What the latest official evidence says about AI use, integration and control
AI use is broadening across UK businesses, but the official evidence points to a second challenge: turning access to tools into a dependable workflow with clear ownership, data boundaries, human approval and measurement.
This evidence hub brings the newest ONS and UK Business Data Survey findings into one transparent view without blending unlike surveys. It then turns the practical gap between use and integration into an eight-part workflow readiness test that a business can inspect, download and apply to one recurring process.
Published 20 August 2026. Evidence checked against the linked official sources on the same date. Next scheduled review: 20 November 2026, or sooner if a material new release appears.
Adoption percentages are moving faster than extensive use, system integration, workforce training and documented governance. For an SME, the useful next question is therefore not just “are we using AI?” but “which workflow can we improve safely, visibly and measurably?”
Five evidence signals worth knowing
AI use among businesses with 10+ employees
ONS reported around 35% in June 2026, up from around 12% in late 2023.
Source: ONS/BICS; UK businesses with 10+ employees.Described their AI use as extensive
Only one in ten AI-using businesses with 10+ employees reported extensive use.
Source: ONS/BICS; AI-using businesses with 10+ employees.Integrated AI into existing systems
UKBDS found that about one in five businesses already using AI reported system integration.
Source: UKBDS; 1,870 AI-using businesses.Had any AI policy or guidance
This consisted of 5% with a formal written policy and 12% with informal guidance.
Source: UKBDS; 1,870 AI-using businesses.Were uncomfortable with external AI training on their data
Among businesses handling digitised data, 73% were somewhat or very uncomfortable with their own data being used to train an external organisation's AI. This makes data boundaries and supplier checks part of practical adoption, not a technical afterthought.
Source: UKBDS; 4,090 businesses handling digitised data.Why the headline adoption figures differ
The ONS and UKBDS figures answer related but different questions. Both are useful, but neither should be used as a correction factor for the other.
| Feature | ONS: AI in UK businesses | UK Business Data Survey 2026 |
|---|---|---|
| Headline used here | Around 35% used at least one AI technology. | 41% used AI for at least one purpose. |
| Population | Businesses with 10 or more employees. | Businesses handling digitised data. |
| Reference period | June 2026. | Fieldwork from October 2025 to January 2026. |
| Method | Business Insights and Conditions Survey estimates. | 4,450 CATI and online interviews, weighted by size and sector; 4,090 handled digitised data. |
| Best use | Track change, extent, sector and workforce patterns in the stated population. | Examine business-size, integration, governance, data attitudes and regulatory awareness. |
| Do not do | Do not average the two headline percentages or present either one as “all UK SMEs”. | |
What changes by business size
UKBDS reported higher AI use, integration and formal policy rates among larger businesses. The denominators still matter: AI use is a share of digitised-data businesses, while integration and formal-policy figures are shares of businesses already using AI.
| Business size | AI use | Integrated into systems | Formal written policy |
|---|---|---|---|
| Sole trader | 40% | 18% | 3% |
| Micro | 41% | 27% | 8% |
| Small | 51% | 31% | 17% |
| Medium | 58% | 31% | 22% |
| Large | 82% | 57% | 56% |
Source: UK Business Data Survey 2026. “AI use” denominator: businesses handling digitised data. “Integrated” and “formal written policy” denominators: businesses already using AI. Differences describe survey estimates; they do not prove that business size caused adoption or better outcomes.
What the evidence supports — and what it does not prove
| Supported reading | Unsupported leap |
|---|---|
| AI use has broadened in the measured UK business populations. | Every business should automate immediately. |
| Extensive use and reported integration remain minority patterns. | Every non-integrated use is unsuccessful or wasteful. |
| Larger businesses report higher rates across several measures. | Business size alone causes successful adoption. |
| Policies, training and data confidence are practical adoption factors. | A policy document by itself proves responsible or effective use. |
| A controlled pilot can produce evidence about one workflow. | A readiness score guarantees savings, compliance or return on investment. |
The Ostina workflow readiness test
Score one specific, recurring workflow from 0 to 2 across eight checks. Use 0 where the element is unclear, 1 where it is partly defined, and 2 where it is explicit and usable. This is a practical design aid, not an official benchmark or certification.
| Check | Question | Two-point evidence |
|---|---|---|
| Frequency | Does the workflow occur often enough for improvement to matter? | Frequent and reasonably predictable. |
| Ownership | Who owns the outcome and can approve change? | One accountable owner is named. |
| Trigger and outcome | What starts the work, and what proves it finished correctly? | Both can be stated plainly. |
| Data boundary | Which data is permitted, prohibited and retained? | The boundary is documented. |
| Human approval | Which outputs or actions need a person? | A named approval gate covers consequential or unusual cases. |
| Exception route | What happens when information is missing or the result looks wrong? | A clear route reaches the right person. |
| Baseline | Can current time, volume, error or delay be measured? | A pre-pilot baseline exists. |
| Fallback | Can essential work continue if the assisted process stops? | The existing route is known and tested. |
- 0-5: clarify the workflow before introducing AI.
