Firm Data on AI

Firm Data on AI (NBER Working Paper 34836, October 2026 revision)

TL;DR

An NBER working paper by seventeen economists at central banks and universities, led by Ivan Yotzov (Bank of England), Jose Maria Barrero (ITAM) and Nicholas Bloom (Stanford). They put the same AI questions to senior executives through central-bank business surveys in seven countries. This is the third version (“February 2026, Revised October 2026”), and its headline figures differ from the first two.

1. 78% of firms use some AI technology (seven-country average, May-July 2026; p. 15). At the start of 2026, in the first wave, the figure was 69% (p. 4). Country figures run from 95% (US) to 56% (South Africa).
2. AI spending is very concentrated. US firms expect to spend about $2,000 per employee on AI in 2026. 1.8% of US firms and 1.3% of UK firms account for half of all AI spending (p. 19). The abstract rounds this to “just 2% of firms accounting for over 50% of total spending” (p. 2). This is a US and UK finding only.
3. Executives report almost no employment effect so far. 87% report no impact over the past three years, and the average is about zero (p. 21). They expect a small fall, -0.4%, over the next three years, but more than 40% expect some change either way: churn rather than net loss (pp. 21-22).
4. Executives report a small productivity effect so far, and expect more. AI raised their firms’ labour productivity (sales per employee) by about 0.5% over three years, or “0.2% a year” (pp. 2, 5, 23; Table 4, p. 50: 0.46%). They expect 1.7% over the next three years, about 0.6% a year (p. 23). These are self-reports.
5. The paper explains why adoption figures disagree. Using UK data, it shows that firm size mix, weighting scheme and the respondent’s position each move measured adoption by “between five and twelve percentage points”, and that they compound (pp. 16-17).

Key data

Measure	Figure	Definition	Page (PDF)
Firms using any AI technology	78%	Sample-size-weighted average of seven surveys, N = 8,879, May-July 2026	15, 33
By country	US 95, UK 88, Canada 82, Germany 78, Australia 68, Mexico 64, South Africa 56	US, UK, Germany, Mexico employment-weighted; others equal-weighted	15, 33
Same, first wave (Nov 2025-Jan 2026)	69% (four countries)	As published in the February and March 2026 versions	4; March version p. 12
Generative AI as the use	63% of firms	Most common category	15
Firms expecting to use AI within 3 years	82%	N = 8,610	17, 59
US AI spending per employee	$1,363 (2025) → $2,043 (2026)	Employment-weighted; software, subscriptions, hardware, training, IT support	18
Aggregate US AI spending	~$280 bn (body); ~$300 bn (abstract)	Grossed up from the SBU	2, 18
Share of firms making half of all AI spending	1.8% (US), 1.3% (UK)	2026	19
Executives using AI themselves	80%; 2.1 hours a week (1.5 in early 2026)	Own use in a typical working week	2, 19-20
No AI effect on own-firm employment, past 3 years	87% of firms	Five-category scale	21
Expected employment effect, next 3 years	-0.4% (US and UK -1.3%)	Categories coded ±7.5%, ±2.5%, 0	21
No AI effect on productivity, past 3 years	83% of firms (89% six months earlier)	Sales per employee	5, 23
Realised productivity effect	+0.46% cumulative (≈0.2% a year)	Table 4, all firms	2, 50
Expected productivity effect	+1.7% over 3 years (≈0.6% a year)		23
UK growth accounting, past 3 years	+0.17 pp a year on GVA per worker; +0.13 pp on TFP	Table 3	24-25, 49

Method and definitions

• Who answered. Senior executives (CEOs, CFOs, senior finance managers) in existing central-bank and university panels: the Atlanta Fed’s Survey of Business Uncertainty (US), the Bank of England’s Decision Maker Panel (UK), the Bundesbank Online Panel (Germany), Macquarie University’s BOSS (Australia), the Bank of Canada’s Business Leaders’ Pulse, Banco de México’s EMAER and the South African Reserve Bank’s survey (pp. 8-14). Respondents are unpaid and identity-checked, which the authors set against paid online executive panels (p. 10).
• Fieldwork. Wave 1: November 2025 to January 2026 (US, UK, Germany, Australia). Wave 2: May to July 2026, all seven countries (pp. 4, 9). Germany’s employment and productivity figures still come from January 2026 (pp. 20, 22). A Bank of Korea survey of 749 firms (89% adoption) is reported apart, in Appendix C (p. 84).
• The adoption question. “Which of the following artificial intelligence technologies, if any, does your business currently use?” Firms tick any of: data processing using machine learning, visual content creation, generative AI, image processing, robotics and autonomous vehicles, other (pp. 15, 33). The question asks about current use, not use in production.
• Category change between versions. The first two versions asked about “text generation using large language models” and listed robotics and autonomous vehicles separately (March version, p. 12). This version asks about “Generative AI” and merges the two.
• Effects are executives’ own estimates on a five-point scale, coded to ±7.5%, ±2.5% and 0 (pp. 20, 23). Productivity means sales per employee.
• Funding: ESRC, Smith Richardson Foundation, Stanford Impact Labs (p. 1).

