Around 35% of UK businesses with 10 or more employees were using artificial intelligence in June 2026, up from around 12% in late 2023. The figure comes from the Office for National Statistics, published on 20 July 2026. The same release carries a second number that has attracted far less attention. The average adopter uses about 1.6 AI technologies, up from 1.4 across the same two and a half years, and only 10% of AI-using businesses describe their own use as extensive.
Adoption has spread across the economy without deepening inside the firms doing it. Three times as many British businesses now use AI, and they are using it for almost exactly as much as they were when they started.
That gap matters more to the UK's productivity arithmetic than the headline adoption rate does. Buying software is not the same as changing how work gets done, even as the cost of running the models themselves collapses, and the ONS data is the clearest public evidence yet that British firms are further along the first than the second.
Adoption has widened, usage has not
The ONS breakdown is worth reading closely. Large businesses with 250 or more employees report the highest adoption at 49%. The smallest firms, with fewer than 10 employees, report 28%. On the surface that is the expected pattern, with resources tracking scale.
Reverse the question and the pattern inverts. Among businesses that use AI at all, 17% of the smallest firms describe their use as extensive, against 9% of the largest. The organisations spending the most are using it the least intensively. A small firm that adopts an AI tool tends to rebuild a process around it. A large firm tends to add it to a department and leave the process alone.
Sector variation is wider still. Almost three-fifths of information and communication businesses use AI. In construction the figure is 13%. Coverage of the release framed the trend as widening rather than deepening, which is the fair reading of the numbers.
Barriers are not the obvious ones either. Some 41% of businesses report no barrier at all to adoption. Cost is cited by between 7% and 14%. The constraint that shows up most consistently is expertise, named by up to 18% of firms with 100 to 249 employees. Among medium and larger businesses, training existing staff is the most common integration approach, used by 40%.
Most projects never leave the pilot
The ONS measures use. It does not measure return. For that, the most useful recent evidence is an annual survey published by the workforce analytics firm Orgvue on 14 May 2026, covering 1,163 senior decision makers across the United Kingdom, Ireland, the United States, Canada, Australia, Hong Kong, Malaysia and Singapore.
Some 78% of those organisations had AI projects that failed or stalled at the pilot stage. Of those, 43% reported projects stuck in pilot and 35% reported outright failure. Nearly a third, 32%, said they did not understand how to implement AI effectively.
The reason given for the rush is the part worth dwelling on. Some 57% of leaders said they deployed AI because competitors had done so. A further 34% said they lacked the expertise to manage the workforce change that came with it.
A separate study for the Department for Science, Innovation and Technology, carried out by IFF Research and Technopolis Group across 3,500 UK businesses and published on 17 February 2026, puts a figure on the outcome. Among businesses that had adopted AI, 75% reported improved workforce productivity. Only 12% reported increased revenue. Operational gain is being felt. It is not reaching the accounts.
Deployment driven by competitive anxiety produces tools without process change, which is precisely the shape the ONS data describes. The direction of travel is not in doubt: the ONS business insights bulletin of 8 January 2026 recorded around a quarter of businesses using AI in late December 2025, rising to 44% among firms with 250 or more employees, with a further 15% planning to adopt within three months. The line goes up every quarter. The depth measure does not move with it.
Where the shortfall shows up first
The gap is most expensive in commercial functions, because that is where a judgement error converts directly into money. Pricing, procurement and sales negotiation all involve a large analytical component that AI handles well, sitting on top of a judgement component that it does not handle at all.
| Preparation task | What AI does reliably today | What it still needs from a person |
|---|---|---|
| Assembling a fact base on a supplier or customer | Fast, broad, good at surfacing published data | Judging which facts the other side already knows |
| Modelling scenarios and trade positions | Generates options quickly, including ones a team would miss | Deciding which options are credible in this relationship |
| Drafting an opening position | Competent, well structured, sounds authoritative | Knowing what the position signals about your intent |
| Benchmarking a price against the market | Strong where data is public, weak where it is not | Distinguishing a comparable deal from a superficially similar one |
| Reading the other party's likely behaviour | Poor. It has no access to the relationship history | The whole of it |
| Deciding what to concede and when | None. It will produce a confident answer anyway | The whole of it |
The negotiation consultancy The Gap Partnership makes a similar argument in its analysis of AI and commercial growth: the tools have become good at the half of preparation that can be written down, while the half that decides the outcome, whether the other party will actually move, sits outside what a model can see.
