Where AI Value Lands

oecd-sme-ai-adoption-2025

10 min read

OECD — AI Adoption by Small and Medium-Sized Enterprises (G7 Discussion Paper, 2025)

Citation

  • Title: AI Adoption by Small and Medium-Sized Enterprises
  • Authors: Flavio Calvino, Marco Bianchini, Marguerita Lane, Jose Montegu, Vincent Verger, Slavina Ancheva (with contributions by Celine Caira, Alice Holt, Maximilian Reisch).
  • Year: 2025
  • Venue: OECD Secretariat discussion paper, prepared at the request of the 2025 Canadian G7 Presidency, in support of the proposed G7 SME AI Adoption Blueprint for the Industry, Digital and Technology Ministerial Meeting, Montreal, 7–9 December 2025.

Central claim

AI is diffusing across G7 economies but unevenly: large firms adopt at roughly 3.4× the rate of small firms, the gap is wider for AI than for any other digital technology, and it persists after controlling for sector, age and asset composition. Closing it requires four enablers operating simultaneously — connectivity, AI-enabling inputs (data + compute), skills, finance — calibrated to where each SME sits on a buyer-side taxonomy (Novices → Optimisers → Explorers → Champions). G7 policy approaches converge on operational instruments even where they diverge in framing.

Key arguments and findings

  • OECD-wide AI adoption (firms 10+ employees) rose from 5.6% to 14% between 2020 and 2024; the largest jump (8.5% → 14%) is 2023→2024, the generative-AI inflection. Still well below cloud (>50%).
  • In 2024 OECD-wide: 40% large firms (250+), 20.4% medium (50–249), 11.9% small (10–49) used AI.
  • AI adoption in core business functions in G7 ranges from 1.9% (Japan) to 6.1% (US) in 2024 — depth lags headline adoption.
  • Generative-AI individual usage outruns firm adoption: ~40% of US adults 18–64 used generative AI by late 2024 (22% weekly at work); ChatGPT hit 700M WAU by July 2025 with 2.6B daily messages, 27% work-related.
  • Country trajectory 2023–2025 (Figure A.1): UK 9.9→16.1→20.5%; Germany 11.6→19.7%; Japan 16.9→18.4%; Canada 6.1→12.2%; France 5.9→9.9%; Italy 5.0→8.2%; US 4.9→6.2→9.2%.

2. Adoption gaps

  • Gap holds across all G7 countries, all AI applications, and persists after controlling for sector, age, and asset composition — a structural size effect.
  • Gap is largest for capital-intensive applications (autonomous robots: 7.2% large vs 0.7% small, ~10×) and smallest for low-friction generative tools (NLG: 16.7% vs 4.6%, ~3.6×).
  • Country adoption ratios large:small (Figure A.2): Italy 4.7× (widest), France 3.85×, Germany 2.85×, Japan 2.0×, Canada 1.9×, UK 1.85×, US 1.65× (narrowest).
  • Only 29% of SMEs using generative AI deploy it in core business activities; most use is peripheral (marketing, content drafting, customer service).
  • Younger firms adopt more regardless of size; US AI-hub start-ups outpace incumbent SMEs.

3. Sectoral patterns

  • 2024 OECD averages: ICT ~44.5%, Professional services ~26%, vs Construction 7.2%, Accommodation & food 7.8%, Transportation 9.2%. ICT pulled away from manufacturing by ~15 pp between 2021 and 2024.
  • Calvino et al. (2024) multidimensional AI-intensity index puts Media (58–60), Telecom (61), IT services (62–63), Computer & electronics (26), Scientific R&D (72) in the top quartile across human capital, innovation, exposure, and use.
  • All sectors roughly doubled adoption 2021–2024 but absolute gaps widened.
  • AI adopters cluster in the top productivity decile: +40% France, +120% Germany, +240% Italy vs. bottom decile.
  • Within-cohort labour-productivity premium for AI adopters: >4% in most G7 countries, sometimes >15% — but shrinks once cloud, ICT specialists, training and complementary digital capabilities are controlled. AI is a multiplier of complementary digital capital, not an independent driver. J-curve drag (McElheran et al. 2024).
  • Macro projection (Filippucci et al. 2025): 0.2–1.3 pp annual labour-productivity gain across G7 next decade — Japan/Italy 0.2–0.8 pp, UK/US 0.4–1.3 pp. Reference: US ICT boom of 1990s delivered 0.5–1.5 pp.

