Robots can do 74% of physical tasks. They are cost-competitive for… 0.3% of work
The clickiest number is that robots can already perform 74% of physical tasks in the U.S. economy. The more important number is very different: Anthropic’s new research estimates that robots are currently cost-competitive with human labor for only about 0.3% of job tasks. That gap is a useful reality check on claims about the imminent “end of work.”
Anthropic researchers built a robot exposure index designed to ask not “what will robots do in ten years?” but “which tasks can current robot technology perform, and in what environment?”. For RoboMorrow, the value is precisely the separation between technical feasibility and economic automation.
Seventy-four percent of physical tasks does not mean 74% of jobs
The paper estimates that physical work accounts for roughly 46% of U.S. working time. Robots can perform about 74% of those physical tasks in at least some setting. Across all work, that corresponds to roughly 34% of working time.
This is not a forecast that 34% of jobs disappear. Occupations consist of many tasks, and some automated steps require purpose-built environments. The authors also model costs, human preferences and regulation. Exposure measures capability, not layoffs.
Four exposure levels show where a robot can actually operate
Anthropic defines four levels. E0 means a current robot cannot perform the task. E1 means it can in a purpose-built robotic environment, such as a production line. E2 means it can in a structured human workplace, such as a warehouse. E3 covers unstructured environments such as a city road.
That classification cools down the headline. About half of physical tasks fall into E1, meaning automation depends on a specially prepared environment. Only a small share reaches E3. This is far from a general humanoid walking into any workplace and taking over without integration.
The largest barrier is economics
The authors estimate that robots are currently cheaper than human labor for only about 0.3% of job tasks. Their historical cost trend scenario suggests it could take roughly 40 years for the cost-competitive share to reach 10% if past price declines continue.
That is a scenario, not a countdown clock to 2066. Hardware costs could fall faster or slower, AI could improve utilization, and wages and regulations will change. But 0.3% is a powerful reminder that a capability demonstration is not a business case.
Method: roughly 900 occupations and 19,000 task descriptions
The analysis uses O*NET, covering around 900 occupations and about 19,000 task descriptions. Claude helps classify physical, cognitive and interpersonal requirements and assess the environment in which present-day robots can perform each task. The authors publish methodological detail, prompts and supporting data.
That is both a strength and a limitation. Model-assisted classification enables scale but can introduce judgment errors and depends on the quality of evidence available for individual robots. The index does not replace a process study at a specific company.
Warehousing and driving are more exposed than nursing and repair
The results show high exposure in transportation and warehouse work. Nine of the ten most exposed occupations in one ranking are vehicle operators. By contrast, nursing and general repair contain more interpersonal work, unpredictable manipulation and operation in variable settings.
The paper also finds a demographic skew in the U.S. data: workers in more robot-exposed occupations are more likely to be male, less educated and lower paid. That describes the U.S. dataset; it should not be treated as a universal profile for every labor market.
A historical backtest is not a forecast
The researchers test whether occupations more exposed to robots available in earlier decades later experienced weaker wage and employment outcomes. Across roughly 50 years of data, they find that higher prior exposure is associated with later declines, even after accounting for some industry trends.
That does not establish a simple causal mechanism for every occupation and cannot supply a precise timetable for the next decade. Trade, technology, demand and regulation move together. The backtest supports the index as a signal, not as an election-style forecast of winners and losers.
What this means for the robotics market
The best near-term automation opportunity may not be where the robot looks most human. It is where tasks are frequent, environments are sufficiently structured, integration can be repeated and labor or downtime costs justify the machine. Warehousing, material movement, cleaning and selected production processes therefore remain natural deployment zones.
For humanoids, the paper creates a tougher test. If their advantage is using human environments without major redesign, they should move tasks from E1 toward E2 or E3 at an acceptable total cost. That would be much stronger evidence than a long list of laboratory skills.
The number to remember is 0.3%, not 74%
The 74% figure explains the technical opportunity. The 0.3% figure explains the economic gap. Both are necessary. The first helps explain why capital is flooding into robotics; the second helps explain why most workplaces are not yet staffed by robot fleets.
RoboMorrow will use that distinction as a filter for launches: not only “can the robot do it?”, but “in what environment, with how many interventions and at what total cost?”.
For buyers, capability is not ROI
The most useful conclusion from the study is straightforward: a robot being able to perform a task does not mean the automation project is economical. Anthropic measures task exposure and required work environments, then evaluates cost separately. That is why 74% and 0.3% can both be true. The first number asks whether present-day robots can perform physical tasks under some conditions; the second reflects how little total working time is currently spent on tasks for which a robot is cheaper than human labor.
For companies, this means the shortlist should begin with the process rather than the robot form factor. Stable pallet transport may fit an AMR, a repetitive station may fit a cobot or fixed automation, and a humanoid becomes interesting only when flexibility and use of human-built infrastructure compensate for higher cost and operational risk. A pilot should measure cost per completed mission, intervention time, robot utilization and the amount of human work genuinely removed from the process.
Featured image: official Anthropic hero artwork for the “What work can robots do?” study. It is not product photography or a RoboMorrow metric graphic. Image source.
Sources and methodology
This article uses primary sources plus editorial cross-checking and was verified on 3 October 2026. Manufacturer claims are kept separate from independent evidence. This is not a RoboMorrow hands-on test.