and the latter produce ambiguity. The consequence is that the decisions with the greatest structural consequence receive the least deliberative capacity. Predicted signature: attention allocation inversely proportional to decision irreversibility.

Memory evaporation at the boundary. Organisational knowledge resides disproportionately in individuals rather than artefacts, and departure produces loss the organisation cannot detect, because detecting it requires knowing what was known. The loss surfaces later as unexplained degradation in judgement quality within the affected domain, which is then attributed to other causes. Predicted signature: decision quality degradation in domains following key departures, with a lag, and without contemporaneous attribution.

These eight are a starting position. Some will survive contact with data and some will not, and several will require decomposition. The claim being advanced is not that this particular list is correct but that a list of this kind is constructible, that it has not been constructed, and that constructing it would do for organisational judgement approximately what the heuristics and biases programme did for individual judgement.

6. Open problems

The following are stated as research questions rather than propositions.

1. What are the dimensions of enterprise intelligence? Psychology developed measurable constructs for individual intelligence with substantial predictive validity and substantial controversy. Whether analogous dimensions exist at organisational scale, whether they are universal across sector and geography, and which predict resilience rather than short-term performance, are open.

2. How does collective judgement emerge? Woolley et al. established a collective intelligence factor at small-group scale (2010). Whether this extends to organisational scale, what mediates it, and whether it can be modelled computationally, are unresolved.

3. What is organisational understanding, and can it be measured? Distinguishing an organisation that accurately models itself from one that has merely accumulated information about itself requires a construct that does not currently exist.

4. What governs organisational learning rate? Given identical signal, why do organisations update at different speeds, and what structural conditions determine this?

5. Can organisations possess metacognition? Can an organisation observe its own reasoning, recognise flawed patterns, and correct them, and can artificial systems support this without substituting for it?

6. How does intelligence scale with organisational size? An enterprise of ten behaves differently from one of a hundred thousand. Whether the relationship is monotonic, where it breaks, and what fails first, are unknown. Ashby's requisite variety suggests a formal treatment is possible.

7. Which judgements belong to humans, which to machines, and which require both? This is treated in the literature as a question of capability. It is more usefully treated as a question of accountability transferability, irreversibility, and whether the decision turns on inference or on values.

8. How does trust between human and machine judgement develop, and how should disagreement be resolved? A system that is right and disbelieved is worthless; a system that is wrong and believed is dangerous. The governing dynamics are unstudied.

9. What is the expression gap? If organisations contain more intelligence than they express, the ratio between the two is quantifiable in principle. Establishing it, even approximately, would be the single most consequential measurement in this field.

10. Are there invariant patterns across industry, culture and time? Whether organisational behaviour admits of law-like regularity, and what the limits of prediction are, determines whether this becomes a science or remains a discipline of well-informed description.

7. What would count as progress

A research programme requires standards of evidence. Four are proposed, in ascending order of difficulty.

Descriptive adequacy. A model characterises an organisation's decision patterns in terms its leadership recognises as accurate and did not previously possess. Achievable, useful, and insufficient.

Generality. Constructs derived from one organisation reproduce in others under the same instrument. This is the threshold at which observation becomes science, and it is the reason instrument standardisation matters more than instrument sophistication.

Predictive validity. A system states in advance where a specific organisation's judgement will fail, names the structural condition responsible, and is correct at a rate substantially exceeding chance. The prediction must concern cognition rather than outcome: that a decision of a given class, held at a given level, under a given authority arrangement, will be delayed past usefulness or resolved on a basis the organisation would reject if it were visible.

Interventional validity. A structural change identified by the system produces the predicted effect in the predicted direction and approximate magnitude. This is the threshold at which the field becomes experimental rather than observational, and it is the standard against which the programme should ultimately be judged.

8. Failure modes and constraints

Four risks are inherent to this work and are stated here as commitments rather than caveats.

Surveillance drift. An instrument capable of observing organisational cognition is capable of monitoring individuals. The distinction between recording conditions and evaluating persons must be enforced structurally, in what the instrument is capable of producing, rather than by policy. An instrument that could be repurposed for personnel assessment will be, and the corpus will be destroyed in the process along with the trust that produced it.

