Most measures of AI's reach into the labor market score what language models can do, which leaves out physical work. This report asks which physical tasks robots can already perform, on the premise that jobs are more likely to be affected soon when robots already do their work somewhere. Anthropic's Economics team published it on 30 September 2026 as a web page and PDF titled What work can robots do? (indexed on the research page as "Can we predict the jobs robots will do?"), by Russell Legate-Yang and Maxim Massenkoff, with a separate 56-page appendix and a data release on Hugging Face.
This note follows the 44-page PDF and the appendix. The PDF in the library's vault folder (named as a proof) is byte-identical to the one the page links. The data release is the robot_exposure folder of the Anthropic EconomicIndex dataset, added on 30 September 2026; its README is a source for counts below.
What the report does
Which tasks count as physical
The report starts from O*NET 29.3, the US Department of Labor database of about 900 occupations and 19,000 task statements (Measuring exposure). Claude rates every task on how much of its time is physical, cognitive and interpersonal work, the skill level each component needs, and whether the task could be done without it (Appendix A.1). A task counts as physical if at least half its time is physical, if it needs physical skill that takes months or more to build, or if it could not be done without physical work. "Teach dance students" qualifies under the second rule despite being 35% physical by time. About 7,600 tasks qualify; the README gives 7,594.
As validation of these ratings, the appendix correlates them at the occupation level with indices built from the Bureau of Labor Statistics Occupational Requirements Survey, a survey of employers: 0.76 for the physical time share against physical demands, and 0.89 and 0.78 for cognitive and interpersonal skill levels against problem-solving and social demands (Appendix A.2, Appendix Figure 2).
The exposure rubric
A robot is defined as an autonomous physical machine that senses and acts; a car wash that scans cars counts, a surgical system fully controlled by a surgeon does not (Measuring exposure, footnote 10). Each physical task is placed in one of four tiers by the least controlled environment in which a robot can do it today:
- E0: no robot can do it.
- E1: a robot can do it in a purpose-built robotic environment, such as an assembly line.
- E2: in a structured human workplace, such as a warehouse or hospital.
- E3: in an unstructured environment, such as a city road or construction site.
Because O*NET statements are terse, Claude first describes representative instances of each task and how often each occurs, then rates each instance with web search; the task's tier is the least structured environment in which robots can do at least half of its instances by weight (Measuring exposure; Appendix A.3). In the report's example, "Dig trenches" is split between cutting linear trenches in open ground (rated E3, citing an autonomous excavator retrofit) and digging by hand around buried pipes (E0); more of the instances resemble the latter, so the task is E0. Claude must name a specific robot and quote sources showing a deployment, commercial sale or demonstration, and the robot must do the task about as well as a person, counting reliability, errors and speed. Across the project, including the historical ratings, Claude made about 650,000 web searches and cited over 275,000 quotes from over 90,000 web pages (Appendix A.3). The release records 56,933 cited sources for the present-day ratings: 20,877 deployed, 25,743 commercial and 10,313 demonstrations (README).
An occupation's robot exposure index averages its tasks' tiers on a 0 to 3 scale, weighted by the share of working time each task takes. Those time shares are themselves Claude's estimates (Figure 2 notes). Employment weights come from BLS data for 2025.
Which model rated, and how the ratings were checked
Appendix F gives the prompts and says they were used "with Claude Opus 5", with responses constrained to an output schema; the README adds that Opus 5 rated with web search. Claude therefore does most of the measurement in the pipeline: classifying tasks, estimating time shares, rating exposure, estimating robot costs, and, for the supporting analyses, describing groups of unexposed tasks and rating barriers to automation (Appendix F.1 to F.7). The prompts tell Claude its ratings "will be used in economic research and audited by human reviewers", but the report gives no results from such an audit and no comparison of exposure tiers with human experts' ratings. The validation it offers for the exposure measure is indirect: the occupational-requirements correlations above (for the task taxonomy, not the tiers) and the historical backtest below.
