Robots can already do about three-quarters of the physical tasks in the US economy. They’re cheap enough to replace a human on only 0.3% of them. That’s the main finding of a new robot exposure index from Anthropic, published September 30. It’s one of the first serious attempts to measure what physical AI can do today, rather than what people fear it might do someday.
Anthropic reports that the tasks robots can technically handle add up to about 34% of US working hours. The catch is that most of that capability only works in tightly controlled settings. Having the ability isn’t the same as being worth deploying.
🔬 How the Study Worked
Anthropic’s researchers started from O*NET, a US government database that links about 900 occupations to roughly 19,000 job task descriptions. Claude scored each task on a rubric covering its physical, cognitive and interpersonal requirements.
For robots, the key question at the task level was simple: can a robot today perform this task, and if so, under what circumstances? That second part matters a lot. A robot that can sort packages on a fixed conveyor line is a very different thing from one that can fix a leaking pipe in someone’s cramped basement.
The team then put a cost lens on top of the capability data. They asked where robots actually beat human labor on price, which is the test that decides whether any business will buy one.
📊 The Key Numbers
- ~75% of US physical job tasks can be done by currently available robots, at least in limited settings
- 34% of US working hours are covered by those tasks
- 0.3% of job tasks are ones where robots are cost-competitive today
- 40 years is how long it would take that share to reach 10%, if robot prices keep falling at their historical rate
- ~80% of job tasks by working time are exposed to either robots or LLMs
The gap between 75% and 0.3% is the whole story. Engineering progress has moved much faster than economics.
👷 Who’s Exposed, and Who Isn’t
According to Anthropic, workers in robot-exposed jobs are more likely to be male, less educated and lower paid. That’s close to the opposite of the profile for LLM exposure, which leans toward white-collar, knowledge-heavy work.
The examples make it concrete:
- Highly exposed: driving and warehouse jobs, where today’s robots can already do much of the work
- Barely exposed: nursing and general repair, where present-day robots can do very little, even in controlled environments
The roughly 20% of work that neither robots nor LLMs touch has two things in common. It’s either deeply interpersonal or it needs physical skills that robots don’t have yet.
💡 Why This Matters
What stands out to me is that Anthropic is separating “can a machine do it” from “will anyone pay for a machine to do it.” Most automation headlines blur those two questions together. That’s how you end up with predictions that half of all jobs are gone in a decade, followed by a decade where that doesn’t happen.
The study also fills in the other half of a picture Anthropic has been building. Its earlier labor market research looked at language models and office work. With robots added, you get a combined view: roughly four out of five working hours are now within reach of some kind of AI, but the speed of change depends heavily on the cost curve.
⚠️ Caveats Worth Keeping in Mind
A few points from the study’s own design deserve attention:
- The 40-year estimate assumes past price trends hold. A breakthrough in humanoid robots or a big jump in manufacturing scale could shorten it a lot.
- Claude did the scoring. It’s a consistent, scalable method, but it relies on how a model reads task descriptions, not on watching robots work in the field.
- “Limited settings” is doing heavy lifting. Most of that 75% figure depends on controlled environments that many real workplaces don’t offer.
🧭 What to Do With This
If you’re planning workforce strategy, don’t treat robot exposure and LLM exposure as the same risk. They hit different workers on very different timelines.
- Operations leaders: watch the cost curve, not the demo videos. Capability already exists for warehouse and driving tasks, so price is the trigger.
- Workers in exposed roles: skills that are interpersonal or need hands-on problem-solving in messy environments are the most durable.
- Policymakers: the robot-exposed workforce is lower paid and less educated, so reskilling plans for this group should look very different from those aimed at office workers.
My recommendation is to treat physical automation as a slow, price-driven change, not a sudden cliff. Keep an eye on robot unit costs, because that’s the number that will tell you when the 0.3% starts to climb. The full methodology and occupation-level data are available in Anthropic’s original report.