If robots can do much physical work, why are so few tasks economical to automate?
Anthropic researchers rate 7,594 physical job tasks and estimate that robots can perform 74% in some setting, yet are currently cheaper than people for only 0.3% of all work. The gap between technical exposure and economic adoption is the result that matters for workers.
Editorial responsibility: The Impact of AI Editorial Desk · Report a factual concern
Research topic
How current robot capability, deployment environment and cost change the meaning of occupational exposure
At a glance
- 1The study rates 7,594 physical tasks drawn from O*NET and estimates that robots can perform 74% of physical tasks in at least one setting, equal to 34% of all US working time.
- 2After estimated deployment costs are compared with labour compensation, robots are currently cost-competitive for only 0.3% of all tasks; a capability estimate is therefore not an employment forecast.
- 3The analysis finds that highly exposed workers are disproportionately male, Hispanic, less degree-qualified and lower paid, but the ratings and task-time estimates rely heavily on Claude-assisted judgements and US data.
Living evidence record
Impact record IAI-0L1S3JE
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent support
Present
Record status
Monitoring
Last checked
1 October 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what Anthropic published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
The study separates what a robot can do from what it pays to automate
Anthropic researchers Russell Legate-Yang and Maxim Massenkoff published a US robot-exposure index on 30 September. Their starting data are O*NET descriptions covering roughly 19,000 tasks across about 900 occupations. They identify 7,594 tasks as physical, then use a four-level rubric: a robot cannot do the task; can do it only in a purpose-built robotic environment; can do it in a structured human workplace; or can do it in an unstructured setting such as a public road.
On that definition, robots can perform 74% of physical tasks in at least one setting, representing 34% of all working time. That does not mean a robot can take over 34% of jobs. Many ratings describe a machine working on one task in a factory, warehouse or other constrained environment, while an occupation contains many tasks and exceptions. The report's most useful contribution is to keep technical exposure, workplace deployment and cost as separate questions.[1]
Cost reduces the near-term exposure estimate dramatically
For each exposed task, the researchers estimate the annualised cost of the relevant robot, including hardware lifespan, installation and the output required to match a worker. They compare this with Bureau of Labor Statistics compensation scaled to the share of a worker's time spent on the task. On those assumptions, robots are cheaper than people for only 0.3% of all tasks today. Roughly 300,000 workers are in jobs for which robots are estimated to handle 95% of tasks cost-competitively, but this is a modelled exposure group rather than a count of layoffs.
The paper estimates that robot costs would need to fall about 70% before automation became cost-competitive for 10% of current work. If quality-adjusted prices continue their historical decline of about 3% a year, that threshold would take roughly 40 years. A faster scenario reaches half of physical work by 2050, but only by assuming costs fall up to four times faster and capabilities expand twice as quickly. These are conditional scenarios, not dates by which jobs are predicted to disappear.[1]
The denominator and weighting choices are substantial
The headline percentages are not simple counts of task statements. Each task is weighted by occupation employment and by an estimated share of working time. Claude generates detailed task examples, searches for demonstrated or deployed robots, estimates how often examples occur and helps score the least structured environment in which a robot can perform at least half of the task's examples. The study requires citations showing a deployment, commercial sale or demonstration, and says results remain similar when demonstration-only ratings are excluded.
Those safeguards improve traceability, but the system still makes many judgement calls. O*NET descriptions can be terse, working-time shares are estimated rather than directly observed and web evidence is uneven across industries. Robot cost estimates also combine public price information with assumptions about utilisation, maintenance and equivalent output. The authors release reasoning and sources for rated tasks, which makes scrutiny possible, but publication by the company that built Claude is not the same as independent validation.[1]
Exposure is concentrated among workers already facing disadvantages
Using 2020–2024 American Community Survey data, the study reports that workers in the most exposed fifth of occupations are 20 percentage points less likely to be women and 16 points more likely to be Hispanic than unexposed workers. They are 55 points less likely to hold at least a bachelor's degree, earn about $30 less per hour and have more than double the unemployment rate. Driving, warehouse and material-handling work ranks high; nursing, general repair and highly interpersonal or dexterous tasks rank lower.
A historical back-test from 1977 links greater robot exposure to later wage and employment declines after controlling for industry trends and other measured confounders. That strengthens the case that the index contains information, but it does not establish that today's relationships will repeat under AI-powered robotics. New tasks, safety rules, unions, procurement choices, workplace redesign and regional labour shortages can change adoption. Workers need occupation-specific transition plans, not a single national automation percentage.[1]
What would change our assessment
Confidence would rise if independent teams reproduce the task ratings and cost estimates, if employers publish observed utilisation and total deployment costs, and if longitudinal data show predicted high-exposure occupations adopting robots before low-exposure ones. Updates should report how much a robot assists a person rather than replaces a task and whether productivity gains affect pay, injuries, hours and staffing. Comparable research outside the United States is essential because wages, regulation and work organisation change the economics.
The near-term displacement assessment would strengthen if robot costs fall much faster than the historical rate while dexterity and reliability improve in unstructured settings. It would weaken if integration, supervision, insurance and maintenance costs remain high, or if people and regulators continue to reject automation in care, education and safety-critical work. The present result is not that physical jobs are safe or doomed. It is that capability has moved much further than economical, socially acceptable deployment.[1]
What this means for people
- Drivers, warehouse workers and other lower-paid physical workers may face earlier task redesign, so training and income support should follow observed deployment rather than distant capability headlines.
- Care, repair and service workers may see assistive tools before replacement because dexterity, trust, regulation and interpersonal contact remain binding constraints.
Global context
The index uses US occupations, pay and regulation. Lower wages can make robots less cost-competitive in many countries even when the same machines are technically capable; labour shortages, higher wages or industrial policy can accelerate adoption elsewhere. Global comparisons therefore require local cost, safety, infrastructure and worker-protection data rather than applying the US 0.3% estimate unchanged.
What the evidence does not yet show
- The analysis is company research and is not presented as a peer-reviewed paper; Claude helps classify tasks, estimate time shares, find evidence and estimate costs.
- O*NET, BLS and American Community Survey data make the conclusions US-specific; different wage levels, regulations and workplace designs can change cost competitiveness elsewhere.
- Exposure describes task capability under stated conditions, not observed adoption, job loss or a causal forecast of future employment.
What to watch next
- Independent replication of the 7,594 physical-task ratings and annualised robot cost estimates.
- Observed deployment, utilisation, maintenance and supervision costs in driving, warehousing and production.
- Whether new robotics capabilities improve manipulation in unstructured environments rather than only controlled demonstrations.
Evidence trail
Sources used for this report
Links checked 1 October 2026
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Continue the story
Related reporting
Work & Skills
Who is workplace AI leaving behind? PwC finds an access gap—but not its cause
PwC surveyed 49,364 workers across 48 countries and regions. AI use rose, while reported access to learning fell; the cross-sectional responses show an association, not proof that AI caused the divide.
5 min · 2 sources
Work & Skills
Will AI take my job? What the evidence says about exposure, hiring and skills
The ILO estimates task exposure, the OECD examines skills and IMF staff study hiring. Together they show uneven risks and practical questions, not a prediction for one worker.
6 min · 4 sources
Work & Skills
IDB projects an AI growth dividend for Latin America, with wages dependent on mobility
Preliminary scenarios put regional output between 0.3% and 5.1% higher after a decade. Workers' ability to move into expanding roles determines whether wages rise or fall in the model.
6 min · 2 sources
Reader discussion
Add evidence, experience or a question
No account is required. Reader notes are published after a brief civility, relevance and safety check; disagreement is welcome.
Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.