What it finds
How much work robots can do
Weighting tasks by time and employment, 54% of US working time is cognitive or interpersonal and 46% physical (Figure 3). As shares of all working time, about 12% is physical work no robot can do (E0), 23% is E1, 10% E2 and 1% E3. So 74% of physical work, or 34% of all work, can be done by robots in some circumstances. Within physical work alone, E1 is about half, E2 22% and E3 2% (Measuring exposure, Figure 2 notes). In the task file, 2,419 physical tasks are E0, 3,967 E1, 1,103 E2 and 105 E3 (README).
That headline depends on two judgment calls the appendix tests (Appendix A.4):
- Dropping tasks whose ratings rest only on demonstrations lowers the exposed share of physical work by 1 percentage point, with a job ranking almost unchanged (rank correlation 0.99).
- Dropping ratings in which Claude counted a robot that does a related task (wheel-aligning robots in car factories for mechanics' alignment work, for example) lowers the exposed share of physical work "from about three-quarters to a half." The authors defend these ratings as showing what robots could likely do, or nearly do, with today's technology.
Which jobs are most exposed
Nine of the ten most exposed occupations with at least 20,000 jobs are vehicle operators; taxi drivers lead with an index of 2.2 (Figure 4; 2.2235 in the released occupation file). The report's taxi-driver employment comes from BLS occupational data and so excludes independent ride-share contractors (footnote 24). Other examples: stockers and order fillers score 1.0, tapers who finish drywall 1.6 and recycling and reclamation workers 1.7, the last two reaching similar scores by different routes (some tasks highly exposed against most tasks moderately exposed).
Workers in the top fifth by exposure differ from unexposed workers, using American Community Survey data for 2020 to 2024 (Figure 5): they are 20 percentage points less likely to be female, 16 points more likely to be Hispanic, 55 points less likely to hold a bachelor's degree, earn around $30 less per hour, and face an unemployment rate more than twice as high. The report notes that this is in many ways the opposite of the pattern found for LLM exposure.
Robots and LLMs together
Combining robot exposure (E1 or higher) with the LLM exposure scores of Eloundou et al. (2024) raises the share of exposed work from around half to 81% (Robot and LLM exposure, Figure 6). In transportation and moving occupations, LLMs alone expose under 15% of tasks and LLMs plus robots about 90%. No broad occupation group is under 40% exposed. Personal care and service, installation and repair, healthcare support, and community and social service are the least exposed groups; Claude's grouping of the unexposed task statements describes them as hands-on, face-to-face and sometimes regulated (Appendix C.1).
Barriers to automation
Claude was asked which barriers would each, on their own, stop robots from doing a significant share of each physical task (Robot and LLM exposure; Appendix C.2). On these ratings, missing capabilities block about 70% of physical tasks, with manipulation the binding skill for about half of all physical tasks; planning and reasoning, the skill AI seems most likely to improve, is the main limit for 8%. Regulation blocks 14% and human preferences about a quarter, together about 30%. Cost is a barrier for nearly every task. These are Claude's judgments of what would happen if a barrier were removed, not observations.
Costs
For each exposed task, Claude estimated, with web search, the annual cost for the cited robots to produce what a human worker produces in a year on that task, including hardware, integration, maintenance, software, energy, supervision and other line items, with fixed costs annualized (Robot costs and adoption; Appendix D.1). A task is exposed to cost-competitive robots if that cost is below the worker's total compensation times the task's share of time; occupation totals were again estimated by Claude, net of shared robots.
- Robots are cost-competitive for 0.3% of all working time, which includes roughly 300,000 workers for whom robots can do 95% of their tasks (Robot costs and adoption, footnote 44).
- Packers and packagers are the largest cost-competitive occupation: a set of robots estimated at over $2 million replaces about 14 workers, about $45,000 a year per worker replaced against about $49,000 in compensation for work on which they spend 97% of their time (Figure 7). The report notes that the occupation's employment fell 22% since 2015 and that the BLS projects it to lose the 11th-most jobs of any occupation by 2035 (footnote 41).
