Where AI Value Lands

ai-skill-shortage-as-diffusion-bottleneck

4 min read

AI skill shortage as diffusion bottleneck

Across the OECD G7 evidence base, skills are the most consistently cited barrier to AI adoption among SMEs — more consistent than finance, more consistent than connectivity, more consistent than data access. This makes the skills constraint the single largest practical brake on AI diffusion into the long tail of firms.

The headline evidence

  • 50% of SMEs in a four-G7-country OECD survey report that their employees lack the skills to use generative AI (OECD, 2025).
  • Over 50% of ~1,000 SME respondents to the 2025 D4SME survey (six G7 countries) cite lack of knowledge about how to use generative AI as a barrier — with stark cross-country variation: 80% Japan, 50% Canada, 40% UK and Germany (Bianchini & Lasheras Sancho, 2025).
  • Roughly one third of SMEs report worker shortages, and a similar share cite insufficient skills/experience among staff (Figure 9; see ai-skill-shortage-sme-share).
  • Among SMEs using generative AI in Canada, Germany and the UK, twice as many report increased skill needs as report decreased ones (Figure 10). Japan is the outlier (5.1% increase, 4.1% decrease, 90.8% neither).
  • Under 30% of SMEs using generative AI report any AI-related employee training — ranging from 11.3% in Japan to 29.4% in Canada (Figure 12).

Why skills, not capital, is the binding constraint

Three independent mechanisms compound:

  1. Selection effect. Firms with a high share of high-skilled workers are more likely to adopt digital tech and show higher returns to adoption (Calvino et al., 2022). Smaller firms with skill-heavy leadership over-perform on both adoption and ROI.
  2. Operational effect. ICT specialists are correlated with AI adoption across OECD countries even after controlling for firm size, age and sector (Calvino & Fontanelli, 2023). ICT engineers in particular are pivotal (Fontanelli et al., 2025).
  3. Training-access effect. SMEs have fewer staff to spare for training, higher per-worker training cost, and face poaching risk if they invest in scarce skills (OECD, 2023). Time constraints are the most-cited barrier to non-formal job-related learning across the OECD.

The result: skills shortages compound rather than dissipate. A firm that lacks the skill to adopt also lacks the slack to train.

What kind of skills

The paper is emphatic that both technical and non-technical skills matter. AI vacancies are concentrated (only ~1% of vacancies require complex AI skills), but a third of all vacancies are highly AI-exposed (Green, 2024). For exposed jobs, the OECD finds generative AI has raised the importance of:

  • Data analysis and interpretation (46.4% “more important”)
  • Creativity and innovation (41.9%)
  • Programming and coding (39.0%)
  • Communication and collaboration (35.8%)
  • Clerical and administration (34.0%)
  • Critical thinking and problem-solving (33.5%)
  • Customer service and sales (31.1%)

This is the empirical anchor for fusion-skills — adoption demands a blend of technical fluency and judgement/creativity/communication.

Implications for the working hypotheses

  • diffusion-vs-innovation: Strong support. If the binding constraint on AI diffusion is human capital, then a country’s diffusion capacity is governed by its labour market and training system — the Ding thesis, validated at SME scale.
  • H2_u-curve-of-value: Thins the bottom-right tail. Without skilled workers, SMEs cannot reach the application layer where value supposedly lands. Value can only accrue to small firms whose owners/managers are already digital-literate; the rest stall at “novice” (see sme-policy-pathway-novice-to-champion).
  • scaling-gap: A direct mechanism. Pilots fail to scale not because the model is wrong but because there is no internal workforce to operate and adapt it.
  • taker-shaper-maker: SME skill levels gate the climb from taker to shaper to maker.

Open question

Why is Japan such an extreme outlier — 80% reporting a knowledge barrier, 90.8% reporting no change in skill needs, lowest training rate? Either (a) Japan’s SMEs are using gen-AI for narrower, more turnkey use cases that genuinely don’t require skill changes, or (b) Japanese respondents systematically under-report. The OECD paper does not adjudicate. Worth flagging as a research gap.

Seller-side mirror

The skill shortage at the buyer side creates the demand for what ai-deep-tech-researcher startups supply (frontier AI capability that the customer cannot build in-house). It is the demand-side reason the Weber pattern exists. See ai-startup-business-archetypes-weber for the full four-archetype context.

See also

Graph View

Backlinks