From Shubhi K | Product & Market Analysis
NBER Asked 6,000 Executives About AI: 89% Saw No Gain, Revision Says 0.2%
On this page
Nearly 6,000 senior executives in the US, UK, Germany and Australia were asked whether AI had changed labor productivity at their firm over the past three years. 89% said it had not. The NBER AI productivity survey behind that number is the largest representative read on enterprise AI so far, and it is also easy to misquote. Here is what it measured, what the October 2026 revision changed, and how it compares with vendor ROI surveys.
Key takeaways
- 89% of executives reported no AI impact on labor productivity over three years. The figure comes from NBER Working Paper 34836, fielded November 2025 to January 2026 through central-bank business surveys in four countries.
- "No impact" is not "negative impact". It means the gain, if any, was too small for a CFO to see at firm level, in a sample where about 31% of firms were not using AI at all.
- The revised paper puts the realised gain at 0.2% a year. Nearly 9,000 executives in seven countries expect 0.6% a year over the next three years, a 3 times step-up that is still a forecast.
- Vendor surveys report success rates several times higher because they ask different people a different question. Google Cloud's 74% ROI figure sampled only firms already deploying generative AI and counted ROI on any use case.
The short answer. NBER Working Paper 34836 surveyed nearly 6,000 executives through central-bank business panels. 89% reported no AI effect on their firm's labor productivity over three years. The revised version, with nearly 9,000 executives, estimates a realised gain of 0.2% a year and an expected 0.6% a year. That supports "small so far", not "AI does nothing".
What NBER working paper 34836 actually asked
The paper is titled "Firm Data on AI". Its authors include Nicholas Bloom and Steven J. Davis of Stanford, alongside economists at the Bank of England, the Federal Reserve Bank of Atlanta, the Deutsche Bundesbank and Macquarie University. It was first issued in February 2026 and is published jointly as Atlanta Fed Working Paper 2026-3.
The design is the reason it carries weight. These are not opt-in panels recruited by a vendor. Each country ran the same questions through an existing business survey that a central bank or university already uses to track the economy.
The project dashboard names the panels. The US data come from the Atlanta Fed's Survey of Business Uncertainty, a monthly survey of CEOs, CFOs and senior executives. The UK used the Bank of England's Decision Maker Panel. Germany used the Bundesbank Online Panel for firms, and Australia used Macquarie University's Business Outlook Scenarios Survey.
According to the NBER Digest summary, the teams fielded identical questions between November 2025 and January 2026. Respondents were nearly 6,000 CEOs, CFOs and senior finance managers, drawn from stratified firm samples.
Executives were asked what AI had done to labor productivity at their own firm over the past three years, and what they expected over the next three. They were also asked about employment over the same windows. The exact wording for each country is published as screenshots on the project dashboard.
That framing matters more than any other detail. A worker who saves an hour a week on drafting email does not change a CFO's view of firm output per employee. The survey measures what has reached the level of the business, which is the level your budget is judged at.
The 89% figure, read carefully
The headline is accurate. It is also routinely stretched into claims the survey does not make. Three readings are worth separating before you quote it in a board deck.
No impact is not a negative impact
The 89% did not report that AI made them less productive. They reported no change they could attribute to it. The remaining 11% is the ceiling on firms reporting any effect at all, positive or negative.
That is a statement about detectability. Executives answering from finance roles see sales, headcount and costs. A gain smaller than ordinary quarterly noise would register as zero, and the survey has no way to tell "nothing happened" from "too small to see".
Almost a third of the sample was not using AI
The Digest reports that 69% of businesses across the four countries used AI in some form. US firms led at 78%, and Australia trailed at 59%. The most common use was text generation with large language models, at 41% of firms.
So the 89% includes firms that had not adopted anything. Here is the arithmetic the headline skips. If every firm reporting an effect was an adopter, then at most 11 of every 69 adopters, about 16%, saw a change. That is an upper bound I computed from the published shares, not a figure in the paper, and it still describes a small minority.
