Anthropic's Economic Index measures how AI is actually being used across the economy by mapping real usage onto occupational tasks, rather than asking people in a survey what they think they might do with it. It is one of the few published measures built on observed behaviour instead of stated intent, which is what makes it worth building on.

It is also, by construction, shaped around the United States. We are working with TAC ECONOMICS on an adaptation to the EU27, and this post is about why that is a research problem rather than a data-processing one.

The occupational map does not transfer

The index rests on a task taxonomy: a structured description of what occupations actually consist of. The American one is O*NET, maintained by the US Department of Labor, and the index inherits its categories, its granularity and its assumptions about how work decomposes.

Europe does not use it. Occupations are classified under ISCO-08 and described through ESCO; sectors run on NACE; the labour force data lives at Eurostat and in twenty-seven national statistical institutes. None of these is a relabelling of the American equivalent. They cut the same reality at different joints, and a crosswalk between them is an analytical choice with consequences, not a lookup table.

Twenty-seven labour markets, not one

The second difficulty is that "the EU27" is a legal entity, not an economy with a single labour structure. The sectoral composition of Ireland and that of Romania have little in common; the share of an occupation exposed to AI means something different in each. An aggregate figure for the Union that is not built up from comparable national figures is a number without a referent.

Language is the third. Usage data in a European setting is multilingual by default, and any mapping from what someone asked to the task it corresponds to has to hold across languages rather than degrade quietly in the smaller ones.

What we are publishing

The work is joint: TAC ECONOMICS brings the economic framing and the labour market side, we bring the data and model engineering. The output will be the methodology and the results together, including where the adaptation is well-founded and where it rests on assumptions a reader should be able to disagree with.

We will publish it here when it is ready. If you work on this and want to argue with the approach before it is fixed, we would rather hear it now.