Scope and reliability

• Self-report by executives. Adoption, spending and effects are what executives say about their own firm. The paper finds CEOs and CFOs report adoption four to seven points higher than other respondents, and cannot tell whether they see more or claim more (p. 17).
• The headline moved between versions for two reasons. New data (wave 2, three more countries) and a coding change. In the first version, US firms that skipped the adoption question counted as non-users, giving 78%; this version treats them as missing, giving 94% (November 2025) and 95% (June 2026) (footnote 13, p. 15). So the US figure in the Adoption Monitor (77.5%) and this paper’s US figure (95%) describe the same survey under different coding.
• Weighting is mixed. Four surveys are employment-weighted, three equal-weighted, and the seven are then averaged by sample size (p. 33). The 78% is not a population-weighted share of firms. On UK data, employment weighting gives 88%, business-count weighting 83% (p. 16).
• The sample favours larger firms than official statistics do. The UK panel covers firms with ten or more employees; Mexico’s covers establishments with more than 100 workers (pp. 8, 13).
• The employee survey has gone. The March version contrasted executives with US employees, who expected AI to raise employment by 0.5% (March version, p. 2). That comparison is not in this version.
• Working paper, not peer-reviewed (p. 1).

Dynamic-capabilities reading

• digital-sensing/digital-scouting: a standing, twice-yearly scan of how many firms use AI, how much they spend and what executives expect, across seven countries.
• digital-transforming/improving-digital-maturity: the paper measures how far firms have got (use, spend per employee, executive use), and finds the intensity highly concentrated: a small share of firms makes most of the investment.

How it connects

• enterprise-ai-adoption and the thread ai-maturity-measurement-comparison: an executive-survey adoption figure (78%) far above official firm statistics, with the paper’s own account of the gap.
• The Fed’s FEDS Note reports the same US survey’s November 2025 wave as 78% employment-weighted, against the Census BTOS’s 18% of firms. The two differ on how the US non-response is coded.
• The Adoption Monitor reports this paper’s first wave. Its 68.6% (four countries) and this version’s 78% (seven countries) are different waves, samples and codings.
• Rendtorff-Smith (OECD) gives 20% of firms in 2025, counted by firm number. The paper’s size and weighting analysis bears directly on that gap.
• ai-employment-effects: executives report near-zero net employment change and expect churn. Humlum and Vestergaard find null effects on earnings and hours in Danish register data.
• micro-productivity-trap: a realised self-reported gain of about 0.2% a year. Hatzius (Goldman Sachs) puts AI investment’s GDP contribution at about 0.1 point, and the Transformation Tracker rates productivity growth neutral.
• Eurostat, Use of AI in enterprises: official EU statistics put firm adoption at 19.95% in 2025, counting every firm with 10 or more persons once.

What was actually ingested

The October 2026 revision (86 pages): abstract, all of the body (pp. 2-24), the conclusion, Tables 3 and 4, the notes to Figures 4A, A4, A7 and A8, and Appendix C. Figure values were read from the PDF text layer, not from the chart images. For the reconciliation with the Adoption Monitor, the March 2026 version (70 pages) was also read for its abstract, introduction and section 4.1. The February version was checked only for its abstract and section 4.1 figures, which match the March version.

Linked entities and concepts

• Concepts: enterprise-ai-adoption, ai-employment-effects, micro-productivity-trap, technology-adoption-theories.
• Entities: Stanford Digital Economy Lab (Bloom is its featured affiliate on the Adoption Monitor).
• Dangling (single-source mention, deferred): Nicholas Bloom (also named on the Adoption Monitor, the Canaries Dashboard and the RaboResearch table, but in author: only here), Ivan Yotzov, Jose Maria Barrero, Steven J. Davis, Brent H. Meyer, the other twelve co-authors, National Bureau of Economic Research, Federal Reserve Bank of Atlanta, Bank of England, Deutsche Bundesbank.