Sources & methods
Thirteen source groups. Different populations, periods and questions. A guide to what the evidence can—and cannot—tell us.
Updated with OECD technology co-use and US Census BTOS. All 304 earlier observations retain their values and periods. 335 ledger observations across 13 source groups; the new OECD entries include 28 country rates and their published mean.
Full evidence library: tables, reports, downloads and methods →
01 / The research
Where is generative AI used? Estimated user shares across 147 economies, H1 2025–Q2 2026, for people aged 15–64. Separate US estimates cover Q1 2026.
Adjusted telemetry and modelled local estimates measure use, not proficiency or productivity; global and US methods differ.
What happens inside AI conversations? Claude chat and Cowork geography and purpose: May 2026. Output analysis: 10 April–10 June 2026.
Classified conversations from one provider are not population adoption; the country index compares conversation share with working-age population share.
Which businesses use AI? EU enterprises with 10+ persons employed, 2021–2025; dataset snapshot updated 15 June 2026. JRC adds strategic-sector analysis.
Broad AI includes more than generative AI. Finance and several public services are excluded; strategic-sector coverage varies substantially.
How far has organisational deployment progressed? 1,719 respondents in 97 nations; fieldwork 4 May–8 June 2026, weighted by national GDP.
A self-report survey, not a firm census. Attributed profit effects are not causal estimates; annual samples differ.
What do workers experience? 15,482 US employees, surveyed 26 January–24 March 2026. Effects questions cover 7,845 AI users.
Reported speed, quality and responsibility are perceptions; effects among users cannot be read as outcomes for the whole workforce.
How does student use compare with institutional provision? 1,054 full-time UK undergraduates; weighted Savanta survey, December 2025, published in the 2026 edition.
Provision is student-reported. The 2026 use question adds wellbeing options, complicating comparisons with earlier editions.
What supports critical AI use? Early-2026 surveys: 40,543 pupils aged 13–18 and 2,567 teachers answering AI questions. A separate study covers 3,073 parents.
School recruitment and a pupil sample largely in England limit generalisation. Routed questions have different bases; parent fieldwork dates were not verified.
How does guidance relate to students’ own preparation? 1,003 students at US four-year colleges and universities; College Pulse weighted online panel, 10–23 September 2026. Released 30 October 2026.
Separate, overlapping agreement measures, not complements or an adoption rate. Weighted online-panel self-reports; working paper, peer review not verified. Design effects were not verified. No causal learning outcomes or direct comparison with HEPI. The three figures record agreement or strong agreement, with N=1,003 for each item.
Which UK businesses use AI, and what limits adoption? UK private-sector businesses with at least five employees in SIC A–N and R–S. 3,503 telephone interviews with technology-responsible respondents; Market Location commercial sampling frame; weighted by business size and sector; 5% response rate. Fieldwork: 7 January–21 April 2026; published 29 October 2026.
Broad AI, including pilots, not only generative AI. Selected sectors and a five-employee threshold differ from Eurostat; no direct UK–EU ranking. Self-reports, not a census or causal productivity evidence. Low response creates nonresponse risk; weighting does not establish its absence. The report compares 32% with 16% in the prior wave. These are repeated cross-sections, not a panel; questionnaire and achieved samples changed. The 78% skills-barrier figure covers nonusers planning adoption (unweighted n=236), not current users. Real-estate adoption is omitted: executive summary/figure 2 reports 45%, while the main text reports 42%; unresolved discrepancy.
How often do people use standalone AI, and encounter AI in search? Generative AI and News Report 2026, published 29 October 2026. YouGov online questionnaire; nonprobability samples with quotas and weighting for age, gender, region, education and political leaning. Combined-respondent estimates use YouGov weights, not an unweighted average of country percentages. 2026 fieldwork: 14 August–28 September. Six-country weekly use: Argentina, Denmark, France, Japan, United Kingdom and United States; n=12,565 in 2025 and n=12,774 in 2026. Nine-country search exposure adds Australia, Finland and Norway; n=18,921.
Self-reported use and exposure, subject to recall and interpretation; online samples underrepresent people offline. No conventional sampling margin of error; not a worldwide estimate or directly comparable with Microsoft usage estimates. Different country sets and questions mean the weekly-use and search-exposure figures must not be subtracted as an exposure gap. Weekly-use comparison covers the same six countries, not a longitudinal panel. The prompted standalone-tool list changed from 13 tools in 2025 to 14 in 2026; this qualification accompanies the trend. The search measure includes deliberate and incidental exposure within search engines only.
