Data study · 10 September 2026 · 6 min read

The 11 jobs AI is actually doing

Anthropic measured AI use across 756 occupations. The median is zero and only eleven clear 50 percent. The list, the figures, and what a zero does not mean.

Most writing about AI and work describes what could happen. Anthropic's economists went and measured what is happening, across 756 occupations, using their own platform traffic. The result is one of the few numbers in this field that describes the present tense rather than a forecast.

It is also much smaller than you would expect, and far more concentrated.

Across all 756 occupations in the dataset, the median observed exposure is zero. Not low. Zero. The average is 7.7 percent. Only 86 occupations clear 25 percent, and only eleven clear 50 percent.

The eleven

This is the top of the published list, with Anthropic's own figures for the share of each occupation's tasks appearing in automated, work related use on their platforms.

Figures from Labor market impacts of AI by Maxim Massenkoff and Peter McCrory, published 5 March 2026. This dataset is one of the sources behind our own index, so these occupations already carry it in their scores.

Look at what the list has in common. Every one of them is a job where the work product is text, code, a record or a number, produced to a specification, at volume. Nothing on it requires hands, a room, or a person in front of you.

Now look at what is missing. No trades. No nursing. No teaching. No logistics. Not because those jobs are safe forever, but because nobody is currently handing those tasks to a model in a work context, which is what this measure counts.

What "observed" means, and why it matters

Most exposure studies ask a question of the form: could a model do this job's tasks? Anthropic asked a different one: is it doing them? Their definition is that jobs are exposed to the extent their tasks are both theoretically feasible for language models and observed on their platforms in automated, work related use.

That distinction changes what the number is good for. A theoretical score tells you where the pressure will land if adoption arrives. An observed score tells you where adoption already is. The two families of exposure research barely correlate, and this is a third thing again: not capability, but capability actually being used.

Two limits worth holding onto. It measures Claude, so a task being handled by a different assistant somewhere else does not appear. And it measures usage, so an occupation with few users on the platform reads as a zero whether or not it is genuinely untouched. A zero here means not observed, not proven safe.

That is exactly why our own index does not run on this data alone. It aggregates this source with others across 961 occupations, so a role gets a score built from several independent research groups rather than one platform's traffic.

What to do with your own number

If your occupation is on that list, the honest read is that the tools are already doing a meaningful share of the task types your job is made of, today, in real work settings. That is not the same as your job disappearing, and this same research found no systematic rise in unemployment among highly exposed workers. It does mean the change is not hypothetical for you.

If your occupation is not on the list, the useful question is not whether you are safe but which parts of your week resemble the work on it. Almost every job contains some share of produce-a-document, answer-a-query, reconcile-a-number work. The occupation average hides how much of that share is yours.

That share is what our assessment actually measures. It takes about two minutes, asks about your seniority, your sector and how much of your week is routine, and returns a score for your task mix rather than for your job title. It is free, and there is no signup.

You can also look up your occupation's research baseline directly in the index, or start from Computer Programmers (59) and Customer Service Representatives (68) to see how the two most exposed roles in the Anthropic data score once every other source is weighed in alongside it.

Figures come from the AI Job Risk Index and were current when this was published. Scores change as new research is ingested, so the index is always the live version. See how scoring works. Informational guidance based on published research, not professional career or financial advice.