Three years starts before most firms deployed
The window ran back to late 2022 or early 2023, depending on the country. Most enterprise deployments of generative AI started well inside that window. A three-year look-back averages a long period of zero with a short period of use, which mechanically pulls the answer toward "no impact".
My view is that this is the strongest single reason not to read the 89% as a verdict. It is a snapshot of adoption timing as much as of AI's effect.
The Atlanta Fed summary of the paper reports that over two-thirds of top executives used AI regularly, but averaged 1.5 hours a week, and one quarter reported no use. Leaders who spend 1.5 hours a week with a tool are not well placed to see what it does on the floor. That cuts both ways, and the next section shows how.
The October revision: 9,000 executives and a 0.2% answer
The paper was revised in October 2026, and most coverage still quotes the February version. The current NBER abstract describes a Summer 2026 survey of nearly 9,000 CFOs, CEOs and senior executives. The country list grew to seven, adding Canada, Mexico and South Africa.
From a share to a rate
The revision replaces the "how many saw an effect" framing with a size estimate. It reports that AI adoption raised annual labor productivity growth by 0.2% over the last three years. Executives expect a 0.6% annual boost over the next three.
The two framings agree. If roughly nine in ten firms saw nothing, an economy-wide average gain of 0.2% a year is what you would expect. The revision does not overturn the 89%. It converts it into a number you can put in a model.
Use grew, and spending is concentrated
The revised abstract reports that 78% of firms now use AI, and that 80% of executives report regular use. Average executive use rose to 2.1 hours a week, up from 1.5, a 40% rise in about six months.
The spending line is the most useful new fact for buyers. US firms expect to spend about $2,000 per employee on AI in 2026, roughly $300 billion in aggregate. Just 2% of firms account for over half of that spending. Averages across a market that concentrated tell you very little about the typical firm, which is the firm most readers here run.
Why executives expect AI productivity gains 3 times larger
Both waves show the same shape. Realised gains are small, and expected gains are larger. In the winter wave, executives forecast a 1.4% productivity gain over three years, with output up 0.8% and employment down 0.7%. Annualised, 1.4% over three years is about 0.46% a year, close to the revised 0.6%.
The employment piece deserves attention. The Digest reports that about two-thirds of the expected workforce reduction comes from reduced hiring, not layoffs. If that holds, you will see it in slower recruiting before you see it in any productivity statistic, which is the pattern covered in what payroll data shows about AI job losses.
The people closest to the work expect less
The paper also surveyed about 3,000 US workers, and the gap is telling. US executives expected AI to raise productivity by 2.3%, while workers expected 0.9%. On jobs, executives expected a 1.2% cut and workers expected a 0.5% increase.
Workers also reported more hands-on use, about 1.8 hours a week. I would weight the worker number more heavily. The people using the tools are more sceptical about output than the people approving the budget, and that ordering usually reflects who has seen the rework.
The NBER AI productivity survey against vendor ROI surveys
If you buy enterprise software, you have seen a different number. Google Cloud's 2025 study reported that 74% of executives achieved ROI within the first year. That looks like a contradiction. It is not, once you line up who was asked what.
The Google Cloud release surveyed 3,466 senior leaders at enterprises that already had generative AI deployed. Its sharpest figure, 88%, came from the 13% it classed as agentic early adopters, and referred to ROI on at least one use case. Productivity was cited by 70% as a top driver of value, which is a reason, not a measured gain.
