Does AI Break the Pattern?

The meeting was at nine and he was in the room at ten to, which is not a habit of his, and the reason for it that morning was that he wanted to see the screen before there were people sitting in front of it.

Eleven were on the invitation and nine came. It is the room on the first floor with the long table and the window onto the yard, and the blind on that window has been broken at one end since before he arrived, so on a bright morning half the table gets the sun and the people who know sit on the other half. He sat on the sunny side, because he had come in early and it had not occurred to him, and then spent the first part of the meeting unable to read his own screen.

Six items. His was fourth.

Quality Systems went second. He had seen the deck before — not this version, but a version, in October, and probably one before that. The programme is nineteen months old now. The woman presenting it is competent and was not the person who scoped it, which she has never said in a meeting and does not have to, because everybody in that room can do the arithmetic between what the roadmap slide showed in October and what it shows now. Two milestones have moved a quarter to the right. One has been renamed.

The roadmap slide is still very good.

His own item took four minutes. He said there was a discovery report now, monthly, that the first one had found sixty-one people on the site using generative AI tools through company machines with nothing behind any of it, and that he was proposing two things — a licensed platform, and a communication to line management. There were no names in the deck and he read none out. He said, because it is true, that this was not about catching anybody. The line on the slide said establish visibility of unmanaged AI usage, and he read it in the flat voice you use for a slide you wrote yourself and have already stopped hearing.

Nobody objected. Somebody from IT asked which platform and he said that was the next question rather than this one. Somebody asked whether the communication should come from him or from the plant manager, which is the sort of question that means a thing has already been agreed. It was minuted as an action with his name against it and a date in October.

He had thought about the whiteboard twice since Monday — once on the walk to the car, and once at about half past six that morning standing in his own kitchen — and both times he had got as far as understanding that a line about establishing visibility produces, in the ordinary course of events, a cloth. Both times the next thought did not arrive. There is no sentence available to him that means leave that board alone, there is no meeting at which such a sentence would be received, and he is not a man who says things in rooms without knowing what they cost. So the line stayed where it was and he read it out, and his four minutes ended, and the small unfinished feeling he had about it went wherever those go.

Item five was the one the invitation had been about. Please come prepared to share what your team is working on — a sentence somebody in the group function writes into a calendar entry with no idea what it does to the people who read it. It ran the way those items run. Two people described things that were being procured. One described a pilot that had also been described at the last one.

So he asked the room the ordinary question, because it was his item in everything but name and because the round was thinning out. Was anybody actually doing anything with any of this. Not in a pilot. In their week.

Maria had been in the room since five past nine, on the shaded side, third seat from the door, with a printed agenda in front of her that she had not written anything on.

What she felt when the question came was not fear, and it is worth being exact about that, because fear is the word that gets installed afterwards by whoever tells the story and it is the wrong one. She has been afraid in that building. At twenty-six, waiting on a decision about a sort order that never came, she was afraid for about three weeks and then stopped, and what replaced it has lasted eleven years and does not feel like anything at all. This was flatter than fear and much faster. It was the sensation of arriving at a calculation you have run so many times that you can no longer catch yourself running it.

She had decided on Sunday evening. Then again on Monday, and again in the corridor at two minutes past nine, and each time the decision came out the same and each time it presented itself as fresh.

So she said she had a way of doing the deviation reports now. That it was quicker. That it pulls the week together and gives her a summary on a Friday morning, and that it saves her a fair amount of time.

Every word of that is true and none of it is the thing.

She did not say that it clusters by failure mode across suppliers rather than by supplier, which is the whole of what it does and the entire reason it works. She did not say that it caught the seal defect five weeks early in the spring, which sits on the record as an ordinary catch by an experienced engineer and has never been attached to anything. She did not say my own account, or eighteen euros, or nine months. None of that was concealed in the sense of a decision taken not to say it. There was a version of the sentence that fitted the room and a version that did not, and the one that fitted arrived first — the way it had arrived first in the canteen on Friday, in front of a colleague with no authority over her whatsoever.

It took about forty seconds.

The response was warm and it was not empty. Two people said versions of that sounding useful. The man from logistics, sitting almost opposite her, the one who made the remark about pallet labelling four years ago, said he had wondered whether something like that was possible. And the man with excellence in his title said it sounded like exactly the kind of thing they wanted more of, and that the licensed platform, once it was in, would let people do that sort of thing properly.

He meant it as encouragement, and it is encouragement, and he is right about the platform. Read from her chair it is also a description of what would happen to her tool if she put it in front of the thing he has just been given an action to buy — that it acquires a scope and a sponsor and somebody’s name, which is Chapter 8 arriving in a sentence offered kindly across a table by a man trying to be helpful.

She said that would be good.

