Why Now

Five years before Maria built anything, she sat in a room and watched a man from a consultancy explain her own plant to her.

He was good, and that was the uncomfortable part. He had the diagrams, the process maps, the failure taxonomies. He knew which supplier codes mapped to which line, and he knew it because he had spent four months building the model that mapped them. He was thirty-one and he had been in the building since March.

Maria had been there nine years.

She knew things he didn’t. She knew the seal supplier’s quality dipped every August, that the third shift logged deviations differently to the first, that the downtime on line two had been logged against the wrong press since a system migration nobody ever fully unpicked. None of that was in the deck, though, because the deck was the official version. And when the programme ended and the man went back to wherever men like that go, the deck stayed, and the model stayed, and the understanding of how the model actually worked left the building in his laptop bag.

Two years later somebody asked Maria why the downtime on line two looked strange.

She said she’d raise a ticket.

What it feels like to work in a hollowed-out company

I want to spend some time on that feeling, because it is the part the research handles worst and the part every engineer in a large company recognises instantly.

It is not incompetence. Maria is not incompetent, the plant is not badly run, and the consultancies were not con artists. Every individual decision in the sequence was defensible, which is a sentence you should get used to, because it is the shape of every mechanism in this book.

What it is, I think, is something more specific and more corrosive: being accountable for outcomes you no longer have the means to produce.

You are the quality engineer. Quality is your name on the org chart. But the system that measures quality was built by a vendor, is maintained under a support contract, and can only be changed by raising a request that goes into a queue managed by a team in another country who will ask you to justify the business impact of the change in a form with eleven fields. You know exactly what the change should be. You have known for a year. The knowledge in your head has no route into the system that runs your job.

So you develop the habits. You keep the spreadsheet. You learn which of the three official reports is the one that is actually right. You build a private mental map of where the data lies to you, and you carry that map around as tacit knowledge, and you never write it down, because writing it down would take a week and nobody asked for it and there is no field for it in the eleven-field form.

Without ever deciding to, you become a person who works around their own systems.

And here is what that does over nine years, which is the thing I want you to sit with before we get to the evidence. It teaches you that building things is not what people like you do. Building is what the vendor does. What people like you do is specify, request, escalate and wait. You become the customer of your own operation.

That belief is the ground this book is planted in. It is not a skills gap. It is a category belief about who gets to make things.

How the knowledge left

The path by which companies arrived here was not stupid either. At every step it was financially sensible.

Through the nineties and two-thousands, large organisations worked out that a great deal of what their internal IT departments did was undifferentiated. Running a mail server was not a competitive advantage. Maintaining a data centre was not a competitive advantage. And the people who did it were expensive, hard to hire and hard to retain.

So it went out. First infrastructure, then applications, then whole functions. The SaaS wave completed the logic, because why maintain a system at all when you can subscribe to one? Each decision reduced cost, transferred risk, and freed the organisation to focus on its actual business. Each one was correct on its own terms.

What went out with the work, though, was the learning.

This is the part that never shows up on the business case, because it is not a line item and it does not arrive for years. When a vendor runs your system, the vendor’s engineers accumulate the experience of running it. They learn what breaks, why it breaks, what the workarounds are, how the thing actually behaves under load on a Tuesday in November. That learning compounds, on their side of the contract.

Meanwhile your side accumulates a different competence: procurement, contract management, vendor governance. Real skills, and useful ones, and none of them the skill of building the thing.

Ten years of that and you have an organisation full of capable, senior, well-paid people who are extremely good at commissioning work and have not personally made anything in a decade.

Then somebody stands up at a town hall and says the company must move fast on AI.

What the research actually establishes

The academic version of this has been around since 1992, which by now you will recognise as the book’s usual pattern.

Richard Bettis, Stephen Bradley and Gary Hamel published “Outsourcing and Industrial Decline” in the Academy of Management Executive. Their argument was that outsourcing decisions, individually rational, could cumulatively hollow out a firm’s ability to compete, not through any single bad choice but through the steady transfer of capability to suppliers who then owned it.

The finding has been sharpened since. Masaaki Kotabe and Michael Mol studied manufacturing businesses and found the relationship is not a straight line at all but a curve with a top to it. Outsource nothing and you carry costs you did not need to carry and lock yourself out of things other people simply do better. Outsource more and performance improves, for a while. Past a certain point it turns and comes back down, and it comes down for a different reason than it went up — not because any particular decision was wrong, but because the firm has cumulatively handed away things it turned out to need. Somewhere in the middle there is a peak, and the curve is worse on both sides of it.

Two properties of that shape matter more than the shape itself. The first is that the descent is built out of the same decisions as the climb. Each one resembles the last, each arrives with a business case that is honestly argued and probably correct in isolation, and none of them is the mistake. The mistake is the sum, and nobody’s job is the sum. The second is that the peak is not marked. Nothing announces the crossing. Performance does not fall the following quarter, no committee is informed, and the plant that has just outsourced the last person who knew how to modify a line does not feel any different on the Monday morning. You discover where the top was by standing some way down the far side of it, years later, when somebody asks the organisation to do a thing it used to be able to do and the answer comes back that this is something we buy.

