SMEs and AI in 2026: Ten Sources, and the Question of What They Measure

A café owner in Florence, a former chef who moved into running the business full time after the pandemic, was asked on stage at an OECD roundtable this year how his two cafés use AI. He answered with a sort of cheerful embarrassment: they were doing everything wrong. ChatGPT and Gemini for tightening up menus and internal procedures, for replying to online reviews, for knocking together graphics that were, he said, pretty decent. Then he did some arithmetic. Small restaurants and shops often run on margins somewhere between 3 and 11 percent, and an ERP or CRM package with AI built in can cost the equivalent of 20 to 35 percent of earnings. For owners like him, that kind of software feels like a luxury, and a risky one.

The OECD survey launched at the same event would place him among the “AI Novices”, the bottom level of a four-step taxonomy that runs up to “AI Champions”. The label implies a starting point. His numbers suggest something closer to a sensible place to stay, at least for now. That mismatch sent me through the rest of this year’s research on small businesses and AI, and there’s plenty of it, with headline adoption figures anywhere from about 16 percent to 74 percent. What follows lines up ten sources, sorts out where they agree and where they split, and then follows three questions the splits raise, including where outside advisers can usefully fit.

The ten sources, and what each one counts

Most of the disagreements below trace back to who was asked, how, and what counted as “using AI”, so each source is worth a sentence.

Three come from official statistics or economic research. The US Census Bureau’s Business Trends and Outlook Survey (BTOS) is nationally representative and asks whether firms use AI in a business function. Eurostat’s annual enterprise ICT survey covers EU firms with ten or more employees, which leaves micro-businesses out entirely. A January 2026 European Investment Bank working paper by Aldasoro and colleagues uses data on more than 12,000 EU and US firms to estimate what AI adoption actually does to productivity, rather than asking owners how they feel about it.

Five are business surveys. The OECD’s 2026 D4SME survey reached just over 2,000 SMEs through platforms including Amazon, Rakuten, Kakao, Sage and Wolt; roughly two-thirds of the responses came from Japan, and the authors say plainly that the sample isn’t representative. Bitkom, the German digital industry association, published a phone survey of about 600 German companies with 20 or more staff in September. Goldman Sachs surveyed some 1,250 participants in its 10,000 Small Businesses programme early in the year. Homebase, a scheduling and payroll software company, polled 750 US small business decision-makers. CPA Australia’s annual Asia-Pacific Small Business Survey covers firms with fewer than 20 employees across eleven markets, including around 300 in Hong Kong.

The last two are different in kind. Chase tracks the share of its small business customers who pay for AI tools on a business card, which measures spending rather than self-reported use. And Jason Wade, a consultant, has published a practical guide for owners arguing that they should build workflows rather than tool stacks. Alongside these sit the voices from that OECD panel: the Florentine café owner, a sales manager for two small Danish hotels, and a freelance e-commerce copywriter based in the Netherlands.

Where they agree

The most consistent finding is that AI use is wide and shallow. In the OECD sample, 61 percent of firms use some form of AI, yet three-quarters of those users are Novices relying on off-the-shelf tools for isolated tasks. Among the Goldman Sachs respondents, 93 percent report a positive impact while only 14 percent have fully integrated AI into core operations. Bitkom puts it most bluntly: of the German firms already using AI, none said they were fully exploiting its potential, and 59 percent said they weren’t exploiting it at all. Wade reaches the same diagnosis from practice. Access stopped being the obstacle a while ago; what’s rare is the operating discipline that turns a tool into a better business number.

The second point of agreement is a size gradient that appears however the data is sliced. Eurostat’s 2025 figures show 17 percent of small enterprises using AI, against 30 percent of medium and 55 percent of large ones. In the US, BTOS found that between December 2025 and May 2026 AI use rose among firms with at least 20 employees but didn’t change significantly below that line. The official OECD data cited in the D4SME report show the small-versus-large gap widening from 23 percentage points in 2023 to about 35 in 2025, even though small-firm adoption more than doubled over the same period. Homebase splits by revenue instead and finds 54 percent adoption among businesses under $500,000 a year against 81 percent above $5 million. The methods differ and the shape doesn’t: small firms are moving, and large ones are moving faster.