5. Four-quadrant adopter taxonomy

Two axes (complexity of AI use × scope of AI application), with digital maturity as diagonal third dimension:

  • AI Novices — off-the-shelf LLMs, isolated/functional scope. SF coffee roaster (ChatGPT for SEO/copy); Hamburg freelance photographer.
  • AI Optimisers — off-the-shelf, cross-functional. Sèvres artisan bakery (ChatGPT + Midjourney + Shopify); Brescia bag manufacturer.
  • AI Explorers — custom/frontier in narrow scope. Tokyo wholesale B2B with custom multilingual sales-mediation agents.
  • AI Champions — custom/frontier across functions. Calgary healthcare-tech SME (LLMs + NLP + CV for clinical notes); Cambridge biotech with proprietary knowledge-graph + LLM stack for rare-disease drug discovery.

6. Four enablers

  • Connectivity. Small-firm high-speed broadband 41.86% (2020) → 59.97% (2024); large-firm 69.40% → 87.59%; gap ~27.5 pp throughout — diffusing but not converging. Median fixed broadband speed 53 → 178 Mbps (Q4 2019 → Q4 2024); metropolitan-vs-remote mobile gap widened from 5 to 45 Mbps. Japan fibre share 79% (G7 leader); Germany 12.2% (below OECD average). FWA y/y growth: US +39%, UK +30.4%.
  • AI-enabling inputs. Data interoperability and access to compute. Country compute pledges include UK GBP 2bn Compute Roadmap, Canada CAD 300m compute, France EUR 109bn AI pledge, Japan ~USD 65bn by 2030.
  • Skills. The most consistently cited barrier: 50%+ of SMEs in four G7 countries report employees lack the skills to use generative AI. D4SME lack-of-knowledge barrier: Japan 80%, Canada 50%, UK 40%, Germany 40%. Worker/skill shortages among gen-AI-using SMEs: Japan 44.7%/41.4%, Germany 43.9%/33.1%, Canada 31.1%/32.7%, UK 29.1%/37.3%. AI training participation: Canada 29.4%, UK 24.0%, Germany 23.2%, Japan 11.3%. Generative AI raises importance of both technical and judgement skills: data analysis 46.4%, creativity 41.9%, programming 39.0%, communication 35.8%, clerical 34.0%, critical thinking 33.5%, customer service 31.1%. AI vacancies +33% across 14 OECD countries 2019–2022; only ~1% require complex AI skills.
  • Finance. SME finance structurally and cyclically constrained; credit conditions tightened 2022–2024, shifting SMEs toward short-term lending — the wrong shape for multi-year AI integration.

7. Policy approaches (G7 country profiles)

Four operationally distinct G7 SME-AI policy models that converge on instruments even as they diverge in framing:

  • Market-led (US): SBIR/STTR, CREATE AI Act.
  • State-coordinated industrial policy (France/Italy/Germany): France EUR 109bn AI pledge + EUR 360m IA-Clusters + EUR 200m Osez l’IA; Italy EUR 6bn Transition 5.0 + EUR 350m SME digital; Germany EUR 5bn AI strategy + ~EUR 30m Gen-AI-for-SMEs.
  • State-coordinated + soft-law (Japan/UK): Japan ~USD 65bn by 2030 + ¥500bn Catalog-Type; UK GBP 2bn Compute Roadmap + GBP 1.2bn/y skills by 2028–29.
  • Federally-funded regionally-delivered (Canada): CAD 2.4bn (5y) + CAD 300m compute.

Germany is the only G7 country where client disapproval ranks as a top barrier among gen-AI non-users — a trust-driven, EU-rights-led signal.

8. Policy recommendations

The OECD proposes a policy pathway matrix mapping the four adopter stages to different enablers and instruments: Novices need awareness and off-the-shelf access; Optimisers need integration and skills; Explorers need compute and data infrastructure; Champions need R&D and frontier-talent policy. This is the buyer-side mirror of the supply-side L0–L7 substrate ladder and a buyer-side analogue of taker/shaper/maker.