The item moved on. Somebody asked about a supplier portal. The meeting finished eight minutes early, which is the only thing anybody in it remembered about it by Thursday.

Nothing was suppressed in that room. Nobody was punished, nobody was overruled, nobody was so much as discouraging, and the two people who between them held both halves of the answer sat four seats apart for fifty-two minutes and neither of them did a single thing wrong. He asked the question this book says he should ask. She answered it truthfully. The answer was smaller than the thing, and he heard it at the size it arrived in, because there is no way for a person to hear a thing at a size other than the one it arrives in.

A reader who has come through twelve chapters has been braced for a meeting that goes badly, and this one does not go badly, and that is deliberate. The version where somebody gets punished is the version everybody already believes and it is the rarer one. Twelve chapters of machinery, and what the machinery produces at the end of it is fifty-two minutes in which nothing happens and nobody is at fault.

The tool ran on Friday morning, the way it has every Friday for nine months.

The question, asked properly

Chapter 2 ended on a question and then spent ten chapters not answering it, which was not an oversight. It could not be answered until the machinery had been laid out, because the answer turns on which parts of the machinery a technology can actually reach.

Is AI just the sixth wave of a pattern that has survived all five before it, or is it, for the first time, the thing that changes the pattern itself?

This is the most conceptual chapter in the book and I would rather say so at the top than have it inferred from the thinness of the citations. What follows is argument. The evidence in it has all appeared earlier and is doing a second job here; nothing new is being introduced that would settle anything, because there is nothing available that would. The honest description is that I am reasoning forward from a pattern, and reasoning forward from a pattern is the thing this book has spent thirteen chapters warning about when other people do it. So I will give the other side its strongest form first, and I will not be giving it a straw version, because I think one of its three moves is dangerous to my position and I have not seen anybody make it.

The case that it breaks

Three arguments, in ascending order of how much trouble they cause me.

The gate’s price collapsed. Chapters 4 and 5 established that an increment does not fail a business case so much as fail to be worth one. The approval cycle has a fixed cost — the form, the numbers, the slot in a calendar, the fortnight of preparing to be asked a question — and the increment is smaller than that cost by construction, which is why the scale threshold discards things for being small rather than for being wrong. Now watch that arithmetic when the artefacts cost an afternoon instead of a fortnight. The case and the risk note write themselves, and the ratio that made the whole exercise irrational inverts. On this reading the submarine was never a response to hostility so much as a response to paperwork, and the paperwork just got cheap.

Adulthood got cheap with it. Chapter 12 said an improvement needs different conditions to survive being born and to survive being twenty, and that the adulthood conditions are documentation, a second user, a place it lives and somebody accountable for it. Of those, documentation has always been paid for in the currency people have least of, which is why the continuous-improvement literature has spent decades watching local gains evaporate.

Maria’s tool leaves in a laptop bag next spring because nobody wrote it down. Writing things down is the single task this technology is best at. A submarine-to-surface transition that used to cost a project could now cost a Thursday afternoon, and if that is true then the whole of Chapter 12 is a solution to a problem that is about to solve itself.

And the nursery may no longer be needed. This is the serious one. The entire justification for the submarine, from Criscuolo onward, is delayed assessment — the embryo gets to become what it is going to be before anybody judges it, and there is no path from the first version to the approvable version that runs through a meeting. That argument has a hidden premise, which is that the distance between the two versions is measured in months.

Collapse the cost of building and you collapse the distance. If the sixth version of a thing can exist by the end of a weekend, then the window during which the idea is unapprovable shrinks toward nothing, and the person can arrive at the defensible version before anybody has had the opportunity to judge the indefensible one. There is nothing to protect, because the vulnerable phase is over by Tuesday.

Put together, that is a coherent and optimistic account in which the mechanisms in Parts II and III remain entirely real and stop mattering, because the technology routes around them from below. I do not think it is right. I do think a reader could put this book down at the end of Part III, read the same evidence, and arrive there, and I would not be able to call them careless.

The case that it is the sixth

Every previous wave lowered the cost of building too. That is what a technology wave is. The mainframe put computation inside the firm, the personal computer put it on the desk of the person with the problem, client-server put the data where they could reach it, the web removed the need to own any of it, and mobile removed the need to be anywhere in particular. Each was described at the time as the thing that would let people just get on with it.

The Durability Ledger from Chapter 2 is the instrument for exactly this, and it returns the same answer every time you run it: the shape is older than the tool, and the tool did not move it. Schön has every element of Maria’s Tuesday in a paper published in 1963, when the newest of the five was a machine in a room most of the company was not allowed into.

