Capable, but rarely cheaper.

The report's two numbers

On Sept. 30, two researchers at Anthropic published an estimate of how much physical work today's robots can already do. They define a robot as an autonomous physical machine that senses and acts. They estimate that robots can do 74% of the physical tasks in American jobs, which together make up 34% of all working hours.

The same report has a second number. Robots are cheaper than paying people for just 0.3% of job tasks. If robot prices keep falling at their past pace, the authors estimate it would take 40 years for that share to reach 10%.

One headline led with the first number: robots "can replace 74 percent of manual labor," with the cost caveat after the comma. The report says only that robots can do those tasks, "mostly in limited settings." Whether a robot can do a task and whether it costs less than a person are separate questions. The report measures both and finds a wide gap between them.

On Sept. 24, the International Federation of Robotics said a record 5 million industrial robots were operating in factories worldwide in 2025. Factories are built around the machine, with the surprises engineered out, and that is where robots do best. The estimate depends on how controlled a workplace is.

How you grade a robot

The study starts from O*NET, a database sponsored by the U.S. Department of Labor that describes around 900 occupations through around 19,000 job tasks. Each task is one plain line. For hand packers and packagers, one reads: "Seal containers or materials, using glues, fasteners, nails, and hand tools."

Claude did the first pass. It scored each task description for physical, cognitive and interpersonal demands. "Dig trenches" counts as physical; so does "Teach dance students," which also takes thought and people skills. Then, for each physical task, Claude was told to search for specific robots that could do it, check where those robots operate, and quote its sources.

The rating that matters is about place. Each task gets one of four. E0: no robot can do it. E1: a robot can do it in a purpose-built robotic workplace, like a factory assembly line. E2: in a structured human workplace, like a logistics warehouse. E3: in an unstructured setting, like a city road. Assembling hospital meal trays to match patients' diets is E1, because food-prep robots already pack meals on conveyor belts. Delivering medicines to nursing stations is E2. Driving a taxi is E3. Dyeing a client's hair is E0.

The authors rate the place because, they argue, robots find unpredictable surroundings hard, so most robots at work today operate in environments engineered for them. In the report's terms, an E1 task is possible in purpose-built environments but not in wider settings.

That is how the 74% breaks down. Robots can do half of all physical tasks, but only in purpose-built settings. They can do another 22% in structured human workplaces, and only 2% in the open world. The remaining 26% (100% minus 74%) no robot can do yet.

A task robots can already do is what the report calls exposed. Exposure measures what is technically possible today. The authors note it's "difficult to predict exactly" how this will affect jobs. By the same yardstick, around half of all work is exposed to large language models, the technology behind chatbots, and that rises to 81% once robots are added.

Robots are cheaper than paying people for just 0.3% of job tasks.

The packer's arithmetic

To see why ability and cost come apart, take the report's own example. Packers and packagers are the largest occupation where robots already beat people on cost. O*NET's task list for hand packers runs from inspecting products and weighing them to lining cartons, sealing and labelling containers, and moving finished goods onto conveyors or loading docks.

No single robot does all of that. The authors count several: machines that pack containers, move materials around warehouses, inspect goods visually, put together cardboard boxes, print and apply labels, and seal packages. Together, they estimate, those robots cost over $2 million to buy and install, and replace the yearly work of around 14 workers.

To compare the two, the authors spread the machines' up-front cost over their working life, allowing for the cost of money over time. A task counts as cost-competitive when the robot is cheaper than the slice of a worker's total pay and benefits spent on that task. Packers spend 97% of their time on tasks robots can do, and each costs around $49,000. On those numbers, the robots come out about $2,500 a year cheaper.

That is the report's showcase for robots on cost, and the margin is thin next to a purchase of more than $2 million. It rests on estimates of what the robots cost, how long they last and how much of the work they cover. Across almost all other job tasks, the person is still the cheaper option or the robot can't do the work.

Claude and the authors split the work. Claude read the task descriptions and searched out the robots behind each rating, quoting its sources. The authors note that their rubric "requires many judgment calls." The frame was theirs: deciding that place matters more than the robot, choosing how to price a machine, and testing the method against history.

What the numbers don't say

"Robots can do three-quarters of manual work." They can do three-quarters of physical tasks somewhere. Half of all physical tasks are possible only in purpose-built settings, and just 2% in the open world. Tasks are also only parts of jobs.

"Exposure is just theory." The authors checked their method against the past. From 1977 to today, jobs more exposed to the robots of their time saw wage and employment declines in later decades. Capability moves, too. In 1977, robots couldn't do 62% of physical tasks; today they can do all but 24% of those same tasks. That works out to robots picking up about 2% of the tasks they couldn't do, each year.

"Cost is the only thing in the way." It's one of four barriers, and they overlap. Capabilities and cost are the biggest, the authors say. By their estimates, today's regulations wouldn't allow robots to do 14% of physical tasks, and people's preferences hold robots back from a quarter.

"A model did the rating, so it's objective." Claude applied a rubric the authors wrote, and they say it needed many judgment calls. Robot costs were estimated with web searches, starting from the robots cited for each task. Treat the result as a careful first map.

"This is about machines." The report also describes the people whose tasks are exposed. In its comparison, highly exposed workers are 55 percentage points less likely to hold a bachelor's degree or higher, earn around $30 less per hour, and face an unemployment rate more than twice as high.

What to watch, what to do

Watch the price of robots. Industry and government data suggest robot prices have fallen roughly 3% a year since the 1990s, the authors write. For robots to be cheaper than people for 10% of today's work, costs would need to fall about 70%. At 3% a year, that takes around 40 years; a 20% fall takes around seven. A break in that trend would move the picture faster than any new capability.

Watch the middle tier. Warehouses are the report's example of a structured workplace, and they are already buying robots. The IFR said on Sept. 30 that transportation and logistics stayed the top use for professional service robots, with 117,500 sold in 2025. Full-size humanoid robots sold about 7,000.

Watch the open world. Of the 10 most exposed occupations in the report, nine are vehicle operators, and taxi drivers top the list with an index of 2.2. Self-driving cars haven't replaced drivers at scale, the authors note, but the capability is there. If robots start working reliably on roads and in homes, the 2% is the number that moves.

Then read your own job the way the study did. On O*NET OnLine, look up your occupation and read its task list. For each task, ask whether a machine does it anywhere today, in what kind of place, and whether it would be cheaper than a person there. Desk work gets the same treatment, with a chatbot in place of the robot.

When a report, demo or headline says AI or robots "can do" a share of work, ask which tasks, in what setting and at what cost.