Ford has 900 AI cameras watching its quality operation, trained to catch defects as parts move through the line. Over the last three years, it hired 350 veteran engineers — many of them former employees, others brought in from suppliers. Inside the company they are called the gray beards.
They were not brought back to replace the cameras. They were brought back to retrain them.
“Mistakenly, we thought that by just introducing artificial intelligence and ingesting the design requirements that we had, that would produce a high-quality product,” Ford’s vice president of vehicle hardware engineering said on a press call last month, in remarks reported by Bloomberg. The company had let its most experienced people go before any of what they knew had been loaded into the systems that replaced them.
That last sentence is the whole story, and it isn’t a story about cars.
The boomerang
This is not one embarrassed carmaker. It is now common enough to have a shape.
Robert Half surveyed nearly two thousand US hiring managers. Of those who had eliminated a role primarily because of AI, 32% have since hired someone back into the same job or one very like it. In finance it was 44%. Orgvue, surveying senior leaders last year, found that 39% had made people redundant following an AI deployment — and of those, 55% judged the decision wrong. Gartner expects that by 2027, half of the companies that cut customer service jobs for AI will be rehiring for similar work, often under new titles.
The bank that counted the wrong thing. Last July, Commonwealth Bank — Australia’s largest — moved to cut 45 customer service roles after installing an AI voice bot, having told the union the bot was taking 2,000 calls a week off the queue. The union checked. Call volumes had gone up. The remaining staff were on overtime and team leaders were being pulled onto the phones to cope. By late August the bank had conceded its assessment was flawed — the roles, it admitted, “were not redundant” — apologised, and reversed the cuts.
Nobody in these stories bought a product that did nothing. The AI did real work; it just didn’t do all of it. Each company measured a genuine capability, drew a straight line from it to a headcount number, and got the arithmetic wrong in the same place.
What the 6% actually is
IBM has been unusually specific about where that line sits. Its internal HR assistant, AskHR, handled more than eleven and a half million interactions in 2024 — leave balances, payroll questions, the small administrative traffic of a very large company — and settled 94% of them without routing to a human. The other 6% went to specialist HR partners. IBM’s chief human resources officer is blunt about what lives in that remainder: the complex matters, and the things she says AI “will never have the judgment, wisdom and experience” to handle.
Almost every automation project produces a version of this number. The mistake is reading it as a staffing ratio.
The 6% is not 6% of the value. It is the appeal against the decision, the customer whose situation doesn’t match any form, the defect that looks fine to a camera and wrong to someone who has seen that supplier’s work for twenty years. It is, reliably, the part of the job where being wrong is expensive. A support queue where 94% is handled instantly and 6% is handled badly is not a 94% success. It is a business with a failure nobody is watching and a smaller bill.
The routine work was never just the cost. It was the training ground — and nobody has ever run an apprenticeship on the exceptions alone.
Where the 6% people come from
Here is the part the spreadsheet cannot see.
Nobody is hired able to handle the 6%. There is no course for it and no way to recruit directly into it. A senior person is simply someone who has worked through several thousand instances of the 94% and can now recognise, quickly and often without being able to explain why, the one case that isn’t. Judgment is compressed repetition. The boring work was the mechanism that produced it.
So when you automate the routine tier, you remove the tasks and the ladder. The effect is invisible for a year or two, because your existing seniors are still there, still running on experience they accumulated under the old regime. It shows up later, when they leave and nobody behind them has done the reps.
The people who watch this closely have started naming it. Martin Fowler, writing up a retreat of senior practitioners this summer, lists five headline findings. One: organisations are “colliding with a real apprenticeship crisis.” Another: in their own field, producing the work is no longer the bottleneck — checking it is. That is the same finding twice. The machine produces; humans decide whether what it produced is right; and deciding is the expensive skill nobody is being trained in.
Ford’s engineers didn’t come back to inspect parts. They came back to teach the system what a defect looks like — which was only possible because they had spent careers looking at parts. The knowledge had to exist in a person first. It always does.
Before you cut
Three questions worth answering while the decision is still reversible.
Find your 6% first. Before approving any automation, get the person who does the work today to describe the cases the tool will not handle. Not the volume — the consequences. If nobody can name them specifically, the pilot has not run long enough to cut anyone.
Ask what the routine work was teaching. For every role you’re reducing, name where its replacement will come from in five years. “We’ll hire senior people” is not an answer; every company saying it is bidding for the same shrinking pool, and the pool shrinks precisely because everyone stopped growing their own.
Redeploy before you make redundant. IBM automated hundreds of HR roles and its overall headcount went up — it moved the money into engineers and salespeople. That is the version that works. The reversals all share one feature: the company treated an efficiency gain as a headcount cut rather than as capacity it now had to spend somewhere better.
The technology in these stories did what it promised. The error was never about AI at all. It was a company looking at the least interesting part of its work, seeing only a cost, and not noticing it was also the nursery.
You can automate the 94% this quarter. The 6% takes fifteen years to grow — and it starts in the work you just deleted.