Are AI Layoffs a Strategic Mistake?

In May 2026, American employers announced 38,579 job cuts and gave the tracking firm Challenger, Gray & Christmas the same one-word explanation: artificial intelligence. It was the highest monthly total the firm has recorded since it started tracking the category, forty percent of everything announced that month, and the third consecutive month AI outranked every other stated reason. Through May, AI had already been cited in 87,714 cuts for the year — more than the 54,836 recorded across all of 2025 [1].

Sit with one word in that paragraph: cited. These figures come from employers explaining their own decisions. Whether the software actually did the work the departing employees used to do is a separate, much harder question, and it happens to be exactly where the best available science lands somewhere nobody expected — not on replacement, but on amplifying the very people a lot of these announcements just let go.

Corporate America ran a version of this experiment before, under the banners of "downsizing" and "reengineering" — same promise, different technology: cut the people, keep the output, pocket the difference. Wayne Cascio's studies of major US corporations found that downsizing alone did not produce higher returns, and that the productivity, quality, and morale of the employees who remained took a documented hit alongside it [3]. A 2024 longitudinal meta-analysis of the wider downsizing literature landed in the same uncomfortable place: results mixed at best, with short-term costs that most announcements never priced into the decision [4].

The mechanism behind those findings has a name in the literature — institutional knowledge: the undocumented judgment about how things actually work, which walks out the door with the person, not the headcount line. The 1990s lesson was never "don't ever cut." It was that cuts justified by a technology's promised productivity, made before that productivity has actually been measured, fail in a specific and well-documented way. Which is what makes the next section's finding worth pausing on.

The single best piece of evidence about what generative AI actually does inside a company is a field study of 5,179 customer-support agents at a Fortune 500 software firm, published in the Quarterly Journal of Economics. Giving agents a generative-AI assistant raised productivity by about 15 percent on average — but that average hides the number that matters here: novice agents improved roughly 34 percent, while the most experienced agents barely moved, with a small dip in quality [2].

The researchers' own explanation is that the tool works by capturing the patterns of the best, most experienced workers and handing them to everyone else. It is, in effect, bottled institutional knowledge. Read that against the layoff numbers above, and the implication gets sharp fast: the measured gains depend on the continued presence of the experienced people whose judgment the system learned from. A company that cuts its veterans isn't just spending down its own expertise — it's cutting off the training signal for the tool that was supposed to replace them. The strongest evidence for AI's value and the strongest argument against cutting people to get it turn out to come from the same study.

That's the established science. Whether these are even AI layoffs at all is where credible people start disagreeing. Competing hypothesis A serious body of reporting on the announcement wave finds experts skeptical that most of it reflects genuine replacement: companies have an obvious incentive to dress up ordinary cost-cutting, and pandemic-era overhiring corrections, as forward-looking AI strategy for investors — and the coverage notes that demonstrated substitution is the part almost nobody actually shows [5]. Read this way, the Challenger figures [1] partly measure how popular a narrative has become, not how much substitution has actually happened. Then there's whether any of it pays. A reported survey of executives at AI-deploying firms found that the heaviest cutters and the lightest cutters showed nearly identical financial results Emerging evidence [7] — a single, unreplicated finding we label accordingly, though it rhymes with the 1990s literature above [3][4]. The other side answers that it's simply too early: deployment takes years, coding functions already show real substitution, and CFOs surveyed on the subject expect substantially deeper AI-driven cuts ahead even while admitting the productivity case is still thin [6]. Both camps agree the announcements are running well ahead of the measured returns. What they can't agree on is whether that gap is a lie or simply a lag.

These layoffs don't stay contained to a single balance sheet. Institutional knowledge is the input the QJE study shows the tool actually depends on [2], which means every headcount decision is also a capital-allocation decision about a nonfinancial asset most companies don't measure. That decision also reshapes the labor market a step downstream: the same AI-exposed occupations experiencing these cuts show the youngest workers already losing ground fastest [8], and a companion investigation on this site (Why AI Is Changing How Companies Hire) documents how the hiring machinery on the other side of that pipeline — résumés, screening, entry-level postings — is absorbing the resulting pressure.

Nothing below is a prediction — each path traces a different answer to the same open question: does the amplification effect from the field study generalize, or does this settle the way the 1990s downsizing wave eventually did? The data to tell them apart hasn't fully arrived yet.

This investigation's title doesn't get a yes-or-no answer today. AI is now the most-cited reason for announced US job cuts [1]; the best field evidence says the technology amplifies workers, especially the least experienced, by encoding what veterans already know [2]; and the youngest workers in the most AI-exposed occupations are measurably losing ground while senior colleagues hold steady [8]. Public, trackable indicators worth watching over the coming months and years:

• Layoff & rehiring data — Whether the Challenger AI category holds its record pace, and whether companies start reporting realized savings rather than expected ones [1]. Rehiring announcements at firms that made large AI-attributed cuts, and boomerang-hire rates as a rough proxy for knowledge debt.
• Corporate disclosures — Margin and productivity disclosures from heavy-cutting versus light-cutting AI adopters.
• Academic research — An independent replication of the returns-parity finding, with sturdier methodology [7]. Peer-reviewed follow-ups to the QJE study in other functions — coding, legal, finance — testing whether the augmentation pattern holds or substitution starts to appear [2].
• Long-term performance verdict — Whether 2025–26's AI-attributed cuts eventually correlate with outperformance or underperformance, the same test that settled the 1990s downsizing debate [3][4].

If you've watched an AI-attributed layoff up close — did the work actually go to the software, or to the people who were left to absorb it? Tell us what you saw. We read every message, and firsthand accounts shape where we look next.