The most pointed result for our purposes is Carmen Weigelt’s, in the Strategic Management Journal. Outsourcing buys a firm access to a new technology. It does not buy the ability to integrate that technology with the way the firm already works, or to build on it afterwards — and Weigelt’s term for that ability, integrative capability, is internal by definition. Access can be purchased. Absorption has to be grown.

Her setting is banks adopting the internet, not anything with a factory in it, and I would rather say that than have you find it out later. What she found there was that heavier reliance on outsourcing carried a performance cost, and — this is the part worth carrying out of the paper — that the cost was smaller for firms that already had experience with a related technology. Prior internal capability did not become unnecessary once the work went outside. It became the thing that decided whether going outside hurt.

Read that twice, because it inverts the intuition the whole outsourcing era ran on. You do not save the capability by buying the service. The capability is what determines whether buying the service works at all, and the conditions under which it is grown are precisely the ones you have just handed away.

Sean Handley documented what follows as the perilous effects of capability loss: firms that lose internal competence manage their outsourcing relationships worse, which produces worse outcomes, which typically produces more outsourcing. The loop closes.

Underneath all of it sits the resource-based view of the firm, which holds that competitive advantage comes from resources that are valuable, rare and hard to imitate. Capability is exactly such a resource, and it is not a thing you own like a building. It is a thing you maintain by using.

The town hall

The research I have just cited stops at the edge of the company. Firm capability, firm performance, competence measured in numbers that hold at the scale of the whole organisation. It never comes down the stairs, onto the floor, into the particular Tuesday when the plant manager stands up in front of two hundred people and says the words everyone has been waiting for, which are that we need to move fast on AI.

Watch Maria in that room.

She is not excited. She notices that she is not excited, and it bothers her, because five years ago she would have been. Around her some people are nodding the way you nod at a slide. A few are typing. Nobody looks the way the plant manager plainly hopes they will look.

What Maria feels, sitting there, is closer to being asked to run a race after a decade of being driven everywhere. She has spent nine years becoming excellent at a particular thing, which is knowing this plant, its suppliers, its lies, its August dip, and almost none of that excellence is the kind you can act on directly anymore. When something needs building, someone is hired to build it. When something needs changing, she raises a request. She has become very good at describing what should happen and completely out of practice at making it happen. And now the manager wants speed, and boldness, and initiative, from a room full of people the company spent ten years teaching to wait.

She could tell him why the enthusiasm isn’t there. She won’t. But she knows.

That much I will state without a footnote, because you can walk into any hollowed-out company and find it in the third row of the town hall: the capable, senior person who has quietly concluded that building things is not what people like them do anymore. It is not rare and it is not subtle once you know to look for it. In my experience it is the single most common expression in the room when the AI announcement lands.

What I cannot tell you, and what Maria herself couldn’t tell you if you stopped her on the way out, is exactly why the excitement drained away. Whether it is the habit of specifying instead of making. The slow discouragement of a hundred tickets raised into a queue. An honest reading of skills that have gone quietly rusty. Or something nearer to identity, some sense of what her job is for that shifted so gradually she never caught it moving. Probably all of them together, braided, pulling harder on some people than on others.

I have a guess, and I will give it as a guess. But the guess is not the point. The point is that the dread is real and observable and everywhere, and that it arrives at the exact moment the company most needs its opposite. You do not need to resolve the why to see that the what is going to be a problem.

Because here is what Maria does after the town hall. She goes back to her desk. She does not propose anything. She does not raise her hand about the tool she has already built, the one that works, the one that is exactly what the plant manager stood up and asked for.

She opens the browser tab, and she gets back to work.

The second thing that changed

Now the other half, and this one moves fast.

For most of the sixty years we traced in Chapter 2, running a submarine was expensive. Schön’s champions in 1963 needed political capital and a laboratory. Lockheed’s Skunk Works needed a building, a budget and an executive who would look away. The engineer with the unauthorised spreadsheet in 1995 needed weeks of evenings and a real command of the tooling.

The cost of hiding was high, and the cost of building while hidden was higher, which is why the fleet was historically small. It took a certain kind of person: stubborn, senior enough to survive being caught, technical enough to do it alone.

That constraint is gone.

The MIT numbers from Chapter 1 are the measurement of its absence, with around ninety per cent of employees using personal AI tools for work against forty per cent of companies with sanctioned ones. That ratio is not really a story about enthusiasm for a new technology. It is a story about a barrier that fell.