- Robotaxis are estimated at only about $7,000 a year more than taxi drivers, while welding robots cost about five times as much as welders, and cleaning robots several times as much as dishwashers and janitors (Figure 7 and text).
- Industry and government data suggest robot prices have fallen about 3% a year since the 1990s (Robot costs and adoption; Appendix E). At that rate a 20% decline takes about seven years and would make robots cost-competitive for the physical work of 2.8 million workers, 0.8% of working time; reaching 10% of working time needs about a 70% decline, around 40 years (Figure 8 and text).
- Ranking occupations by an exposure measure that accounts for cost gives nearly the same order as the main measure (rank correlation 0.95, Appendix D.2).
The authors call the cost comparison "more suggestive" than predictive because of adoption frictions and cost uncertainty, and they note that cost parity need not mean rapid job loss, since automation raises productivity mainly once machines are much cheaper than people (Robot costs and adoption).
The historical backtest
To test whether exposure predicts disruption, the authors rate task descriptions from the Dictionary of Occupational Titles (1977, 1991) and O*NET (2002, 2008) against the robots of each time, and relate exposure to later changes in occupational wages and employment from the Census and American Community Survey, comparing occupations within the same industry and period (Appendix B). Because models trained on modern data "know" which tasks were later automated, Claude is told to consider work as it existed in the target year and to cite sources that date each robot to that time; the authors call this a mitigation of look-ahead bias, not a removal (Appendix B.2).
- Pooling three overlapping periods of about 20 years from 1977 to 2024, an occupation whose tasks could all be done by robots had wages 7.1% lower 20 years later than an unexposed occupation in the same industry (95% interval 5.4% to 8.9%) and 34.2% fewer jobs (16.4% to 52.0%) (Appendix B.3, Appendix Figure 5).
- Wage effects are about 7% in each period; employment effects are about 50% in 1980 to 2000 and about half that in the two later periods (Appendix Figure 6).
- A placebo test finds that exposure as of 1977 does not predict wage changes from 1970 to 1980; the employment placebo is also null but imprecise.
- Adding controls for baseline wages and demographics attenuates the effects; wage effects stay significant, while the employment interval "narrowly includes zero," regaining significance when unionization and offshorability controls are added (Appendix Figure 7).
- Historical robots worked mostly in E1 settings, so, the authors note, the historical data "say little" about the E2 and E3 exposure that adaptable AI-controlled robots may create.
The main text summarizes these results as giving "some confidence" that the measure can predict job disruption.
How fast capabilities have grown
Rating 1977 tasks against today's robots, 62% of physical tasks that no robot could do in 1977 has fallen to 24% of those same tasks, which the authors convert to robots becoming able to do about 2% of the remaining physical work each year (Robot costs and adoption; Appendix B.4). Combining that rate with 3% annual cost declines, robots become cost-competitive for half of today's physical work only in 2085; in a "fast adoption" scenario with cost declines up to four times faster and capability growth twice as fast, half is reached by 2050, and automating 90% still takes 53 years. The appendix gives the business-as-usual date for 10% of work as the 2060s (Appendix E). These scenarios hold tasks and wages fixed, apply uniform improvements to every task, and are described as "partial equilibrium"; the authors say they "are not meant to predict job loss" (Appendix E).
The prompts used
The measure is produced by written instructions to Claude, printed in full in Appendix F: task-composition, time-share, exposure, cost, occupation-cost and barrier prompts. The exposure prompt defines each tier in words, with the "mark" of each (for E1, "the setting is built around the machine, not around people"), and the barrier prompt adopts the International Organization for Standardization definition of a robot. The report does not test how ratings change with different wordings of these prompts. Its sensitivity checks vary which ratings are counted (demonstrations, related robots, instance shares), not the instructions.
Limitations and open questions
- No human check of the tiers. Every exposure rating, time share and cost estimate is Claude's, from a single model; the report gives no inter-rater comparison with humans, though its prompts say ratings are audited (Appendix F).
