Let’s talk about the job that came back.
Ford, earlier this year. Quality inspection. The company had put automated systems onto work that used to belong to people, and according to Bloomberg’s reporting in June, it started hiring experienced engineers back — hundreds of them — to deal with quality issues the automated systems couldn’t catch.
Charles Poon, Ford’s VP of vehicle hardware engineering, said the quiet part to the BBC: AI is a fantastic tool, but it’s only as good as the information you train it on. Which is a polite way of saying the model was confident and the cars were wrong.
Now here’s the part I actually want to sit in. When those roles went, there was a slide. There’s always a slide. Efficiency, simplification, the future of the function. Somebody presented it to a board and the stock did a small pleasant thing.
When the roles came back, there was no slide. There was a paragraph in a wire story that most people scrolled past, and a line item in a hiring plan.
That asymmetry is the whole story.
Picture the version of this that happened at a smaller company, because it happened ten thousand times this year and almost none of it got reported. A firm cuts eleven junior staff in a support or ops or finance function. The announcement is internal, the savings are real, the CFO gets a win. Fourteen months later a contract recruiter is calling the same names back, offering more money than they were on, because the tickets the juniors used to close are now sitting in a queue that two senior people are working through at night.
And the seniors are tired. And the queue is a month deep. And the number that would have told you this was coming was never on any dashboard, because nobody built a dashboard for it.
Let me give you the actual scale, because the vibes on this are running way ahead of the data in both directions.
Challenger, Gray & Christmas — the outplacement firm that has tracked announced US job cuts for decades — reported at the start of October that AI had been cited in 120,136 announced job cuts through September. That’s roughly 21% of all cuts this year, and it’s the single leading stated reason. Technology alone announced 165,925 cuts, about 29% of the national total.
Now the part that gets left out of the panic thread: total announced cuts through September were 573,195 — down 39% from the same period in 2025. Layoffs are not surging. They’re falling. What’s surging is the share of them that get attributed to AI.
That matters, because “AI” is a very convenient word on a WARN notice. It’s a growth story for the board, a defensible reason for a cut that was really a demand problem, and it costs nothing to say. Some of those 120,000 were genuinely displaced by a model. Some were displaced by interest rates and got a better excuse.
But the reversals are real, and they’re specific. Commonwealth Bank of Australia replaced more than forty customer service staff with a voice bot, watched call volumes go up instead of down, and reversed the cuts. IBM put AI on HR requests, got to about 94% handled automatically, and found the remaining 6% was where the actual job lived — ethical dilemmas, edge cases, the stuff that requires someone with a spine and a name. IBM is now tripling US entry-level hiring.
Robert Half’s data, given to CNBC: 32% of US hiring managers eliminated a role primarily because of AI, then rehired for the same or a similar position. Orgvue found 39% of business leaders made redundancies over AI deployment, and 55% of that group now admit the decision was wrong.
Fifty-five percent. That’s not a rounding error or a learning curve. That’s a strategy with a coin-flip hit rate, sold as inevitable.
Here’s what I think people are missing, and it’s not the failure rate. Failures are cheap to admit when they’re framed as iteration.
The thing that got cut wasn’t a task. It was a training function.
Junior work has always been mispriced on purpose. The salary was never the real cost of a junior — the supervision was. A senior spends two hours a week correcting their work, explaining why the reconciliation is off, why the client reacted that way, why the clause matters. That supervision is expensive, and it produces almost nothing this quarter.
What it produces is a person who in four years can be handed something ambiguous and not panic.
So a junior role is a subsidy. The firm pays below the market value of the output, and in exchange the person accumulates reps under supervision. Both sides get something. That’s the deal.
Automate the task in the middle of that deal and you haven’t removed a cost. You’ve removed the mechanism that converts effort into judgment — and the salary line disappears while the conversion stops. Nobody notices for about three years, because the seniors who already have the judgment are still there, absorbing the work quietly and not logging it.
IBM’s CHRO, Nickle LaMoreaux, said it plainly at a summit in New York: if we don’t continue to invest in entry-level hires, what happens in three to five years? There’s no pipeline. The well simply dries up.
The well drying up is the second-order effect nobody prices. Cut juniors across an industry and you don’t just save money — you send a signal to nineteen-year-olds that this career doesn’t want them. They go and do something else. Supply shrinks. Then when every firm discovers at the same time that it needs people who can actually judge the output, they’re all bidding for the same thin cohort, at a higher price, with a worse selection.
You borrowed against your own future capability, at a rate nobody wrote down, from a lender who didn’t ask for collateral because they knew you’d be back.
And the reversal stays quiet for a reason that has nothing to do with humility. Announcing the rehire is announcing that the original announcement was theatre. That’s a narrative problem before it’s an accounting problem, and narrative problems get handled with a paragraph in a wire story.
So what do you actually do with this, if you run something.
Put one line in every automation business case: who is learning this work now. Not who does it. Who is learning it. If the answer is nobody, you haven’t automated a task, you’ve closed an entry point — and you need to decide, deliberately, whether you’re going to pay for the replacement reps somewhere else.
Three things that follow from that.
Name the training function per role, in writing, before the cut. If the role exists partly to produce a future senior, say so and price it. A business case that treats a junior as pure cost will always win the spreadsheet, because the thing it destroys has no cell.
Budget senior hours as a cost line, not as goodwill. If two hours a week of correction is what makes the next senior, then it’s a line item with a number on it, and it competes with everything else for money — which is exactly the accountability it deserves. Unbudgeted mentorship is the first thing to go when a quarter gets tight, and it goes silently.
And keep a rep log. Not tasks completed. Decisions, with the person’s name on them, and what happened. That log is what lets you promote on evidence instead of vibes, and it’s the only artifact that survives a model change — which, if you’ve been paying attention this month, is not a hypothetical.
If you’re the one on the other side of this, the correction is real but it arrives late and on somebody else’s schedule. The people who get called back are the ones whose work is legible — who can point at three decisions they made and what it cost when they were wrong. Collect those. Nobody is going to collect them for you.
And the principle underneath all of it, because there is one: apprenticeship is a consent arrangement. Cheap labour now, in exchange for reps and a shot at judgment later. When a firm takes the cheap labour, skips the reps, and books the saving, the deal was never consent. It was extraction with a start date.
A company that stops teaching has decided to rent its future judgment from a market it helped thin out. It’ll get away with it for a while. Then it’ll pay retail, late, for someone else’s judgment — and call that a hiring plan.