AIEL Monitor · monthly brief · 2026-09
AI and the labour market, September 2026
A monthly snapshot from the AI-Econ Lab: international, with Sweden in depth, on public data. In focus this month: Firm size and sector: who adopts?
Are the laggards catching up?
Which firms have actually started using AI, and which have not? Adoption is measured by a survey of firms. Every industry uses far more AI than it did in 2021, so the interesting question is no longer who has started but whether the ones that started late are catching up.
Catching up, and falling behind
And by industry, on the same survey and the same year.
Adoption climbs steeply with firm size: 31% among small firms (10–49 employees) against 72% among large ones (250+) in 2025, and every size class has risen since 2021. The industries that adopted least are growing fastest in proportional terms and falling further behind all the same: construction multiplied its adoption by 7.5 since 2021, against 2.8 for information and communication, and yet the distance between highest and lowest industry widened from 27 to 76 percentage points.
Using AI and hiring for it are different decisions, and the same industries do not make both. In our own advertisement data, manufacturing asks for a named AI skill 5 times as often as the other service industries, 2.13% of its advertisements against 0.46%, even though SCB records the service firms as the heavier users, 51% against 33%. One possible explanation: manufacturing more often builds its own AI tools and must hire for them, while service firms use tools that already exist.
Source: SCB, ICT usage in enterprises (NV0116), 2025. Full method at ai-econlab.com/monitor/#method. * No 2021 figure is published for these rows.
Use is not intensity
These numbers are what firms report in a survey: whether they use AI at all, not how much they use it. The job-advertisement data measure something else again: whether firms want to hire for specific AI skills. The two should be compared with care.
From the lab
DAIOE is now an open, citable dataset. The lab's Dynamic AI Occupational Exposure measure (occupation-year AI-exposure scores for 2010–2023 with a 2024 refresh, on five occupational classifications including Swedish SSYK) is published on Zenodo with a DOI (10.5281/zenodo.21873968), together with all the code that builds the measure and the full documentation on GitHub. Quick starts in Python and Stata get a ranked list of occupations in a few lines. This is the measure behind several of the Monitor's exposure indicators; anyone can now use, check and cite the same numbers.