AI and the labour market, August 2026

The question

Capability is no longer the obvious constraint. Frontier AI agents now finish tasks of up to 12–17 hours of human-expert work about half the time, a length that has been doubling roughly every four months, and the computing power behind notable models has grown about 5.2-fold a year since 2020. Studies of individual tasks find the gains are real: 15% more customer-support issues resolved per hour in one firm (Brynjolfsson, Li and Raymond, 2025); 40% less time and 18% higher quality on professional writing (Noy and Zhang, 2023); 12% more consulting tasks, completed 25% faster, but 19% less accurately on one task chosen to sit outside the frontier (Dell'Acqua et al., 2026). Yet in 2025 only 20% of EU enterprises used AI, 35% in Sweden, and AI skills were required in a median 1.9% of job postings across 22 countries, 2.8% in Sweden. So why is a technology that demonstrably helps still so little used? Firms were asked.

What the data show

The obstacle firms name most often is people, not money: 5.3% of Swedish enterprises say a lack of relevant expertise stops them using AI, against 1.2% who say AI is simply not useful to them. Cost ranks low, at 1.8%.

Source: Eurostat, isoc_eb_ain2, 2025. Full method at ai-econ-lab.github.io/monitor/#method.

What it does not show

Read this as how widespread each obstacle is, not as a division of non-adopters between causes. Eurostat reports the shares as a percentage of all enterprises with ten or more employees and allows several answers, so they do not sum to a total. They also cannot be rebased onto the 65% of Swedish firms not using AI, for a reason worth stating plainly: since 2023 the Swedish question is put only to firms that considered AI and decided against it, so most non-adopters are never asked. The ranking survives that, since every reason shares a denominator, but the levels describe a narrow group and the silence of the rest is a routing artefact rather than a finding.