What work can robots do?

Anthropic research uses Claude to score ~19,000 O*NET task descriptions and constructs a robot-exposure index separating what robots physically can do from what is economics. Headline finding: robots 'can already perform 74% of physical tasks in the US, making up 34% of working hours,' and robots plus LLMs together 'expose all but one-fifth of employment' — yet robots are cost-competitive with human labor for just 0.3% of job tasks, with a projected 40 years for that share to reach 10% at historical price-decline rates. A 50-year backtest (1977–present) validates the approach: occupations more exposed to existing robots suffered greater wage and employment declines in subsequent decades, and robots add roughly 2% of previously-undoable physical work each year.

Gives executives a quantified, capability-vs-cost way to separate near-term from long-horizon physical-work automation — robots can do the work but almost never at competitive prices — sharpening workforce and capital-allocation planning that LLM-based forecasts ignore.

Use the exposure index to sequence which physical-work roles (driving, warehousing) face pressure first versus LLM-exposed roles, and pressure-test that against the cost-competitiveness 0.3% baseline before automating physical operations.