Us, Them and AI: Integrating AI Well Starts Before the Decision Feels Settled

Under pressure the pull is to leap straight into categorising. I keep being reminded lately that all of this starts with confusion.

Most organisations have already decided to adopt AI. The decision feels settled and the problem feels named, and that is usually when people stop listening for new signal in the noise.

The work the model learned from and the work it is being pointed at may not be the same work, and nobody in the room can tell yet. The words are doing their usual job. "Integration." "The AI." "This workflow." Each one covers a gap it was never built to reveal.

The confusion that comes with this is worth slowing down for. It is an accurate first reading of a situation where the familiar words have stopped pointing reliably at the things they used to point at. The more useful move is to notice it, understand what it is made of, and let that understanding shape where you go next.

What it is usually made of is a question about knowledge. AI systems learn from what gets written down. Process manuals, ticket logs, formal communication, whatever the curators kept. The model becomes fluent in the version of the work that survived governance. The version that actually keeps the organisation running is mostly elsewhere, in the corrections an experienced operator makes without thinking, the back-channel conversation that resolves an ambiguous specification before anyone files a ticket, the workaround a team has used for three years because the documentation step never made the priority list. This is the informal layer of an organisation, the part that fails the audit trail and keeps the business functioning.

So the question sitting under the confusion, before any technology decision, is about what the model was never shown. What has it learned from? Whose work is in the training data? Whose work is invisible? And in the kind of work you are introducing AI into, how much does the difference actually matter?

Some Work Hides From the Record

In some kinds of work the documented version is most of what there is, and AI trained on it performs well. Payroll, data validation, rules-driven processing. The work has been deliberately stripped of judgement, so the manual and the practice are nearly the same thing. This is the native habitat of AI, and where most organisations have their easiest start. Little of the real work is missing from the record, and the return is clear.

Other work only looks like that from the outside. A senior analyst follows the documented procedure and then, at the point that matters, overrides it on an instinct built from years of cases that never made it into any manual. The documentation is a floor. The real work happens above it, in the judgement the procedure cannot capture. AI trained on the floor is fluent in the procedure and has never seen the override.

There is a well-documented case from the North Sea oil industry where this turned costly in a way nobody designed. Geophysicists began deferring to the AI's read because the institutional cost of disagreeing with it had grown too high. They could not be faulted if the AI said there was oil and it was wrong. If the AI said no and they overruled it and were wrong, the personal exposure was unbearable. So the override quietly stopped happening. The model had been trained on the documented decisions, never on the abductive insight that produced the good ones, because that insight is by its nature absent from the record. The AI did something subtler than replacing the experts. It made exercising expertise too risky to bother with. The people most able to catch the error were the ones with the most reason not to.

Further out still is work where the documentation is barely a trace. Most organisational change, most strategy, most of what brings leaders here. The record shows what was decided and says nothing about what was sensed in the deciding. What the model never sees is most of what matters, and AI's honest role shrinks to surfacing signals a team might otherwise miss while the interpretation stays with the people who carry the context. At the far edge is genuine crisis, where the situation has never happened in this shape and there is nothing to train on at all.

There is a framework, Cynefin, that maps this spectrum properly and names its domains. The point worth carrying out of it is simpler than the map. Work has different relationships to what gets written down, and an AI trained on the written-down version will either succeed quietly or mislead confidently depending on which kind of work it meets. Reading the confusion is, in large part, working out which kind you are looking at before you decide anything. Work rarely announces itself, and it often turns out to have been somewhere other than where you first put it.

When the Confusion Walks Into the Room

A merger or an acquisition surfaces all of this with unusual force, and it is worth watching closely, because it shows what happens when the confusion gets skipped rather than read.

Two organisations meet. The named function exists on both sides, so the acquirer assumes the work is the same work and the tooling will carry across. Both sides want the integration to succeed and they agree on that completely. What they have never had is a shared sense of what the words point to. "The workflow" meant one thing in one building and something quietly different in the other, and nobody felt confused about it, because each side was sure it already knew.

What transfers is the documentation. What gets left behind is the informal layer that made the workflow function. If the AI on the acquirer's side has learned only from one organisation's records, applying it to the other's work is a category move dressed as a system upgrade. The target's tacit knowledge was never part of what the model learned. The target's people watch their informal layer vanish into a tool that does not know it exists and was never told it was there. The integration plan calls this AI integration. The lived experience is two understandings of the same word meeting, and only one of them surviving the audit trail.

What the plan does not price is how fast the target's people form their read of all this, and how slowly that read reverses. The judgement lands early, often in the first weeks, and it lands negative by default. Once a team has decided the new system does not understand their work, every later misfire confirms it and every success gets explained away. The cost shows up downstream as the things that quietly stop. The overrides that no longer happen, the workarounds nobody volunteers, the bottlenecks that form where informal knowledge used to flow. By the time it reaches a status report it reads as slippage with no obvious cause.

The acquirer who arrived confident was the one carrying the risk. Competence narrowed the view. The merger looked like a transfer of documented function, so that is what got planned for, and the part that actually ran the place was never on the map. The confusion the target's people felt, watching their work disappear into a system that did not recognise it, was the accurate reading. It arrived too late to be useful, because the certainty upstream had already made the decision.

Staying In It a Little Longer

There is a small thing worth doing before the confusion gets resolved too quickly, and it works anywhere.

Take one workflow where AI is being discussed, tested, or quietly used. Get three people who touch that workflow from different angles to describe, separately, what it actually does and what doing it well looks like. Keep the accounts apart until everyone has answered.

Then read the spread. When the three accounts line up, the work probably holds still, the documented version is close to the real one, and the model has a fair chance of having learned something resembling the truth. When the accounts diverge, and they often diverge in ways that surprise the people in the room, the discovery is larger than any question about which tool to buy. The organisation never shared a single description of this work to begin with. A model cannot have learned a thing the people doing the work never agreed on. That disagreement was there long before AI arrived. AI only made it expensive.

The discomfort in that room is the data. The point is not to resolve it quickly. The point is to understand what it is telling you before you decide anything.

The Opportunity

AI is going to keep accelerating, and in the parts of the organisation where it should accelerate, it will. Where the documentation is nearly the whole truth, it is already changing how the work gets done. Where expertise and machine analysis combine, they reach results neither would alone. These are real gains and worth pursuing.

The quieter opportunity sits alongside them. It is the ability to notice the confusion a new deployment brings, to understand what it is made of, and to know which parts of the work depend on knowledge the model has never been shown. That noticing has a short window. The weak signal is loudest early, before the confident plan hardens and before the negative read sets, and it gets harder to hear with every week that the certainty upstream goes unquestioned. Noticed early, it costs little to act on. Left until it surfaces in a status report, it has already become the thing nobody can quite explain. The technology decisions follow from the noticing, and they go better when they do.