AI readiness starts before the rollout

The short version

About a third of UK firms with ten or more staff now say they use some form of AI, nearly three times the share in late 2023. Adoption is relatively shallow: only 10 per cent of firms using AI report using it extensively. In earlier ONS analysis, the barriers firms named most often were difficulty identifying where AI would be useful in the business, cost and a lack of expertise. The Office for National Statistics publishes the survey as official statistics in development and calls its new extent-of-use results early indications.

The failure rates quoted for AI projects, 80 or 95 per cent, turn out not to be measurements. The stronger evidence, from interviews with practitioners, the National Audit Office and the government's own review of its technology, points to a short list of reasons AI projects stall: a problem that was never clearly defined, information that is missing or not fit for use, and systems that cannot easily connect.

Government evaluations of the same AI assistant show why a baseline matters. Across government, staff reported saving 26 minutes a day, an estimate built from survey answers. One department that also tested people on set tasks found 72 per cent of the users who completed its diary study satisfied, but no evidence that the time saved had improved productivity, and it had no measurement from before the pilot to compare with. The preparation that settles questions like these, a clear task and owner, an honest baseline and agreed rules for stopping or extending, is worth doing whether or not an AI project follows.

Where UK organisations actually are

The Office for National Statistics has asked UK businesses about AI since late 2023. In June 2026 about 35 per cent of firms with ten or more employees said they used at least one AI technology, up from about 12 per cent. Use rises with size, from 28 per cent of firms with fewer than ten employees to 49 per cent of those with 250 or more, and it varies widely by sector: almost three-fifths of information and communication businesses report using AI, against 13 per cent in construction.

The ONS describes the use as relatively shallow. The average number of AI technologies per adopting business has risen only from about 1.4 to 1.6, and only 10 per cent of firms using AI report using it extensively. The ONS calls its extent-of-use results early indications and the whole series official statistics in development, and it counts AI through a list of named technologies, not a single definition. Its figures record what firms say they use, which is a different thing from what they have built into their work.

Other official surveys give different levels, for good reasons. The same survey put AI use at 25 per cent of businesses in late December 2025, on a different base. The UK Business Data Survey 2026 found 41 per cent, but only among businesses that handle digitised data, and with a definition that names everyday generative AI tools and counts any use, free or paid. None of these contradicts the others. Each figure means something only beside its base.

Asked what had delayed or prevented adoption in the previous three months, 41 per cent of firms with ten or more staff said nothing had. A lack of expertise was reported by around 18 per cent of firms with 100 to 249 staff, and cost by between 7 and 14 per cent, depending on size. In earlier analysis the ONS found the most commonly reported factors were difficulty identifying business use cases, cost and a lack of expertise. The first of those is a question about the work, not the technology, and an organisation can start on it without buying anything.

Why first projects stall

Two failure rates for AI projects circulate widely: 80 per cent, attributed to RAND, and 95 per cent, attributed to MIT. Neither is what it is presented as. RAND's 2024 report quotes the 80 per cent figure from other sources, "by some estimates", and its own evidence is interviews with 65 experienced data scientists and engineers. The 95 per cent comes from a 2025 report its authors called preliminary findings, whose figures they describe as directionally accurate and based on individual interviews, and critics have pointed out that it presents no data for the number.

The other figures in circulation are of the same kind. Gartner predicted in 2024 that at least 30 per cent of generative AI projects would be abandoned after proof of concept by the end of 2025, and said in January 2026 that at least half had been, without publishing a method. S&P Global Market Intelligence reported in 2025 that the share of companies abandoning most of their AI initiatives before production had risen from 17 to 42 per cent in a year, from a survey of 1,006 IT and business professionals in North America and Europe. These are forecasts, assertions and self-reports: useful as signals, weak as measurements.

Where the stronger evidence agrees is on the reasons. RAND's interviewees traced failure to five root causes, among them a problem that was misunderstood or miscommunicated, data the organisation did not have, and infrastructure that could not manage its data or deploy a finished model.

The National Audit Office found in 2024 that limited access to good-quality data was a barrier to implementing AI across government, and that updating legacy systems would take time. The government's review of its own technology, published in January 2025, found that 70 per cent of public-sector respondents said their data was not well co-ordinated or interoperable, and only 27 per cent said their data infrastructure gave a comprehensive view of their operations.

Understand the work before choosing a tool

The question of where AI would help is best answered by looking closely at one piece of work. That means describing a single workflow as it actually runs: what arrives, who handles it, which systems it passes through, where it waits, and where a person makes a judgement that matters.

Information is the part most often underestimated. Missing or unusable information is one of the reasons that recur across the evidence, in industry and across government. Before any tool is considered, it is worth knowing where the information for the chosen workflow lives, who can grant access to it, how accurate it is, and what the organisation is not allowed to do with it.