The third is what holds firms back. OECD respondents most often name maintenance costs and lack of time for training, each cited by close to 40 percent. Among German firms not yet using AI, 85 percent told Bitkom they lack the know-how to implement it. Seventy-three percent of the Goldman Sachs owners said more training and resources would help. And in the OECD data, two-thirds of firms that hadn’t benefited from public digitalisation programmes didn’t know those programmes existed. Money and tools matter, but the more common gap is knowing what to do and having the hours to figure it out.

Where they disagree

How many SMEs actually use AI?

The headline numbers fall into two clusters. At the low end sit BTOS at 17 to 20 percent, Eurostat’s small enterprises at 17 percent, and Chase’s card data showing around 16 percent of its small business customers paying for AI tools in July 2026. At the high end are Homebase at 74 percent, the OECD’s platform sample at 61 percent and Bitkom at 57 percent.

These are answers to different questions. Census asks about AI used in a business function. Chase counts payments, so a café owner on free ChatGPT never shows up. Homebase and the OECD reached firms that already use a software platform. Bitkom and Eurostat both exclude the smallest businesses. An owner who occasionally asks a free chatbot for a promotional post counts as an adopter in some of these surveys and not in others. Wade makes the same point with a four-layer breakdown that runs from employees using AI on their own, through owner-reported use and AI embedded in a defined process, up to full integration with governance. Before quoting any adoption rate, it’s worth knowing which layer it describes.

Who gets more out of it?

This disagreement is more interesting. In the OECD’s Japanese sample, the smaller the firm, the more likely it is to say AI has had a significant or transformational effect: a third of self-employed respondents say so, against just over a fifth of medium-sized firms. The EIB paper points the other way. It estimates that AI raises labour productivity by about 4 percent on average, with the gains concentrated in medium and large firms and dependent on complementary investment in software, data and training.

Both can hold at once, because they measure different things. The OECD figure is how owners feel; the EIB figure is productivity estimated from company financials. A one-person business that suddenly has something to draft, translate and summarise for it may feel enormous relief, but turning that relief into measurable productivity usually means cleaning up data and redesigning processes, and those fixed costs are easier to absorb across a larger operation. I ran into a similar pattern in Two Studies, Two Opposite Answers: studies that seem to contradict each other often reconcile once you look at what each one was measuring.

A positive impact, but untapped potential

Goldman Sachs and Homebase both find roughly nine in ten AI-using owners reporting a positive effect. Bitkom asked something else, whether firms are exploiting AI’s potential, and found almost nobody willing to say yes. “Positive impact” is a low bar; one email written faster clears it. “Exploiting the potential” asks the respondent to imagine how much more there might be. That the first number is high and the second low suggests most owners already sense they’ve only touched the surface.

Where to go next

The recommendations diverge most. The OECD taxonomy sorts firms by how widely they use AI and how sophisticated the tools are, and the implied direction is upward, supported by financial aid and more accessible training. Wade argues nearly the opposite: start with one boring, frequent, measurable workflow, put a named person in charge of approving anything with financial or legal consequences, and don’t expand until that first workflow shows a documented improvement. Most small businesses, in his view, should buy tools and adapt their workflows rather than build systems of their own. The Florentine owner added a third view that neither framework quite captures. Support schemes, he said, tend to assume growth is the goal, but family restaurants and neighbourhood shops often don’t need to grow. They need ways to keep what makes them distinctive sustainable.

Three questions worth staying with

With the agreements and disagreements side by side, I think three questions matter more than any adoption rate.

Is the yardstick measuring tools or discipline?

The OECD taxonomy has two axes: how widely AI is used and how complex the tools are. Wade’s framework adds a dimension the OECD doesn’t measure, namely whether a workflow has an owner, a trusted source of information, a human checkpoint and a metric someone tracks. On the first yardstick, a café that drafts review replies with ChatGPT, has someone check each reply and tracks response time is still a Novice. On the second, it’s already implementing. A company that has bought several custom tools with no owner and no metric scores higher on the first yardstick and looks like a textbook failure on the second.

