On September 30, 2026, Anthropic economists Russell Legate-Yang and Maxim Massenkoff published “What work can robots do?”, which builds a robot exposure index for the US economy. They used Claude to rate 7,594 physical job tasks from the O*NET occupational database on four tiers, E0 to E3, according to how controlled an environment a robot needs to do the task, from a purpose-built factory cell up to unstructured places such as city streets.
The headline is a gap between capability and cost. By the index, robots can perform 74 percent of physical tasks, which account for 34 percent of all working hours, though mostly only in controlled settings. Combined with exposure to language models, roughly 80 percent of job tasks face some automation exposure. Yet robots are cost-competitive today for only 0.3 percent of job tasks, and at the historical price decline of about 3 percent a year it would take about 40 years to reach 10 percent. Looking back 50 years, jobs with higher robot exposure saw larger declines in wages and employment. Workers in the most exposed occupations are disproportionately male, less educated, paid about 30 dollars an hour less, and face unemployment more than double the average; taxi drivers (index 2.2) and warehouse workers are among the examples.
Why it matters: most AI labor research measures exposure to language models, which mainly touches office work. This extends the same method to physical work and adds the missing variable, price, which explains why technically possible automation has not happened.
What it does not show: the ratings come from a model, not from observed deployments. The authors note the index assumes today’s demonstrated robot capabilities predict future disruption, and that AI-driven robots could leapfrog the scale by gaining abilities never shown in controlled settings, in which case the 40-year projection would be far too slow.