The AI Boomerang: Why the Companies Cutting Jobs Are Also the Ones Rehiring
Across 2025, the story was that AI would replace workers. In 2026, a stranger pattern has emerged: many of the firms that cut jobs are quietly rehiring, and the cost of running AI is turning out to be far higher than the savings it promised. The lesson is not that automation failed. It is that judgment was mispriced, and that using these tools well is a skill few have actually bothered to learn.
Earlier this year, Microsoft began cancelling most of its internal licences for Claude Code, the AI coding tool it had rolled out to thousands of its own engineers only six months earlier. The reason was not that the tool failed. The reason, as Fortune reported, was that it worked too well: employees used it so heavily that the bill became difficult to justify, and the company began moving its engineers towards GitHub Copilot instead. Microsoft was not alone. Uber’s chief technology officer told The Information in April that the firm had burnt through its entire AI budget for 2026 in just four months, driven above all by its engineers’ enthusiasm for the same category of coding tools. Meta and Amazon, meanwhile, have been pushing staff to consume as many AI “tokens” as possible, even as those tokens carry a usage-based price that climbs with every prompt.
This is not the story most of us were told.
For most of the past two years, the dominant narrative was simpler: AI would absorb the work, and the people who did that work would be let go.
Sam Altman, who sits closer to the centre of this shift than almost anyone, told a Federal Reserve audience in 2025 that some areas of the job market would be “just like totally, totally gone.” For a while, the layoffs appeared to prove him right.
Then the same man said almost the opposite. In an interview with CNBC in June 2026, Altman observed that the companies adopting AI most aggressively are also the ones hiring the most, while the firms blaming layoffs on AI tend to be using it the least. He admitted he had underestimated how jagged these models would be. They do some things remarkably well, he said, and other things, particularly complex, long-horizon work, not well at all.
I work in the part of the economy where this reversal is most visible, and what interests me is not that the original prediction was wrong.
It is that the correction tells us something the prediction never could.
The Layoffs Are Counted. The Rehiring Is Estimated.
Both halves of this story are true at the same time, which is exactly why it is so often misunderstood.
The layoffs are real and still accelerating.
By The New York Times’ count, more than 150 technology companies cut at least 115,000 employees in the first five months of 2026, with Meta, Coinbase and Block each shedding a tenth of their workforce while pointing to AI. Yet a large share of these cuts appears to be reversing. A survey of 600 HR professionals by Careerminds found that roughly a third had already rehired between a quarter and half of the roles they had eliminated, and another third had rehired more than half. Gartner expects that by 2027, half of all companies that cut customer service roles for AI will be staffing those functions again. Forrester found that 55 per cent of employers who made AI-driven cuts already regret them.
A note of honesty about those numbers, because they deserve the same scrutiny we are about to apply to the layoffs. The firings are counted: tracking sites log each announcement, and the outplacement firm Challenger, Gray & Christmas, attributed nearly 55,000 job cuts directly to AI in 2025 alone. The rehiring, by contrast, is estimated. As The Washington Times conceded in its own reporting on the trend, no reliable national count exists of workers replaced by AI who later regained their jobs. None of this makes the rehiring imaginary; the pattern recurs across independent sources, and the Klarna reversal below comes from the company’s own chief executive. But it does mean the boomerang is, for now, better evidenced in direction than in magnitude. And if AI makes a convenient story for firing, it is worth asking whether “AI regret” is becoming an equally convenient story for consultancies with forecasts to sell.
The layoff labels deserve the same doubt. These figures look contradictory only if you assume the cutting and the rehiring share a single cause, and they do not. As The New York Times noted, Meta’s cuts followed an $80 billion retreat from its metaverse bet, Coinbase was weathering a crypto slump, and Block had simply tripled its workforce during the pandemic and grown too large. An analyst quoted in the piece called AI “a nice excuse.” Wall Street rewards an AI story, and a layoff framed as a transformation reads better on an earnings call than one that admits a strategic mistake.
Much of what gets labelled an “AI layoff,” in other words, is ordinary cost-cutting.
Where Automation Underdelivered
Set the disguised layoffs aside and look at the genuine ones, where firms truly tried to replace people with software. Here, the pattern is consistent, and Klarna is its clearest illustration.
In 2023, the Swedish fintech announced that an AI assistant built with OpenAI was handling the workload of 700 customer service agents, and reported significant savings, including around $10 million on marketing alone. Its chief executive, Sebastian Siemiatkowski, declared that AI could already do all the jobs humans do. Within two years, he had reversed course. Customer satisfaction had fallen on the interactions that mattered most: the disputes, the complaints, the cases that required discretion. Cost, he admitted, had become “a too predominant evaluation factor,” and the result, in his own words, was lower quality. Klarna began rebuilding its human support function.
The error was precise. As the Forbes contributor TerDawn DeBoe puts it, companies assumed AI would replace people when, in fact, it replaces tasks. A customer service representative does more than answer questions. They sense when a frustrated caller has a real problem rather than a bad mood. They know when to escalate. They remember that a particular client has already been put on hold three times.
AI manages the answering; it struggles to recognise when answering is not what the moment requires.
The same gap showed up elsewhere: marketing teams that cut copywriters found AI-generated content converting worse, because it did not grasp the company’s voice or what customers actually responded to. The portion of any role that resists automation turns out to be the part that was hardest to see, and hardest to replace.
