In February 2026, NBER published working paper 34836, “Firm Data on AI,” by more than a dozen economists including Ivan Yotzov, Nicholas Bloom, Steven J. Davis, Brent H. Meyer, Paul Mizen, Gregory Thwaites, and Ben Zhe Wang, spanning the Federal Reserve Bank of Atlanta, the Bank of England, the Deutsche Bundesbank, and Macquarie University. Coverage by AI News on February 20 pushed it into the tech conversation, and its core finding is blunt to the point of discomfort: AI adoption is broad, and the measurable productivity impact is close to zero.
What makes the survey worth taking seriously is the method. Nearly 6,000 phone-verified, unpaid executives — predominantly CEOs and CFOs — across the United States, United Kingdom, Germany, and Australia, with the data cross-checked against ten years of macro output and employment figures from national statistics agencies. Against the usual vendor surveys and Twitter polls, this is some of the firmest evidence we have on what AI is actually doing inside companies.
Wide Adoption, Meager Measured Impact
The numbers set up a sharp contrast. 69 percent of surveyed firms actively use AI in the business, with adoption notably higher among younger and more productive firms, and over two-thirds of the executives themselves use AI tools regularly. The leading use cases are LLM text generation (41 percent), visual content creation (29 percent), and ML-based data processing (28 percent). UK firm-level adoption climbed from 61 to 71 percent during 2025 alone.
But looking back over the past three years, roughly nine in ten executives say AI has had no measurable effect on their firm’s employment or productivity. Usage intensity is also thin: executives average about 1.5 hours of AI use per week. Adoption is wide; depth is shallow — that is the whole story in six words.
The Expectations Gap Between Bosses and Workers
The more interesting half of the paper is expectations. The same executives, looking forward three years, on average expect productivity to rise 1.4 percent, output to gain 0.8 percent, and employment to fall 0.7 percent. US executives are the most optimistic, projecting 2.25 percent productivity gains; UK firms expect 1.86 percent. The hoped-for headcount cuts are modest too, with two-thirds of the expected UK adjustment planned through slower hiring rather than redundancies.
Compare that with workers, and the gap opens up. US employees (via the Survey of Working Arrangements and Attitudes) expect employment to rise 0.5 percent while their executives expect a 1.2 percent cut; employees expect 0.92 percent productivity gains, less than half their bosses’ 2.25 percent. Same technology, opposite narratives — the boss sees fewer people, the worker sees a raise, and at least one side will be disappointed.
Why the Numbers Come Out Flat
Several structural reasons are worth listing:
- Complementary investment lags: process redesign, data cleanup, and training typically trail tool purchases — the familiar J-curve from every productivity paper.
- Usage is too shallow to matter: 1.5 hours a week does not move a company’s cost structure.
- Measurement definitions diverge wildly: McKinsey reports 88 percent of firms using AI in at least one function, while the US Census Business Trends and Outlook Survey went from roughly 9 percent in early 2024 to 18 percent by December 2025 — a fivefold spread depending on the questionnaire, which is a reminder not to treat any single adoption rate as gospel.
- Selection cuts both ways: adoption concentrates in firms that were already younger and more productive, which makes it hard to attribute gains to AI itself.
What It Means for Product Teams
First, the “adoption” pitch is saturated; the next procurement argument is measurable output change, which effectively endorses every ROI-obsessed enterprise AI product on the roadmap. Second, depth of use is the new north-star metric: a product used 1.5 hours a week and one embedded in the daily workflow are commercially different animals. Third, expectation management is real: with executives banking on an average 1.4 percent productivity gain, any pitch promising tenfold returns will look silly in front of a CFO. For now the evidence sides with the incrementalists — and if the next three years prove them wrong, this paper is the baseline everyone will cite.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.