The Thing You're Best At Is the Thing You Can Least Prove

Ford went looking for 350 engineers.

Not fresh graduates. Experienced ones, the kind who've been around vehicles long enough to know something is wrong before they can even tell you why. Over three years the company hired, rehired or promoted 350 of them onto vehicle quality work, after an approach built around AI fell short of what it was supposed to deliver. Charles Poon, Ford's vice president of vehicle hardware engineering, put it about as plainly as a serving executive ever puts anything. The company, he said, had "mistakenly believed it could swap in AI and still produce a high-quality product."

Swap? You only swap two things once you've decided they're the same thing or equivalent.

So why did it fail? Experienced people had left before anybody learned what they knew. Because none of it had been written down, none of it was in the training data either, so the AI checked the designs without knowing that a whole class of fault existed. The system wasn't stupid. It had been taught everything the company had written down, and what the company had written down turned out to be a lot less than the company thought.

Ford went on to take the top mainstream-brand position in J.D. Power's 2026 initial quality study, at 152 problems per 100 vehicles, its first in sixteen years. The fix was people.

Mind you, this is one company, one function, one executive talking to the press after the recovery, which is the safest possible moment to admit anything. I'm not going to build a theory of the economy on Ford's quality department. But I recognise the mechanism from the other side, because I've helped build the documents that cause it.

When you restructure, you don't look at people. You can't. There are too many of them, and the decision has to stand up to a board, multiple lawyers, a regulator, and so you build a list instead. Every role becomes the things that role does. Tasks, outputs, a scope, a band. That list is genuinely the fairest instrument available, because what's the alternative? Somebody's opinion about who seems important? Obviously not.

But here's the problem with it. The list can only hold what somebody wrote down.

Think about a recipe. The recipe is the part of cooking that can be written down, and it's genuinely useful. You can hand it to anyone. What you can't hand over is the cook: the one who knows this oven runs hot, who can tell from the smell that it needs four more minutes, who has ruined enough batches to recognise ruin before it arrives. Give someone the recipe and they'll produce the dish. Give them the recipe instead of the cook and you find out slowly, over months, which parts were never in it.

Ford wrote down the recipes and let the cooks go.

Every job has that same shape. One half is the written half: the procedures, the deliverables, the parts of the work that survive being typed into a form. The other half lives in the person, and it's the thing you spot on a drawing at four in the afternoon that nobody could have told you to look for, the customer you know will call in a fortnight, the certainty that this supplier's numbers are optimistic, the reason a colleague walks over to your desk instead of opening the process document.

That half has never had a line on the list. Mind you, it's not that somebody decided to leave it out. There's just no way to put it in.

And this is exactly the material that AI is now being priced against.

Look at the arithmetic for a second, you'll see that it's neat and also broken. A model is trained on what has been written down. A job is judged on what has been written down. Both sides of the comparison are made of the same stuff, which is why the swap looks so obvious, and the half that's missing is missing from both sides at once. Nobody in that room is hiding anything. The measurement is careful, documented and defensible. It's just measuring only one half.

Now think about the person sitting on the other side of the table, because this doesn't show up in anybody's cost model. They're told, correctly, that the tasks listed against their role can largely be automated now. They can't argue with the list, because the list is accurate. What they want to say is that the list was never them, that the reason the place works is a hundred things nobody ever asked them to write down. And they can't prove a single one of them. If it could be proved it would have been written down, and if it had been written down it would be in the training data, and if it were in the training data it would already be automated.

The thing you're best at is the thing you can least prove.

So being repriced isn't only a financial event. You're being judged, with real consequences, against a document that never described you, and the case for your own worth is made of material that doesn't count as evidence anywhere. It comes very close to being told you're not who you spent twenty years becoming, by a process that's being scrupulously fair about it.

Ford's story ends well, which is the only reason any of us get to read it. Somebody noticed the quality numbers, somebody connected them to the people who had gone, and the company spent three years and 350 hires buying the missing half back. That's an expensive way to discover that your list was incomplete, and Ford could afford the tuition.

What I can't stop thinking about is how that knowledge used to move. The handover of those things that can't be written down has never been a formal process anywhere I've worked. It happens by proximity. Somebody sits near somebody else for two years and absorbs a way of looking at things that neither of them could write down if you asked them to. It's slow, it's invisible, it costs money now for a return years later, and the seat it happens in is the junior seat, which is the first seat anybody automates.

Ford found out it had stopped, and could still go and buy the people back.

I'm less sure about the places where it's stopping right now, quietly, with nobody counting. Where do you buy it in ten years, when the last people who had it have retired and the only record of what they knew is the part they wrote down?