Britain's businesses doubled their AI use. The structure of that use is the story.
The government's second AI adoption survey interviewed 3,503 UK businesses in early 2026. Read against the three arguments of Making AI Work for Britain, it shows demand that is fragmented and largely invisible, supply that has concentrated on one provider, and barriers that sit in capability and trust rather than technology.
Read the original report and data tables on GOV.UK →
Demand is fragmented, informal and mostly off the books
The book argues that GDS succeeded by consolidating demand through common platforms, standards and spend controls, and that for AI the UK has let demand fragment again.
The survey covers businesses, not government, so it does not test Whitehall's buying directly. What it shows is the same pattern repeated across the economy: many small, fast, uncoordinated decisions, made with free tools or with AI features that arrive inside software firms already pay for.
How AI-using businesses use AI
% of businesses that use AI (unweighted base 1,224). More than one answer allowed.
Show as table
Source: survey question B11, data table 50.
What would make non-adopters more likely to adopt
% of businesses neither using nor planning to use AI (unweighted base 2,041). Prompted, more than one answer allowed.
Show as table
Source: survey question C5, data table 64.
Reading it against the book
The things non-adopters say they need most are shared goods: proven business cases in their industry, a clearer picture of what AI can do, evidence from peers, clearer rules. No single small firm can produce these for itself. A coordinated buyer and standard-setter can.
The spend figures matter for the same reason. When nearly half of AI activity costs nothing and a quarter of firms cannot say what they spend, there is no point in the system where demand is visible, let alone aggregated. That is the private-sector mirror of the missing AI spend controls the book identifies in government.
Only about 2% of businesses meet the survey's definition of intensive adoption. Use is widespread but shallow: most firms are experimenting rather than scaling.
Supply has concentrated on a single provider
The book argues that AI supply has reconcentrated into a handful of US foundation-model providers, and that exit-by-design and multi-vendor architecture are the defence against silent lock-in.
Generative AI models in use
% of AI-using businesses (unweighted base 1,224) and of intensive adopters (unweighted base 101). More than one answer allowed.
Show as table
Source: survey question E6, data table 76. "Other models" 10% (intensive adopters 15%) not shown.
OpenAI models are used by 92% of AI-using businesses, and that share barely changes by firm size (85% to 94%). Google follows at 38%, Anthropic at 14% and Meta at 10%. Mistral, the only European provider named in the survey's list, is used by 2%. No UK-based model provider appears in the list of options at all.
Reading it against the book
The concentration the book describes at national level is visible inside individual businesses. The direction of travel matters most. Firms are moving from standalone tools towards embedding AI in their existing systems, which is where dependencies become hard to see and expensive to unwind. That is the book's silent lock-in, happening one firm at a time.
Two findings point the other way. Intensive adopters use a much wider mix of providers (Google 53%, Anthropic 31%, Meta 20%, DeepSeek 15%) and are far more likely to build their own solutions (51% plan bespoke builds against 24% overall). Deeper adoption goes with more diverse supply. And in follow-up interviews, firms worried about data protection asked for local or on-premises hosting and contractual protections, which is a demand signal for the audit-capable, exit-friendly options the book recommends.
The tables do not show how many firms rely on OpenAI alone. Responses sum to about 1.8 model families per AI-using business, so multi-provider use is common, but its distribution is not published.
The barriers are capability, trust and clarity, not technology or cost
The book argues for a third road between US market speed and EU-style comprehensive regulation: build capability where it matters, set standards others adopt, and use UK institutions as proving grounds for accountable AI.
What prevents businesses adopting AI
% of all businesses (unweighted base 3,503). Prompted, more than one answer allowed.
Show as table
Source: survey question C2, data table 59.
The four most-cited barriers are about knowing how to use AI and whether it pays: skills, understanding of use cases, a business case, and accuracy. Regulation, intellectual property and trust in product safety form a second tier, each cited by around two in five firms. Cost of tools or finance comes near the bottom.
AI adoption by sector, 2025 and 2026
% of businesses using at least one AI technology. Unweighted 2026 bases from 101 (Finance) to 640 (Business services).
Show as table
Source: data table 27 (2026); report Figure 2 (2025). Finance and real estate were reported together in 2025, so no comparison is shown.
Reading it against the book
The adaptive path puts its weight on skills, assurance and standards. That is where the survey puts the barriers. Businesses are not asking for a lighter touch, and they are not asking for heavier rules: they are asking for clarity they can act on and evidence they can trust.
