Anthropic measures which jobs robots can perform

Robots can already perform 74% of the physical tasks carried out in the United States, but that does not mean they are about to replace most workers. The main obstacle is not just what they can do, but how much it costs to put them to work.

An Anthropic study examines which tasks autonomous robots can perform today. These are physical machines that can perceive their surroundings and act without constant human control. The results suggest that drivers, warehouse workers and packers face greater exposure than nurses, repair technicians or caregivers.

Robots can do a lot, but only in very specific places

The study reviews nearly 19,000 tasks across roughly 900 occupations in the United States. For each one, it assesses how far a current robot can go:

• In an environment designed specifically for robots, such as an assembly line.
• In a structured workplace, such as a warehouse.
• In an open and unpredictable environment, such as an urban road.

The difference matters. A robot may be able to sort packages in a facility built for that purpose, but struggle if it has to pick up disorganized objects, move among people or react to an unexpected situation.

That is why, although robots can cover 74% of physical tasks, those tasks account for around 34% of total working time. Only 2% of the physical tasks analyzed can currently be performed in completely unstructured environments.

Driving is one of the most exposed cases. Autonomous vehicles can already operate without a driver in certain cities and conditions. By contrast, robots still struggle with tasks such as handling delicate objects, working around buried pipes or adapting to repairs that never turn out exactly the same way.

Drivers and warehouses appear first

Taxi drivers top the list of the most exposed occupations, with an exposure score of 2.2 out of 3. Nine of the ten most exposed occupations involve operating vehicles.

Warehouses also concentrate risk. Mobile robots already transport goods, load materials and work alongside people in logistics centers. Packers and order pickers are an especially relevant example: the study estimates that robots can perform 97% of their tasks and that, in some cases, they are already slightly cheaper than hiring workers.

That group employs around 560,000 people in the United States. Employment has fallen 22% since 2015, although the study does not attribute all of that decline to robots.

Exposure also reflects labor inequalities. Workers in highly exposed occupations have, on average, less education, earn around 30 dollars less per hour and have an unemployment rate more than twice as high as people working in unexposed occupations. They are also more likely to be men and to hold physically demanding jobs.

Artificial intelligence broadens the reach

When robots and language models, such as those that generate text, are combined, around 81% of working time is exposed to one of the two technologies. Language models mainly affect digital and cognitive tasks. Robots add physical jobs that AI alone cannot perform.

Even so, some work remains more resistant to automation. This includes tasks that require face-to-face interaction, physical contact, manual dexterity or constant adaptation. Caregiving, parts of healthcare, repair work and social services fall into this group.

That does not mean these occupations are protected forever. It means that, with current capabilities, they are harder to automate reliably.

Cost is the main obstacle

The difference between being able to perform a task and being able to perform it profitably is enormous. According to the study's estimates, robots are cost-competitive for only 0.3% of current work tasks.

In many cases, the robot needs sensors, installation, maintenance, software, human supervision and an adapted workspace. A system capable of welding may do the job well, but it still needs help positioning large parts, climbing stairs, checking the result or correcting errors.

If robot prices fall at the rate observed since the 1990s, it would take around 40 years for them to become competitive in 10% of current work. This is a projection, not a guaranteed date. New designs, cheaper manufacturing or advances in humanoid robots could speed up the process, while regulations and customer preferences could slow it down.

The analysis also finds that robots have become capable of performing approximately 2% more of the physical tasks they previously could not do each year. And when reviewing the past 50 years, the authors observe that jobs more exposed to robots tended to experience larger subsequent declines in wages and employment.

The study's signal is concrete: physical automation will not reach every job first. It will reach the jobs robots can already perform in controlled environments. It makes sense to watch transportation, logistics and packaging before imagining robots suddenly replacing nurses, mechanics or caregivers. The future of work will depend as much on machines' capabilities as on their price, the rules and the tasks humans continue to do better.