The £78bn gap is real. The diagnosis is wrong.

The number is real. The story attached to it is not. Britain's small and medium businesses have not failed to adopt AI. Depending on which survey you believe, somewhere between a fifth and four fifths of them already have. What almost none of them have done is put it near the work that makes them money.

Depth is the number of business processes that would stop working if you switched the AI off tomorrow. For most UK SMEs that number is zero, and no adoption survey asks the question.

Six organisations, nine numbers

The Office for National Statistics reported 29% of UK businesses using at least one AI technology in June 2026, up eight percentage points on the year (ONS Business Insights and Conditions Survey, reported July 2026). The ONS's own longer series puts it nearer 35% among businesses with ten or more employees (ONS, Artificial intelligence in UK businesses, 20 July 2026).

Enterprise Nation surveyed 1,320 micro and small businesses and found 21% using AI regularly, with 6% describing it as embedded in daily work across the business (Enterprise Nation, 9 June 2026). A different wave of Enterprise Nation's own research, run with Opinium across 1,000 SME decision-makers, found 81% using AI in some form (Enterprise Nation, 25 March 2026).

IONOS and YouGov asked 1,026 organisations of up to 250 employees and got 37% using AI in operations (IONOS/YouGov via TechRound, May 2026). The British Chambers of Commerce and Atos got 54%, on a sample that was 94% SMEs (BCC, 18 March 2026). The Federation of Small Businesses got 55% (FSB, June 2026). The OECD got 61% (OECD, December 2025).

Nobody here is wrong and nobody is lying. The broad measures, the ones that ask whether anyone in the business has used an AI tool at all, cluster between 50% and 80%. The narrow ones, which ask about regular or embedded use, cluster between 6% and 25%. Same country, same year, same firms. A 75-point spread means the word "adoption" is carrying two incompatible jobs, and the headline always takes the flattering one.

The depth numbers have barely moved

Once you stop counting doors and start counting rooms, the picture goes quiet. Businesses using AI run an average of 1.6 AI technologies, up from 1.4 in 2023. Only 10% report extensive use. Only 11% have trained more than half their workforce (ONS, 20 July 2026).

The OECD sorts AI-using SMEs into maturity tiers and finds 76% of them are novices, running simple tools on isolated tasks. Champions, meaning firms with AI integrated across the business, come to 3.6% (OECD, December 2025).

If I had one number to put in front of a board, it would be 1.6. It says that the median adopter bought a chatbot licence, used it to draft job ads and marketing copy, and stopped. That is a real thing, and it is worth something. It is not what anyone means when they talk about an economy transformed by AI.

Where the £78bn sits

The £78bn figure is a counterfactual: what the UK would gain if smaller firms adopted at large-enterprise rates (Harri Digital, 2026). The FSB models roughly the same idea and lands above £42bn a year (FSB, June 2026). The two disagree by more than the GDP of a small country, which tells you how soft this class of estimate is. Both are directionally right. There is a large pile of unearned value sitting in British SMEs.

The problem is the variable each model turns. Both are built on adoption rates, and adopting at large-enterprise rates would mostly mean buying more licences. Big firms are not deep either. McKinsey's global survey finds around two-thirds of organisations have not begun scaling AI across the enterprise, 39% report EBIT impact at enterprise level, and in any given business function no more than 10% say they are scaling AI agents (McKinsey State of AI, reported by Forbes, March 2026).

So the target the models point at is a company that has bought more tools and changed nothing. Very little of the £78bn is sitting with the firms who never started. Most of it is sitting with the 76% who started and stopped one step in.

A wrong diagnosis buys the wrong intervention

Diagnose this as an awareness problem and you get awareness programmes: campaigns, webinars, free trials, a discount on a licence somebody already has. The barrier data does not support that. The leading barrier the ONS records is difficulty identifying use cases, not ignorance that AI exists (ONS, 20 July 2026). Enterprise Nation finds 46% citing lack of skills and 53% citing cost. IONOS and YouGov found 63% believing they can stay competitive without AI for the next two years, which is a judgement about value rather than a gap in awareness (IONOS/YouGov via TechRound, May 2026).

The advice channel is the tell. 64% of UK SMEs take their technology advice from their accountant (Enterprise Nation, 9 June 2026). The government's own SME Digital Adoption Taskforce has recommended developing a scalable AI-powered chief technical officer service for small firms (GOV.UK, 26 June 2026). Read that recommendation slowly. It is an admission that the scarce input is senior technical judgement, not software, and that the state can see the shortage clearly enough to propose synthesising it.

What depth looks like in practice

Depth is unglamorous and it is mostly writing. Pick one process that runs every week and costs someone real hours. Write down how it currently works, including the exceptions people handle by memory. Change one step of it. Then agree, in advance, what number would tell you in ninety days that the change held, and who owns that number.

The thing that moved our own delivery was not the model. It was writing down how the work happens in a form a machine can follow. On one platform we build, the durable artefacts are 81 memory files covering what shipped and which trap it hid, plus seven short process files that exist because something went wrong once. Every one of the seven is about attention and approval rather than about code, which is why they ended up in our own AI use policy more or less verbatim. Operating discipline, then. Not capability, and it is the half that transfers to a business nothing like ours. Not that it makes us fast at everything: there are tasks we still take back off the agents every time.

Depth also brings duties that shallow use does not, which is a reason people quietly avoid it. Once a process depends on AI, somebody has to own what happens when it is wrong, and the AI literacy obligation has bound every deployer since February 2025 regardless of risk tier, a point worth checking against whether the EU AI Act reaches a UK company at all.

Our own caveat: we are a software studio, so the processes were already written down in a repository, which is an unfair advantage over a plumbing merchant carrying everything in one person's head. The method still transfers. Depth is what moves a company out of the 76%, and buying another licence never has.