Anthropic Report: Robots Can Perform 74% of Physical US Tasks but Are Cost‑Competitive for Only 0.3% of Jobs
TL;DR
- What was announced: Anthropic released a new “robot exposure index” that quantifies how many U.S. job tasks robots can perform today and evaluates cost‑competitiveness.
- Why it matters: The index shows robots can already do 74% of physical tasks (34% of total work hours) but are economically viable for only 0.3% of jobs, implying a long‑run, capability‑driven path to physical automation.
Key Findings at a Glance
- Robot exposure: 74 % of physical tasks (34 % of all work hours) are doable by today’s robots in some environment (purpose‑built, structured, or unstructured).
- Cost‑competitiveness: Robots beat human labor costs for just 0.3 % of tasks; a 70 % cost decline would be needed to reach 10 % coverage, which would take ~40 years at historic 3 % annual price drops.
- Job‑level impact: Jobs with higher exposure (e.g., taxi drivers, warehouse packers) have historically seen larger wage and employment declines; exposure has risen ~2 % per year over the past five decades.
- Demographics: Highly exposed workers are disproportionately male, less educated, lower‑paid, and more likely to be Hispanic.
- Barriers: Capabilities (especially manipulation) limit ~70 % of physical tasks; cost is the second‑largest obstacle; regulations and human preferences each block ~14 % and ~25 % respectively.
Measuring Robot Exposure
Exposure rubric
Anthropic rates each of the 7,594 physical O*NET tasks on a four‑tier scale:
| Tier | Definition |
|---|---|
| E0 | No robot can perform the task. |
| E1 | Robot can perform the task only in a purpose‑built environment (e.g., factory line). |
| E2 | Robot can perform the task in a structured human workplace (e.g., warehouse). |
| E3 | Robot can perform the task in an unstructured environment (e.g., city streets). |
"Robots must also do a task similarly well to humans, factoring in reliability, error rates, and speed." – Anthropic methodology footnote 13.
Claude was prompted to search for concrete robot deployments, cite sources, and assign the lowest‑structured environment where at least half of a task’s examples are feasible. The exposure index for a job is the time‑weighted average of its task scores (0–3).
Validation against history
Using O*NET data from 1977 onward, the authors re‑rated tasks with the capabilities available in each year. Jobs with higher historic exposure experienced larger wage and employment declines, even after controlling for industry trends. This back‑test supports the index as a predictor of future automation risk.
How Much Work Is Already Robot‑Ready?
Physical‑task coverage
- E0 (unexposed): 12 % of all tasks (≈ 25 % of physical‑task time). Example: hair‑dyeing, scaffolding erection.
- E1 (purpose‑built): 23 % of physical‑task time. Example: meal‑tray assembly on conveyor belts.
- E2 (structured workplace): 10 % of physical‑task time. Example: medication delivery robots in hospitals.
- E3 (unstructured): 2 % of physical‑task time. Example: autonomous driving.
Overall, 74 % of physical tasks (34 % of total work hours) are robot‑exposable at some tier.
Occupation‑level exposure
Figure 4 (in the original report) lists the ten most exposed occupations. Nine of the ten are vehicle operators; taxi drivers have the highest index (2.2). Most of their tasks are rated E3, while auxiliary tasks (e.g., interior cleaning) are E2.
Cost‑Competitive Automation
Anthropic estimated the annual cost of deploying a robot for each exposed task, including hardware, integration, maintenance, energy, supervision, and capital costs. A task is cost‑competitive when the robot’s annual cost is lower than the human labor cost for that task (derived from BLS total compensation weighted by task time).
| Occupation | Share of tasks exposed (E1‑E3) | Robot cost vs. labor | Cost‑competitive? |
|---|---|---|---|
| Packers & packagers | 97 % | $45 k/yr per robot replaces $49 k/yr labor | Yes (0.3 % of all jobs) |
| Welders | 68 % | Robot cost ≈ 5× labor cost | No |
| Dishwashers/cleaners | 55 % | Robots several times more expensive than $25‑30 k/yr labor | No |
Only 0.3 % of all job tasks are cost‑competitive today, representing roughly 300 000 workers. A uniform 20 % price drop would raise this to 2.8 million workers (0.8 % of total work time). At a historic 3 % annual price decline, a 70 % drop—and thus 10 % coverage—would require about 40 years.