The gaps are by sector and size more than by place. Adoption ranges from 23% in hospitality to 61% in information and communication, and from 27% of micro firms to 52% of large ones. Across English regions it varies only between 30% and 35%. Regional skills programmes will need a sector and SME lens to reach the firms furthest behind.
On work, firms mostly expect AI to change tasks rather than remove jobs: 58% of users automate parts of jobs, 7% whole jobs. Among intensive adopters, 38% plan to automate whole jobs in the next year. The leaders show where the rest may follow.
What the survey says about the tracked recommendations
UK AI Watch rates government delivery of the book's 25 recommendations. This survey cannot measure that delivery, because it excludes the public sector. It does test whether the problems the recommendations address are real in the wider economy. Status shown is the September 2026 review.
| Recommendation | Status | Survey evidence | Read-across |
|---|---|---|---|
| 1. Establish an AI Coordination AuthorityGovernance & Institutions | Partial | 47% of firms that could estimate spent nothing on AI in 2025; 26% could not say what they spent; only 2% are intensive adopters. | Indirect Shows how invisible and fragmented AI demand is without a coordinating point. Says nothing about government's own structures. |
| 7. Require exit-by-design provisionsProcurement & Contracts | Not started | 92% use OpenAI models; 63% buy ready-made software; embedding in existing systems planned by 54%, up from 40% today. | Strengthens the case Dependency is deepening as AI moves into core systems. |
| 9. Sovereignty by default for sensitive dataInfrastructure & Sovereignty | Not started | 47% cite data protection as a barrier (62% of large firms); interviewees asked for local or on-premises hosting. | Strengthens the case Shows demand for controllable hosting beyond government. |
| 10. Mandate multi-vendor, exit-friendly architecturesInfrastructure & Sovereignty | Not started | Intensive adopters use a wider mix of providers and plan bespoke builds at twice the overall rate (51% vs 24%). | Supports The most capable users already behave multi-vendor. |
| 12. Data residency clauses in major AI agreementsInfrastructure & Sovereignty | Partial | Privacy is the barrier most often named as the single most significant by large firms (22%). | Supports Residency is a live commercial concern, not only a public-sector one. |
| 14. Regional digital skills consortiaSkills & Workforce | Partial | Skills barrier at 65% overall, 78% among planners; English regional adoption varies only 30% to 35%. | Supports, with a caveat The need is real, but the gap is by sector and firm size more than by region. |
| 15. Introduce an AI skills levy with SME ring-fenceSkills & Workforce | Not started | Micro firms adopt at 27% against 52% for large; 41% plan reskilling, 66% of intensive adopters. | Strengthens the case The SME ring-fence targets the firms furthest behind. |
| 19. Named human accountability for algorithmic decisionsEthics & Accountability | Not started | 81% apply at least some human checking, down from 84%; no checking up from 2% to 5%; policies mostly informal. | Indirect Oversight is eroding in practice as use grows, which argues for making accountability explicit. |
| 24. Position AI assurance as a UK export industryInternational Positioning | Partial | 37% don't trust AI product safety or transparency; 50% of non-adopters want greater trust; 48% worry about bias or accuracy. | Strengthens the case There is a domestic market for assurance before any export market. |
| 25. Algorithmic transparency as a UK standardInternational Positioning | Partial | 43% cite concerns about regulation or industry standards; interviewees describe AI as a "black box". | Supports Firms want a clear standard they can point to. |
Status from UK AI Watch, September 2026 review. The other 15 recommendations concern public-sector governance, procurement and civil-service capability that a business survey does not touch.
Method and limits
- Source: Artificial intelligence adoption research 2026 (October 2026), with its published data tables, available on GOV.UK. Fieldwork by BMG Research: 3,503 telephone interviews, 7 January to 21 April 2026, plus 50 follow-up interviews and 7 case studies.
- Coverage: UK businesses with 5 or more employees. Public administration and defence, education, and health and social work are excluded, so the survey says nothing direct about public-sector AI.
- Results are weighted by business size and sector. All responses are self-reported, and most questions allow more than one answer, so percentages do not sum to 100.
- Some bases are small: large firms (130 interviews), intensive adopters (101), Finance (101). Treat figures for these groups as indicative.
- Figures here are taken from the data tables. In a few places the report text differs slightly because it uses a different base (for example, 65% expecting AI budgets to rise excludes "don't know"; the table figure including it is 62%) or, in one case, appears inconsistent (real estate adoption is given as both 42% and 45%; the table shows 45%).
- "Intensive adopters" is the report's own definition: firms that build bespoke AI or embed it in their processes, rate it as highly important, and use it constantly.
- The readings against the book are the author's interpretation. The survey describes business behaviour; it does not test the book's recommendations, which concern government.