The Choice, and the Fleet

The meeting invitation arrived on a Tuesday, and Maria read it twice.

Quarterly review — AI use cases. Please come prepared to share what your team is working on.

She had something to share, and that was exactly the problem.

Eight months earlier she had been drowning in supplier deviation reports. Every non-conforming part that arrived at the plant generated one, with a description, a photograph, a disposition and a root-cause note, and her job was to read them, categorise them, spot the patterns and escalate anything systemic. On a good week there were forty of them. On a bad week there were two hundred, and by Friday evening the patterns she was paid to notice had simply disappeared into the volume.

So one evening, at home, she pasted an anonymised report into a chatbot and asked it to summarise the failure mode in one line. It did. She tried ten more, and it got them right — not perfectly, but better than she did at six on a Friday. Over the following weekends she built a rough workflow: strip the identifying data, batch the reports, extract the failure mode and the supplier, cluster them, flag anything appearing three times in a fortnight.

It took her about six hours to build and it saved her roughly a day a week. After two months she caught a recurring seal defect five weeks earlier than the old process would have. She didn’t tell anyone how she caught it. She just caught it.

Nobody asked her to build this and nobody approved it. It runs on her own account and she pays for it herself, eighteen euros a month, which she has never expensed, because expensing it would mean explaining it.

Now her manager wants to hear about AI use cases.

Maria has a good one, and she is sitting at her desk with the invitation open, doing the arithmetic that this whole book is about.

Here is what that arithmetic looks like from where she is sitting.

If she raises her hand, the tool becomes visible, and visibility means assessment. IT will want to know which model it uses, where the data goes, and whether an anonymisation script written by a quality engineer on a Sunday is sufficient control for supplier information. The honest answer is probably, but I’m not certain, and in a governance review “probably” is a no.

Legal will want to know whether an unapproved consumer account processing supplier data constitutes a policy breach. It does. She knows it does. She has known it did for eight months, and it is a thing she has tried not to think about too precisely.

And Quality Systems will hear about it. Quality Systems has a programme for exactly this, running for eighteen months now, with a budget she has heard described in terms she found hard to believe, involving two consultancies, and it has not yet shipped anything to her plant. She has sat through three of its update presentations. The roadmap slide is very good.

Maria’s tool works today. Their programme will work eventually. Put those two facts side by side in a meeting and they do not read as how useful. They read as how embarrassing.

If she says nothing, none of that happens. She keeps her day a week, the seal defects keep getting caught early, and everyone carries on assuming Maria is simply good at her job, which she is, though not by the margin they think. The tool stays alive exactly as long as she does. When she moves to another plant next year, and she will, it goes with her, undocumented, and the reports start piling up again, and nobody will ever know that for two years this plant had a capability it has now lost, because the capability was never written down anywhere except in Maria’s head and a folder on Maria’s laptop.

That is the choice. Surface it and probably lose it, or hide it and definitely lose it, later, quietly.

I want to leave her there for now. We will come back to her, and by the end of the book we will know what she should do. At this point neither of us does.

The fleet

The temptation is to read Maria as an anecdote, but she is closer to a census.

In 2025, MIT’s Project NANDA published a study of how generative AI was actually being used inside companies, as opposed to how it was being reported upward. The number that mattered was the gap between two others. Around 40 per cent of the companies studied had bought official large-language-model subscriptions. Around 90 per cent of employees were using personal AI tools for work.

Nine in ten people, using tools their employer has not sanctioned, to do the work their employer is paying for. That is not a rebellious fringe in the engineering department. That is most of the building.

The same study found something more uncomfortable. Roughly 95 per cent of formal corporate GenAI pilots — the funded ones, the ones with a steering committee and a name — produced no measurable impact on profit and loss, while the unsanctioned personal use happening on private accounts at kitchen tables was delivering faster adoption and better returns.

The official fleet is in dry dock and the real one is already at sea, and nobody in the building has a map of it.

That 95 per cent figure gets over-read, including by people who should know better. The comparison is not clean. People using personal AI tools have self-selected for enthusiasm and are mostly doing easy, well-bounded tasks, while the formal pilots are attempting hard integration work under real compliance constraints. It is not a fair fight, and “shadow beats official” is not a conclusion the data can carry.

But a weaker conclusion survives, and it is more than enough for what follows: an enormous amount of real value is being created below the waterline, and the organisation cannot see any of it.

That is the submarine fleet. Maria is one boat.

This is not new. It is only newly cheap.

The instinct, when you first see the shape of this, is to treat it as an AI story. AI is what made it visible to me, and it is certainly what makes it urgent now, but the pattern is older than the technology by about sixty years.

In 1963, Donald Schön published an article in the Harvard Business Review about how radical inventions actually got developed inside corporations. His answer annoyed people. They got developed by champions who worked around the official process, often in secret, frequently in defiance of explicit instruction. He gave the behaviour a name that stuck: bootlegging.