The last row is the one that costs money. An AI tool asked what to concede will return a plausible, well-argued and entirely unaccountable recommendation. A negotiator without the experience to challenge it will take it into the room.
What the firms getting a return do differently
Three patterns separate the organisations that convert AI spend into commercial outcome from those still stuck at pilot.
They establish a baseline first. That means knowing the current state of their data, their processes and their people's capability before any tool is introduced, so that improvement can be attributed to something. Firms that skip this step cannot later distinguish a tool that worked from a market that moved.
They sequence human capability ahead of the tooling. The practical version of this is unglamorous. People learn to prepare properly, and then get a tool that makes preparation faster. Done in the other order, the tool becomes a substitute for a skill that was never built, and the output looks fine right up to the point where someone challenges it. Guidance on how to use AI in negotiation makes the same point about sequencing. The Gap Partnership's own analysis of AI and negotiation makes the same argument from the practitioner side.
They measure commercial outcome rather than engagement. Tool adoption rates, licences activated and queries run are all easy to report and none of them is evidence of value. Margin retained, price increases resisted and contract terms improved are harder to attribute and are the only measures that answer the board's question. A practical worked example of that discipline, applied to supplier negotiations, is set out in this guide for procurement teams.
None of this is an argument against buying AI. It is an argument that the ONS numbers describe a sequencing failure rather than a technology failure, and that the firms fixing the sequence will pull away from the ones adding licences.
Frequently asked questions
How is artificial intelligence changing the way organisations approach commercial negotiation?
It is changing preparation rather than the negotiation itself. AI compresses the research, benchmarking and scenario modelling that used to take a commercial team days into something closer to an hour, which means more negotiations get prepared properly rather than prepared at all. What it has not changed is the part of a negotiation that happens in the room, where the variables are the other party's behaviour, the relationship history and what each side is willing to risk. Organisations reporting the largest gains use AI to widen the range of options they consider before a meeting, not to decide which one to take.
What new skills do commercial negotiators need as AI tools take over more of the analytical and preparation work?
The scarce skill is now the ability to challenge an output that looks authoritative. AI tools produce fluent, confident recommendations regardless of whether the underlying data supports them, so negotiators need enough domain judgement to ask what the model did not know. Alongside that, the traditional behavioural skills become more valuable rather than less, because if every party in a market has access to the same analysis, the difference in outcome comes from execution under pressure. The ONS finding that expertise, not cost, is the main adoption barrier points the same way.
How do you assess whether your organisation is ready to use AI effectively in commercial negotiations?
Readiness has three components and all three have to be present. The first is data: whether past deal terms, pricing history and supplier performance exist in a form a tool can use, which in most organisations they do not. The second is process: whether the commercial team has a repeatable preparation method that a tool can accelerate, because a tool applied to an inconsistent process produces inconsistent output faster. The third is capability: whether the people using the output can tell a good recommendation from a confident one. An honest readiness assessment scores all three before any procurement decision is made.
What are the most common reasons organisations invest in AI negotiation tools and then fail to see commercial benefits?
The most common reason is that the tool was bought to answer a competitive worry rather than a defined commercial problem, which the Orgvue finding that 57% of leaders deployed AI because competitors had done so illustrates directly. The second is that no baseline was captured, so no improvement can be demonstrated even where it occurred. The third is sequencing: the tool arrives before the capability to use it, and gets adopted by the people who least need it while the teams losing margin carry on as before. In each case the technology works and the return does not appear.
