Enterprises Cannot See Themselves

Enterprises are among humanity's most complex forms of intelligence. How do we build systems that understand them?

The question is often received as a philosophical one. However it is also an engineering one. Three maxims were carved into the forecourt of the Temple of Apollo at Delphi, where Greek citizens, generals and entire city-states came to consult the oracle before going to war, founding a colony, or committing to anything. The first was two words: know thyself.

Whether it meant knowing your limits or knowing your own nature, the placement was the instruction. Before you ask anyone about your future, understand what you already are. It has survived two and a half thousand years because it is at once obvious and nearly impossible to obey, and in all that time it has only ever been addressed to individuals. It has never been possible to put it into an organisation.

That is the difficulty this essay concerns. It is a difficulty of instruments rather than of will, and it is one we have now solved.

Every significant advancement in artificial intelligence has rested on a corpus. Computer vision advanced because someone assembled fourteen million labelled images. Language models advanced because the internet had already recorded, in text, a substantial portion of what human beings know how to say. Protein structure prediction advanced because crystallographers had spent fifty years depositing solved structures into a shared database. 

There has never been a corpus of enterprise cognition. That absence, rather than any limitation of modelling, explains why organisations have remained opaque to the systems built to serve them. It is the ground our research and systems stand on. 

Enterprise cognition

Ask a person how they reached a decision and they can tell you. The reasoning happened in one place, over minutes or days, and they were present for all of it.

An organisation cannot answer the same question. Its thinking is real, and it sits in its people, in what they know, notice and judge. But it sits in many of them at once, across functions and levels, over weeks and sometimes months. Each person holds part of it. Nobody holds the whole.

So when a decision arrives late, or arrives changed, or does not arrive at all, the organisation has no reliable way of establishing why. It can see who was involved. It can see what resulted. What it cannot see is everything in between: where the decision sat and for how long, what it had to be cleared through, which information reached the person who needed it and which did not, and how often the authority to act was held by someone other than the person who could see what needed doing.

That is where enterprise intelligence is lost. Not in the quality of the people, which is usually higher than the organisation's results suggest, but in the conditions the judgement has to travel through before it can become an action.

Those conditions have never been visible to anyone. Our systems make them visible.

Instrumentation precedes modelling

A field like this does not begin with a model. It begins with an instrument.

Astronomy did not begin with a theory of celestial mechanics. It began with a lens ground well enough to resolve the moons of Jupiter, and the theory followed from what the lens made visible. Genomics did not begin with a model of gene expression. It began with sequencing. In both cases the intellectual advance was downstream of an apparatus that made a previously invisible phenomenon legible.

The apparatus required here has to capture enterprise cognition in structured form, in real time, in a way that is comparable across organisations. Comparability is the demanding part. Bespoke observation of a single company produces a case study, and management literature has produced tens of thousands of those without accumulating into a science. What is needed is the same instrument applied identically across many organisations, producing data in a common structure, so that the constructs can be tested for generality rather than asserted from a single interesting example.

Five classes of signal are necessary, and they are the design specification for any such instrument.

Decision provenance. The formal record of a decision as a first-class object: its class, the authority that held it, the basis on which it was taken, the confidence declared at the time, and the alternatives considered.

Latency and its location. Not simply how long decisions take, but where they wait. A decision that takes three weeks because the question is genuinely hard is a different phenomenon from one that takes three weeks because it was waiting on an authority that had no bearing on it, and the aggregate figure conceals the difference entirely.

Authority variance. The divergence between designed authority and exercised authority. Every organisation has a formal structure and a real one, and the distance between them is one of the most informative quantities available about how the system actually functions.

Unusable knowledge. What an organisation knows but cannot act on, and the conditions producing that gap. This is the hardest signal to obtain and the most valuable, because it marks the boundary of the organisation's usable perception. Knowledge that cannot reach the point of decision has the same effect as knowledge the organisation does not hold.

Outcome linkage. Whether the judgement proved sound, connected to the specific decision and the specific confidence declared at the time it was made. Without this, everything above is a description and with it, the data becomes capable of supporting a theory of judgement quality.

The observer problem

Every science that studies human beings confronts the same difficulty, which is that the act of observation changes what is observed, and while in physics this remains a curiosity of measurement at the smallest scales, in the study of human systems it becomes the central obstacle, because the subject knows it is being observed and holds an interest in what the observation concludes.

