The sales pitch was simple: bring in artificial intelligence, streamline operations, do more with fewer people. A growing body of research now suggests the “fewer people” part happened on schedule while the “do more” part has not.
Two working papers from the National Bureau of Economic Research, drawing on U.S. Census Bureau survey data and a separate multi-country executive survey, find that AI adoption is spreading fast among large firms but delivering only modest, localized productivity improvements. Meanwhile, workforce reductions have followed AI deployments at a striking rate. The pattern across the NBER research and the global executive survey summarized in the NBER Digest is consistent: companies are cutting workers to fund AI experiments that have yet to prove their worth.
AI Is Everywhere. The Payoff Isn’t.
The first NBER paper, authored by Kristina McElheran, J. Frank Li, Erik Brynjolfsson, Zachary Kroff, Emin Dinlersoz, Lucia Foster, and Nikolas Zolas, is titled “The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks.” It uses a special AI supplement to the Census Bureau’s Business Trends and Outlook Survey to map adoption at the firm, function, and task level. It is one of the few studies built on federal survey data rather than consultant polls or tech-company press releases, which gives it unusual weight.
What the researchers found is a story of lopsided adoption. Large firms are far more likely to have deployed AI tools than small or midsize businesses. According to the paper, roughly 5.4% of U.S. firms reported using AI overall, but adoption rates among firms with 250 or more employees were several times higher. Even inside adopting companies, AI use clusters in a handful of business functions rather than spreading across operations. The task-level data is the most revealing part: many workers at firms that have officially “adopted AI” report that their actual day-to-day responsibilities have changed only modestly. The technology is being applied in narrow slices, not across entire workflows.
The second study, authored by Saku Aura, Timo Kuosmanen, Markku Maula, and Mikko Packalen and linked to NBER Working Paper 34836, draws on a multi-country executive survey to assess how businesses around the world are deploying AI. Executives report rising AI adoption rates, but the output gains they anticipated have not kept pace with the spending and organizational disruption involved, according to the researchers’ findings. Respondents describe pilot projects in marketing, customer service chatbots, and coding assistants, yet few report transformative changes in overall productivity. The gap between enthusiasm and results is wide.
Staff Cuts Came First. Returns Have Not Followed.
Taken together, the two papers establish a troubling sequence. Companies adopted AI, cut staff, and then waited for efficiency gains that, as of mid-2026, remain largely unrealized. The workforce reductions were treated as a down payment on future productivity. But the research suggests that for most firms, the future has not arrived yet, and there is no guarantee it will.
This is not an abstract problem. When a company eliminates positions to fund an AI deployment that produces only marginal improvements in a single business function, the net result is fewer workers, the same or slightly better output in one area, and no change everywhere else. The employees who lost their jobs subsidized an experiment, not a proven transformation.
The pattern is especially stark because of who is doing the cutting. The NBER data shows that AI adoption is concentrated among larger, better-resourced firms, the same companies that tend to set industry norms for staffing levels and compensation. When these firms cut headcount and cite AI as the reason, smaller competitors often feel pressure to follow, even if they have not adopted the technology themselves.
What the Data Cannot Yet Answer
Several important questions remain open, and honest reporting requires flagging them.
Neither study provides longitudinal employment data that directly ties a specific AI deployment to a specific number of layoffs at an individual firm. The Census Bureau survey captures which firms are using AI and for what, but it does not track headcount changes over time in a way that lets researchers draw clean causal lines. The executive survey relies on self-reported assessments, which carry inherent limitations: managers may overstate adoption to signal innovation to investors or understate gains that have not yet appeared in quarterly earnings.
There is also a timing problem that cuts both ways. AI tools deployed in 2024 and early 2025 may simply need more runway. Enterprise software rollouts historically take years to reach full productivity, and AI systems that require retraining workers or restructuring workflows could follow the same slow curve. The current data cannot distinguish between “AI does not work as promised” and “AI has not worked yet.”
It is also worth noting that some task-level studies, including research on AI coding assistants and writing tools, have shown meaningful productivity gains for individual workers in controlled settings. The disconnect may be less about whether AI can improve specific tasks and more about whether those task-level gains translate into firm-wide efficiency when layered onto complex organizations with legacy systems, retraining costs, and integration challenges.
Why the Burden of Proof Still Belongs to the Companies Making the Cuts
For workers who have already lost their jobs, the distinction between “AI failed” and “AI has not succeeded yet” is academic. They are out of work either way. And for policymakers weighing how to regulate AI-driven workforce changes, the NBER findings raise a pointed question: should companies be allowed to justify mass layoffs on the basis of projected AI gains that have no track record of materializing?
The research does not answer that question directly, but it reframes the debate. The dominant narrative from Silicon Valley and corporate boardrooms has been that AI-driven layoffs are an inevitable, even rational, response to technological progress. The NBER data suggests something less flattering: many of these cuts look less like strategic adaptation and more like speculative bets made with other people’s livelihoods.
Until longitudinal data links AI adoption, employment changes, and firm performance more directly, the claim that AI-driven staff reductions are economically necessary should be treated as a hypothesis, not a settled fact. The burden of proof belongs to the companies making the cuts, and so far, the evidence is not on their side.
Warren Cohen is a finance writer based in Phoenix, Arizona, covering personal finance topics including credit, banking, and beginner investing. He earned his degree in business administration from Arizona State University and began his career working in consumer finance, where he gained direct experience with lending and credit systems. He now writes for personal finance websites and fintech platforms, focusing on clear, practical content that helps readers make informed financial decisions.
