Can robots do 74% of physical work? Watch out for that headline.

I've frequently seen AI and robotics studies turn into headlines that are far more bombastic than the results themselves allow. Something like "robots can already perform 74% of physical work" can quickly turn into "robots can eliminate 74% of physical jobs."

That's not what Anthropic's new study, "What work can robots do?", shows. I read the paper carefully because it makes a distinction I consider essential to any serious discussion about the future of work. A profession's exposure to robotics does not mean the professional will be replaced. The study analyzes tasks within occupations and estimates under what conditions robots available today can perform them.

The methodology starts from roughly 900 occupations and 19,000 tasks in the American O*NET database. About 7,600 tasks classified as physical are evaluated at four levels. At E0, robots cannot perform them. At E1, they can only do so in environments specifically prepared for robots. At E2, they can operate in structured human environments, such as warehouses. Only at E3 can they perform the task in unstructured environments, such as a road.

That's where the number likely to generate headlines comes from. About 74% of physical tasks, weighted by work time and employment, show some degree of exposure to robotics. They represent approximately 34% of the total work time analyzed in the US economy.

But the detail completely changes the interpretation. Most of that exposure depends on controlled environments. Roughly half of physical work falls under E1, 22% under E2, and only 2% under E3, where the robot can operate in an unstructured environment.

There is, therefore, a huge gap between "a robot can perform this task under some condition" and "a robot can replace today the person who holds this job."

And a second, even more important gap appears: the economic one. Although tasks corresponding to 34% of work time are technically exposed to robotics under some condition, the study estimates that robots are currently cost-competitive for only 0.3% of work. That doesn't mean 0.3% of jobs, but an estimated share of work weighted by tasks and the time devoted to them.

Technical capability and economic viability are different things. Robotics requires hardware, integration, maintenance, energy, supervision and, in many cases, changes to the environment itself. The study estimates that costs would need to fall by roughly 70% to make robots competitive for 10% of work. Keeping the historical trend considered by the authors, close to 3% in annual price reductions, that would take about 40 years. They themselves warn that technological advances could substantially accelerate that trajectory.

Another number requires the same caution. By combining exposure measures for LLMs and robots, the study arrives at roughly 81% of work exposed to at least one of these technologies. But that doesn't mean 81% is automatable or that 81% of jobs are at risk. The exposure metrics used for LLMs and robots themselves measure slightly different concepts.

There are still bottlenecks that don't disappear simply by adding AI. Capability limitations affect about 70% of physical tasks, according to the methodology. Object manipulation appears as the main obstacle. Regulation is considered a constraint for about 14%, and human preferences for roughly a quarter of tasks.

And there's a fundamental distinction that often gets lost in these discussions. A task is not a profession. A welder doesn't just weld. They position parts, inspect quality, prepare materials, work in difficult locations and handle unforeseen situations. The study shows robots can already weld autonomously under certain conditions, but it estimates that automating the full set of exposed tasks for a welder would cost roughly five times as much as human labor.

I would also add a few methodological caveats to these numbers. Claude takes part in important stages of the study. It is used to classify tasks, estimate how much time workers spend on them, research evidence of robotic capabilities, and assist with cost estimates. The authors also acknowledge that O*NET descriptions can be short and omit aspects of the activity that are trivial for humans and difficult for machines. The numbers should therefore be interpreted as structured estimates of exposure, not as direct measurements of automation in the real economy.

And there's a particularly important caution for us, in Brazil. This is essentially a study about the American labor market. Occupations and tasks come from O*NET, employment and compensation data are American, and the economic comparison uses labor costs and deployment parameters associated mainly with the reality of the United States.

I wouldn't transpose these percentages directly to Brazil. Here, wages, cost of capital, taxation, company scale, sector composition, informality, infrastructure, regulation, availability of maintenance and the cost of acquiring or importing equipment all change. Lower wages can make economic substitution less attractive for certain activities. In others, worker shortages, safety, precision, 24×7 availability or quality gains can justify robotization even before simple parity between robot cost and wages is reached. Technology can be global. The economics of automation is deeply local.

My final takeaway is far less cinematic than the narrative that "robots are coming to take our jobs." Robots can already perform significant parts of the work in many professions and can certainly transform occupations, redistribute tasks, increase productivity and reduce demand for certain types of work. The study offers evidence that technological exposure can have real effects on employment and wages.

But between demonstrating a task, executing it reliably in the real world, making it economically competitive, integrating it into the production process, gaining regulatory and social acceptance, achieving adoption at scale, and finally eliminating an occupation, there are several steps.

Mixing all of them into a single percentage makes for an excellent headline. It doesn't necessarily make for a good forecast about the future of work.