Analysis · 10 September 2026 · 7 min read

What AI can do vs what it does

In computing, 94 percent of tasks are feasible for a model and 33 percent are actually being done. That three to one gap is your runway.

There are two numbers for every job, and almost every article about AI and work quotes only the first one.

The first is what a model could do. In computer and mathematical occupations, Anthropic's economists put that at 94 percent of tasks being theoretically feasible for a language model. That is the kind of figure that becomes a headline.

The second is what is actually being done. In the same occupations, measured on Anthropic's own platform traffic in automated, work related use, it is 33 percent.

Roughly three to one. The distance between those two numbers is the most practically useful thing in this research, and it is where your planning actually happens.

The gap is not a rounding error, it is the timeline

A capability that exists and is not being used is not a capability that has changed anyone's job yet. Everything that fills the gap is unglamorous and slow: someone has to decide to change a process, buy the thing, integrate it with systems that already exist, work out who is liable when it is wrong, retrain the people who stay, and get the cost per task under what the work currently costs.

Those steps run at different speeds in different industries, which is why the same task can automate years apart at a bank and at a startup. It is also why capability announcements are a poor guide to your own timeline. Our post on timing covers the published projections; the short version is that adoption, not capability, sets the date.

The Massenkoff and McCrory paper, published on 5 March 2026, is unusually direct about what has and has not shown up in the labour market so far.

What has actually happened to employment

No systematic increase in unemployment among highly exposed workers since late 2022. That is the paper's own finding, and it deserves to sit next to every alarming projection you have read.

One real signal, and it is at the entry level. For workers aged 22 to 25 in exposed occupations, the job finding rate dropped about 14 percent compared with 2022. The authors describe this result as just barely statistically significant, and it is a decline measured against a monthly hiring rate of around 2 percent, so it is a small effect on a small base. We report the caveat because they did.

It is worth noting how well that lines up with independent evidence. Stanford's payroll study found a 16 percent relative employment decline for workers aged 22 to 25 in the most exposed occupations while their senior colleagues held steady. Two different teams, two different datasets, the same shape of finding: the pressure lands on the entry rung first.

And the 2026 layoff wave gives the third data point. AI was cited in US job cuts covering 205,000 workers through August, concentrated in customer service, data operations, entry level software and finance back offices. Those are precisely the occupations at the top of the observed exposure list, which is the gap closing in the places it was always going to close first.

What this means for reading your own score

Here is the frame worth carrying away.

A theoretical exposure score is a ceiling, not a forecast. It tells you how much of your job is technically within reach. It does not tell you how much has moved, and on the evidence above, most of it has not.

An observed score tells you how far along your field already is. Eleven occupations out of 756 have crossed 50 percent on that measure, and the median across the whole dataset is zero.

The gap between them is your runway. It is the time you have to become the person who directs the tools rather than the one producing the output they now produce. That is a real window and it is not evenly distributed: if your work sits high on both measures, yours is shorter than most.

Our index carries both kinds of evidence. It aggregates this Anthropic dataset alongside independent research across 961 occupations, weighted so no single source decides a score, which matters because studies measuring different things get published under one label. You can see the sources behind any occupation on its own page, for example Financial and Investment Analysts (56) or Medical Records Specialists (60).

Your number is not your occupation's number

The one thing neither figure can tell you is where you sit, because both describe the typical version of a role and nobody works the typical version. Two people with the same job title and different weeks have genuinely different exposure, and that difference is usually larger than the difference between two occupations.

Our assessment closes that gap. It asks about your seniority, your sector and how much of your week is routine, then returns a score for your actual task mix rather than your title, plus where you sit against the research baseline for your occupation. It takes about two minutes, it is free, and it needs no signup or card.

If you read only one number about AI and your job, make it that one rather than a 94 percent from a headline.

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.