Part of our guide: AI strategy
Most writing about UK AI adoption is built on vendor surveys, and in July and August 2026 three UK institutions published on it in quick succession. The three are the Office for National Statistics, the Bank of England's Agents and a Bank of England staff analysis, and none of them is selling anything.
Read together they describe a country that has adopted AI broadly and shallowly and that is using it to defend existing operations rather than build new ones. They also describe a country that cannot yet demonstrate that any of it has moved productivity, and for a leadership team deciding what to do next, that is more useful than another adoption curve.
Adoption is wide
The ONS release of 20 July 2026 puts self-reported AI use among UK businesses with 10 or more employees at around 35 per cent, up from roughly 12 per cent in late 2023.
Among businesses with 250 or more people it is 49 per cent, a genuine shift in under three years and the number most coverage stopped at.
And it is an inch deep
The same release, read further, sets out the depth:
- Only 10 per cent of adopting businesses describe their use as extensive
- The average adopter uses 1.6 AI technologies, up from 1.4
- Only 15 per cent say more than half their employees use AI
- Only 11 per cent say more than half their workforce has had AI training
The definition of "use" matters here, because a business qualifies by using any one listed technology, and one person with a ChatGPT licence puts a 2,000-person firm in the 35 per cent.
The headline does not say that a third of UK businesses have adopted AI, and what it does say is that a third have started. So one in ten of those has gone deep, and nine in ten have not trained most of their people.
The training figure is the one we would put in front of a board, because an operating model change does not come out of a workforce that has not been taught the tool.
It is being used defensively
The ONS finding that gets least attention is about intent, and improving business operations is the most reported use across every size band. That use sits at close to 60 per cent of businesses, and developing new products or services and exploring new markets are reported by a much smaller proportion.
On this evidence UK businesses are using AI to make the current business cheaper, not to build a different one.
That is a defensible choice in a tight market, and it is also a choice with a ceiling that is worth making deliberately rather than by accident. Cost reduction is finite and competitors can copy it, so if your entire AI programme sits in the 60 per cent, you should be able to say why.
The productivity link is not there yet
This is the uncomfortable part of the picture, and it comes from the Bank of England rather than from us.
A staff analysis published on Bank Underground on 6 August 2026 regressed AI adoption against change in productivity contribution across UK industries. The relationship was not statistically significant at the 5 per cent level, with a p-value of 0.065 and an R-squared of 0.103, and the author is explicit that this is suggestive of correlation at most, not causation.
For anyone selling into financial services the specific finding is worse, because financial and insurance activities show a negative contribution to labour productivity growth, despite being large and a strong AI adopter.
The Bank's Agents, reporting on 24 July 2026 from conversations with businesses through to the end of June, describe the mechanism. Cost savings from AI "are being offset in part by rising expenditure on software, cloud services and AI licences." Firms are saving money and then spending it again on the very thing that saved it in the first place.
They also report where gains do appear: "Productivity gains have been greatest where skilled employees are able to validate and refine AI-generated outputs." That is consistent with the peer-reviewed evidence and it points somewhere specific, because the return comes from capable people supervising the tool well, a training and process question rather than a licensing one.
Why the sceptics and the enthusiasts are both right
The wider evidence explains the gap, and a working paper from Bank of England, Stanford, Atlanta Fed and Bundesbank researchers surveyed around 6,000 senior executives across the US, UK, Germany and Australia. Nine in ten reported no impact on employment or productivity at their own firm over three years, even though 69 per cent of their firms were using AI.
BCG's July 2026 CEO survey found nearly nine in ten CEOs seeing cost or revenue benefits in targeted areas, with only 26 per cent having embedded AI in a broader business transformation, against roughly two thirds running pilots. Only 14 per cent had clearly defined P&L impact for all AI initiatives, and high performers were around seven times more likely to have redesigned workflows.
BCG's parallel survey of 11,749 workers found that workflow redesign raised the likelihood of measurable improvement by 24 percentage points. It also found that 66 per cent of AI users get limited or no guidance on how to redirect the time they save and that over half do not redirect it into anything strategic. Nearly half of the workers surveyed now spend more time managing and directing AI than they spend doing the work itself.
That is the whole picture in one line: the tool works and the time it frees is not being collected, because nobody redesigned the process that would collect it.
What separates the firms that get returns
The best UK evidence on what separates them is also from the ONS, published in March 2025 and largely ignored.
Of firms in the top decile of management practice scores, 88 per cent had adopted at least one advanced technology, against 51 per cent in the bottom decile. Of firms that planned to adopt AI in 2023, 48 per cent of the top decile actually did so in 2024, against 17 per cent of the second-lowest.
The gap is follow-through rather than appetite, and well-run businesses do the thing they said they would do, and that is what shows up as adoption.
The same analysis is careful about causation: technology explains only about 5 percentage points of the 43-point productivity gap between top and bottom management quartiles. Buying AI does not make you well run, and being well run is what lets you get anything out of AI.
What we would take from this
Assume you are shallower than you think. If your AI use looks like the ONS average, you have one or two technologies in the hands of a minority of staff, most of them untrained. That leaves you with a starting position, not a programme.
Decide whether defending or building. Around 60 per cent of UK adopters are improving existing operations, and that is fine as a decision and poor as a default.
Fix training before buying more licences. Eleven per cent of adopters have trained more than half their workforce, and the Bank's Agents find gains concentrate where skilled people can validate output, so more seats will not help if nobody can supervise the tool.
Redesign one process properly rather than piloting five. Redesign is the measured differentiator in both BCG surveys and it is also the plain reading of every finding set out above.
Do not expect the aggregate data to vindicate you. The productivity link is not visible in UK industry data yet, and in financial services the sign is currently wrong, so build a business case that stands on your own measured baseline rather than on a macro trend.
If you want to know where your business actually sits against this, that is what a Breathe discovery sprint is for. The sprint gives you a capability map, a data readiness assessment and a straight answer on where AI creates value in your specific business.
We also wrote about which AI statistics survive checking, and that piece is a useful companion to the findings set out here.
Get in touch if you would like to compare notes on your own numbers.
Sources
- ONS, Artificial intelligence in UK businesses (20 July 2026)
- Bank Underground, is AI making us more productive? (6 August 2026)
- Bank of England, Agents' summary of business conditions (July 2026)
- NBER, Firm-level data on AI (working paper 34836)
- ONS, Management practices and the adoption of technology and AI (March 2025)