How the UK, Ireland and Italy actually use AI at work

AI has reached 68% of all occupations, which is two in three rather than all of them. But in the median occupation where it is used at all, it touches just 21% of the tasks. Only 3% of occupations show AI use across more than three quarters of their work: software QA testers, HR specialists and document management specialists. Read those three carefully, because their task catalogues absorb requests from adjacent work, which lifts them up the depth ranking.

Broad, and shallow. Two thirds of occupations touched, a fifth of the tasks inside them.

Anthropic's classifier also estimates how long each task would take a person working alone, and those estimates run well ahead of the time the conversations themselves take. We have deliberately not put a ratio on that, and you should be sceptical of anyone who does. It is a model-generated estimate rather than a measurement, it is bucketed by order of magnitude, and a finished conversation is not the same thing as a finished job. Fewer than one in ten conversations involving complex cognitive work is even an attempt to complete the task end to end. The direction is real. The multiple is not something either dataset can support.

Google's data says 13% of Gemini conversations are work-related. Anthropic's says 43.4%. That is a 3.3x gap, and it looks like a profound finding about how differently people use the two tools.

It is not. ATLAS covers the Gemini App, Google AI Mode and the Gemini API, but the 13% work share is computed on the consumer surfaces, which reach a search audience of over a billion monthly users. Google itself names that audience composition as a possible complete explanation for the difference. Anthropic's figure covers a paid assistant that people open deliberately.

Quote either number on its own and you have described a product's audience, not a workforce.

None of this is a criticism of anyone. It is where the technology sits a few years into public use, and the same pattern appeared in two independent datasets built by two competing labs.

But the distance between what these tools can do and what we currently ask of them is the entire commercial opportunity of the next few years. It closes fastest for people who learn to drive it properly, and it closes first inside organisations that treat this as a skill to be taught rather than a licence to be bought.

We are a training business, so of course we would say that. The numbers are public. Check them.

Claude usage figures come from the Anthropic Economic Index, May 2026 period, covering 121 countries with published detail down to region and subtopic level.

Gemini usage figures come from Google's AI & Economy ATLAS v1.0, published 23 July 2026, covering 14,653,926 de-identified interactions across the Gemini App, Google AI Mode and the Gemini API. The 13% work share we quote is computed on the consumer surfaces, not on the API.

UK business adoption figures come from the Office for National Statistics release Artificial intelligence in UK businesses, 2023 to 2026.

How we read the Anthropic tree. Request data is published as a hierarchy. Individual tasks such as business correspondence sit at the leaf level; groupings such as starting a business or translation sit above them. We label clusters as clusters, and we never rank a cluster against a leaf. Where we group tasks ourselves, as with the careers figures, that grouping is ours and not a category the dataset defines. Anthropic separately maps the same conversations onto the US occupational task catalogue, which is a different classification and cannot be compared line by line with the request tree.

Limitations we think you should know about. Both AI studies describe observed usage of their own product, not AI adoption across the market. Both are single-month snapshots with no trend series, so nothing here shows growth or decline. Neither identifies users, employers or company size, which is why the SME figures come from the ONS instead. Shares are of all classified conversations, so a topic at 2% of everything is a considerably larger share of the working subset. Not published means an entry fell below the privacy threshold and never means a measured zero. Both studies map conversations onto US occupational taxonomies, and Google notes this makes it difficult to judge whether an occupation is over-represented outside the United States.

Every figure was re-verified against the published datasets at source on 29 July 2026. If you spot an error in our reading of these datasets, tell us and we will correct it publicly.