Figure 1 Generative AI use, people aged 15–64, Q2 2026. Microsoft AI Economy Institute.
28.8%Global North
16.2%Global South
12.6point gap, up from 10.6
Executive summary
How this report is made. AI helps gather sources, organise the report and summarise findings. I select the sources, check the summaries against the evidence and make the final editorial decisions. — Alessandro L. Piana Bianco. Sources & methods AI use is spreading, but access, spending and frequent use tell us little on their own about the quality of that use. Twenty-one source groups trace the conditions that turn uptake into capability: organisational change, support for learning, the tasks people delegate, and the authority they are willing to give AI.
- The geographic gap widened. Microsoft estimates that in Q2 2026, 28.8% of people aged 15–64 in the Global North used generative AI, compared with 16.2% in the Global South. Wider access does not necessarily distribute control over the infrastructure behind AI.12.6 pointsNorth–South gap
- Uptake needs organisational change. Large EU firms remain more likely to use AI than small firms. Kyndryl’s leader survey adds a different lens: deployment, redesigned roles and change management need to be examined separately.55% / 17%large / small EU firms
- Spending growth is uneven. In Menlo’s US enterprise sample, average AI spending growth was much higher than median growth. Spending is an input, not evidence of wider adoption or improved productivity.4.7× / 2.2×mean / median spending growth
- Exposure does not confer editorial authority. AI increasingly mediates access to information. In Reuters’ nine-country sample, people were more comfortable with human-produced news using AI assistance than news produced entirely by AI.45% / 12%comfort with human-led / AI-only news
- Workplace effects need separate measures. US AI-using employees report faster work and added responsibility. In Australia, most businesses covered by the ABS question reported no impact on vacancies. Neither finding establishes jobs lost or economy-wide productivity gains.81.8%no reported vacancy impact, ABS
- Student use outpaces university provision. Nearly all UK undergraduates surveyed had used AI. Fewer than four in ten said their university provided generative AI tools. Guidance and preparation also remain distinct questions.95% / 38%use / tool provision
- Science reveals a workflow pattern. Analysis and writing lead Anthropic’s research-session shares. This shows where AI enters research, not how many scientists use it or whether it produces better science.8 taskstwo Claude interfaces
These studies cover different populations, periods and measures. Their figures cannot be combined into a single adoption rate or a causal account of AI’s effects.
01Access & geography
Reach is growing. The divides persist.
1 · Microsoft Adjusted telemetry estimates use, not skill or productivity. The global denominator is people aged 15–64. World user share: 16.3% (H2 2025), 17.8% (Q1 2026), 18.8% (Q2 2026); periods of unequal length. Microsoft estimates that 18.8% of people aged 15–64 used generative AI in Q2 2026, up from 16.3% in H2 2025. But a rising global average conceals a persistent divide: use in the Global North reached 28.8%, compared with 16.2% in the Global South.
The distance between those regions widened from 10.6 to 12.6 percentage points. Diffusion is advancing without becoming evenly distributed.
A different geographic lens. Anthropic’s country index tracks Claude conversation activity relative to population. A value of 1× means proportional activity; it is not a percentage of people. 2 · Anthropic National averages hide local divides too. In Q1 2026, Microsoft estimated average AI user shares of 32.9% in metropolitan counties and 16.2% in rural counties. Access to AI is shaped by the places and institutions through which people encounter it.
Frequency and exposure tell different stories.
10 · Reuters Institute YouGov weighted nonprobability online survey; 2026 fieldwork 14 August–28 September. Six-country bases: 12,565 (2025), 12,774 (2026); nine-country base: 18,921. Self-reports, not directly comparable with Microsoft. The prompted tool list changed from 13 to 14. Figures 1 and 5. The Reuters Institute’s online survey finds weekly use of standalone generative AI tools rising from 34% in 2025 to 51% in 2026 across the same six countries. Separately, 57% across nine countries encountered AI-generated answers or assistants in search engines in the previous week, including incidental exposure within search. These are different measures and country sets, not an exposure gap or a worldwide estimate.
Reach does not establish acceptance.
