AI is everywhere in SMEs. Deep integration is not.
The OECD’s 2026 D4SME Survey puts numbers on a pattern many owners will recognise: adoption is broad, integration is shallow. The value firms report tracks the breadth of use rather than the sophistication of the tool.
Gli Insights sono pubblicati in inglese.
The OECD published the 2026 edition of its Digital for SMEs survey on 13 April 2026. One figure from it has travelled widely: 61% of responding small and medium-sized firms use at least one AI-enabled application. Two figures from the same tables have travelled far less. Among firms that use AI, 5% use customised applications and 3.6% deploy agentic ones, while 76% fall into the category the OECD calls “AI novices” — standard tools applied to isolated tasks.
The distance between those numbers is not a gap between technically advanced and technically weak companies. It describes where firms point the tools they already own — and the survey attaches reported outcomes to that choice.
Sixty-one percent describes firms already trading online, not the SME population
Empowering SMEs in the age of AI: The 2026 OECD D4SME Survey ran between the fourth quarter of 2025 and the first quarter of 2026 and gathered 2,018 responses across twelve OECD countries. The distribution is very uneven — 1,376 responses came from Japan alone, and six countries returned fewer than forty each — so results are reported as country averages, every country weighted equally. Respondents were reached through the digital platforms on which they trade, and the OECD states that the sample is neither randomised nor representative of the SME populations covered.
The paper makes the comparison itself: against the survey’s 61%, the OECD ICT Access and Usage by Businesses database records 17% of small firms and 30% of medium-sized firms using AI in 2025. So 61% is not a national adoption rate; it is what adoption looks like among firms that already sell and operate online — the population many Swiss retailers, distributors and exporting manufacturers now belong to. Even at that leading edge, integration remains thin.
The taxonomy separates the tool a firm buys from the data it points at
The 2026 round applies a classification the OECD developed at the request of the Canadian G7 presidency in 2025. It sorts AI users on two axes: the complexity of the tools — off-the-shelf, customised, agentic — and the scope of their application, from isolated tasks to enterprise-wide deployment.
AI novices, using off-the-shelf tools for isolated tasks, account for 76% of AI users; AI optimisers, applying standard tools across several functions, 15.3%; AI explorers, running customised or advanced systems at limited scale, 5%; AI champions, combining customised or agentic systems with deployment across most of the business, 3.6%.
The two families are pointed at different work. Among off-the-shelf users, 70% apply them to marketing content and 56% to drafting documents and reports. Among the much smaller group running customised systems, the leading uses are customer demand prediction (39%) and process automation (34%); agentic deployments run multi-step workflows such as bookings and invoicing. The OECD notes that the sample’s heavy retail representation partly explains the pattern.
The first family works on inputs a competitor can buy on identical terms. The second works on a company’s own order history, quotation records and service logs.
Reported value rises with the breadth of deployment
Asked how AI currently contributes to the business, 33% of respondents report moderate benefits, 15% a significant effect on efficiency, productivity or decision-making, and 6% a transformational one — 54% in total reporting at least moderate value, against 44% reporting minimal or none.
The Japanese sub-sample is large enough to break down, and it shows where the difference sits. A transformational impact is reported by 2% of firms using AI in isolated contexts, 4% for function-specific use, 11% for cross-functional use and 23% for enterprise-wide deployment. The OECD is careful about what this establishes: these are perceptions, not measured performance, and the analysis reports associations, not causal effects. Depth is not a matter of size, either — in the same sample, 33% of self-employed respondents and 31% of micro-enterprises report significant or transformational effects, against 22% of medium-sized firms.
Adoption is climbing on both sides of the size divide; the distance is climbing faster
The structural context comes from official statistics, not the survey. In the OECD ICT Access and Usage by Businesses database, retrieved in February 2026, AI use among small firms rose from 7.1% in 2023 to 17.4% in 2025, while the gap with large firms widened from 23.4 percentage points in 2023 to 28.5 in 2024 and 34.6 in 2025.
Eurostat’s figures, published on 11 December 2025, show the same diffusion: 20.0% of enterprises with ten or more employees used AI technologies in 2025, up from 13.5% in 2024, with national rates ranging from 42.0% in Denmark to 5.2% in Romania.
Swiss evidence comes from a different kind of source. The SME labour-market study published by AXA Switzerland on 8 October 2025 — Sotomo research institute, 300 SMEs in German- and French-speaking Switzerland, fieldwork in March 2025 — found 34% of respondents deliberately integrating AI into work processes, up from 22% a year earlier, with a further 37% testing it. The most common applications were translation (52%) and correspondence (47%). The Swiss profile matches the international one: broad, and concentrated in language tasks.
European law has moved the obligation from outcome to effort — and it reaches Swiss firms
Two dates this summer matter for any company whose AI output reaches the European Union. The Artificial Intelligence Act, Regulation (EU) 2024/1689, became generally applicable on 2 August 2026. Six days earlier, on 27 July, the Digital Omnibus on AI — Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026 — entered into force and amended it.
Two amendments matter here. Obligations for high-risk systems were deferred — to 2 December 2027 for the standalone use cases in Annex III, and to 2 August 2028 for AI embedded in regulated products. And Article 4 on AI literacy was replaced: providers and deployers must now take measures to support the development of AI literacy among staff and others operating AI on their behalf, without being required to guarantee any specific level for any individual. The recital explaining the change states that AI literacy should be a strategic priority regardless of regulatory obligations.
The reach matters here. Article 2(1)(c) of the AI Act extends it to providers and deployers established in a third country where the output produced by the AI system is used in the Union: a Swiss firm whose product descriptions, quotations or forecasts are used in an EU market is a deployer within scope. Switzerland’s own framework is still being drafted: the Federal Council decided on 12 February 2025 to ratify the Council of Europe’s AI Convention and keep legislative change sector-specific, with a consultation draft due by the end of 2026.
Moving from novice to explorer is a set of ordinary management decisions
Nothing in the survey suggests that firms reporting deeper value bought better technology. What distinguishes them is a short list of decisions available to any management team.
- Identify the process that runs on data no competitor holds: order history, quotation win-and-loss records, service logs, delivery variance. Customised AI concentrates on demand prediction because that is where proprietary data sits.
- Write the data rule before buying the next tool — what may be entered into an AI system, what may not, who authorises an exception. In the Swiss study only about a third of AI-using SMEs had such rules — 23% among firms with five to nine employees.
- Budget time as well as licences. The two most-cited barriers in the OECD survey are maintenance costs (39%) and lack of time for training (38%).
- Choose the scope deliberately: in the Japanese sample, transformational effects were reported by 2% of isolated users and 23% of enterprise-wide ones. One function carried end to end teaches more than one tool per department.
- Treat security as part of the same decision: 46% of surveyed SMEs report no or only minimal security measures, and 22% have already experienced a digital security breach.
Two dates frame the next eighteen months: the Swiss consultation draft due by the end of 2026 and the European high-risk obligations arriving on 2 December 2027. The annual D4SME survey is where any shift in that share would show. The question worth putting on a board agenda is narrower than the technology debate: which of our processes runs on data no one else holds, and what would it take to point a system at it.