- Related-robot ratings. Without them, exposed physical work falls from about three-quarters to a half (Appendix A.4); the headline "three-quarters" includes them.
- Look-ahead bias in the backtest. The historical ratings were made by a model that has seen what happened, mitigated only by instructions and dated sources (Appendix B.2).
- Backtest scope. It covers E1 factory robots and relative changes within industries; effects of more adaptable robots are outside its range (Appendix B.3).
- Costs are Claude's estimates from web sources, compared with average compensation; the authors call the comparison stylized and list omissions on both sides, judging them "roughly balanced" (Appendix D.1).
- Weighting. With Census rather than BLS employment weights, physical work is 40% of tasks rather than 46%, which the appendix attributes partly to how workers and employers describe jobs (Appendix B.2).
- Leapfrogging. AI-powered robots could "leapfrog" the scale and do work no robot does today, which a measure built on current capabilities would miss (Discussion).
- Self-report and interest. This is Anthropic's measure built with Anthropic's model; the release of Claude's reasoning and cited sources for every task (README) allows outside checking of individual ratings, though the released quotations are cut to their first five words, so checking a rating means reopening its source.
Research questions
- How well do Claude's exposure tiers agree with human robotics experts? A random sample of tasks rated independently by experts, with agreement reported by tier, would give the validation the report does not include and show whether disagreement concentrates in related-robot ratings.
- Does the backtest survive a model without hindsight? Rating 1977 to 2002 tasks with a model trained only on earlier text, or with human raters restricted to period sources, and re-running the regressions would show how much of the predictive power could come from look-ahead knowledge.
- Which headline should be used: three-quarters or a half? Reporting the occupational ranking, the LLM-plus-robot share and the backtest with related-robot ratings excluded would show how much each result depends on that judgment.
- Do E2 and E3 exposures predict outcomes as E1 did? As warehouse robots and autonomous vehicles spread, tracking employment and wages in occupations newly rated E2 or E3 against those rated E1 would test whether the tier ordering carries information beyond any exposure at all.
- How accurate are the cost estimates? Comparing Claude's per-task costs for a sample of deployed systems with published deployment costs or vendor quotes would show whether the 0.3% cost-competitive share is biased in either direction.
- Are packers and taxi drivers changing first? The report names them as exposed and near cost parity. Annual tracking of their employment, hours and pay, against similar occupations with low exposure, would test its near-term prediction.
How it connects
- The LLM side of the combined measure is the theoretical-capability measure that Labor market impacts of AI: A new measure and early evidence (Massenkoff and McCrory, 2026) built on Eloundou et al. (2024); the report cites that note's authors for the contrast between robot and LLM exposure (footnote 32).
- Scenarios for our Economic Future (September 2026), by colleagues in the same economics group, models only cognitive work and says it does not allow for rapid advances in robotics; this report measures the physical channel that model leaves out. The connection is the note's own; neither item cites the other.
- How well do job retraining programs work? (August 2026), with Massenkoff as second author, discusses drivers as workers whom short training programs might help if automation reaches them (its §8.1); this note makes the connection.
- Anthropic's own experiments with Claude controlling robots, such as Project Fetch: Can Claude train a robot dog?, test a different question (what a model can do with a robot) and are not cited by the report.
References
- Legate-Yang, R., and Massenkoff, M. (2026, September 30). What work can robots do? Anthropic. https://www.anthropic.com/research/what-work-can-robots-do. PDF (44 pages): https://assets.claude.com/401a473469db99fd39bba1ca6d9a5653a70e2f12.pdf. Appendix (56 pages): https://assets.claude.com/d27288375b0ac486cb9da0a30a94423b36ff0443.pdf. Section headings and figure numbers in this note follow the PDF; appendix sections follow the appendix. record
- Data: https://huggingface.co/datasets/Anthropic/EconomicIndex, folder
robot_exposure(commits 430f214 and 4e5f69f, 30 September 2026), CC BY 4.0. - Works cited above only as the report describes them (Eloundou et al., 2024; Webb, 2020) are listed in the report's references and were not read for this note.