Nobody Trains Their Replacement for Free

Bill Erwin worked the repair bay at Lansing Grand River, and he did the one thing repair men almost never do. He told the floor what he found. For years afterward I could not recall his name, and I once wrote about him without it.

Repair men are notoriously guarded about what they find: the root cause, and what they did to fix it. I saw it at Buick City and again at Mercedes. Keeping it to themselves kept a steady stream of work coming, gobs of overtime, and, honestly, power over the floor.

So I was braced when we launched the Cadillac CTS at Grand River. It was the relaunch of Cadillac, and the car came down the line riven with defects that needed extensive repair. Nothing went to the shipping area without the certification of men like Bill.

Bill broke the mold. Why, or how, could the line possibly send me this crap? What are they doing out there? He became a sort of taskmaster to the team leaders whose areas the defects came from. At the General Customer Audit meeting, a team leader would carefully offer his solution in front of the whole chain of command, then glance at our assistant plant manager, Ken Knight, a demanding man at that quality meeting. Ken would turn and look at Bill. What Bill said next made or killed that team leader in public. So the team leaders learned to consult Bill first. He was patient, creative and collaborative, and he made the floor better at the very thing that kept his bay busy.

George Sivulka, who runs the AI company Hebbia, has a sentence for why companies cannot get their AI agents to work: nobody trains their replacement for free. The repair men at Buick City would have nodded along. Bill is why I do not believe it is human nature. Hoarding is what people do when what they know is the only thing the system pays them for, and handing it over pays nothing. Change what the knowledge earns its owner and you change whether it moves. The choice in front of every company deploying AI is extract or form. That’s the argument.

I learned what hoarding costs at Buick City. For a new supervisor like me, the drill went like this. I would get called to the end of the line for a repeat defect. Usually it was seat inop, a power seat that would not move, and it failed in many ways. The call usually came from Simon Boag, the assistant plant manager, looking at me like: and you’re a PhD? Then I would go to the repair bay to find out what the repair had been, for a clue to the root cause and the fix back at my own workstations. They had no record of the repair they made. The vehicle’s already shipped. Then they would ping-pong me across shifts. When I sent my team coordinators to get the root cause, the repair men gave them grief: stay in your area and don’t build s**t. Only a few took pity on me and gave me enough to cut the defect off at the source. It was painful.

The machine that loops

Sivulka is the best witness to this failure, because he is standing on the other side of it. Hebbia sells AI to BlackRock, KKR and the Air Force, and in July of this year he published the sharpest autopsy of the agent era anyone has written. Companies deployed millions of AI agents and got almost nothing back, and the reason, he says, isn’t weak models. Maybe one employee in a hundred knows how to give an AI context. Hand an agent to the other ninety-nine and it loops: it calls itself over and over, trying to brute-force a task no human ever put into words. He has a brutal phrase for it: you are spending tokens on spending tokens.

Looping is a machine begging for context that lives in someone’s head and hands and was never formed into speech. The repair bay at Buick City knew the root cause of every defect it fixed, and the line never heard it. That was a loop too.

His own reporting hands you the tell. One firm told UBS that its spending with a single model provider went from about twenty thousand dollars in December to nearly a million in July. Ford hired back some 350 veteran engineers, the graybeards, to work alongside its new systems. Read those together and one fact organizes them: the moment the machines arrived, the market paid a premium for the human judgment it had spent a decade trying to automate away.

So Sivulka is right about the wound. Where he grabs the wrong end is the cure, and he tells on himself in one sentence.

Nobody trains their replacement for free

He drops it as a law of nature. Of course workers hoard their context; who would teach the machine that is coming for them? He reaches back to the medieval guilds keeping their methods secret and treats the history of tacit knowledge as one long act of self-protection.

But look at where he grabbed the guild. The master did teach the apprentice, reliably, for centuries, because his name and his standing were bound up in producing the next journeyman. The secrecy pointed outward, at the market. Inside the workshop, formation was the point. Sivulka saw guilds hoard and missed that guilds formed.

The hoarding is a response to a structure, and the structure has a name. We run business on extraction. In an extraction model, what a worker knows flows one way, out of the worker and into the firm, and nothing flows back. None of it is banked to him; it is absorbed, expensed and credited to no one. The most valuable judgment on the floor appears nowhere on the books.

Now ask that worker to hand it to a machine. He is being rational. Teaching the system is self-liquidation: you surrender the one thing that made you scarce, and the ledger does not even record that you gave it. The Buick City repair men understood that perfectly. Their overtime was the only receipt their knowledge ever earned, and they were not going to tear it up.

What Bill was paid in

In September, in an essay called He Does Not Drink Blood, I wrote about a retired repair man back on the floor at Delta Township on Yazaki’s payroll, and I said I had lost his name