Challenger, Gray & Christmas has now published the same finding five months running: artificial intelligence is the single most-cited reason American employers give when they announce job cuts. In July it accounted for 10,970 of them. Through July, employers have announced 477,033 cuts in total this year.
Set that against the largest cross-country survey of what executives say AI has actually done inside their own companies. NBER working paper 34836, Firm Data on AI, put identical questions to nearly 6,000 CEOs, CFOs and senior finance managers, fielded by research teams at the Federal Reserve Bank of Atlanta, the Bank of England, the Deutsche Bundesbank and Macquarie University. More than 90% reported that AI had no effect on employment at their firm over the previous three years. Eighty-nine percent reported no effect on labour productivity.
These are not abstainers answering. Sixty-nine percent of the firms surveyed said they were already using AI in some form — 78% in the US. They are using it, and telling researchers it has not moved their headcount.
Mind the window, then mind the gap
The honest caveat first: the survey was fielded between November 2025 and January 2026, covering the preceding three years. It cannot see spring or summer 2026. A good-faith reading is that the curve bent after the questionnaires went out.
The scale of the divergence makes that a stretch. AI was cited in 87,714 announced cuts through May alone — 22% of everything announced in 2026, and already well past the 54,836 attributed to it across the whole of 2025. For the survey to be reconciled with the press releases, AI would have had to go from touching essentially nobody’s headcount to being the leading stated cause of job losses in roughly two quarters. Technology deployments inside large organisations rarely move that fast. Narratives do.
Sam Altman has said the quiet part out loud: almost every company doing layoffs is blaming AI, “whether or not it really is about AI.”
Our take: The reason cited in a layoff announcement is not a data point about technology. It is a message to shareholders. “We are restructuring around AI” reads as a company investing in its future; “we hired too many people and revenue slowed” reads as a company that made a mistake. Same 300 people, same severance charge, very different multiple. Treat the stated reason as marketing copy until an operating number backs it up.
What the announcement actually buys
Rapid7 is a clean specimen of the modern form. The Boston cybersecurity firm’s board approved a restructuring on 7 August cutting roughly 310 roles — 12% of a workforce of about 2,600 — announced alongside Q2 results by Wael Mohamed, the former Forescout chief who had been in the job around two months. The company framed it as shifting investment toward an AI-first platform, and booked $10m to $11m in charges, almost all of it severance and notice-period cash.
It also attached a number: a 20% operating margin target. That is the useful disclosure. Whatever the AI framing does for the story, the margin goal is what the cuts are underwritten against, and it is the thing that can be checked in two quarters.
What to watch
- Revenue per employee, not the press release. A real AI restructuring shows up as output holding while headcount falls, within two to three quarters. A cost-driven one shows up as margin expansion with flat revenue.
- Whether the roles come back. Rehiring into the same functions in 2027 is the clearest evidence the automation never arrived.
- The next NBER wave. The same four central-bank research teams can rerun the questions on a 2026 window. That is the cleanest available check on whether the announcements were describing reality.
- Challenger’s monthly split. If AI keeps leading the stated reasons while “cost-cutting” and “market conditions” fade from the list, the labelling is drifting faster than the behaviour underneath it.
None of this means AI is doing nothing to employment. The same executives expect it to cut headcount 0.7% and lift productivity 1.4% over the next three years — real, and modest. The problem is a gradual effect being invoked to explain sudden, discrete events. When the explanation is that much bigger than the cause, ask what it is covering.