Half of UK SMEs Now Use AI. Only One in Ten Has Built Anything of Their Own.

In March 2026 the British Chambers of Commerce (BCC) and Atos reported that 54% of the businesses they surveyed were actively using AI, up from 35% a year earlier. About 94% of respondents were SMEs. The headline is that adoption has crossed the half-way line. The more useful detail sits underneath it: most of that use is generic tools, and only around one in ten firms has built something bespoke.

This guide sets out what the numbers say, what they do not, and how to decide whether your own business should stay with off-the-shelf tools or take the next step. If you are weighing AI consultancy in the UK or custom AI tools for a UK SME, this is the context to have first.

What the data actually says

The figures come from the BCC’s Future of Work: AI in the Workplace report, produced with Atos and analysed with the University of Essex’s ESRC Centre for Micro-Social Change. The Essex team also published a working paper, AI adoption and workforce change in SMEs (ISER Working Paper 2026-01, March 2026), using the same BCC survey data. Together they give a consistent picture.

Two things stand out. Adoption more than doubled in three years, and the sharpest rise came in the last twelve months. And firms already using AI are far more optimistic about productivity than firms that are not, which tells you that experience, not marketing, is what changes minds.

A word of caution about the numbers

Not every survey agrees. A YouGov poll of 1,000 UK SME decision-makers, published in August 2025, found 31% using AI-powered tools and a further 15% planning to. The gap is mostly down to timing, sample and wording: “actively using AI” and “using AI-powered tools” are not the same question, and the BCC sample leans towards larger SMEs and professional services. The Essex researchers note that adoption is uneven, with larger SMEs and business-to-business professional services ahead, and smaller, consumer-facing and manufacturing firms behind.

The sensible reading is that the direction is not in doubt, and the exact percentage depends on how you ask. Do not quote “54%” as though it describes your sector.

Generic versus bespoke: the line that matters

The Essex paper makes a distinction that most commentary skips. Most SMEs using AI rely on generic tools such as ChatGPT or Microsoft Copilot. A minority, roughly one in ten, have built bespoke implementations. The two groups behave differently.

• Generic tools are quick to start, cheap per user and good for individual work: drafting, summarising, searching. They know nothing about your approval steps, your pricing rules or your data handling requirements.
• Bespoke tools are built or configured around one workflow in one business. They read your documents, call your systems and follow your rules, and they can keep an audit trail. They also cost more to start and need an owner.

Neither is better in the abstract. Copilot versus custom AI comes down to where the work happens: inside Microsoft 365 and scoped to a person or team, Copilot is usually enough. Across several systems, at volume, or facing customers, it usually is not.

What the headcount finding does and does not say

The reassuring number is that 95% of SMEs using AI saw no change in headcount in the past year. The Essex paper adds a more careful point. Firms with bespoke AI were roughly three times as likely to have restructured job roles, and about a fifth of them attributed staffing reductions to AI. Firms investing in AI training were, surprisingly, more likely to expect headcount reductions, and the authors flag that as unexpected.

This is association, not proof. A firm that has already decided to restructure may be the kind of firm that builds bespoke AI. What it does suggest is that the deeper AI is built into a workflow, the more the workflow, and the roles around it, tend to change. Plan for that conversation with your team before the build, not after.

When generic is enough, and when it is not

Four questions settle most cases:

1. How often does the task repeat? A weekly task rarely justifies a build. A task done fifty times a day often does.
2. How many systems does it touch? One mailbox or one document library suits Copilot. Email, a finance package and a case-management system together usually need integration work.
3. Whose rules does it follow? If the right answer depends on your own pricing, your approval limits or your regulator, a general tool cannot know it unless you build that in.
4. Do you need to prove what happened? A customer-facing or decision-making workflow needs logs and human review you can show. See AI automated decisions after the Data (Use and Access) Act for the legal side.

If the answer to all four is “rarely, one system, generic rules, no”, stay generic and spend your effort on training and policy. If two or more point the other way, a bespoke build is worth scoping. The same logic appears in our look at why AI automation stalls in UK businesses: most stalls come from picking the wrong shape of tool, not the wrong vendor.

A 90-day plan for moving beyond generic tools

Whatever the survey says about everyone else, your decision should rest on your own workflows. This is the order I would work in.

1. Days 1–15: find out what is already in use. Staff are likely using AI whether or not you have approved it. Establish which tools, for what tasks and with what data. Our shadow AI guide covers the method.
2. Days 10–30: put a policy and basic governance in place. Decide what may and may not go into AI tools, who approves new ones and who is accountable. Our free AI policy template and the AI governance framework give you a starting point.
3. Days 20–40: check your data is ready. Copilot and any custom tool will surface whatever your permissions allow. Fix oversharing first. The AI readiness checklist lists what to look at.
4. Days 30–50: pick one workflow and baseline it. Record how long it takes now, how many errors occur and who does it. Without a baseline you cannot show a return. See how to measure AI ROI.
5. Days 45–75: build or configure a small pilot. Use the simplest tool that meets the four tests above, with a human reviewing the output.
6. Days 75–90: measure, decide and tell the team. Compare against the baseline, count the real running cost, and decide whether to widen, change or stop. Be open with staff about what changes in their roles.

Where a consultant helps, and where you do not need one

You do not need outside help to write a policy, run a pilot of Copilot or train staff on prompts. Outside help earns its fee when the work crosses systems, needs security and data-protection judgement, or needs building. In the YouGov poll, 13% of AI-using SMEs already used external consultants and a further 26% were likely to, which fits that picture. If you do hire, ask who owns the code, where the data goes, how you can switch supplier and what happens when the model changes. Our buyer’s guide to AI consultancy sets out the questions, and the document-processing case study shows what a scoped build looks like.

FAQ

Sources: British Chambers of Commerce and Atos, Future of Work: AI in the Workplace (reported March 2026); Bharier, Etheridge and Morais, AI adoption and workforce change in SMEs, ISER Working Paper 2026-01 (March 2026); YouGov UK SME poll (August 2025).