Two of the three currencies do not touch the cost of building at all. Capital lumpiness is a property of plants, coordination and irreversibility, and a cheaper way to write a proposal does nothing to it — you still cannot half-retool a line, and the machinery built to guard that decision will still be the machinery a small reversible thing is routed through, because that is the only machinery there is. Careers select the same way regardless of what tools cost, because Chapter 8’s mechanism runs on visibility of events and a six-hour tool is not an event.

And the third currency gets worse rather than better. Attention is zero-sum, which is the property that makes Bouquet and Birkinshaw’s finding bite, and cheap building increases the number of things competing for a fixed pool. Ten people who can each produce a small working thing in a weekend have not made it easier for any one of those things to be seen. They have made it harder, and they have done it to each other, while the large programme in the room retains every gram of structural weight it had before. Maria’s tool was never rejected. It was never seen, and forty more like it in the same building does not improve the odds of anybody turning to look.

Territory does not care how good the case is. Chapter 4’s second killer was never about evidence. An increment gets defended against because it sits in somebody’s marked water, and the reaction to infringement is a reaction to the marking. A flawless business case produced in an afternoon buys a faster and better-documented no.

Then the asymmetry, which is the one that decides it for me.

Every previous wave lowered the cost of building without lowering the cost of watching. The personal computer handed an engineer a spreadsheet and handed her manager nothing comparable; reading her afternoon still took a person sitting down to read it, and no organisation has ever had that person spare. That asymmetry is the reason the underground survived five technologies, and what protected it was never policy but the cost of looking.

This wave is the first one to lower both, in the same purchase, in about two years, as a side effect of buying something else. And of the two capabilities that arrived together, only one of them has a budget holder, a governance argument and a line on a slide behind it. Chapter 11 is the whole of that story and I will not run it again. What matters here is what it does to the fork: the technology that makes the submarine nearly free is the first one that also makes hunting it feasible, and those two effects are not independent, because the second is a reaction to the first.

Which leaves the optimistic case’s third move, and I think it fails for a reason that has nothing to do with speed. The nursery was never protecting the idea from the calendar. It was protecting it from the observer. Chapter 10 established that observation does not read a value off the thing, it fixes the thing at whatever is legible on the day of looking, and legibility has almost nothing to do with maturity. Maria’s tool on Tuesday morning was nine months old, worked, and had a catch on the record. She still described a faster report, in forty seconds, to a room full of people who wanted to hear something good.

Speed of building does not touch that. You can arrive at the sixth version by Sunday night and still have exactly the same conversation on Tuesday, because what shrinks the thing has nothing to do with its age and everything to do with being looked at by somebody who could act on what they see.

And the tool that clusters by failure mode did not come from nine months of building. It came from nine months of running against real reports, which is a different resource and one that does not compress.

Where I land

Neither branch, and the reason is more useful than either.

AI is the first of the six waves to change the cost on both sides of the disclosure decision at the same time. Building got nearly free, which is the optimistic case and it is real. Watching got nearly free, which is Chapter 11, and it is also real. Every previous wave moved one lever and the pattern absorbed it. This one moves both, in opposite directions, and the outcome is not determined by the technology, because the technology supplied both of the things that would determine it.

So the honest answer to Chapter 2’s question is that the fork is real and the branch is not chosen by the tools. It is chosen by wiring, and the wiring is a small number of unglamorous arrangements that Chapter 12 spent itself on: whether the report comes at function level or with names in an appendix, whether anybody who sees individual output also writes reviews, whether a person may change their own work without asking, whether there is a route from a working thing to a supported thing that does not pass through a business case.

That is a smaller claim than either side of the fork and it is stranger than both. For sixty years the persistence of this pattern was a property of the system — nobody chose it, nobody could have unchosen it, and the technology of the day was not capable of moving it in either direction. That is no longer the case.

The pattern’s survival over the next decade is going to be a decision, and it is going to be taken mostly by people who will not experience themselves as taking one. It will be taken in the default configuration of a product that arrives with individual-level reporting switched on because that was the easier thing to build, and in the ninety minutes somebody spends writing a covering note to go with it. It will be taken by a man with a good record and no bad intent, reading out a line he wrote in forty minutes because he could not think what to put in its place.

This is my argument and not a finding, and the next section says how far it runs past the evidence.

The choice Chapter 1 opened with

I said at the start that we would come back to her and that by the end we would know what she should do, and that at that point neither of us did. So here is the closing of it, and it is not the shape I expected when I wrote that line.

Chapter 1 put the choice as a binary. Surface it and probably lose it, or hide it and definitely lose it, later, quietly. Two branches, one object, one decision, taken on a Tuesday.

Every part of that turned out to be wrong except the arithmetic.