Dependencies come next. A workflow rarely lives in one system. It may rely on a supplier's platform, a spreadsheet one person maintains, a contract that limits how data can be used, or a system due for replacement. The government's review estimated that about 28 per cent of central government systems were legacy in 2024, and around 15 per cent of respondents could not estimate the size of their legacy estate at all. Knowing which systems a change would touch, and who owns each connection, holds its value whatever is decided.

Before any tool is chosen, it helps to have four things written down.

• The workflow in one sentence, and the person who owns it.
• Where its information comes from, who controls access, and what is known about its quality.
• The systems and suppliers it depends on, with an owner for each connection.
• The commitments that constrain it: contracts, retention rules and data protection duties.

Prepare people and responsibility

Many staff already use AI, whether or not their employer has decided to. In a survey of 2,003 UK employees commissioned by Microsoft in October 2025, 71 per cent said they had used AI tools at work that their employer had not approved, and 28 per cent said their company offered no work-approved option, which Microsoft presents as a reason for unapproved use.

A separate study by the University of Melbourne and KPMG found that 39 per cent of UK employees asked about their use of AI at work had uploaded company information into a public AI tool. Both surveys were funded wholly or partly by companies, and neither is an official measure, but they point the same way.

Policy has not caught up. The UK Business Data Survey 2026 found that 17 per cent of businesses using AI had any policy or guidance on it, and only 5 per cent had a formal written policy. Formal written policies fall from 56 per cent of large AI-using businesses to 22, 17 and 8 per cent of medium, small and micro businesses, and 3 per cent of sole traders.

The National Cyber Security Centre, writing about unapproved AI in September 2026, advises reducing the risk rather than assuming it can be eliminated: integrating approved AI tools securely into the workplace, and building a culture of open communication about security, which makes staff much less likely to turn to services the organisation has not approved.

Skills are the next constraint, and the evidence suggests capacity, more than appetite, is the problem. Research published by Skills England and the Department for Work and Pensions, drawing on workshops with around 150 organisations and a survey of 536 employers, concluded that "the main issue is not lack of interest. It is capacity", pointing to limited time, staff pressure, cost, unclear provision and fear of failing in technical areas. It also found that informal learning, through trial and error, peer support, videos and built-in prompts, can help people start but creates uneven and risky practice.

A working paper on about 25,000 Danish workers found that employer encouragement of AI chatbots was associated with more regular use and higher reported benefits.

Responsibility needs a name. The ICO's guidance on AI, now under review, says organisations cannot delegate AI's data protection issues to data scientists or engineering teams, that senior management are also accountable for them, and that most uses of AI involving personal data will need a data protection impact assessment.

The National Audit Office's 2026 good practice guide asks whether there is clear accountability for safe, ethical and effective AI use, including where AI is embedded in a supplier's product. Acas advises employers to consult staff and their representatives before introducing AI, and notes that expecting a role to use it could mean a change to terms and conditions.

A first step small enough to learn from

Official guidance on piloting AI points the same way. The National Audit Office's 2026 good practice guide says the best practice is to start with small, clearly defined, low-risk use cases, with strong senior ownership, and to treat pilots as learning exercises rather than routes to rapid automation. It warns against pilots drifting into live use without proper controls, and it asks whether there is a clear understanding of how the decision to stop, scale or redesign will be made on evidence.

HM Treasury's guidance on evaluating AI, an annex to the Magenta Book, adds two things: establish a baseline early, using data that already exists where possible, and evaluate in stages, so that each stage decides whether to move to the next and what to change. It also warns that the measured effects of AI may change substantially as the technology evolves, which argues for fixed review dates, not a single verdict.

Two of the government's evaluations of an AI assistant show what happens without a baseline. The cross-government trial ran from September to December 2024 with 20,000 licences. Staff reported saving an average of 26 minutes a day, a figure calculated from the midpoints of the ranges they chose in a survey, and the report says it was not possible to identify how the time saved was spent.

The Department for Business and Trade ran its own evaluation of the same tool with 1,000 licences. Of the users who completed its diary study, 72 per cent were satisfied, but the evaluation "did not find evidence that time savings have led to improved productivity". In a small set of blind-assessed tasks, 11 sessions in all, users analysed spreadsheet data more slowly and less accurately than colleagues without the tool, contrary to the diary study's reported time savings for data analysis; the evaluation says to treat these findings with caution. The department collected no data from before the pilot, so it could not compare before and after.

A third evaluation, from HM Revenue and Customs in July 2026, used a randomised cohort design; its participants reported saving 2 to 3 per cent of their working week.

None of these reports disproves the others, and the department says its findings are not comparable with other departments. Together they show that satisfaction and reported time savings are not the same as a measured change in the work, and that telling the difference starts with measuring the work before the tool arrives.

Before the pilot starts, agree four things.

• How the work is measured today, and how it will be measured again.
• Where a person reviews the output, and what happens to the exceptions.
• The evidence that would justify stopping, changing or extending, written down before the start.
• A date for the review, and who takes the decision.

