Thirty-five per cent of UK businesses with ten or more employees used at least one artificial intelligence technology in June 2026. That is close to triple the figure of around 12 per cent recorded in late 2023, and on its own it reads as a technology moving decisively into the mainstream.
The same release contains a second number that complicates the first. The average adopting business uses 1.6 AI technologies. In September 2023 it was 1.4. Adoption has spread across the economy far faster than it has deepened inside any individual company.
What the Figures Actually Say
The Office for National Statistics release, published on 20 July 2026 and covering a survey period of 15 to 28 June, breaks adoption down by size. Businesses with 0 to 9 employees report 28 per cent. Those with ten or more report 35 per cent. Those with 250 or more report 49 per cent.
By technology type, large language models lead at 18 per cent. Visual content creation follows at 16 per cent, data processing using machine learning at 12 per cent, image processing at 6 per cent, and robotics at 2 per cent.
The ordering is worth pausing on. The two leading categories, text and image generation, are the two that require the least integration. They can be adopted by an individual employee opening a browser tab. Data processing using machine learning, which sits at 12 per cent, is the first category on the list that typically requires a company to connect a model to its own systems and its own data.
Wide, Not Deep
Three measures in the release point the same way. The average adopter uses 1.6 technologies. Only 10 per cent of adopting businesses describe their use of AI as extensive. Only 15 per cent report that more than half their employees use AI daily.
Put together, the typical UK adopter is a company where some staff use one or two generative tools, some of the time. That is a real change from three years ago, when almost nobody did. It is not the same thing as a business that has redesigned a process around the technology.
The distinction matters commercially because the returns are not evenly distributed between the two. A tool that helps individuals draft faster produces diffuse gains that rarely show up in a margin. A model connected to a company’s own pricing, scheduling or inventory data changes a process, and process changes are what appear in accounts.
The Sector Gap Is Enormous
Adoption ranges from 58 per cent in information and communication down to 13 per cent in construction. That is a spread of 45 percentage points between two parts of the same economy.
How firms get hold of AI splits along similar lines. Manufacturing, wholesale and retail, accommodation and food services, and administrative services rely most on free-to-use software. Construction, information and communication, and professional, scientific and technical activities are more likely to buy external software or ready-made services.
That difference is more consequential than it looks. Free tools are adopted by individuals and leave no organisational trace: no contract, no integration, no data governance, and no institutional knowledge when the person who used them leaves. Purchased services involve a procurement decision, which means somebody has specified a use case and someone is accountable for whether it works.
Training Is the Missing Piece
Only 11 per cent of businesses with ten or more employees report that more than half their workforce has received any AI-related training. Around 40 per cent of medium and large businesses say their main approach to AI skills is training or retraining existing staff, which is the sensible answer, but the coverage figure shows how early that work still is.
Set that against the barriers data and a pattern emerges. Some 41 per cent of businesses report no barriers to adoption at all. Lack of expertise is cited by around 18 per cent of businesses with 100 to 249 employees, and cost by somewhere between 7 and 14 per cent depending on size band.
A large majority saying nothing is stopping them, combined with a small minority having trained their people, suggests the constraint is not obstacles. It is that most firms have not yet decided what they want the technology to do. Difficulty identifying business use cases appears in the release as a barrier, and it is the one that best explains a 1.6 average.
What It Is Being Used For
More than 60 per cent of larger businesses using AI apply it to improving business operations. The second most common purpose is providing or personalising products and services, which is notably more common among the smallest firms, cited by 31 per cent of those with 0 to 9 employees.
That split is consistent with the size data. A large business has internal processes big enough that a small efficiency gain is worth pursuing. A very small business has no such scale, but it does have a product, and generative tools let it offer something it could not previously afford to produce.
On headcount, roughly half of adopting businesses reported no change. Around 6 to 7 per cent reported reductions, with the highest share among medium-sized businesses. Neither the displacement story nor the hiring boom shows up in these figures.
The Compliance Clock Has Started
One development sits outside the survey period but changes the calculation for any UK firm selling into Europe. On 2 August 2026 the remainder of the EU AI Act started to apply, under Article 113. That date also brings in obligations on providers of AI systems, including general purpose systems, that generate synthetic audio, image, video or text. From the same date, according to the European Commission, the AI Office and the authorities of member states became responsible for implementing, supervising and enforcing the Act, with the AI Office holding enforcement powers over general purpose models including the ability to request technical documentation and evaluate models.
For the 35 per cent already adopting, that turns an informal tool choice into something with a documentation requirement attached. A business that cannot say which AI technologies it uses, in which processes, and with what data, is a business that cannot answer a transparency question. On a 1.6 average with 11 per cent training coverage, a good number of firms are in exactly that position.
The Free-Tool Blind Spot
The reliance on free-to-use software in manufacturing, retail, hospitality and administrative services creates a specific exposure that does not appear anywhere in the adoption percentage.
When a tool is free and adopted individually, the company usually has no record of it. There is no contract, so no supplier obligations. There is no procurement step, so nobody has assessed what data goes into it. There is no licence, so there is no agreed position on who owns the output. And because the adoption sits with a person rather than a process, the capability leaves when they do.
None of that is an argument against free tools, which are genuinely how a great many firms should start. It is an argument for knowing what is in use. A short internal audit, asking which teams use which AI services and what information they put into them, costs an afternoon and produces the register that both a transparency question and an insurance renewal will eventually require.
It also tends to surface the use cases. Firms that run this exercise usually discover that staff have already found two or three genuinely valuable applications informally, which is a considerably cheaper way to identify where to invest than commissioning a strategy.
What to Take From It
The useful reading for a business owner is that the adoption race is largely over and the depth race has barely started. Being in the 35 per cent is no longer a differentiator, because a third of the economy is there.
The measures that still separate firms are the ones the survey shows almost nobody has: more than half the workforce trained, more than half using the tools daily, and AI connected to the company’s own data rather than accessed through a free browser tab. Those describe roughly one business in ten.
The practical first step is unglamorous and is implied by the barriers data. Write down which processes are actually costing time, then ask which of them a model could touch. The firms stuck at 1.6 technologies are not short of tools. They are short of a decision about what to point them at.
About The Author
More Stories
Mountain Warehouse Passed £500m and Its Founder Handed Over the Same Week
The Construction PMI Jumped to 44.7 and That Is Still Contraction
Wales Grew Its Investment 530% and London Still Gained Twenty Times More