Figures interpreted

  • Figure 1 (p. 9) — OECD AI adoption trajectory 2020→2024: 5.6 / 7.3 / 8.1 / 8.5 / 14.0%.
  • Figure 2 (p. 11) — country × firm-size adoption: Korea ~30% all-firms (~63% large), OECD avg ~14%, Australia/Japan/Israel/US ~3–6%.
  • Figure 3 (p. 12) — EU AI applications by firm size: widest gap autonomous robots (10.9×), narrowest NLG (3.6×).
  • Figure 4 (p. 13) — OECD sectoral adoption 2021–2024: ICT ~25 → ~45%, Professional services ~13 → ~26%, Construction/Accommodation stay below 8%.
  • Figure 5 (p. 14) — sectoral AI-intensity heatmap across 4 indicators × 22 sectors; top-quartile Media, Telecom, IT services, Computer & electronics, Scientific R&D.
  • Figure 6 (p. 18) — taxonomy framework: scope-of-application × complexity, digital maturity as diagonal shading.
  • Figure 7 (p. 20) — four-quadrant taxonomy plot.
  • Figure 8 (p. 29) — persistent ~27 pp small-vs-large broadband gap 2020–2024.
  • Figure 9 (p. 31) — worker / skill shortages by G7 country.
  • Figure 10 (p. 32) — gen-AI impact on skill needs stacked (UK 38.6/50.5/10.9; Japan outlier at 5.1/90.8/4.1).
  • Figure 11 (p. 33) — skill-importance shift across seven categories.
  • Figure 12 (p. 34) — AI training participation by country.
  • Figure A.1 (p. 60) — G7 adoption 2023–2025 from national sources.
  • Figure A.2 (p. 61) — large-to-small adoption ratios across G7; Italy 4.7× widest, US 1.65× narrowest.

Hypothesis touches

  • H1 — H1_L0-L7-ladder: Neutral-supporting. The OECD adopter taxonomy is a two-axis decomposition (complexity × scope) on the buyer side, structurally orthogonal to substrate ownership. Most Champion case-study SMEs sit at L0–L2 regardless of quadrant (Cambridge biotech orchestrates rented frontier models). It adds a third independent dimension to existing maturity / autonomy / positioning frames rather than testing the ladder.
  • H2 — H2_u-curve-of-value: Contradicting (specifically the bottom-right). The long-tail application surplus the U assumes is thin where the population is thickest: 11.9% small-firm adoption, 29% of gen-AI-using SMEs deploying in core, ~7% in SME-heavy lagging sectors, persistent 27.5 pp broadband gap. Without dedicated diffusion policy, the bottom risks collapsing into an L-shape. Adds a third axis of doubt alongside circular-ai-economy and divergent-value-stack-optima.

Open questions

  • Do the scaling-gap (intra-firm, pilot-to-scale) and the sme-ai-adoption-gap (inter-firm, small vs large) compound, and by how much? OECD does not stratify the BCG number by firm size.
  • Will the Novice → Optimiser → Explorer/Champion progression happen, or will most SMEs settle at Optimiser (wide-but-shallow)? The 29% core-activity share is the diagnostic statistic.
  • How much of the firm-level AI productivity premium is causal vs selection? The macro 0.2–1.3 pp range remains optimistic relative to task-based-framework’s <0.71% cap.
  • Where does the surplus from SME adoption land — at the SME, the SaaS/model vendor, or the bundling platform (Shopify, Microsoft, Google Workspace)? OECD is silent on capture.
  • Why is Japan such an outlier on skill-needs reporting (90.8% “no change”)? Turnkey use cases or systematic under-reporting?
  • Is the constant 25-pp firm-size broadband gap a steady state or chronic underinvestment? Implications for AI diffusion differ sharply.
  • Italy has the widest large-vs-small AI adoption divide (4.7×) despite ambitious Competence Centre and Transition 5.0 policy — lag, structural firm-size distribution, or policy reach failure?
  • Convergence of G7 instruments alongside divergent rhetoric: does this weaken the digital-empires tripartite frame as a predictor of real capital flows?
  • The OECD frame is silent on China — how would the four-enabler quartet apply to a state-directed AI economy?
  • What is the size distribution within “AI Champions”? Case studies span very different scales — the taxonomy may be too coarse at the top-right.

See also

Concepts created

Ideas created (case patterns)

Data created

Existing wiki touchpoints

Referenced by

Auto-generated reverse-link index — pages in the wiki that link here. Maintained by /lint and reverse-link sweeps.

Hypotheses

Search

Concepts

Sources

Ideas

Data

Graph View

Backlinks