The anticipatory layoff: how CEOs are firing for an AI that hasn't arrived yet

The number that should stop us is not 92,000. It is the gap between that number and the productivity data that is supposed to justify it.

Layoffs.fyi's running tally has crossed 92,000 tech workers cut in 2026, and the year is barely a third old. Meta took 8,000 in April. Oracle's reductions, depending on whose count you accept, sit somewhere between 20,000 and 30,000. Atlassian, Snap, Pinterest, each with their own announcement, each with the same phrase threaded through the press release: AI efficiencies, AI-driven restructuring, focusing the company on an AI-first future. The framing is so consistent across so many companies that it reads less like a series of independent business decisions and more like a shared script.

And it is a script. That's the part I want to spend this essay on.

The finding that reframes everything. In a piece for Harvard Business Review earlier this year, Thomas Davenport and Bala Srinivasan looked at the wave of AI-attributed layoffs and asked an obvious question that very few people in the discourse have been asking: are the productivity gains real? Their answer, based on what companies have actually disclosed and what independent researchers have been able to measure, is that the layoffs are running well ahead of the evidence. The cuts, they argue, are being driven by CEOs' expectations of what AI will deliver, not by measured gains it has already delivered.¹

That sentence deserves to be read twice. Companies are not cutting workers because AI has made those workers redundant. They are cutting workers because executives believe AI is about to make those workers redundant. The redundancy is forecast, not observed.

I want to call this what it is: the anticipatory layoff. And I want to argue that it is a different category of event than the ordinary churn of restructuring, that it has distinctive distributional consequences, and that the way we are currently talking about it, as if it were simply the technology working its way through the labour market, gets the causation almost exactly backwards.

The redundancy is forecast, not observed. The worker is gone; the productivity gain that justified the cut is still a slide in someone's deck.

What the ordinary framing elides. When a company announces an AI-driven layoff, the press coverage tends to treat the AI part as load-bearing. The headline writes itself: machines replace workers. The implicit causal chain is that a capability arrived, that capability was deployed, the deployment made certain roles redundant, and the redundancy showed up as a reduction in headcount. Cause, effect, regrettable but rational.

But that chain has a missing link, and Davenport and Srinivasan put their finger on it. In most of these cases, the deployment hasn't actually demonstrated the redundancy yet. The tools are being piloted. The workflows are being redesigned. The integration work, and anyone who has done enterprise software integration knows this is the hard part, is mostly still ahead. What has happened, concretely, is that a CEO has decided to act as if the future deployment will produce the savings the vendor promised, and to bank those savings now by removing the workers in advance.

This is not a small distinction. It changes who bears the risk of the bet being wrong.

Who pays when the forecast misses. Consider what happens in each scenario. If the AI deployment delivers what the CEO expects, the company has captured the productivity gain ahead of schedule and the laid-off workers are, retrospectively, redundant, though it's worth noting they were made redundant by an executive prediction, not by an actual replacement. If the deployment underdelivers, which the early evidence on enterprise AI suggests is more common than not, the company has a problem: it has lost institutional knowledge, lost the workers who understood the old workflows that the new tools are supposed to augment, and has to either rehire (often at higher cost), contract the work out, or absorb the degraded output.

Notice who carries the risk in either branch. In the first branch, the company gets the upside and the workers get the downside slightly earlier than they would have anyway. In the second branch, the company eats some operational pain, but the workers are still gone, still looking for jobs in a market saturated with people whose CEOs made the same bet. The worker bears the cost of the executive's prediction being wrong. The executive does not.

There is a name for this kind of arrangement in other contexts. We call it moral hazard.

The herd, and what's driving it. One company doing this would be a story about that company. Dozens of companies doing it within the same eighteen-month window, all citing the same justification, is a story about something else. It is a story about how a particular narrative about AI has hardened inside C-suites to the point where not announcing AI-driven cuts has started to look, to certain audiences, like a failure of strategic seriousness.

The audiences matter here. The most important audience for a tech-company layoff announcement is not the remaining workforce; it is the capital markets. Wall Street has been rewarding AI-attributed cuts with share-price bumps on the theory that they signal disciplined cost management and credible AI strategy in a single move. CEOs have noticed. The announcement is the asset. Whether the underlying productivity story bears out over the next two years is a separate question, and one that, by the time the answer is in, will be obscured by a dozen other variables.

This is what I mean when I say the causation is backwards. The ordinary framing has the technology causing the layoffs. The more honest framing, looking at what Davenport and Srinivasan found, is that the narrative about the technology is causing the layoffs, and the technology itself is somewhere further down the causal chain, perhaps doing some of the work the press releases claim, perhaps not, but in any case running well behind the cuts that have been justified in its name.