PwC's 2026 Global CEO Survey sits between the two. As The Register reported, 56% of 4,454 business leaders saw neither higher revenue nor lower costs from AI, and only 12% saw both. That is secondary coverage of a PwC release, and the citation should be upgraded to PwC's own document.
| Survey | Who was asked | What counted as success | Headline |
|---|---|---|---|
| NBER w34836, winter wave | ~6,000 executives, stratified samples, users and non-users, 4 countries | Any change in firm labor productivity over 3 years | 89% saw none |
| US CFO Survey (Atlanta Fed, Richmond Fed, Duke) | 603 panel firms plus 145 supplemental, late 2025 | Self-reported output per worker gain in 2025 | 1.8%, with revenue-implied gains much smaller |
| PwC Global CEO Survey 2026 | 4,454 business leaders | Higher revenue or lower costs from AI | 56% saw neither |
| Google Cloud ROI of AI 2025 | 3,466 leaders, all already deploying generative AI | ROI on at least one use case within a year | 74% reported ROI |
The vendor survey is not wrong. It answers a narrower question. "One use case paid back" and "the firm got more productive" can both be true at the same company in the same year.
I have argued before that vendor ROI surveys share a methodology gap, and this comparison is the cleanest illustration of it. The 63-point spread between the NBER ceiling and the Google Cloud figure is mostly sample selection and success definition, not disagreement about reality.
How much weight the NBER survey can bear
A representative survey of executives is the best firm-level evidence available. It is still a survey of executives. The table separates what the paper supports from what people use it to claim.
| Claim | Supported? | Why |
|---|---|---|
| Most firms could not see an AI productivity effect through early 2026 | Yes | Direct reading of the 89%, from representative samples |
| Average realised gain is small, around 0.2% a year | Yes, with caveats | Revised estimate, self-reported, averaged over adopters and non-adopters |
| AI does not raise productivity | No | Detectability and adoption timing explain much of the zero |
| A 0.6% annual boost is coming | No | That is the respondents' forecast, and executives forecast more than workers |
| Your firm's AI tools are not paying back | No | An average across thousands of firms says nothing about your deployment |
The macro picture lines up with this reading. The gap between firm-level reports and national statistics is the subject of the AI productivity paradox in GDP data, and the profit-line version of the same gap appears in the gap between AI spending and EBIT impact.
Where this reading of the survey is weakest
I have argued that the 89% is credible and modest. Here is where that position is most exposed.
Self-report cuts in an unknown direction
Every number in this post is what an executive said. The Atlanta Fed's CFO Survey analysis shows the problem from the other side. CFOs self-reported a 1.8% productivity gain from AI in 2025, yet the gain implied by their own revenue and headcount figures was much smaller. The authors call it a productivity paradox and point to delayed output and quality gains that revenue does not yet capture.
So self-reports can run high as well as low. I cannot tell you which bias dominates in the NBER sample, and neither can the paper. That is a genuine hole in the case.
Quality gains are invisible to this question
A productivity question anchored on output per worker misses faster response times, fewer errors and better documents sold at the same price. If AI mainly improves quality, a firm-level productivity question will undercount it for years. The J-curve view of general-purpose technologies predicts exactly this pattern, and the survey cannot rule it out.
The two waves are not a clean time series
The winter wave covered four countries and the revision covers seven, with a different sample size and season. Comparing 1.5 hours with 2.1 hours, or 69% adoption with 78%, mixes real change with sample change. I have quoted both because the authors do, and you should treat the trend as directional.
What the survey means for your AI budget
For a founder or director, the paper is most useful as a calibration tool, not a verdict. If you are planning on the assumption that AI will lift output per employee by double digits this year, you are planning outside every representative estimate published so far.
The concentration finding matters most. If 2% of US firms account for over half of AI spending, then the typical firm is a light user. The evidence on where measurable generative AI return has shown up says the same: returns cluster in narrow, well-instrumented use cases.
My recommendation is to stop asking "is AI making us more productive" at firm level. You will get the same answer the NBER respondents gave. Ask it of one workflow with a recorded baseline, where a 20% change is visible, and let the firm-level number follow. The evidence on AI time savings in work hours shows why task-level gains do not automatically reach the income statement.
Watch the next revision for three things. First, whether the realised rate moves off 0.2% as the look-back window drops pre-adoption years. Second, whether the executive and worker forecasts converge. Third, whether the hiring channel shows up in the employment answers before productivity moves.
Frequently asked questions
What did the NBER AI productivity survey find?
