The Pulse
Anthropic Finds Robots Can Do 74% of Physical Tasks, but Cost Too Much
Anthropic researchers estimate that robots can perform 74% of physical tasks in the United States, but are cost-competitive for just 0.3% of job tasks. The study says robot-cost estimates often include human supervision, exception handling
AI.info Team · Written with AI ·
Anthropic researchers estimate that robots can perform 74% of physical tasks in the United States, but that capability rarely comes at a price businesses can justify. In a study published September 30, Russell Legate-Yang and Maxim Massenkoff estimate that robots are cost-competitive for just 0.3% of job tasks. The distinction separates what a machine has demonstrated it can do from what an employer can afford to automate.
74% of physical tasks, under different conditions
The researchers assess 7,594 tasks classified as physical, drawing on O*NET’s descriptions of occupations and tasks. They use Claude to help classify work and estimate how much time workers spend on each task, then assess whether existing robots can perform it. The task estimates are weighted by estimated time spent and employment, so the headline share is not a count of robots or workplaces.
The study ranks robot capability by the environment a machine needs. A task may be beyond robots today, possible only in a purpose-built setting such as a factory line, feasible in a structured workplace such as a warehouse, or achievable in an unstructured environment such as a city road. Anthropic counts demonstrated capabilities and considers how reliably and quickly a robot performs compared with a person.
That approach produces a broad capability estimate but not a claim that robots can replace workers across ordinary workplaces. Driving and warehouse jobs rank among the more exposed because robots already perform portions of that work. Nursing and general repair rank lower, because present-day robots can do little of that work even in highly controlled environments.
The price gap is larger than the capability gap
“Robots are cost-competitive for just 0.3% of job tasks.” — Russell Legate-Yang and Maxim Massenkoff, the study’s authors
The study compares robot costs with worker compensation and finds a much smaller share of work where the machines can compete on price than where they can perform the task at all.
Anthropic says robot prices have fallen about 3% a year since the 1990s, based on industry and government data. If that pace continues, the share of work for which robots are cost-competitive would take about 40 years to reach 10%, the researchers estimate. The researchers add that cost parity need not mean high unemployment: tasks that workers still perform may become bottlenecks, and their robot-cost estimates often include human supervisory, exception-handling, or repair work.
Capability is another major constraint. The researchers estimate that limits in robot skills could prevent adoption across about 70% of physical tasks, with manipulation—the ability to touch and handle objects—especially important. They also identify human preferences and regulation as obstacles: people may not trust a robot with some work, such as dressing children, or may value the social interaction a person provides, while rules restrict machines in fields such as health care, protective services, and education.
Robots and language models reach different work
Anthropic estimates that robots and large language models together expose about 80% of job tasks by working time. The overlap does not mean either technology can complete those tasks independently or that employers will automate them. Rather, robots can potentially handle parts of physical work that language models cannot, while many tasks still demand interpersonal judgment or physical skills machines lack.
The study’s historical analysis adds context, not a forecast of immediate job losses. Looking back over 50 years, the authors find that jobs more exposed to robots experienced larger declines in wages and employment in later decades. They also estimate that robots have become capable of about 2% of the physical work they previously could not do each year, while stressing that future capability gains and adoption rates are uncertain.
Today’s footholds point to uneven change
The researchers’ estimates suggest that early effects are more likely in work where robots already operate, including driving and warehouse tasks, than in jobs that require adaptable hands-on care or repair. That pattern could change if machines become more capable, cheaper, and easier to deploy beyond controlled settings. But the study’s central constraint is immediate: a robot’s ability to perform a task does not make it an economical substitute for a worker.
Anthropic’s measure is designed to track tasks robots can perform now, not to predict which new abilities future systems will acquire. The researchers say periodic updates to the work can track robot capabilities. For now, their estimate puts cost parity at 0.3% of job tasks.