The name has been reinvented roughly once a decade since. Skunkworks. Pet projects. Friday afternoon work. Creative deviance. Shadow IT. Stealth innovation. Now shadow AI. Each generation of managers discovers it, names it, and treats it as a novel symptom of whatever technology has just arrived, which is a bit like discovering rain.

The academic definition is worth having exactly, because every clause of it does work later on. Bootlegging, in the research literature, is a non-formalised, non-declared, bottom-up innovation process undertaken for the benefit of the bootlegger’s own firm. Non-formalised: it isn’t in the plan. Non-declared: nobody upstairs knows. Bottom-up: it starts with the person holding the problem. And for the benefit of the firm, which is the part that surprises people.

Because the natural assumption about hidden work is that hiding implies wrongdoing. Someone is slacking, or empire-building, or running a side business on company time. The research says otherwise, consistently and across decades. When Masoudnia and Szwejczewski interviewed 55 bootleggers across 34 high-technology firms, the dominant motivation they found was not personal gain at all. It was to benefit the organisation. The secrecy was never the point. The secrecy was the workaround.

Maria is not stealing from her employer. She is subsidising her employer at eighteen euros a month, and hiding the receipt.

There is a distinction in this literature that will run through the whole book. Permitted bootlegging is the sanctioned kind: 3M’s fifteen per cent, Google’s twenty per cent, the officially blessed sandbox. Hardcore bootlegging is the other kind, work that continues after somebody with authority has said stop, and researchers have a name for that one too, creative deviance. We will spend real time on it, because the distance between “we allow this” and “we allow this until it becomes inconvenient” turns out to be where most of the damage happens.

The part that should worry you

If the story were simply that employees do unauthorised work and it is mostly fine, this would be a short book with a sensible conclusion about governance.

The complication is that the hidden work is not merely tolerable. It appears to be better.

The most rigorous study we have is by Paola Criscuolo, Ammon Salter and Anne ter Wal, published in Organization Science in 2014. They studied bootlegging inside a large R&D organisation quantitatively, at the individual level, looking at who did it, how much, and what came of it. Their finding was that bootleg effort was associated with achieving high levels of innovative performance. Roughly nine per cent more time spent underground raised the probability of a top innovation rating by about seventeen per cent.

Their explanation for why is the hinge the entire book turns on, so I want to state it plainly. Working underground gave people two things they could not get above the surface. It let them explore uncharted territory, directions nobody would have approved in advance because nobody could have justified them in advance. And it let them delay the assessment of embryonic ideas.

Delay the assessment. Not avoid it — delay it. An idea in its first weeks is fragile, half-wrong and impossible to defend, and it cannot survive a gate review, not because it is bad but because it is young. So the submarine is less a hiding place from judgement than a nursery for things not yet ready to face it.

Which sets up the cruelty at the centre of the whole thing. The process that would kill the idea early is the same process that would fund it later. You cannot get to the version worth approving without a period in which it is not approvable.

Maria’s tool was garbage for the first three weekends. If she had presented it in month one it would have been rejected, and rejected correctly. By month eight it works. There was no path from the first version to the second that ran through a meeting.

There is also a sting in the Criscuolo findings that we will come back to in Part II. The bootlegging advantage did not hold universally. When the organisation went through periods of formalising its R&D process, with more structure, more gates and more discipline, the people who increased their underground work became less likely to innovate rather than more. The submarine only works in certain waters. Tightening the process did not stop the hiding; it stopped the hiding from working.

Why she can’t just tell someone

Come back to Maria for a moment, because there is a piece of her arithmetic I skipped, and in my experience it is the piece most executives miss entirely.

The obvious question is why she doesn’t simply raise it. Her manager is reasonable. The company says it wants exactly this. There is a slide about it.

The answer is that Maria is not weighing whether her tool is good. She is weighing whether it will be seen, and those are different problems with different mechanics.

Large organisations run on a scarce resource that nobody puts in the budget, which is attention. William Ocasio made this the centre of one of the most cited arguments in strategic management: firms behave the way they do because of how they channel and distribute the attention of their decision-makers. What gets acted on is what gets noticed. Merit does the filtering later, if at all. Notice comes first.

So how does attention get allocated? Cyril Bouquet and Julian Birkinshaw studied 283 subsidiaries of multinational companies to find out how units capture the attention of headquarters, and their answer had two parts. The first they called weight, meaning your structural position: how big you are, how significant your market, how much of the company’s mass sits behind you. Weight pulls attention automatically and never has to ask for it. The second they called voice, meaning initiative-taking, profile-building, the active work of being noticed. Units without weight have to use voice, and it does work, but it costs something and it can be spent.

The property that makes this bite is that attention is zero-sum. Headquarters cannot attend to everything, so every increase in attention paid to one unit is a decrease in the attention available to everything else.

Now look at the Quality Systems programme through that lens. Eighteen months, two consultancies, a budget large enough to be quoted in disbelief. It is not competing with Maria for attention, it is the attention. It has mass. It does not need to be good, or fast, or shipped, to pull the room toward it. It only needs to be big, and it is.