10 · Reuters Institute Nine-country weighted sample, n=18,921; 14 August–28 September 2026. Very or somewhat comfortable. Separate scenarios, not parts of a total or a worldwide estimate.Wider exposure does not mean equal acceptance of every use. Across the Reuters Institute’s nine-country sample, 12% were comfortable with news produced entirely by AI, compared with 45% for mostly human-produced news with AI assistance and 67% for entirely human-produced news. These are attitudes towards production methods, not measures of news consumption, accuracy or trustworthiness.
02Organisations & delivery
Using AI is easier than integrating it.
3 · Eurostat & JRC EU firms: small = 10–49 persons; large = 250+. The survey covers selected activities and broad AI, not only generative AI. In the EU, 55.03% of large enterprises used AI in 2025, compared with 17.00% of small enterprises. Firm size is associated with a substantially different level of adoption.
4 · McKinsey “Large” instead means annual revenue above $1bn. Survey respondents, not a census of firms. A separate McKinsey survey finds a similar divide in agent scaling: 40% of respondents in large organisations report scaling agents in at least one function, versus 22% in smaller organisations. Access to a tool and integration into an organisation are different milestones.
Two questions, two outcomes. Self-reported productivity and attributed financial effects cannot establish a causal return on AI investment. Nor do personal gains automatically establish organisational returns. In McKinsey’s survey, 80% report improved personal productivity; 37% attribute some positive organisational EBIT contribution to AI. These answers point to different levels of change, rather than a measured conversion from productivity into profit.
A separate UK survey finds a size gradient too.
9 · UK Government 3,503 telephone interviews; UK private-sector businesses with 5+ employees, SIC A–N and R–S; 7 January–21 April 2026. Size/sector weighted; 5% response. The skills-barrier base is 236 nonusers planning adoption. Published 29 October. Self-reports, not causal effects. Tables 27 and 61. The UK government’s 2026 survey finds 32% of eligible businesses using AI, including pilots. Use ranges from 27% among firms with 5–9 employees to 52% among those with 250 or more. These are broad-AI measures; the population and sector coverage differ from Eurostat’s, so the figures do not form a UK–EU ranking.
Among nonusers planning adoption, 78% cite limited AI skills, expertise or knowledge. Use and the capability to adopt remain different questions.
Co-use adds another layer.
12 · OECD Figure 3.4; published 5 November 2026. Equal country weight; national definitions and survey coverage vary. European cloud measures cover purchased intermediate or sophisticated services. The OECD’s new outlook finds that around 12% of enterprises with at least ten employees use AI, cloud computing and data analytics together. This is an unweighted mean of 28 covered country rates, mainly for 2025 with older observations in some countries. It shows technologies being used alongside one another; it does not establish integration within the same workflow or causal returns.
13 · US Census BTOS; released 5 November. Collection: 19 October–1 November. Voluntary, weighted business survey. The instrument changed to “Super Intelligence or SI”; its effect on comparability is unknown. No trend is inferred. A separate US snapshot puts AI use in a business function at 24.2% of covered nonfarm employer businesses during 5–18 October 2026. In another question, 26.8% of businesses reported employees using generative AI for work. Both are shares of businesses, not employees. The questions overlap and differ from the UK and EU measures, so they do not support a direct ranking.
Geography matters within countries too.
11 · OECD · Eurostat 2025 data; published 4 November 2026. Figure 1.9. Equal weight per region; limited country coverage. The workbook does not provide regional sample counts or confidence intervals. The OECD’s 2026 regional report adds a spatial view of business adoption. 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.
The analysis uses Eurostat’s existing business survey data. It adds geographic detail rather than an independent survey, and covers broad AI use by enterprises with at least ten persons employed in selected nonfinancial sectors.
Deployment and organisational redesign are different changes.