Where the Tooling Spend Blew Up
Underdelivery is only half the financial story. The other half is that even where AI performs well, the cost of running it has confounded the companies most committed to it, and this is where Microsoft and Uber come back in.
Under usage-based pricing, AI becomes more expensive precisely as it becomes more useful, because every additional prompt adds to the bill.
At Uber, where engineers were actively encouraged to maximise their AI usage and ranked on internal leaderboards, individual engineers were soon running up monthly API costs of between $500 and $2,000 each. For context, Uber’s total research and development spend reached $3.4 billion in 2025, up nine per cent year on year, with AI a key driver; the AI allocation within it still could not hold. Bryan Catanzaro, Nvidia’s vice-president of applied deep learning, has said that for his team, the cost of compute is now far beyond the cost of the employees.
Nor is this a problem that falling prices will quietly solve. Gartner’s research, cited in the same Fortune report, expects the unit cost of AI inference to fall by nearly 90 per cent by 2030, yet predicts enterprise AI will not get cheaper, because agentic systems consume vastly more tokens per task, usage grows faster than prices fall, and providers will not pass the full savings on. Cheaper tokens, bigger bills. The corporate response is already visible: Uber has now imposed a $1,500 monthly cap per employee, per coding tool. The technology that was meant to eliminate a salary line has instead become a metered utility that finance teams must ration, and the savings a company expects can quietly invert into an operating expense no one budgeted for.
The Second Bill: Undoing the Mistake
For the firms that cut too deeply, a further cost arrives when they try to reverse the decision, because rehiring does not restore the old cost structure. The firm pays to recruit and retrain. It absorbs the loss of institutional memory, the long-serving employee who knew which vendor always invoiced late and which client disputed every charge. And it pays more for the role itself, because the role has changed. The position that returns is a hybrid one, requiring someone who can do the work and manage the AI now woven through it. The figure DeBoe cites is a job that paid $55,000, returning at $75,000 or more. According to the Careerminds survey, roughly one in three companies that rehired spent more on restaffing than they had ever saved by cutting.
Put the running costs and the rehiring costs together, and a sharper truth emerges. This is not really a story about AI being too expensive, though it often is.
It is a story about human judgment being re-priced.
Companies learned what that judgment was worth only after they removed it and watched the quality collapse.
The Skill Almost No One Has Bothered to Learn
It would be easy to read all this as reassurance, as proof that the human worker is safe and the moat held. I do not think that is the lesson, and I held a more comfortable version of it for longer than I should have.
The jobs coming back are not the jobs that left.
They demand data literacy, the ability to direct these systems with precision, and the judgment to recognise when a polished, confident output is quietly wrong. Altman’s own framing is useful here: the people getting the most from these models, he said, are the ones who combine them with real expertise. The premium is not paid for being human. It is paid for being the kind of human who can work with the machine without being captured by it.
This is where I will speak from my own experience rather than the reporting. I trained myself on these tools deliberately, through online courses and a good deal of trial and error, and the single most useful thing I learned is that there is no one tool for everything. The model that drafts a sharp piece of long-form copy is not the one that builds a campaign brief, generates on-brand visuals, analyses performance data, or handles SEO research. Treating AI as a single magic box, asking one general-purpose chatbot to do all of it, is precisely how people end up with the bland, forgettable output that gives the whole category a bad name. Using it well means knowing which tool serves which need, and knowing enough about your own craft to judge whether what comes back is any good.
That is the difference between dabbling and skill. Ask a model for a campaign idea, and it will hand you something competent and forgettable, the statistical average of everything done before. Ask the right question, shaped by a real understanding of positioning and audience, and the output changes, not because the tool grew smarter but because you did. The model depends on judgment it does not possess. Whoever supplies that judgment captures the premium the market is now, expensively, learning to pay.
There is an objection here worth meeting head-on.
If part of that premium exists because today’s models are jagged, what happens when the jagged edges smooth? The honest answer is that part of it will shrink.
The share of the premium earned by working around present flaws, catching the confident error, supervising the task the model cannot yet hold together, is rented rather than owned, and every model release erodes it a little. But the deeper share was never about the flaws. Knowing what a brand should sound like, whether an idea will move the audience it is aimed at, which of the five plausible outputs is the right one: that is not a gap in the technology waiting to be closed, it is the definition of the work. The premium on judgment is durable; the premium on babysitting is not, and it pays, for now, to be earning both while understanding the difference.
Which is why I would argue plainly that proper training in these tools is no longer optional for anyone whose work touches them. Not the casual familiarity of having tried a chatbot once, but real fluency: understanding where each model is strong and where it remains jagged, learning to interrogate its output rather than accept it, and treating it as an instrument of your own thinking rather than a substitute for it. Precisely because the tools will keep improving, the fluency has to be maintained rather than acquired once; the person who trained on last year’s models and stopped is only marginally better off than the person who never started.
Every technological shift in creative work has rewarded the same instinct. The photographers who thrived after digital were not the ones who mourned film, but the ones who understood that the eye was still theirs. The advantage went to those who grasped what the technology could not do and built their value there. The companies now rehiring learned this the expensive way, by removing the judgment and watching it break. The rest of us can choose to learn it the deliberate way instead, by treating the next few years not as something to survive but as something to train for. What we choose to develop alongside these tools remains entirely ours to decide.