Ford put AI into its vehicle quality work. The results fell short. It then spent roughly three years hiring 350 veteran engineers back, some of them former employees and some from its suppliers.
This is a June 2026 story rather than a new one, and it is worth being precise about the date because the ending has already happened. After the rehiring, Ford took the top mass-market place in J.D. Power’s 2026 US Initial Quality Study, its first in sixteen years.
So it worked. What is useful is what it cost, and why it was needed at all.
What Ford’s own explanation was
Charles Poon, Ford’s vice president of vehicle hardware engineering, put it plainly:
“Mistakenly, we thought that by just introducing artificial intelligence and ingesting the design requirements that we had, that would produce a high-quality product.”
Read that twice, because the mistake is more specific than it first appears.
Ford had the design requirements. It was not missing documentation. It ingested them, which is to say the information a new engineer would be given on their first day was all present and machine readable.
What was missing was everything the experienced engineers knew that had never been written into a requirement. Which tolerances actually matter and which are inherited from a decision made in 1998. Which supplier’s parts arrive slightly out of spec in winter. What a particular noise means.
None of that was in the documents, because none of it had ever needed to be. It lived in people, and it was retrievable at any time by asking one of them.
Then the people left.
The failure is a sequencing failure
It would be easy to read this as a story about AI not being good enough. That reading is available and it is wrong, or at least it is the less useful half.
The judgement was never captured. It was not captured before the AI arrived either, and nobody noticed, because the people holding it were still in the building and answering questions.
Automation did not destroy that knowledge. It removed the reason to keep the people who held it, and the knowledge left with them, quietly, over a period in which every individual departure looked routine.
That is the transferable risk, and it applies at a scale far below Ford’s. Every team has a handful of people who know why things are the way they are. Very little of that is written down. Most of it has never been asked for, because it has always been available.
The measurement problem this creates
Here is where it becomes a decision you might actually be making this quarter.
If you are buying AI training for your team, you will be sold on completion. Places filled, modules finished, satisfaction scores, a certificate. Those numbers are easy to produce and easy to report, and every provider has them.
They measure attendance.
The harder question is what changes in the work afterwards, including how people judge the output they are given. That is the number nobody offers because it is genuinely difficult to produce, and its absence is not usually deliberate.
There is a second confusion worth separating out. Speed and reliability are not the same measure, and AI makes it unusually easy to mistake one for the other. Work arrives faster. Whether it is right takes longer to establish than whether it is quick, so the fast signal lands first and the accurate signal lands later or never.
A team can get measurably faster and quietly less reliable, and every dashboard will show the first thing.
What to write down before you buy anything
Before the training is booked, write one paragraph describing what will be different in the work three months afterwards.
Not what people will know. What will be different in the output: fewer of a specific kind of error, a decision that stops being escalated, a report that stops needing a second pass.
If you cannot write that paragraph, you are not ready to buy the training, and the problem is not the training. It is that nobody has decided what it is for.
Then a second paragraph, which is the Ford lesson rather than the training lesson. Name the two or three people on your team whose judgement you would struggle to replace, and write down one thing each of them knows that is not documented anywhere.
That exercise takes twenty minutes and it is uncomfortable, because the list is always shorter than you would like and the items on it are always more important than you expected.
Who is telling you this
This framing reached me through a contributed interview with Felicia Tan, Director of Tribe Academy, a Singapore AI training provider. Tribe Academy sells AI training, which is the remedy the piece recommends, and that should be on the table.
Her argument stands or falls on its own, and I think the useful part of it survives: training judged on completion rates measures the wrong thing, and domain judgement is built by doing the work for years rather than by attending anything.
I will declare my own position too, since it points the same way. We sell AI training. An article arguing that training should be measured on changed work rather than attendance is an article I would like to be true.
That is exactly why the Ford facts are doing the work in this post rather than the argument. Ford is a named company that said in public that it got this wrong, explained why, and paid for three years to undo it. You can check every part of that without taking anyone’s word for it, mine included.
The order that matters
Ford’s mistake was not buying the technology. It was assuming the requirements were the knowledge.
Write down what your experienced people know before you need it. That costs an afternoon now, and it is the only version of this that does not involve hiring anybody back.