Maria’s tool has no weight at all. One person, six hours, eighteen euros a month.

So 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. Nobody files a decision. There is nothing to appeal. The idea doesn’t die, it just never registers.

This is the first of three scarce resources we will find bent in the same direction. Attention is one. Capital is another, and in a manufacturing company the capital story is brutal, because the decisions are one-way doors and the machinery guarding them was built for cathedral-sized bets rather than for six hours on a Sunday. Careers are the third, because the people promoted to allocate the first two are, systematically and measurably, the people most drawn to the big visible bet.

Three currencies, three separate mechanisms, three bodies of research that have never really been introduced to one another, and all of them pushing the same way.

That convergence is the argument of this book. Each of the three currencies is well documented on its own and I will cite the research as we go. What I think nobody has done is set them side by side and notice they all push the same way, and that connection is mine to make and mine to defend. You will not have to take it on faith. You will watch all three act on Maria long before the book asks you to agree with me about anything.

A curtain in a factory

I want to plant one thing before we go further, because you will need it much later and it is better if it is already sitting in your mind.

In 2012 a researcher named Ethan Bernstein published a study conducted inside the world’s second-largest mobile phone factory. He was not looking for hidden innovation. He was looking at transparency, the idea, popular then and popular now, that the more managers can see of the work, the better the work gets.

What he found was that the workers on those lines had developed better methods than the official ones — faster and more reliable — and that they hid them. When managers walked the floor the line reverted to the procedure in the manual, and the improvements went back underwater until the coast was clear.

The suggestion boxes on every line were empty.

Then Bernstein did something almost comic in its simplicity. He hung a curtain.

I will leave it there. We will come back to what happened, and it will matter more once you have seen what the rest of the machinery does.

What this book is going to do

Three questions, in order.

Is this permanent? If hidden innovation is a quirk of a particular era or a particular management fashion, then it is a problem you can wait out. If it recurs identically from Schön’s 1963 laboratories to Maria’s kitchen table, then it is wiring, and you have to design for it. That is the rest of Part I, and I think the answer will surprise you, because the age of this research usually gets treated as its weakness. In my opinion it is the opposite. A finding that survives sixty years, four technology revolutions and every management fad in between is not dated. It is load-bearing.

Why does the system do this? Not out of malice and not out of stupidity, which is the disappointing part. Parts II and III show a machine that suppresses its own best small ideas while behaving, at every individual step, entirely reasonably. The capital process is right to fear one-way doors. The gate is right to demand evidence. The executive is right to prioritise. Each mechanism is defensible on its own and together they are catastrophic. And then the harder question, the one I think should have been asked decades ago: if all of this is so well documented, why has nobody fixed it? The answer to that is the darkest thing in this book.

What do you do? That is Part IV, from both seats — Maria’s, and the one across the table from her. I will tell you now that the answer is not “surface everything” and it is not “let a thousand submarines sail.” It is stranger than that, and it runs against the instinct of every executive I have ever watched encounter this problem for the first time. The instinct is to go and find the fleet. That instinct is the single most expensive mistake available here, and I will spend a chapter showing you exactly what it costs.

Where the evidence stops

This chapter rests on four studies and they are not equally strong, so it is worth saying which is which before the rest of the book starts leaning on them.

Schön’s 1963 article is an argument rather than a dataset. Its authority comes from how often the finding has been rediscovered since, by people who were not reading Schön and mostly did not know he existed, which is a different kind of evidence and in some ways a better one, but it is not the kind you can put a confidence interval on. Masoudnia and Szwejczewski interviewed fifty-five bootleggers in thirty-four firms and asked them why they did it. All thirty-four firms were high-technology, and a self-reported motive is the most flattering number a person ever produces about themselves. Criscuolo, Salter and ter Wal is the rigorous one, quantitative and at the individual level, and it is one organisation. The NANDA study is a year old and describes a technology that changes faster than anyone can finish studying it, and I have already said what its headline figure cannot carry.

None of them is the finding. The finding is that four different methods, in four different decades, in industries with nothing much in common, all point the same way, and that after sixty years of people looking, nothing of comparable weight points the other way. That is what the chapter is claiming, and it is a weaker claim than any single one of those studies would let me make on its own.

The place it could break is the fleet itself. I have asserted that the submerged work is large — most of the building, not a fringe — and the only direct measurement of that in this chapter is a survey of self-reported tool use, which counts people who use a chatbot for work and not people who have built something.

So here is the study I would like somebody to run. Go into one organisation of any size, with an amnesty in writing, and count what people have actually built on their own accounts: how many things, how old, how many other people use them, and how many of the builders’ own managers already knew. If the answer is that there are a handful, that they are trivial, and that the managers mostly knew, then there is no fleet and this book is about a rounding error. My expectation is that the last number is the one that will surprise people.

Maria illustrates the mechanism. The research carries it. Those are different jobs and I have not confused them.

Next: Chapter 2 — why almost all the research behind this chapter is old, and why that turns out to be the strongest thing about it rather than the weakest.