Robots Can Do 74% of Physical Work. They Win on Cost for 0.3%

Robots can already do 74% of the physical tasks in American jobs. They are cheaper than the people doing those tasks for 0.3% of all work. Both numbers come from the same Anthropic paper published on September 30, 2026, and the distance between them is the story.

The paper, "What work can robots do?" by Russell Legate-Yang and Maxim Massenkoff, builds what the authors call a robot exposure index. It is the physical-world counterpart to the LLM exposure measures we cover most weeks, and it points at a different set of workers.

How the index is built

[Fact] The authors start from O*NET, about 900 occupations and 19,000 task statements. Claude first sorts out the tasks that could not be automated without a physical machine: 7,594 of them. Then each physical task gets a rating on a four-step scale based on how controlled an environment a robot needs:

• E0: no robot can do it.
• E1: a robot can do it in a purpose-built setting, like an assembly line.
• E2: a robot can do it in a structured human workplace, like a warehouse or hospital.
• E3: a robot can do it in the open, like a city street.

[Fact] Only demonstrated capability counts. Claude has to search for real robots and quote sources showing deployments, sales or demonstrations, and the robot has to match human reliability and speed. The authors say results barely change when demonstrations are dropped.

[Fact] "Dig trenches" shows how strict this is. Cutting a straight trench in open ground is rated E3, because an autonomous excavator retrofit exists. Digging by hand around buried pipes is E0. The second kind of work takes more of the task's time, so the whole task is rated unexposed.

What robots can do today

[Fact] Physical work is 46% of all US working time. Of that, robots can do something in some setting for 74%, which is 34% of all hours. But the split is lopsided: 23% of all work is doable only in a purpose-built environment, 10% in structured workplaces, and just 1% in unstructured places. That last 1% is mostly driving.

[Fact] Nine of the ten most exposed occupations are vehicle operators. Taxi drivers top the list at 2.2 on the 0 to 3 index. Recycling and reclamation workers score 1.7, drywall tapers 1.6, stockers and order fillers 1.0.

[Fact] Tapers show that a middling score can hide two different jobs. Robots cannot press paper tape into wet joint compound; that part is messy and unexposed (26% of tapers' time). But a drywall robot can scan walls, spray later coats and sand them, so 41% of their time is rated E3.

The cost gap is the real finding

[Fact] For every exposed task, Claude estimated what it would cost to deploy the cited robots for a year and compared it with the worker's total compensation for that share of time. Robots win on cost for 0.3% of all work.

[Estimate] Put the two headline numbers side by side: robots can technically do about 113 times more work (34% versus 0.3%) than they can do profitably today.

[Fact] Packers and packagers are the largest occupation where robots already win. A full line costs over $2 million to buy and install and replaces about 14 workers. Annualised, that is around $45,000 per replaced worker, against a worker who costs about $49,000 and spends 97% of their time on robot-doable tasks. The robot saves roughly $2,500 a year.

[Estimate] That margin is thin. $49,000 times 97% is about $47,500, so the robot is only about 5% cheaper. Every robot cost in this paper is a Claude estimate built from web searches. An estimating error of 6% or more in the wrong direction would flip packers back to "not cost-competitive". The paper says as much: it calls the cost measure "more suggestive".

[Fact] Elsewhere the gap is wide. Robots that could cover a welder's tasks would cost about five times what the welder costs. Cleaning robots for dishwashers and janitors are "several times more expensive" even though those workers earn $25,000 to $30,000 less than welders. Robotaxis are estimated at only about $7,000 more than a taxi driver, but regulation holds them back.

[Fact] Robot prices have fallen about 3% a year since the 1990s. At that pace, a 20% decline takes about seven years and would make robots cheaper for the work of 2.8 million people (0.8% of all hours). Reaching 10% of all work needs a 70% decline, about 40 years.

[Estimate] We checked one example from the paper's figure notes. Postal Service mail carriers: robots can do 94% of their tasks, at an estimated $166,000 a year against median compensation of $82,000. Parity needs robot costs to fall about 54%. At 3% a year that is roughly 25 years.

Who is exposed is the opposite of the AI story

Most coverage of AI and jobs, ours included, is about office work. This paper turns that around.

[Fact] Using 2020-2024 American Community Survey data, workers in the top fifth of robot exposure are 20 points less likely to be women, 16 points more likely to be Hispanic and 55 points less likely to hold a bachelor's degree than unexposed workers. They earn about $30 an hour less and have more than twice the unemployment rate. The authors note this is "in many ways the opposite" of LLM exposure.

[Fact] Combining the two measures changes the picture. LLMs alone expose about half of all work. Adding robots raises that to 81%. Transportation and moving tasks go from under 15% to about 90%. No broad occupation group is below 40%.

[Fact] What is left is hands-on and face-to-face work: personal care, healthcare support, installation and repair. Capability is the main barrier for about 70% of physical tasks, and half of physical tasks would need better manipulation, the ability to touch and handle objects. Regulation blocks 14%, human preferences about a quarter.

The backtest, and what it cannot tell you

[Fact] The authors re-rated job tasks as of 1977 and found that occupations more exposed to robots back then saw larger declines in wages and employment in later decades, after controlling for industry trends. They also estimate that robots learn about 2% a year of the physical work they previously could not do: in 1977 they could not do 62% of physical tasks; today's robots can do all but 24% of those same tasks.

[Claim] That backtest supports using today's capabilities as a guide, but it measures a slow technology. Industrial robots spread over decades. If AI-powered manipulation improves faster, the scale can be "leapfrogged", which the authors say directly.

Three limits are worth stating plainly.

• [Fact] Claude does almost all of the measurement: which tasks are physical, how much time each takes, which robots exist and what they cost. A human check of a sample is not reported in the main text.
• [Fact] The paper is written by Anthropic economists and rates tasks with Anthropic's own model. It is a company research post, not a peer-reviewed article. Our site also draws heavily on Anthropic's labour data, so read our coverage with that in mind.
• [Claim] Cost parity is not the same as job loss. The robot estimates often include human supervision and repair. The authors cite work showing productivity gains arrive when machines are much cheaper than people, not merely equal.

[Fact] One data point argues against reading this as "nothing will happen for 40 years": packer employment has already fallen 22% since 2015, and BLS projects the occupation to lose the 11th-most jobs of any occupation by 2035.

What this means for work

• Drivers and warehouse workers are where the paper expects change first. For taxi and truck drivers the barrier is regulation and a modest cost gap, not capability. Watch local robotaxi rules more than technology news.
• Packers and packagers are the one large group where the numbers already favour robots, though only by about 5%. Skills in running, clearing and maintaining packing lines are the part of the job that stays.
• Welders, tapers, cleaners and dishwashers face robots that can do much of the work but cost several times more. The cost curve gives years, not months.
• Nurses and repair workers sit at the bottom of exposure. The work is hands-on, social and often regulated, which is exactly what both robots and LLMs handle worst today.

For occupation-level data, see our pages on taxi drivers, truck drivers, hand packers and packagers, stockers and order fillers, welders, tapers, janitors and cleaners, dishwashers and registered nurses.

Sources

• Russell Legate-Yang and Maxim Massenkoff, "What work can robots do?", Anthropic, September 30, 2026: https://www.anthropic.com/research/what-work-can-robots-do

About this analysis

This article was written with AI assistance. Figures marked [Fact] were read from the full main text of the Anthropic post, not from its summary. Figures marked [Estimate] (the 113-times gap, the 5% packer margin, the 25-year mail carrier timeline) are our own arithmetic from the paper's reported numbers. We did not read the separate appendix. No figures or tables from the paper are reproduced.