Interaction with Large Language Models (LLMs)
When LLM exposure (tasks that an LLM could halve in time) is combined with robot exposure, 81 % of all work is covered by at least one technology. Robots add exposure especially in transportation, moving, and office‑admin support, where LLMs alone cover <15 % of tasks.
"Robots expose more and different kinds of jobs than LLMs alone." – Section Robot and LLM exposure.
The remaining 19 % of work is highly interpersonal or requires fine‑motor skills that neither current robots nor LLMs possess.
Who Is Most Affected?
Using the 2020‑2024 American Community Survey, Anthropic compared workers in the top quintile of the robot exposure index with those in the bottom quintile.
- Gender: 20 pp fewer women among highly exposed workers.
- Education: 55 pp fewer hold a bachelor’s degree or higher.
- Income: $30/hour lower average hourly pay.
- Unemployment: More than double the unemployment rate.
- Physical demands: Much higher incidence of heavy‑lifting, extreme heat, and hazardous‑contaminant work.
These patterns contrast sharply with LLM‑exposed jobs, which tend to be higher‑skill, higher‑pay, and more gender‑balanced.
Main Barriers to Wider Adoption
| Barrier | Share of physical tasks limited | Typical example |
|---|---|---|
| Capabilities (manipulation, planning, mobility, perception) | ~70 % | Fine‑motor tasks like “Bleach, dye, or tint hair” or “Erect scaffolding”. |
| Cost | – (cost‑competitiveness analysis above) | |
| Regulation | 14 % | Autonomous medical transport, protective‑service duties. |
| Human preferences | 25 % | Tasks involving trust (e.g., dressing children) or social interaction (e.g., greeting guests). |
If capability gaps close faster than cost declines, the remaining 30 % of tasks could become widely automated, but the current cost barrier will still dominate near‑term adoption.
Future Outlook & Scenarios
- Historical trend: Robots have added ~2 % of physical task coverage per year since 1977.
- Cost‑decline scenarios: At a 3 % annual price drop, robots become cost‑competitive for half of physical work only after 2085. A “fast‑adoption” scenario (four‑fold faster cost drops and double the capability growth) pushes that horizon to 2050.
- Policy implications: Early monitoring should focus on occupations with the highest exposure index (vehicle operators, warehouse packers). Regulation, safety standards, and labor‑market policies will shape the speed of adoption.
- Complementarity with LLMs: As robots gain manipulation dexterity, they may combine with LLMs to automate tasks that currently require both physical and cognitive skills (e.g., assisted construction, on‑site diagnostics).
Conclusions
- Robots are technically capable of most physical work today, but economic and capability barriers keep widespread automation limited.
- Exposure predicts labor‑market outcomes: historically, higher robot exposure correlates with wage and employment declines.
- Cost is the dominant near‑term hurdle; a 70 % price reduction is needed for robots to be cost‑effective for 10 % of jobs, a timeline of roughly four decades at current price‑trend rates.
- Policy and research should prioritize (1) lowering robot costs through scaling and modular design, (2) advancing manipulation and perception capabilities, and (3) addressing regulatory and preference barriers for high‑impact occupations.
Anthropic’s robot exposure index provides a concrete, data‑driven foundation for tracking the economic impact of embodied AI and for guiding both corporate investment and public policy.
References (selected)
- Legate‑Yang, R., & Massenkoff, M. (2026). What work can robots do? Anthropic. https://www.anthropic.com/research/what-work-can-robots-do
- Eloundou, T., et al. (2024). GPTs are GPTs: Labor Market Impact Potential of LLMs. Science, 384, 1306‑1308.
- Acemoglu, D., & Restrepo, P. (2020). Robots and Jobs: Evidence from US Labor Markets. QJE, 128, 2188‑2244.
- Rivière, Y., & Denain, J. (2026). Where Autonomy Works: Evaluating Robot Capabilities in 2026. Epoch AI.
- BLS (2025). Occupational Employment and Wage Statistics. https://www.bls.gov/oes/tables.htm
Data and code for the robot exposure index are available in the accompanying data release linked in the original Anthropic post.
Sources
- OriginalWhat work can robots do?