- 6-11: design a narrow, staff-assisted pilot and close the missing controls.
- 12-16: the workflow may be ready for a controlled pilot, subject to data, legal and supplier checks.
Download the evidence and readiness test
The workbook includes a read-me, claim-level evidence table, business-size breakdown and editable readiness score. The CSV provides the same source table in an open format.
Version 1.0, 20 August 2026. No email or form is required.How to turn the score into a useful pilot
- Choose one outcome. Avoid “use more AI”. Name a recurring result such as an owned enquiry, approved follow-up or complete internal record.
- Observe the current workflow. Record arrivals, copying, checking, decisions, handoffs, exceptions and completion.
- Close the low-scoring gates. Give the workflow an owner, data boundary, approval point, baseline and fallback before adding more autonomy.
- Design the smallest useful assistance. A draft or connected preparation step with human approval may be more valuable and containable than full automation.
- Test normal and difficult cases. Include missing, duplicate, sensitive, misleading and unavailable-system scenarios.
- Compare with the baseline. Measure useful completion, time, rework, exceptions, reliability and staff experience.
- Expand only with evidence. Add another action, category or system when ownership and control remain clear.
For the surrounding design detail, compare AI tools with integrated workflows, use the AI governance checklist for UK SMEs, and review the seven-gate AI readiness checklist.
Governance and accountability are operational requirements
The ICO AI and data protection risk toolkit helps organisations identify and reduce risks to people's rights and freedoms. The ICO currently states that its AI guidance is under review following the Data (Use and Access) Act, so a live decision should always be checked against the current official page.
The CMA's guidance on AI agents and consumer law also makes the accountability boundary clear: a business remains responsible when an AI agent acting for it breaks consumer law. Human approval, monitoring and stop conditions therefore belong in the workflow, not in a policy file nobody uses.
Frequently asked questions
Why do the ONS and UK Business Data Survey AI adoption figures differ?
They measure different populations, periods and questions. ONS reported around 35% among businesses with 10 or more employees in June 2026. UKBDS reported 41% among businesses handling digitised data, from fieldwork between October 2025 and January 2026. They should be read separately, not averaged.
What does the 21% AI integration figure mean?
In UKBDS, 21% of businesses already using AI said their AI tools were integrated into existing business systems. It is not 21% of all UK businesses, and it is a self-reported measure rather than a technical audit.
Is the Ostina workflow readiness score an official government benchmark?
No. It is an Ostina framework for structuring a practical conversation about one workflow. It is not stated or endorsed by ONS, DSIT, ICO or CMA, and it does not certify compliance or predict return on investment.
Does a small business need an AI policy before using AI?
The right level of governance depends on the use and risk. At minimum, staff need clear approved-use, data, oversight and escalation rules. UKBDS found that only 17% of AI-using businesses had any formal or informal AI policy or guidance.
How often will the evidence hub be updated?
Ostina reviews the hub when ONS or DSIT publishes materially newer UK business evidence. Every figure includes a release date, population, denominator, caveat and official source link so readers can check currency directly.
Does a high readiness score guarantee that AI automation will work?
No. A high score only suggests that the workflow is better defined for a controlled pilot. Privacy, security, consumer law, accessibility, supplier checks, testing, human oversight and a measurable baseline still need to be addressed.
Primary sources and methodology
- Office for National Statistics: Artificial intelligence in UK businesses, 2023 to 2026, published 20 July 2026.
- Department for Science, Innovation and Technology: UK Business Data Survey 2026, published 18 June 2026.
- Information Commissioner's Office: AI and data protection risk toolkit.
- Competition and Markets Authority: complying with consumer law when using AI agents, published 9 March 2026.
Method: figures were transcribed from the official publications above and checked against their stated population or respondent base. Survey estimates are presented as published; no composite index, causal claim or invented benchmark has been added. Ostina's readiness framework is identified separately from official evidence.
The practical next step
Download the workbook and score one recurring workflow with the person who owns the outcome. If the score reveals an unclear trigger, data boundary, approval point or fallback, fix that first. When the workflow is ready for a contained test, Ostina can help map it through a focused workflow discovery and design the appropriate level of AI-supported automation.