17 · Kyndryl Vendor-sponsored survey of 3,700 senior leaders in 21 countries, fielded 10 July–2 September 2026. Half represented firms with revenue of at least $1bn. Different agreement thresholds; not representative national adoption rates or causal effects. Kyndryl’s November survey gives another view of integration. 84% agreed that AI was deployed beyond pilot projects, while 27% completely agreed that roles had been redesigned around AI and 29% completely agreed formal change-management programmes were fully implemented. The questions use different response thresholds, so subtracting these shares would manufacture a measured gap. Their useful contribution is narrower: deployment and changes to roles and routines need to be examined separately.
Read the organisational-redesign evidence
Spending is a different dimension of diffusion.
18 · Menlo Ventures Published 3 November 2026. Morning Consult surveyed nearly 400 US enterprise decision-makers on 8–26 September 2026. Menlo does not disclose the recruitment frame, response rate or representative weighting. Menlo invests in companies discussed in the report.Menlo’s survey reports AI spending at 4.7 times its prior-year level on average, versus a median of 2.2 times. The distance between them suggests spending growth is uneven within the sample. It does not measure how widely AI is adopted or whether it improves productivity.
03Work & job quality
Faster work can mean more work.
5 · Gallup January–March 2026. Effects are reported by 7,845 US employees who used AI at least a few times in the previous year. Answers overlap. These are perceptions, not measured causal productivity gains. Among US employees who use AI, 63% say it helps them work faster. Many also report more creativity and better-quality work. At the same time, 31% say their employer asks them to take on more responsibility.
Speed is only one dimension of job quality. The question is also who gets access, how expectations change, and whether workers gain time for work they value.
Outputs are not outcomes.
2 · Anthropic Model-classified conversations from one provider. Excludes Claude Code and API. Percentages sum to 100.01% because of rounding. Anthropic’s conversation data offers a view of what people ask AI to do. Work accounts for 43.36% of classified conversations, personal uses 40.20%, and coursework 16.45%. That range helps explain why one adoption rate cannot tell us what AI is changing.
Australia adds a recruitment test.
15 · Australian Bureau of Statistics Reference date: 21 August 2026. AI results added 4 November. Excludes firms with fewer than five employees and item nonrespondents. Final survey weights, with no nonresponse adjustment; response bias is possible. No causal inference. The Australian Bureau of Statistics asks businesses how AI has affected their vacancies. For August 2026, 81.8% reported no impact, 3.8% fewer vacancies, 0.6% more and 13.0% were unsure. These are reported effects on recruitment, rather than a measure of AI adoption or jobs lost.
| Reported effect on vacancies | Share of businesses | |
|---|---|---|
| No impact | 81.8% | |
| Fewer vacancies | 3.8% | |
| More vacancies | 0.6% | |
| Unsure | 13.0% |
The four published shares total 99.2%. The source leaves that discrepancy unexplained; the explorer preserves the figures without adjusting them. See the Australian values and qualifications
Productivity gains are still uneven.
16 · Reserve Bank of Australia Graph 2; approximate share as described by the RBA. The liaison program mainly covers medium and larger firms. The prior-year comparability warning applies to Graph 1. Read the context In its 9 November speech, the Reserve Bank of Australia reports that around 35% of firms responding to its AI liaison questions had seen no productivity gains yet. The responses cover June–October 2026. This adds a firm-level view of the gap between using AI and gaining from it.
Liaison respondents are not a representative national firm survey. Their reported experience cannot establish an aggregate productivity effect. The RBA’s adoption comparison with the previous year also uses changed samples and questions.
AI fluency is entering job requirements.
20 · McKinsey Global Institute November 2025 report, Exhibit 7; Lightcast postings and BLS employment. Workers in occupations meeting a threshold, not AI users or vacancies. An older labour-market signal adds a different perspective. US occupations in which at least 5% of job postings requested AI fluency employed about 1.0 million people in 2023 and 7.0 million in 2025, according to McKinsey Global Institute. This indicates the spread of AI-related requirements across occupations, not that all those workers use AI or have acquired the skills.
Creative work raises questions of authority too.
21 · Dubai Future Foundation 2024 qualitative report; workshops and expert consultation, October 2023–September 2024. Not a representative survey. Dubai Future Foundation’s creative-industries consultation foregrounds authorship, compensation and cultural representation. These are not adoption measures. They make another condition of useful diffusion visible: who has a say in how creative work is used, credited and rewarded.