Wade’s approach carries its own assumption, though. Inventorying thirty to fifty documents, naming an owner for every workflow and building a KPI dashboard is a project in itself for someone who, as the Florentine owner put it, still has to know which sponge is for which cleaning job. So I’d add a third yardstick: given the owner’s limited attention, how many judgments can be reliably handed over to a process or a tool? It’s harder to measure than tool complexity, but it fits the café owner’s situation better, and it leaves room for the possibility that staying a “Novice” is, for some businesses, a reasonable choice.

AI stands in for a missing organisation. Who carries the cost?

Self-employed respondents report the highest perceived benefit from AI, and the same OECD survey finds them the most likely to report stress and burnout from digital tools. CPA Australia’s Hong Kong results point the same way from a different angle, describing digitalisation and automation as easing labour constraints. My reading is that for the smallest businesses, AI fills in for colleagues and departments that never existed. The drafting moves to the machine, but deciding which draft goes out and which suggestion fits still lands on the same person, so the workload can shrink while the density of judgment rises. That’s a hypothesis, and it needs finer-grained data to test. It does suggest that counting hours saved misses part of the picture, and Wade makes a related point: saved hours aren’t a return unless they get redeployed to something.

Who teaches small businesses to use AI?

The OECD data show SMEs filling digital skills gaps mainly through internet searches, external consultants and friends and family, with about one in five turning to generative AI directly. Government-funded digital training reached fewer than 5 percent of firms in Greece, Japan and Korea. Two-thirds of German firms told Bitkom their staff’s AI skills were low. When the main way to learn AI is AI itself, people tend to pick up the mechanics more readily than the judgment: whether a tool suits the situation in front of them, and how to notice when its output is wrong. It’s the same concern I raised in an earlier piece on apprenticeship. Tacit knowledge used to travel through everyday contact between experienced and newer people, and the channels now replacing that contact don’t obviously carry judgment with them.

Where outside help fits

Read the other way round, the same data sketch where outside help, whether a consultant, a trade association, an accountant or simply a well-informed friend, can contribute, and where it runs out.

The first role is triage. The Danish hotel manager named uncertainty as one of the biggest barriers: not knowing where to start, which tools to trust, or whether any of it creates value. Someone familiar with the tools and the public programmes can do the first round of filtering. Before that, they can help a business decide which yardstick it should be measured by, whether that’s growth, relief from staffing pressure or simply less time lost to paperwork.

The second is spreading fixed costs. For a small firm, the EIB finding that productivity gains depend on complementary investment in data, software and training is mostly a finding about costs it can’t spread. An adviser who works with many similar businesses can reuse workflow templates, knowledge-base structures and approval rules, sharing some of that cost across clients.

The third is passing on judgment along with the mechanics. Wade dismisses prompt cheat sheets and one-off workshops as low-value help, and I largely agree, since that material is now something AI can teach directly. What’s harder to learn alone is when to trust an output, where a human check belongs, and which steps shouldn’t be handed to AI at all.

The fourth is a security baseline. Close to half of the OECD sample had no or only minimal security measures, and the two Danish hotels were knocked offline for three weeks five years ago when they were caught up in an incident affecting some 600 European hotels. For many small firms, their security is in practice their vendors’ security, and having someone review vendor arrangements and access settings is more realistic than learning it all from scratch.

The limits matter just as much. Outside help costs money, and the café owner’s margins are a hard constraint. Advisers also have incentives of their own, which tend to point clients toward more complexity and more services; the second half of Wade’s guide, after all, introduces his own company’s offering. One fair test of any adviser is whether, having looked at a business, they’re willing to say that what it’s doing now is enough.

Where this leaves us

Laid side by side, the ten sources are most certain about a question rather than an answer. Are we measuring SME AI maturity by tools, by scope, by self-reported benefit, or by the judgments an owner can reliably hand off within limited attention? Each yardstick gives the same café a different score and leads to different advice. I lean toward the last one, even though nobody has a good way to measure it yet. If you run or advise a small business, the cases in front of you are a practical place to start: which decisions you could already hand off with confidence, and which still only you can make.

Sources