Maria’s tool cost six hours and eighteen euros a month. In 1995 the same capability — parse unstructured reports, classify failure modes, cluster by supplier, flag recurrences — would have been a project. A real one, with a budget, a spec and a vendor. It would have gone through the gate we examine in the next chapter, and it would have failed there, and we would never have heard about it.

Now it is a weekend.

Notice what specifically changed, because it is not quite what people say. AI did not make people want to build things they were not allowed to build. They always wanted that, and Chapter 2 is sixty years of them doing it. What AI changed is the economics of deviance. It took an activity that required real resources and made it available to anyone with a browser and an afternoon. The behaviour did not change. The price did.

There is a second property that is easy to miss, and the research on covert AI use names it directly. When Waters-Lynch and colleagues studied hidden generative-AI use inside firms, they identified concealability as a condition shaping whether hidden work happens at all, meaning how easily the activity can be kept out of sight. A skunkworks needs a building, and you cannot hide a building. Maria’s workflow lives in a browser tab on a personal account and produces no artefact anyone else can see. It is not merely cheap to build, it is nearly free to conceal.

Cheap to build, cheap to hide. Both barriers fell at once.

The convergence

Put the two halves together, because separately they are interesting and together they are the reason this book exists.

The capability to build things independently arrived, cheap and fast and available to anyone, at precisely the historical moment when large organisations had spent thirty years training their people to believe that building things was somebody else’s job.

Those two facts are in tension and the tension has to resolve somewhere. What it produces is not a workforce that suddenly starts proposing ambitious projects, because proposing is the part they have learned doesn’t work. What it produces is a workforce that quietly solves its own problems and doesn’t mention it, because the tools are now within reach of the individual while the institutional route remains exactly as closed as it ever was.

That is the fleet, and that is why it is so much larger than it has ever been, and why it grew so suddenly that most executives have not registered it yet.

There is a bleaker way to state it and I think it is the accurate one. The organisation hollowed out its ability to build. Its people are now rebuilding that ability privately, on their own accounts, at their own expense, in their own time, and hiding it, because the organisation’s formal machinery would treat it as a policy violation rather than as the restoration of a capability the organisation itself gave away.

Maria is not going around her employer. She is doing, at eighteen euros a month, the thing her employer stopped being able to do.

The question this leaves open

I want to close Part I by planting something and then leaving it alone for two hundred pages.

We have established that the pattern is old, structural, and survives every technology. We have established that it just got radically cheaper, and that the ground was unusually well prepared for it. What we have not established, and what I do not think anyone can yet establish, is whether AI is simply the sixth wave of a very old thing or the first one that changes what the thing is.

The case for the sixth wave is Chapter 2 in its entirety. Mainframes, PCs, client-server, the web, mobile. Each arrived, each was hailed as transformative, each got absorbed, and the pattern of hidden bottom-up work continued underneath, unbothered. On that reading, shadow AI is what shadow IT was, at greater scale, and it will be governed and absorbed the same way.

The case for different this time is harder to dismiss than usual. Previous technologies changed what employees could use. This one changes what employees can make, alone, without permission, at a cost that no longer requires approval. That is not a difference of degree in the tooling. It is a difference in who is capable of production.

I have a view. I am keeping it until Chapter 13, because you should watch the machinery work before you decide whether anything can break it.

What comes next is that machinery. Part II is the anatomy of how a large organisation, acting reasonably at every step, buries the best small ideas it has. It begins where every idea has to go, and where a surprising number of them stop.

Where the evidence stops

This chapter makes a historical argument, and historical arguments are the easiest kind to make badly, so it is worth marking the joins.

The hollowing-out evidence is narrow where the claim is wide. Kotabe and Mol’s curve comes from manufacturing businesses, Weigelt’s result from banks adopting the internet, and Bettis, Bradley and Hamel’s warning was an argument made in 1992 rather than a measurement taken since. I am reading them together as one story about capability leaving firms over three decades. No single study says that, and nobody has run the study that would.

I am claiming a sequence, and sequences are where hindsight does its worst work. The order I have given — outsource, lose the people who knew, discover the loss when something needs changing — is the order the sources support individually. It is not an order anybody has traced through one company over thirty years, and the reason it reads so cleanly is partly that I have arranged it. Every account of a slow decline looks inevitable once you know it happened.

The convergence is my argument, not a finding. That the hollowing-out and the arrival of cheap capability meet now, in this decade, at this cost point, is the load-bearing claim of the chapter and there is no study behind it at all. It is two well-evidenced trends and a line drawn between them by me. If it is wrong, the most likely way it is wrong is that the capability was never as thin as the outsourcing figures make it look, because the people who knew things stayed and simply stopped being asked.

And this is the chapter with the least direct evidence in Part I. Chapters One and Two rest on mechanisms with controlled work behind them. This one rests on economic history, which is real but is not an experiment, and on a reading of the present that will look either obvious or foolish in ten years with nothing in between.

Next: Part II. The machinery — beginning with the gate.