04Learning & literacy
Use is widespread. Support is uneven.
6 · HEPI & Kortext 2026 edition: 1,054 full-time UK undergraduates, surveyed in December 2025. Provision is student-reported. The 2026 use question added wellbeing options; the editions are not perfectly like-for-like. In HEPI’s 2026 survey edition, 95% of UK undergraduates had used AI for at least one purpose. Only 38% said their institution provided generative AI tools. Both measures have risen, but widespread personal use still sits alongside less extensive institutional provision.
The distinction matters: using AI is not the same as being supported to use it well. The teacher and pupil evidence brings that missing layer into focus.
Data
| edition | Any AI use | Tools provided |
|---|---|---|
| 2024 | 66% | 9% |
| 2025 | 92% | 23% |
| 2026 | 95% | 38% |
Guidance is not the same as preparation.
8 · Dartmouth 1,003 US four-year college students; College Pulse weighted online panel, 10–23 September 2026. Agreement or strong agreement; working paper released 30 October. Different questions and population from HEPI; no direct cross-country comparison. Table 12, PDF p. 22. A new US college survey adds another view. In Dartmouth’s weighted sample, 79.3% say they mostly developed their AI skills independently, while 68.1% report meaningful guidance from professors and courses. These answers can overlap. Only 49.5% say their institution adequately prepares them to use AI in their future career.
Receiving guidance and feeling prepared are different judgements. Neither establishes what students have learned.
Judgement needs attention too.
7 · National Literacy Trust Early 2026 recruited school samples. Teacher and pupil measures have different bases; pupil respondents were predominantly in England. In the National Literacy Trust’s teacher sample, 80.6% of AI-aware teachers had used AI, while 32.6% had received school training or support. Among AI-using pupils, 52.3% said they checked outputs for accuracy and 24.1% said they usually copied AI output for homework.
These figures do not measure learning outcomes. They make a more immediate issue visible: the practices around AI deserve as much attention as the rate of uptake.
05Science & research
Research changes task by task.
14 · Anthropic, science 207,049 sessions; 2 August–30 September 2026. Consumer plans, LLM classification, one provider. Session shares are not researcher adoption or measured productivity. Ideation may be undercounted. Science adds a workflow lens to diffusion. In Anthropic’s research-session sample, data analysis accounts for 33.7% of Chat/Cowork sessions and 48.1% of Claude Code sessions. Writing accounts for 32.0% and 17.6%, respectively. Data gathering and ideation make up smaller shares.
| Research task | Chat and Cowork | Claude Code | ||
|---|---|---|---|---|
| Data analysis | 33.7% | 48.1% | ||
| Writing | 32.0% | 17.6% |
The distribution shows where AI enters research. Whether that work produces more reliable findings, faster experiments or better science remains a separate question.
06Infrastructure & agency
Access is not control.
19 · Stanford AI100 November 2026 expert synthesis, covering 2021–2026; infrastructure discussion pp. 98–105. Not a population survey or a new adoption estimate. Wider access to AI does not necessarily distribute control over it. Stanford’s 2026 AI100 review describes an interdependent infrastructure of models, data, networks, chips and data centres, with frontier development concentrated among a limited set of organisations and countries.
Open-weight models offer different deployment options, but access to a model is not the same as the capacity to operate, adapt or govern it. This adds another question to the geography of diffusion: not only who uses AI, but who can shape the systems they depend on.
What to watch next
Look beyond the adoption rate.
An interpretation of the evidence. These twenty-one source groups answer distinct questions. Read together, they suggest what to investigate, not a single causal account. The useful question is no longer only “How many people use AI?” It is what makes use valuable, for whom, and under which conditions.
- Reach. Are the gaps between places and groups narrowing?
- Capability. Are organisations providing the tools, training and time to integrate AI?
- Agency. Who can operate, adapt and govern the systems they depend on?
- Outcomes. Are changes in work, learning and research actually being measured?
Diffusion describes the spread of a technology.
It leaves the quality of its use to be explained.