Preparation that pays either way

Most of this preparation has value whether or not an AI project follows. A workflow that has been described and measured, with an owner, is easier to improve by any means: a better form, a clearer rule, a connection between two systems, or a decision to leave it alone. Information that is understood and governed is safer to use for anything. A policy that tells staff what they may put into an AI tool protects the organisation now, since the evidence suggests many staff are already using one.

It also makes the eventual decision easier to take well. When a workflow is described and measured, a demonstration can be judged against a baseline rather than an impression, and a supplier's claims can be checked against how the work actually runs.

A note from the founder

Most organisations that come to us do not open with AI. They ask how to make a process cheaper, faster or less dependent on one person, and AI is one of several possible answers to that question. Starting there keeps the conversation about the work rather than the tool.

The clearest sign that an organisation is not ready is that nobody can say what the work costs today. The clearest sign that it is readier than it thinks is a team that already knows which task swallows its week, and has the records to show it.

My own view is that this is about lasting well rather than moving quickly. An organisation that understands its own processes well enough to improve them is future-proofing itself, whether or not AI turns out to be the right tool this year.

What we could not establish

There is no reliable measure of how many AI pilots in the UK reach production. The figures in circulation are forecasts, assertions or surveys that are not specific to the UK.

No official UK statistic measures how many staff use AI tools their employer has not approved. The figures here come from company-funded surveys: one commissioned by Microsoft, which sells AI tools, and one part-funded by KPMG.

We found no rigorous UK evidence that AI training improves the quality or safety of staff's work, as distinct from how much they use AI. The ONS does not track the effect of AI on entry-level hiring, and a June 2026 government analysis of hiring data found entry-level hiring falling broadly in line with the wider market, with no causal evidence yet of AI's effect.

How it was researched

Every figure is traced to its primary source: the Office for National Statistics, government statistical releases and evaluations, the National Audit Office, regulators and the research papers themselves. Where a figure is self-reported, a forecast, or commissioned by a company with an interest, the report says so. The evidence dates from 2025 and 2026, apart from RAND's 2024 study of failed AI projects and the National Audit Office's 2024 survey of government bodies, which remain the best evidence on their questions. Gartner's 2024 forecast and the ICO's AI guidance, last updated in 2023 and now under review, also predate 2025. Every claim was then checked against its source by a separate reviewer before publication.

Sources

Links open the primary source. Dates are publication dates; fieldwork periods are given in the text where they matter.

Adoption and barriers

1. Office for National Statistics, Artificial intelligence in UK businesses: 2023 to 2026, 20 July 2026
2. Office for National Statistics, Business insights and impact on the UK economy: 8 January 2026
3. Department for Science, Innovation and Technology, UK Business Data Survey 2026, 18 June 2026

Why projects stall

1. RAND Corporation, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, 13 August 2024
2. S&P Global Market Intelligence, Generative AI experiences rapid adoption, but with mixed outcomes, 30 May 2025
3. Gartner, press release on generative AI proofs of concept, 29 July 2024
4. Gartner, Why 50% of GenAI projects fail, 26 January 2026
5. Challapally and others, The GenAI Divide: State of AI in Business 2025 (preliminary findings), July 2025
6. Futuriom, Why we don't believe MIT NANDA's weird AI study, 26 August 2025
7. National Audit Office, Use of artificial intelligence in government, 15 March 2024
8. GOV.UK, State of digital government review, January 2025

People and responsibility

1. Microsoft UK, Rise in shadow AI tools raising security concerns for UK (survey by Censuswide), 13 October 2025
2. KPMG UK, UK attitudes to AI (University of Melbourne and KPMG global study, 2025)
3. National Cyber Security Centre, The hidden risks of shadow AI, 7 September 2026
4. Department for Work and Pensions and Skills England, Skills for AI: what works for AI upskilling in the UK, 2026
5. Humlum and Vestergaard, NBER Working Paper 33777, 2025, revised 2026
6. Information Commissioner's Office, Guidance on AI and data protection: accountability and governance, last updated 15 March 2023, under review
7. Department for Science, Innovation and Technology, A snapshot of entry-level hiring in the UK, 8 June 2026
8. National Audit Office, Good practice guide for organisations using AI, May 2026
9. Acas, 1 in 4 workers worry that AI will lead to job losses, 28 April 2025
10. Regulation (EU) 2026/1744 amending the AI Act, Official Journal of the European Union, 2026
11. European Commission, AI literacy: questions and answers, 27 July 2026

Pilots and evaluation

1. HM Treasury and Evaluation Task Force, Guidance on the impact evaluation of AI interventions (Magenta Book annex)
2. Government Digital Service, Microsoft 365 Copilot Experiment: Cross-Government Findings Report, June 2025
3. Department for Business and Trade, The Evaluation of the M365 Copilot Pilot, August 2025
4. HM Revenue and Customs, Evaluation report: phase 3 trial of Microsoft Copilot, 9 July 2026