AI could speed up 94% of computer and maths tasks; Claude covers 33%. The gap is where your role changes next

Fortune reported on 23 October that Pantheon Macroeconomics' Samuel Tombs and Oliver Allen estimate AI is now costing the US more than 10,000 jobs a month. The losses are in accounting, advertising, office administration and legal services. Construction jobs are rising at the same time, because technology companies are building data centres as fast as they can. The headline reads like a judgement on whole professions. Anthropic's own measure of how its model is actually used points somewhere more useful. It looks at the tasks inside a role, and at the distance between what AI could do there and what it is doing.

Capability is a poor forecast. Usage is a better one

Anthropic's economics team built a measure it calls observed exposure. It starts with task-level estimates from Eloundou et al. (2023) of whether a language model could make a task at least twice as fast. It maps those across the roughly 800 occupations in the US O*NET database. Then it checks them against real, work-related Claude traffic.

The gap is wide. In the Computer & Math category, language models could in theory speed up 94% of tasks. Claude currently covers 33%. Office & Admin occupations score 90% on theoretical capability.

Where usage has caught up, it is concentrated. Computer programmers top Anthropic's list at 75% coverage. Customer service representatives come next, because Anthropic increasingly sees their main tasks running through direct API traffic. Data entry keyers sit at 67%. At the other end, 30% of workers have zero coverage. They include cooks, motorcycle mechanics, bartenders and lifeguards.

The usage measure deserves attention because it tracks an independent forecast. For every 10 percentage point rise in coverage, the Bureau of Labor Statistics projection for that occupation's growth to 2034 falls by 0.6 points. Anthropic found no such relationship using theoretical capability alone.

What AI could do predicts little. What it is actually used for predicts more.

The exposed group is also not who the headlines imply. Workers in the most exposed quartile earn 47% more on average than those with no exposure. Graduate degree holders make up 17.4% of the most exposed group and 4.5% of the unexposed one.

What holds the gap open

Anthropic names the reasons usage trails capability: model limitations, legal constraints, specific software requirements and human verification steps. Its example is authorising drug refills. Eloundou et al. rate the task fully feasible for a language model, but Anthropic has not seen Claude perform it.

Look at that list as an operator. Apart from the model's limits, every item is a job someone has to do. A person builds the integration, signs off the legal exposure and designs the check. The gap closes when a capable person does that work, and that person then understands the new workflow better than anyone.

The measure also weights how AI is used. Fully automated use counts in full. Augmentative use, where a person works alongside the model, gets half weight. Anthropic treats automation as the stronger signal of job impact.

The Pantheon number needs one caveat. Anthropic found no systematic rise in unemployment among highly exposed workers since late 2022.

Tasks move first. Hiring slows next. In Anthropic's data, unemployment has not yet followed. That order gives you time, as long as you spend it on the gap rather than the headline.

Your Next Move

1. Sort your week into three columns. List the tasks you did in the last five working days. Mark each one: already done by AI in your team or field; feasible for AI but not yet done; not feasible. The middle column is your agenda for the next quarter, because that is where coverage grows.

2. Close one gap task yourself, with you in the loop. Pick a task from the middle column. Build the workflow so that AI does the routine part and you hold the verification step. Then write down the check you apply. AI does the routine work. You do the thinking. You become the person who knows how the task now runs, which is a stronger position than being the person it was taken from.

3. Own one of the obstacles. Where a task is held back by legal sign-off, a missing integration or a verification step, volunteer to own that piece. If you hire, use the same logic for entry-level roles: rebuild the junior role around checking and judgement before you decide not to fill it.