How does enterprise adoption vary by region? Figure 1.9 provides 91 regional rates across 12 selected countries. Reference year: 2025. Published: 4 November 2026. Enterprises with at least 10 persons employed in covered nonfinancial sectors; broad AI use.
OECD analysis of existing Eurostat isoc_r_eb_ain2 data, not an independent survey. Selected 2025 regional coverage in 12 countries; not Europe-wide. Regional achieved sample counts and confidence intervals are not supplied in the workbook. Employment exposure in column D is excluded: it has a different denominator. In the selected 2025 data, the simple average regional share of enterprises using AI was 11.43% across 43 Central and Eastern European regions, compared with 26.10% across 48 other covered regions. These are unweighted means of regional rates across 12 countries, not enterprise-weighted adoption rates for Europe.
Covered countries: Austria, Belgium, Croatia, Czechia, Denmark, Hungary, Lithuania, Norway, Poland, Slovak Republic, Spain, Sweden.
Which firms use AI, cloud and analytics together? Figure 3.4; published 5 November 2026.
The workbook’s OECD mean is 12.2162086899778%. Recomputing the simple mean of the 28 covered member-country rates reproduces it. Brazil, Croatia, Romania, Bulgaria and the EU aggregate are excluded from that calculation. This is not an enterprise-weighted rate for the whole OECD.
Reference years are mainly 2025: Korea uses 2024, Canada 2023, Japan 2021–2023, and Australia financial year 2021–2022. The population is enterprises with at least ten employees in covered sectors. National survey coverage varies; the workbook supplies no achieved sample counts or confidence intervals.
The source labels analytics “big data analytics”; European Statistical System countries and Canada’s 2023 data use the broader “data analytics” measure. European cloud data concern purchased intermediate or sophisticated services, including business software, database hosting and development platforms. Canada covers only the development-platform use case among sophisticated services. Co-use establishes neither workflow integration nor causal returns.
How do businesses report AI use in functions and employee tasks? Two national estimates, retained as separate measures.
Reference period: 5–18 October 2026. Responses were collected 19 October–1 November and published 5 November (cycle 202622). National.xlsx records 24.2% at Response Estimates E16 (business-function use) and 26.8% at E19 (employee generative-AI use reported by the business). Corresponding standard errors are 0.61 and 0.56 percentage points, not confidence intervals. Both denominators are businesses, not employees; the questions can overlap.
BTOS is a voluntary probability survey covering nonfarm employer businesses in the United States, DC and Puerto Rico. Exclusions include agricultural production, railroads, the Postal Service, central banking, funds/trusts, NAICS 813 organisations, private households and public administration. Approximately 1.2 million firms are sampled annually across six panels of about 200,000; each firm is asked every twelve weeks. Those are sampled invitations, not achieved responses. Business weights adjust for nonresponse by state/size and sector/size. Sampling error, measurement error and residual nonresponse bias remain possible.
The 19 October collection changed the instrument to “Super Intelligence or SI”, including the generative-use question. Its effect on comparability has not been established. An earlier November 2025 change broadened business use from producing goods or services to any business function. No trend is inferred here; populations, size thresholds, reference windows and questions differ from the UK and EU surveys. These two US percentages are not subtracted as an employee-use gap.
02 / Reading the evidence
No pooled adoption score is calculated. People, firms and conversations remain separate. Publication dates are not observation dates; half-year and quarterly windows have different durations. Percentage-point changes subtract comparable published endpoints, and missing values never mean zero.
Self-reported benefits do not establish causal effects on productivity, job quality or learning. The US workforce and US/UK education surveys remain evidence about their respective populations; they are not directly comparable cross-country estimates.
Country comparisons follow the supplied CSV. The Q2 report revises four Q1 values by 0.1 percentage points; original vintages remain available. Conflicting HEPI encouragement and prior-year teacher-training figures are omitted. Eurostat series breaks and category changes remain documented.
Real-estate adoption is omitted: executive summary/figure 2 reports 45%, while the main text reports 42%; unresolved discrepancy. Unlabelled curves are not reconstructed from pixels. Read the full editorial decisions and transcription notes ↗.
Research belongs to its respective creators. Microsoft’s collection carries an MIT notice; JRC uses CC BY 4.0, with exclusions. Anthropic’s dataset specifies CC-BY without a version. Other reports retain publisher rights; numerical facts are transcribed and charts redrawn. The UK government publication uses the Open Government Licence v3.0 except where otherwise stated. Reuters Institute permits reproduction under CC BY with attribution. No single licence covers the whole collection.
Microsoft’s original licence ↗ All attribution and reuse notes ↗
03 / The library
Regional values · CSV · Data and provenance · JSON · OECD workbook ↗