The workers who are gone before the tools arrive. I find it hard to write about this without naming what it does to the people on the receiving end. Eighty-thousand-plus tech workers, many of them mid-career, many of them with mortgages and dependants, many of them having moved cities or countries for these jobs, are now in a labour market where the explanation for their unemployment is a technology that, on the evidence, has not yet done the thing it is being credited with doing.

That matters for how they tell their own story, and it matters for how they are perceived by the next employer. "I was let go because AI made my role redundant" lands very differently from "I was let go because my CEO believed AI was about to make my role redundant, and turned out to be wrong." The first sounds like the worker was simply on the wrong side of a technological transition, which is a thing that happens and is no one's fault. The second sounds like the worker was on the wrong side of an executive bet, which is a different kind of injury.

The workers themselves do not get to choose which framing prevails. The press release does.

The historical rhyme, and where it breaks down. It is tempting to say we have seen this before, that every wave of automation produces overshoot, that the dot-com era had its own version of this, that the labour market eventually absorbs the displaced. There is something to the parallel and I don't want to dismiss it. Technological transitions do produce displacement, and the displaced do, eventually, mostly find other work, and the economy as a whole does, eventually, mostly come out ahead.

But three things about the current moment make me cautious about the easy version of that story.

First, the speed. Previous automation waves moved at the pace of capital expenditure cycles, factories had to be built, equipment had to be installed, workers had to be retrained. The current wave is moving at the pace of executive memos. A CEO who decides on Monday that AI will reduce engineering headcount by twenty percent can announce the layoffs on Friday. The deployment will follow at its own pace, but the cut doesn't wait for it. Workers are absorbing the displacement in real time; the offsetting job creation, if it comes, will arrive on the slower clock.

Second, the breadth. Previous automation waves tended to hit specific occupational categories, manufacturing, then routine clerical work, then certain kinds of logistics. The current wave, at least in its current framing, is being applied across white-collar work broadly. Engineers, designers, marketers, customer service, middle management, HR, legal support, content. When the displacement is concentrated, displaced workers can move sideways into adjacent fields. When the displacement is presented as economy-wide, the sideways move is harder to find.

Third, and this is the one Davenport and Srinivasan's finding really sharpens, the anticipatory character of these cuts means we are not yet seeing the technology's actual labour effects at all. We are seeing executive predictions of those effects, applied as policy. If the predictions turn out to be roughly right, the historical rhyme will hold and we will eventually look back at this as a painful but ordinary transition. If the predictions turn out to be substantially wrong, if AI augments more than it replaces, if the productivity gains are smaller and slower than the press releases claim, if the tools require more human supervision than the vendor promised, then we will look back at this as a wave of layoffs that had no underlying technological justification at all. Just a story that everyone agreed to act on.

The early productivity data, for what it's worth, is not encouraging on the prediction. METR's study of experienced developers using AI tools found a nineteen percent slowdown on real-world tasks, even as the developers themselves believed they were faster. MIT's State of AI in Business report found that ninety-five percent of enterprise AI pilots were producing zero measurable return. These are not the numbers of a technology that is currently making large categories of workers redundant. They are the numbers of a technology that is still figuring out what it is good for.²

What this means for how we should talk about it. I am not arguing that AI will not eventually displace significant numbers of workers. It probably will, in some categories, on some timeline. I am arguing that we should be precise about what is happening now, because the imprecision is doing harm.

Right now, in 2026, the layoffs are running ahead of the technology. The justification offered by companies, that AI has made these roles redundant, is, in most cases, a forecast presented as a fact. The cost of treating the forecast as a fact is being paid by the workers who are gone, and by a labour market that is being asked to absorb displacement that may, in retrospect, turn out not to have been technologically necessary.

The press could be more careful. "Meta laid off 8,000 workers, citing AI" is a very different headline from "Meta laid off 8,000 workers in anticipation of AI productivity gains it has not yet demonstrated." The first launders the executive's prediction into a description of reality. The second names what is actually happening. The first is what we have been getting; the second is what the evidence supports.

Regulators could be more curious. When a company announces large layoffs and attributes them to a specific cause, there is no requirement that the cause be evidenced. A company can attribute a workforce reduction to AI productivity gains it has not measured and will not be asked to measure. This is, on reflection, an odd state of affairs. We require companies to substantiate forward-looking statements about revenue. We do not require them to substantiate forward-looking statements about why the people they just fired are no longer needed.

And those of us writing about this could stop calling it a transition, as if it were something happening to the labour market from outside, and start calling it what it is: a series of decisions, made by specific executives, about who bears the risk of an unproven bet. The technology is not doing the firing. The CEOs are. The technology is the story they are telling about it.

I think that distinction matters. I think it will matter more, not less, as the actual evidence on AI's labour effects accumulates. And I think the workers who have been let go in the meantime, the ninety-two thousand and counting, deserve to have their displacement described accurately, even if the accurate description is less convenient for the people who decided they were redundant.

Footnotes

¶