Anthropic’s Economics team published “What work can robots do?” on Wednesday. It builds an exposure index: how much of the physical work done in the US could today’s robots actually perform? The headline number looks dramatic. The one underneath it is very sober.
Four tiers
The base is O*NET, with around 900 occupations and about 19,000 task descriptions. Claude scored each description for physical, cognitive and interpersonal demands, then searched the web for specific robots behind the physical ones. Deployments, commercial sales or demonstrations had to be documented, and the robot had to do the job about as reliably and quickly as a person.
What matters isn’t whether a robot can do a task at all, but where:
- E0 means it can’t.
- E1 means only in a purpose-built robotic environment, like an assembly line.
- E2 means a structured human workplace, like a logistics warehouse.
- E3 means an unstructured one, like a city road.
So the measure is how much order the environment needs. Robots struggle with the unpredictable, which is why most of them live in engineered worlds.
74%, 34%, 0.3%
Three numbers carry the study. Robots can do 74% of physical tasks in the US, and that comes to 34% of working hours, though mostly only in tight conditions. Together with language models, roughly 80% of all tasks by working time are exposed; what’s left is heavily interpersonal or needs dexterity robots don’t have. Driving and warehouse jobs are highly exposed. Nursing and general repair barely are.
Then the third number: for 0.3% of tasks, a robot is cheaper than a human. If prices keep falling the way they have, it takes 40 years for that share to reach 10%.
The 50-year backtest supports the method: jobs more exposed in 1977 did see bigger wage and employment declines in the decades that followed. The people in those jobs, the study notes, are more likely to be male, less formally educated and lower paid.
Capability isn’t the brake. Price is.
Automation arguments almost always run on capability. This study flips that: the capability is largely there, the price isn’t, and 40 years to ten percent is not a forecast that justifies panic. Worth noting that Anthropic uses the same approach here as in the economic scenario explorer: measure what works today, then compute, instead of guessing at years.
For those of us in software, the subordinate clause is the real sentence. Eighty percent exposed, the cognitive half of it through LLMs. The physical half of that sum is waiting on a price collapse. Ours isn’t.
Sources: