Anthropic's economic research team published a report titled "What work can robots do?" on September 30, written by Russell Legate-Yang and Maxim Massenkoff. The findings split into two layers: on technical capability, robots can already handle 74% of physical work tasks in the United States, tasks that account for 34% of total hours worked across all jobs. On cost, only 0.3% of tasks are currently cheaper to hand to a robot than to a human worker.
Where the Data Comes From
The foundation is the U.S. Department of Labor's O*NET occupational database, covering roughly 900 occupations and 19,000 specific tasks. The team layered on 2025 occupational employment and wage data from the Bureau of Labor Statistics, 2026 employer compensation cost data, and the American Community Survey from 2020 to 2024. Claude was used to score each task's "robot exposure" and estimate automation costs.
The report sorts physical tasks into four tiers based on how controlled the environment is:
- Environments purpose-built for robots (E1), accounting for 23% of labor hours
- Human workplaces with some structure (E2), accounting for 10%
- Unstructured environments like roads (E3), accounting for 1%
- Physical tasks robots cannot do at all (E0), accounting for 12%
The occupations with the highest exposure are taxi drivers (index 2.2), followed by shuttle and chauffeur drivers (2.0), then warehouse, delivery, and packaging workers. The report also layers robots together with large language models, concluding that combined, only about one-fifth of jobs fall entirely outside their reach.
Where the Bottlenecks Are
Among four categories of obstacles, capability gaps affect roughly 70% of tasks, cost constraints touch nearly all of them, regulatory limits apply to 14% of physical tasks, and a preference for human service applies to 25%. The report singles out dexterous manipulation:
"Manipulation capabilities stand out: half of physical tasks wouldn't be automated at scale unless robots become more adept at touching and handling objects."
The 40-Year Math
On cost, the report runs the numbers using historical data: robot prices have fallen by roughly 3% a year in the past. At that pace, the share of tasks that are cost-competitive would climb from 0.3% to 10% — a process estimated to take about 40 years.
At that same rate, robot prices would roughly halve every 23 years, reaching about 30% of today's level after 40 years. At that pace, hardware getting steadily cheaper alone isn't nearly enough to close the gap with human labor — machines would also need to handle more tasks per unit, or deployment and maintenance costs would need to drop significantly.
The report lays out its own assumptions: it applies a single price-decline rate across all tasks, does not yet account for feedback effects from preference and regulation, and calls its cost estimates "approximate." Humanoid robots and embodied AI models have drawn heavy funding over the past two years; if the cost of dexterous hands, training data, and whole-unit systems falls faster than 3% a year, the 40-year figure would shrink substantially — a scenario the report did not model separately.
Who's Most Exposed
Compared with workers who face little exposure, the fifth of the workforce with the highest exposure has a share of women 20 percentage points lower, a share of Latino workers 16 points higher, a share of workers with a bachelor's degree 55 points lower, hourly wages roughly $30 lower, and an unemployment rate more than double.
The study relies solely on U.S. data. The task structure, labor costs, and factory automation levels in China's manufacturing and logistics sectors differ substantially from those in the United States, so the figures of 0.3% and 40 years don't translate directly.
Sources: Anthropic economic research report, CocoLoop, U.S. Department of Labor O*NET database documentation. Task share and hours-worked figures follow the report's own methodology; the price-halving timeline is an estimate based on a 3% annual decline.