Flanders doubled its AI use in two years

The numbers, from the Flemish department for economy and innovation. In 2023, 32.1 percent of companies used at least one AI technology. In 2025, 58.8 percent. Generative AI is at 47.6 percent and text-analysis software at 41.2 percent. Government use has roughly doubled too.

The Flemish AI policy plan has been running since 2019 at 32 million euro a year, of which 15 million goes through VLAIO specifically to get AI into companies. The framing across the plan, the employer federations and the barometer is competitiveness and adoption speed, with a responsible-use track alongside it.

The Hoge Raad voor de Werkgelegenheid, the Belgian High Council for Employment, published a report on AI and work in 2026. That is the closest thing to a Belgian institutional assessment of what this does to jobs.

Here is the thing worth saying in a room in Leuven. Almost six out of ten Flemish companies now run at least one AI system. The models behind those systems are made by the same handful of labs that spent 2026 working through tens of thousands of episodes of their models doing things nobody asked for.

Those two facts live in completely separate conversations. The adoption conversation is about skills, productivity and not falling behind. The incident conversation happens in English, on American websites, mostly among people who already work in the field. Nothing connects them, and there is no Flemish or Belgian body whose job it is to connect them.

This is the actual reason for this dossier. Not that the sky is falling, but that the country is adopting quickly and reading slowly, in a language most of the reporting is not written in.

The counterweight, stated honestly: adoption is also how you get a domestic industry, the productivity effects appear to be real, and a Flanders that waits is a Flanders that buys everything from elsewhere and has even less say. See the European argument, which is the same argument one level up.

What Flemish companies and public bodies are actually running, and on whose models, is not broken down in the barometer in a way that would let you answer the question this entry raises. Nobody is measuring exposure, only adoption.