Goodhart failure. Any measure that becomes a target ceases to measure. Quantities in this domain are unusually vulnerable, since the participants are sophisticated and the measures concern their own conduct. Measures must be designed to be difficult to optimise directly, and their status as measures rather than targets has to be maintained institutionally.

False quantification. Assigning numbers to organisational phenomena produces an appearance of rigour that may exceed the underlying construct validity. This field will be tempted toward premature formalism, and the appropriate discipline is to hold constructs provisionally and publish their failures.

Automation of judgement. The commercial gradient in enterprise artificial intelligence runs toward replacing human judgement, since that is where the visible cost sits. This programme takes the opposite position on empirical rather than sentimental grounds: inside an enterprise, judgement is abundant and the conditions permitting its expression are scarce. A system that supplies judgement addresses a shortage that mostly does not exist while leaving the actual constraint untouched.

9. Conclusion

Enterprises employ most working people, allocate most productive capital, and produce most of what societies depend upon. They are also, by a considerable margin, the least understood intelligent systems of comparable consequence. We can characterise the failure modes of a transformer more precisely than the failure modes of a leadership team, and the second governs vastly more human activity than the first.

The obstacle has been observational rather than conceptual. The faculties are real, the questions are askable, and the theoretical materials have existed since Simon, Ashby and Beer. What has been missing is an apparatus capable of making organisational cognition legible in a form that accumulates across organisations. That is now buildable, and the corpus it produces will determine whether the questions in section 6 admit of answers.

The wager underlying this programme is that organisations are not short of intelligence. They are short of the conditions under which the intelligence they already contain can be expressed, tested and acted upon, and that shortfall is structural, measurable and correctable. If that is right, the return is not incremental improvement in organisational performance. It is a scientific account of one of the most important forms of collective intelligence humanity has produced, and a substantial recovery of capability that already exists and is currently being wasted.

References

• Argyris, C. (1977). Double loop learning in organizations. Harvard Business Review, 55(5).

• Ashby, W. R. (1956). An Introduction to Cybernetics. Chapman and Hall.

• Beer, S. (1972). Brain of the Firm. Allen Lane.

• Coase, R. H. (1937). The nature of the firm. Economica, 4(16).

• Cyert, R. M. and March, J. G. (1963). A Behavioral Theory of the Firm. Prentice-Hall.

• Edmondson, A. (1999). Psychological safety and learning behavior in work teams. Administrative Science Quarterly, 44(2).

• Galbraith, J. R. (1974). Organization design: an information processing view. Interfaces, 4(3).

• Hayek, F. A. (1945). The use of knowledge in society. American Economic Review, 35(4).

• Janis, I. L. (1972). Victims of Groupthink. Houghton Mifflin.

• Kahneman, D. and Tversky, A. (1974). Judgment under uncertainty: heuristics and biases. Science, 185(4157).

• LeCun, Y. (2022). A path towards autonomous machine intelligence. Open Review.

• Levitt, B. and March, J. G. (1988). Organizational learning. Annual Review of Sociology, 14.

• March, J. G. (1991). Exploration and exploitation in organizational learning. Organization Science, 2(1).

• Morrison, E. W. and Milliken, F. J. (2000). Organizational silence: a barrier to change and development in a pluralistic world. Academy of Management Review, 25(4).

• Nonaka, I. (1994). A dynamic theory of organizational knowledge creation. Organization Science, 5(1).

• Polanyi, M. (1966). The Tacit Dimension. Routledge.

• Simon, H. A. (1947). Administrative Behavior. Macmillan.

• Simon, H. A. (1955). A behavioral model of rational choice. Quarterly Journal of Economics, 69(1).

• Walsh, J. P. and Ungson, G. R. (1991). Organizational memory. Academy of Management Review, 16(1).

• Weick, K. E. (1993). The collapse of sensemaking in organizations: the Mann Gulch disaster. Administrative Science Quarterly, 38(4).

• Weick, K. E. (1995). Sensemaking in Organizations. Sage.

• Williamson, O. E. (1975). Markets and Hierarchies. Free Press.

• Woolley, A. W., Chabris, C. F., Pentland, A., Hashmi, N. and Malone, T. W. (2010). Evidence for a collective intelligence factor in the performance of human groups. Science, 330(6004).