SME AI Adoption Gap

Across the OECD, 40% of large firms (250+ employees) used AI in 2024, against 20.4% of medium firms (50–249) and just 11.9% of small firms (10–49). The gap is wider for AI than for any other digital technology — small firms are roughly half as likely as large firms to adopt cloud or IoT, but less than one-third as likely to adopt AI. Source: oecd-sme-ai-adoption-2025.

This is the base-rate fact behind every conversation about AI diffusion in the real economy. The “long tail” of SMEs — the dominant population of firms in every G7 economy — is the part of the diffusion S-curve that has barely started to bend.

What “the gap” actually is

Three findings stack on each other:

1. Adoption-level gap. OECD aggregate: 40% / 20.4% / 11.9% in 2024 across large / medium / small. Gap visible in every G7 country (see sme-vs-large-firm-ai-gap for the country-level figures).
2. Application-mix gap. Eurostat 2024: SMEs lag on every AI application, but the gap is largest for capital-intensive applications (autonomous robots / vehicles: 7.2% large vs 0.7% small — a ~10× gap) and smallest for low-friction generative tools (natural language generation: 16.7% vs 4.6% — a ~3.6× gap). The cost of the underlying compute and integration tracks the gap size.
3. Purpose-mix gap. Among firms that do use AI, SMEs concentrate generative AI on peripheral tasks — marketing, customer service, content drafting — while large firms apply AI to logistics, R&D, ICT security. Only 29% of SMEs using generative AI report deploying it in their core activities (OECD 2025 SME survey).

The third finding is the load-bearing one for this paper. Most SME AI use today is generative-AI surface decoration. The core business processes — where productivity gains would actually compound — remain largely AI-untouched.

What drives the gap

OECD (Calvino & Fontanelli 2023; Calvino et al. 2022) decomposes it. After accounting for sector, firm age, and asset composition, the gap persists — it is not just a “SMEs operate in less AI-intensive sectors” story. Four structural barriers remain:

• Connectivity — disparities in high-quality broadband between urban and rural areas, and between firm sizes.
• AI-enabling inputs — access to quality datasets and compute resources.
• Skills — consistently cited by SMEs as their #1 barrier to AI adoption.
• Finance — short-horizon financial constraints that block long-term complementary investment.

These are the OECD’s “four enablers” — the policy lever framework Section 4 of the source develops in detail.

Why this is load-bearing for this wiki

• For H2_u-curve-of-value: the “bottom of the U” in the long-tail application sense (the broad SME population) is thin. If SMEs cannot adopt, the application layer’s surplus pools at a small number of digitally mature large firms — not at the broad base the U-curve framing implicitly assumes. This is a real tension. The U may have a shallow bottom-right in the SME dimension even where the top is empirically validated (middle-layer-defensibility).
• For diffusion-vs-innovation: SMEs are the literal substrate of the diffusion S-curve. Ding’s claim that diffusion (not innovation) decides power transitions is empirically operational here — the gap is the diffusion bottleneck.
• For scaling-gap: the OECD SME gap and the BCG/Accenture 74% scaling gap are different gaps — one is inter-firm (small vs large), the other is intra-firm (pilot to enterprise scale). They likely compound: most SMEs do not even reach the pilot stage that the BCG number measures.

Counter-pressures

• The gap is narrowing. OECD-wide adoption rose 5.6% → 14% from 2020 to 2024 (see oecd-ai-adoption-trajectory-2020-2024). Most sectors doubled adoption between 2021 and 2024. The SME tail is moving — slowly.
• Younger firms break the pattern. Start-ups adopt AI at materially higher rates than incumbent SMEs of comparable size. The age dimension may matter more than the size dimension once cohorts are isolated (Calvino & Fontanelli 2023; McElheran et al. 2024 for the US).
• Generative AI lowers the entry barrier. Natural-language interfaces reduce the skills threshold for basic uses — the smallest gap (NLG) is the gap most directly compressed by ChatGPT-class tools. The peripheral-use pattern may be a transition phase, not a steady state.

Tensions

• Cross-country comparison is fragile. Definitions of AI, the sampling frames, and survey years differ; the OECD itself flags this caveat on every figure.
• “Adoption” is a coarse signal. A firm that has a single employee using ChatGPT for emails is counted alongside a firm running production ML pipelines. The oecd-sme-adopter-taxonomy is the OECD’s response to this — disaggregating “AI use” into four meaningful categories.

Related

• oecd-sme-ai-adoption-2025 — primary source.
• oecd-sme-adopter-taxonomy — the four-quadrant decomposition that gives the gap structure.
• sme-vs-large-firm-ai-gap — the keystone data block.
• oecd-ai-adoption-trajectory-2020-2024 — the OECD-wide trajectory the gap sits inside.
• sectoral-ai-diffusion-pattern — sectoral concentration that compounds the size gap.
• ai-productivity-firm-level — the productivity link that makes the gap consequential.
• diffusion-vs-innovation — Ding’s diffusion-marathon thesis; SMEs are the literal substrate.
• H2_u-curve-of-value — the gap thins the bottom-right of the U.
• scaling-gap — a different gap (intra-firm pilot-to-scale); the two compound.
• enterprise-adoption-ladder — large-firm parallel ladder.
• foundational-enablers — the four-enabler frame the OECD shares with WEF.
• capital-labor-divergence — the gap is one of the mechanisms by which AI surplus concentrates in large-firm capital pools.
• ai-young-worker-hiring-slowdown — if entry-level workers displaced from large-firm exposed occupations cascade into SMEs, the gap interacts with the absorption capacity of smaller firms.