Nearly 6,000 Executives Were Asked the Same Question: After Three Years of AI, Over 90% Report No Effect on Employment and 89% None on Productivity

3 views

NBER working paper 34836, "Firm Data on AI," surveyed nearly 6,000 senior business executives in the US, UK, Germany and Australia. Adoption is not the problem: 69% of firms across the four countries report current AI use — 78% in the US, 59% in Australia — with text generation via large language models the most cited application at 41%. But asked about actual effects over the past three years, more than 90% report no effect on employment and 89% report none on labor productivity, measured as revenue per employee. The same executives stay optimistic about the next three years, forecasting AI will raise their firm's productivity by 1.4%, output by 0.8%, and cut employment by 0.7%.

Wide Adoption, Shallow Use

Start with the uncontested half. Across the four countries 69% of businesses report current AI use, highest in the US at 78% and lowest in Australia at 59%, and the most cited application is text generation with large language models at 41%. About 75% of firms expect to be using AI over the next three years. Intensity is another matter. More than two-thirds of senior executives say they personally use AI in a typical week, averaging around 1.5 hours. CEOs use it more often than CFOs and other senior executives. An hour and a half a week, set next to the narrative that AI is reshaping work, is itself a finding: in most firms today AI looks like a tool someone opens occasionally, not something wired into a process.

"No Effect" Is Not "No Value"

The paper's conclusion needs reading precisely: it does not say AI lacks value. It says measurable returns at the firm level remain limited. Productivity here is revenue per employee — a coarse measure, but an honest one. If AI were displacing hours or amplifying output at scale, it ought to show up there. The author list includes Nicholas Bloom, Steven J. Davis and Jose Maria Barrero, who have spent years running firm surveys and remote-work research, across a sample of nearly 6,000 firms in four countries. The methodological boundary matters: these are self-reported perceptions from executives, not figures derived from administrative records.

Reading It Against Other Evidence

Corroborating evidence points the same way. PwC's 2026 Global CEO Survey of 4,454 chief executives across 95 countries found 56% reported neither revenue nor cost benefits, with only 12% reporting both. Commentators have reached for the Solow paradox — the 1987 line about seeing the computer age everywhere except in the productivity statistics. Keep it separate, though, from the Stanford study on this site a few days ago, which used ADP payroll records and found employment 19% lower for 22-to-25-year-olds in AI-exposed roles. That one uses administrative data; this one uses executive self-reports. They do not contradict each other — one measures the entry point for a specific age band in specific roles, the other measures totals across a whole company, and the first can vanish inside the second. The genuinely awkward comparison is a third thing: Oracle, Salesforce, Lufthansa, Accenture and Standard Chartered have all invoked AI in job-cut announcements, while executives in this survey say AI has not affected employment. The practical implication: if you are building the business case for AI spending, aggregate measures like revenue per employee will probably not produce the number you want in the near term. What is more likely to hold up is process-level measurement — handling time for a specific step, rework rate, cost per task — rather than dividing a whole company's output by its headcount.

via: NBER working paper 34836, NBER Digest summary, People Matters, Inc.