The Missing Half of the Ford Lesson

Olli Takala wrote a piece this week that deserves to be read twice. His argument, drawn from Ford’s decision to replace quality inspection with AI and cameras — and then hire the experienced engineers back — is one of the clearest statements I’ve seen of a truth this industry keeps relearning: expert judgment is not a set of tasks you can isolate and automate. It is the ability to decide when information is incomplete and the cost of error is high.

He’s right. And I want to build on it, because his article names the problem with unusual precision but stops at the threshold of the answer. The question he leaves open is the one I spend my days on: if tacit knowledge is strategic capital, how do you actually transfer it into a system — without removing the human whose judgment made it valuable in the first place?

Where Ford actually went wrong

Ford’s mistake was not using AI. It was the assumption underneath it — that a camera plus a model could replace an inspector, rather than extend one. When the system couldn’t interpret quality the way an experienced engineer could, the tacit knowledge it was supposed to capture wasn’t in the system at all. It had walked out the door with the people who were let go.

Takala’s framing is exactly right: that experience never sat on the balance sheet, but it was the very thing the AI’s conclusions depended on. Lose the experts too early and you don’t just lose performance — you lose the ability to build competitive AI at all, because you’ve thrown away the training signal.

The part that’s now possible

Here is what has changed, and why I think Ford’s lesson points forward rather than back. We no longer have to choose between “human does it” and “machine replaces human.” There is a third design — a system that works alongside the expert, learns from how they reason, and keeps the decision in their hands.

This is the principle we build on, and it has a name: present, don’t decide. An agent that reasons through a fault with the technician — laying out evidence, weighing possibilities, showing its work — but never taking the decision, because the person accountable for the machine has to own the call. The human stays exactly where Takala says they must: at the point of highest uncertainty, where the cost of error is greatest.

But working alongside the expert is only half of it. The deeper move is what happens to the reasoning afterward. Every fault worked through this way — the symptoms, the reasoning that held up, the resolution that proved out — becomes part of the system’s accumulated case memory. Not a document nobody reads. A living record the agent reasons over the next time a similar fault appears.

That is the mechanism Takala gestures toward when he writes that the most valuable expert may be the one whose way of thinking can be transferred — to other people, and gradually to AI. It doesn’t happen by ripping the expert out and hoping the model absorbed enough. It happens by putting the expert and the system in the same loop, so the reasoning is captured as it’s used, and the expert’s judgment compounds instead of evaporating.

Why the discipline matters more than the capability

There is a catch, and it’s the reason most “AI captures your expertise” promises fail. A system that learns from its own sessions is only as good as what it records. One confident, wrong conclusion — absorbed into memory and reasoned from later — quietly corrupts every future case that draws on it. Ford’s inspectors would never have signed off on a bad call with false confidence. A careless AI will.

So the memory has to be built from earned conclusions, not assumed ones: grounded in the real asset, honest about what it cannot verify, willing to say “I don’t know — let’s look at the machine” instead of guessing. That discipline is what turns accumulated sessions into trustworthy institutional knowledge rather than a growing pile of plausible mistakes. It is the difference between a system that preserves expertise and one that dilutes it.

The shift Takala is pointing at

His closing line is the one I’d underline: competitive advantage won’t come from who has the most automation, but from who best combines automation and human judgment in the situations where neither is enough alone. That is precisely the design goal — not an AI that replaces the expert, but one that multiplies them, and keeps their judgment alive inside the system after they’ve gone home for the night or retired for good.

Ford’s reversal isn’t proof that AI failed. It’s proof that we were asking it to do the wrong job. The right job — the one that’s finally buildable — is to keep the human at the center of the hard decision, and to make sure that every time they make it well, the organization gets a little smarter and never has to learn it from scratch again.

That’s the path forward from what Ford discovered. And with the right tools, it’s no longer a whiteboard idea.