The Systems No One Wrote Down

Ask five people in a company to describe the same core process — request, approval, delivery — and more often than not you will get five accounts that contradict each other in several places. Who signs off, when something counts as commissioned, at what point a query becomes necessary: the answers diverge. The work gets done anyway. Nobody is really mistaken here. The process everyone has in mind often does not exist as text at all — and where it does, it is not always carried out the same way. People get used to how things run; they are not walking process libraries.

This is not a shortage of order. This IS the order.

The Documented Is Not the Only Thing That Works

What holds a company together is rarely captured in full by its records. It sits in routines that grew out of occasions nobody remembers; in responsibilities that depart from the job description because one person happens to be better at something; in what people know about each other and therefore never have to negotiate. Michael Polanyi coined the phrase for this kind of knowledge: we can know more than we can tell.¹ In organizations this holds not only for individuals but for the working order itself, which operates without ever being stated.

This invisible structure is no less real than the org chart. On the question of what actually happens, it is often the more reliable account. The org chart says who reports to whom. It does not say whom you call when something is pressing.

Why Implicit Order Now Has Practical Consequences

As long as only people worked inside this order, the unstated part transferred well enough. You picked it up by being present, by asking, by watching. A new colleague needed a few months and then knew how things run here, without anyone having to write it down.

An AI system that forms a picture of a company from its data has no such route. It knows what was recorded. The gap between the written and the working order used to be an inconvenience. It turns into a verdict once a system answers from what it holds, without knowing where its edges are.

That gap is unevenly distributed. Bloor Research put the point in July 2026 as follows: bias enters at the point of capture, in what a system was allowed to see in the first place.² A body of data represents coordination, onboarding, and keeping relationships intact less well than it represents cycle times and unit counts — one leaves traces in the system, the other happens between people. What matters most stays invisible.

This work is not a side duty attached to particular roles. Someone who brings a new colleague up to speed, catches a conflict early, or gets two departments talking keeps the operation running as surely as someone who delivers unit counts. The only difference is that one can be counted and the other cannot. Where a company records only what is countable, half of its own work disappears from its picture of itself — and nobody decided it should.

Recording Is the New Designing

This is where taking stock turns into an act of design. Writing down how something works changes it. Not as a side effect, but because a documented process has a different standing than a practiced one. Before, there was established practice that people oriented themselves by. Afterward there is a standard to depart from — and a departure now calls for a reason. The same move that counted as experience before counts as an exception after.

Recording the unstated is therefore the most demanding part of designing an intelligence architecture, not merely its preparation. It has to distinguish between two cases that look identical from outside: some things nobody wrote down because the occasion never arose; others went unstated because that is the only way they work.

Niklas Luhmann described the second case early on. Organizations depart from their own formal rules, and not only to their detriment: some departures are what keep the operation working at all, because a formal order cannot anticipate every situation.³ Recording such a practice does not make it visible — it ends it. It worked because it was off the record. What remains is a very orderly operation with less room to maneuver than before.

This objection does not dissolve. It is the price, and anyone who records should know they are paying it.

To What End Do We Map, and for Whom?

At some point the purpose of all this recording has to be settled. A system that knows who does what and when can take work off people’s hands: it retrieves past cases, flags deadlines, prepares what someone would otherwise assemble by hand. The same system can supervise them: it compares, rates, reports anomalies. The difference does not lie in the data. It lies in how the record was cut — what was captured, at what resolution, and who gets to read the analysis.

Calling a system humane here means something specific: that it supports people’s work rather than assessing it. This is not a property you add to a finished system, because in a finished system the intent behind the original cut is no longer legible. Two systems with the same interface may have come out of quite different decisions, and nothing on the surface tells you which.

Anyone who does not choose the cut deliberately still chooses. Recording also happens more readily when something is easy to record: clicks, cycle times, ticket counts. Coordination work drops out, not by intent, but because it is harder to count. And whatever falls out of the record does not exist for any system that learns from it.

What Should Be Made Explicit?

Concluding from all this that the safest course is to document nothing and leave the grown order undisturbed is not an available option. It would mean keeping the company illegible and opaque to AI agents, with the consequence that agentic systems stay some distance off target and turn out barely usable. Agents would then answer questions only from whatever happened to be within reach — tickets, minutes, calendars, shared drives that nobody assembled for this purpose. The mapping is already well under way; it is simply unsupervised.

The choice is therefore not between recording and not recording. It is between deliberate and incidental recording, and incidental recording is happening everywhere right now. A careful yet deliberate mapping of what a company does implicitly is essential if its AI systems are to be of any real use.

The Open Question

Companies have established ways of describing themselves. Process modeling sets down how a workflow is supposed to run. Knowledge management collects what staff are able to articulate. Data governance settles ownership, quality, and access for bodies of data. All three were built for their own purposes: for control, for continuity, for evidence and retrievability.

None of them was designed to capture a company’s tacit order. Build AI systems on that documentation alone, meaning to support people’s work, and the best case is that the systems deliver no real benefit. The worst case is that a NEW invisible working order takes hold — one that can lead to a loss of control, because neither where a piece of information came from nor what it means can still be judged.

This is a central challenge for companies. As things stand, it is being answered by systems that nobody assigned to answer it.

Footnotes

1. Michael Polanyi, The Tacit Dimension, Doubleday, 1966, p. 4: “we can know more than we can tell.” Polanyi develops the concept for individual skill; the extension to an organization’s working order is made by this article. ↩

2. Bloor Research International, Intelligence Architecture, Not Artificial Intelligence, July 3, 2026, following a session of UNITAR’s Gender Dialogue Series in June 2026. The piece is published without a named author. Accessed September 3, 2026. ↩

3. Niklas Luhmann, Funktionen und Folgen formaler Organisation, Duncker & Humblot, 1964, especially chapter 8 on “brauchbare Illegalität” (useful illegality). ↩