NBER Working Paper 34836 surveyed nearly 6,000 executives in the US, UK, Germany and Australia between November 2025 and January 2026. 89% reported no AI impact on labor productivity at their firm over the prior three years, and more than 90% reported no employment effect. A revised version with nearly 9,000 executives estimates a realised gain of 0.2% a year.
What is NBER working paper 34836?
It is "Firm Data on AI", by Ivan Yotzov, Jose Maria Barrero, Nicholas Bloom, Steven J. Davis and co-authors at the Bank of England, Atlanta Fed, Bundesbank and Macquarie University. It was first issued in February 2026 and revised in October 2026. It pools identical questions asked through existing central-bank and university business surveys in several countries.
Does the NBER survey prove AI does not improve productivity?
No. It shows most executives could not see a firm-level effect over a three-year window that began before most firms adopted generative AI. About 31% of the winter sample was not using AI at all. Small, recent or quality-based gains would register as no impact, so the result supports "small so far" rather than "no effect".
How much do executives expect AI to raise productivity?
In the winter wave, executives forecast a 1.4% productivity gain over three years, with output up 0.8% and employment down 0.7%. The revised paper puts the expectation at 0.6% a year. US workers expected less than US executives, 0.9% against 2.3%, so treat the executive forecasts as an upper estimate.
Why do vendor AI ROI surveys show much higher success rates?
They usually sample only firms already deploying AI and count success on any single use case. Google Cloud's 2025 study, for example, reported 74% achieving ROI within a year among 3,466 leaders at firms with generative AI deployed. The NBER survey asks a broader sample about firm-wide productivity, which is a much harder bar.
How should a company measure enterprise AI productivity?
Measure one workflow at a time against a baseline recorded before deployment. Track output per hour and error or rework rates for that task, plus hiring in the affected team. Firm-wide productivity is too noisy to show a small AI effect, which is exactly why 89% of executives in the NBER survey reported none.
Where to start this week
Take the single AI tool you spend the most on and write down the one task it was bought to speed up. Find the throughput figure for that task from before rollout. If no baseline exists, record one now from a team not yet using the tool, and compare in 60 days.
Then reread your 2027 plan for any line that assumes firm-wide AI productivity gains above 1% a year. Every representative estimate published so far sits below that. Keep the assumption only if you can name the workflow that will deliver it.
Related in this series
For the macro view, read why AI productivity is not showing in GDP. For the vendor side, see the methodology gap in AI ROI surveys.
References
- NBER, Firm Data on AI, Working Paper 34836, Yotzov, Barrero, Bloom, Davis et al., issued February 2026, revised October 2026. Used for the 9,000-executive revision, 0.2% and 0.6% figures, spending concentration and executive hours.
- NBER Digest, Global Evidence on Business Use of AI, May 2026. Used for the 89% figure, adoption shares, forecasts, fieldwork dates and the US worker comparison.
- Federal Reserve Bank of Atlanta, Firm Data on AI, Working Paper 2026-3, 24 March 2026. Used for executive AI use and author affiliations.
- Firm Data on AI project, survey dashboard, accessed 11 November 2026. Used for the names of the national survey panels.
- Federal Reserve Bank of Atlanta, How Might AI Change the Workplace? Evidence from Corporate Executives, 25 March 2026. Used for the CFO Survey 1.8% self-reported figure and the revenue-implied comparison.
- Google Cloud, Google Cloud study reveals 52% of executives say their organizations have deployed AI agents, 4 September 2025. Used for the 74% and 88% ROI figures and sample.
- The Register, Majority of CEOs report zero payoff from AI splurge, 20 January 2026. Secondary coverage of PwC's 2026 Global CEO Survey, used for the 56% and 12% figures.
Weakest point in the source base: every productivity figure here is self-reported by executives, the exact per-country question wording was read only from summaries, and the PwC figures come from press coverage rather than PwC's own release. Figures last verified 11 November 2026.
Related reading