The enterprise is the acute case, since the quantities that matter most in enterprise cognition are precisely those a participant has reason to misreport under observation: confidence, uncertainty, dissent, dependency. These are not incidental features of the phenomenon but constitute the phenomenon itself, and each of them is distorted by the awareness of being measured, which means that an instrument built without regard for this will produce a dataset in which the most important variables are the least reliable.

The scale of the distortion is illustrated by Ford in 2006, where every executive of a company losing billions of dollars reported operational status as green, not through dishonesty but because the instrument had been correctly understood as an evaluation, and so collected what was safe to declare rather than what was true. The failure was total, and it belonged entirely to the design of the instrument rather than to any deficiency in the people using it.

This sets the hardest condition in the field, which is that an instrument for observing enterprise cognition must be structurally incapable of assessing individuals, since participants infer an instrument's real function from what it does rather than from what it claims and they infer it within days, so that any apparatus capable of being turned toward individuals will be treated as though it already has been, at which point the signal degrades into performance and the data becomes a record of what the organisation was willing to say about itself.

The requirement is methodological rather than ethical, and the distinction matters, because an instrument that cannot be trusted does not collect a compromised version of the phenomenon but fails to collect the phenomenon at all.

What understanding would look like

Suppose the instrument exists and the corpus accumulates, and we are then obliged to answer what it would actually mean for a system to understand an enterprise, since the word is used loosely enough in this field that it has come to mean very little.

The weakest version is description, in which a model characterises an organisation's decision patterns in terms its leadership recognises as accurate, and while this is achievable and genuinely useful, it should not be mistaken for understanding, since a description that is recognised is only a claim about the organisation's willingness to accept it rather than a claim about the world.

The stronger version is prediction of a particular kind, and here the distinction matters more than it might appear. A system understands an organisation when it can state in advance where that organisation's judgement will fail, name the structural condition responsible, and prove correct at a rate substantially better than chance. The prediction cannot be that revenue will decline, which is a forecast about outcomes and can be arrived at by many routes, but that a decision of this class, held at this level, under this arrangement of authority, will be delayed past the point of usefulness or resolved on a basis the organisation itself would reject were it visible at the time. This is a falsifiable claim about cognition rather than about consequences, and it is the standard the field ought to hold itself to, since anything weaker permits a system to be right about the enterprise for reasons that have nothing to do with having understood it.

The strongest version is interventional, and it is the one that decides whether this becomes a science. A system understands an organisation when a structural change it identifies produces the effect it predicted, in the direction anticipated and to approximately the magnitude claimed, which is the point at which enterprise intelligence ceases to be an observational discipline and becomes an experimental one, subject to the same demands of replication and disconfirmation that any other experimental field accepts.

The role of the machine

One final point, because it determines the shape of everything above.

The purpose of such a system is to model the conditions under which human judgement operates in collaboration with artificial intelligence. AI is not to supply the judgement.

This is a research position rather than a reassurance. Inside an enterprise the judgement is the valuable material, and it is already present in quantity. What is scarce is not intelligence but the conditions permitting intelligence to be expressed, harnessed, tested and acted upon. A system that substitutes machine judgement for human judgement attends to a shortage that mostly does not exist, while leaving the real constraint untouched. A system that makes visible where judgement is being suppressed, delayed, misallocated or overridden attends to the constraint itself, and leaves the enterprise more capable of thinking rather than less required to.

Which judgements should sit with people, which with machines, and which require both is among the open questions of this field, and I doubt it has a general answer. It has an answer for each class of decision and behaviour, determined by reversibility, by whether accountability can be transferred at all, and by whether the decision turns on values rather than on inference. Establishing that properly requires deep research in its own right.

Which returns us to Delphi. The instruction was never to be told who you are. It was to come to know it. An oracle answers; self-knowledge is arrived at, and the distinction was understood to matter enough to carve into stone. A system that hands an organisation its conclusions offers the oracle. What organisations have never had is the older and harder thing: the means to see themselves clearly enough to reach their own.

Building artificial intelligence that understands enterprises begins with being able to see them. That is the order we worked in. Our system is the instrument, capturing decisions, moves, enterprise learning, authority, latency and silence as they occur, built so that it observes conditions and never assesses people. What it produces is the corpus that has never existed.