There was never one object. Chapter 10 took that apart: the thing that exists in the dark and the thing that exists in the light are different objects, and the act of moving between them is what makes the second out of the first. And there was never a moment of decision, because the decision was made at twenty-six, in front of an alphabetical queue that is still alphabetical, and what happened on Tuesday morning was recall rather than deliberation — which Chapter 9 said in as many words, and which I did not fully understand the consequences of until I had to write the meeting.

What she actually did was both branches at once, and neither of them properly. She surfaced, in the only form available to her, which was a true sentence about a faster report. She stayed submerged, because the thing that makes it worth having was still on her laptop when she sat down. The plant received neither the risk nor the benefit. And nobody in that room could have told you which of the two things had just happened, including her.

Chapter 1 also predicted the ending and I did not notice at the time that it had. It will not be rejected, because rejection would require somebody to look at it first. It will simply never be seen, which is a quieter fate and one that leaves no record at all. Twelve chapters later, that is exactly what happened, and it happened in a meeting convened specifically to look, with the right question asked out loud by the right person.

So: should she have raised her hand?

She was right, and being right is not the same as being well served. Every term in her arithmetic was bent by machinery nobody built on purpose, and the three things that make a hidden improvement survivable in daylight — a rate she can defend, a second person who has used it, an event it has already survived — she has one of, and two of them require somebody else. Nobody in that building is building them. Call none of that a failure of nerve or a failure of goodwill; it is what one operating system and no route produce in a building where everybody involved is reasonable.

She transfers in the spring. The tool goes with her, and the deviation reports start piling up again in the summer, and the seal defects go back to being caught five weeks later, and nobody connects any of it to anything because nobody knew there was anything to connect. There will be no incident and nothing will appear in a report. The plant will simply be slightly worse at something it was previously good at, for reasons that are not written down anywhere.

That is the default ending. It is happening now, in most large organisations, at a scale nobody can measure, and the only thing standing between it and a different ending is a small number of people deciding to change the wiring before they buy the instrument.

Where the evidence stops

This chapter is argument standing on evidence presented earlier, and the argument runs further past its supports here than anywhere else in the book. The whole of it should be read at a lower confidence than Parts II and III, and lower than Chapter 12, which was already the provisional half.

The six waves are a framing, not a dataset. Nobody has measured rates of hidden innovation across technology generations, because nobody could — the phenomenon is invisible by construction and the baseline does not exist for any decade, including this one. The Durability Ledger is a habit for noticing that a problem is older than its name. It is not an instrument, it produces no number, and I have used it here as a structuring device rather than as evidence.

The fork is not falsifiable in the short run. Both branches predict a great deal of what is currently visible. Rising official adoption is consistent with the pattern breaking and equally consistent with the fleet relocating and the compliance number improving as the estate goes dark, which is Chapter 11’s whole warning. Anybody who tells you in 2026 that the data already settles this is reading their preferred branch into an ambiguous signal, and that includes me if I ever do it.

The both-costs-fell asymmetry is mine and it is unquantified. The claim that this is the first wave to lower the cost of building and the cost of watching in the same purchase is an argument about relative magnitudes that I cannot put numbers to. Somebody could reasonably say that the web lowered the cost of watching too, through logging and email retention, and that the difference here is one of degree. I think the degree is large enough to be a difference in kind, because summarising judgement work at scale was not previously possible at any price. But that is a judgement about a threshold and I have not measured where the threshold is.

And the optimistic case may be right. Its third move is the one I would attack this chapter with. If building becomes cheap enough that the version shown in month zero is already the good version, the observation principle loses most of its bite, because there is no better hidden version for the visible one to be a diminished copy of. My answer is that Maria’s tool got good by being run against nine months of real reports rather than by being built for nine months, and that running time does not compress the way building time does. That answer is an argument about where the value in an increment comes from. It is testable and it has not been tested, and if it is wrong then the pessimistic half of this chapter goes with it.

Here is the study that would settle it, and it is the one I would most like to see run. Take a cohort of employee-built AI tools now, identified in a way that does not require anybody to confess to anything — through a voluntary registry with no evaluative consequence, or through a third party with genuine anonymity. Record for each one the date it was built.

Then, five years from now, measure two proportions: how many were voluntarily shown to somebody with authority within their first year, and how many survived being shown in a recognisable form. If both rise substantially while the cost of building stays low, the pattern broke, the optimistic case was right, and this book is a period piece about the last years of a sixty-year phenomenon. If they hold where they are, it is the sixth wave.

The problem with that study is the one that has attached itself to every study I have asked for in this book, and by now it should be familiar. There is no baseline for the first number, in this decade or any of the previous five, and there is no baseline because the thing being measured hides from measurement, and the reason it hides from measurement is the subject of the book. A phenomenon that cannot be counted is real anyway; it is only unfalsifiable at the speed anybody would like.