AI's labor debate often focuses on office work. Physical work has a different constraint: software has to become a machine that can sense, move and manipulate the real world reliably enough—and cheaply enough—to compete with people.
Known: technical exposure is already broad
Anthropic researchers estimate that robots can perform 74% of physical job tasks in the United States in at least some circumstances. Because physical tasks make up less than half of total work time, that corresponds to about 34% of all working hours. Most of this capability exists in purpose-built or structured environments. Only about 2% of physical tasks were rated doable by today's robots in unstructured environments.
Driving and warehouse work rank among the more exposed areas. Nursing, general repair and work requiring flexible interpersonal or physical interaction remain much less exposed. Read the primary research →
The same analysis estimates robots are currently cost-competitive with human labor for only about 0.3% of job tasks. Being technically possible and being worth deploying are very different thresholds.
Known: cost is a major brake on deployment
The researchers estimate that robot costs would need to fall roughly 70% for robots to become cost-competitive for 10% of human work under their assumptions. If historical robot-price declines of about 3% per year continued, they estimate that threshold would take roughly 40 years.
That is not a forecast. Humanoid manufacturing, cheaper hardware, better AI control or new production methods could change the cost curve. Regulation, financing, reliability and customer preference could slow it.
Uncertain: exposure is not destiny
The index uses Claude to help rate task exposure and estimate how workers divide time across tasks. The authors validate the measure against historical labor outcomes, but the future may not follow the historical relationship. A job is also a bundle of tasks: automating one part can increase the value of the remaining human work rather than remove the occupation.
The historical analysis finds that occupations more exposed to available robots later experienced greater wage and employment declines, after accounting for several confounders. That makes exposure worth watching, but it does not establish that a specific worker or occupation will be displaced on a particular timeline.
Speculation: physical AI could become the next major automation frontier
If AI improves robot perception and high-level control while hardware becomes cheaper, physical automation could accelerate beyond historical trends. Anthropic's separate robotics experiments this year found frontier models improving at navigation and manipulation when paired with pretrained controllers, while direct low-level control and difficult unstructured tasks remained unreliable.
A faster robotics curve is therefore plausible, but today's evidence does not justify claims that humanoids are about to replace most physical workers.
What to watch
- Quality-adjusted robot cost per productive hour, not demo prices.
- Performance in messy, unstructured workplaces rather than controlled demos.
- Dexterity, exception handling and the amount of human supervision still required.
- Commercial deployments in logistics, driving, food preparation and manufacturing.
- Whether robot capability expands faster than deployment costs fall.
How to prepare
Physical workers
Learn the automated systems entering your field and strengthen the tasks that remain hard to mechanize: troubleshooting, judgment, customer interaction, coordination and work in unpredictable environments.
Students & career changers
Don't classify a career as simply “safe” because it is physical. Ask which tasks happen in structured environments, how expensive automation is, and whether the work requires human trust, dexterity or adaptation.
Employers
Measure total deployment economics: equipment, integration, supervision, downtime, safety and maintenance—not just whether a robot can complete a demo task.
Our read
The most useful distinction is capability versus economic deployment. Robots can already perform far more physical work than their current economic footprint suggests. That gap argues against both extremes: “physical jobs are safe” and “robots are about to replace everyone.” Watch the gap itself. If capability broadens while cost and supervision requirements fall quickly, the labor implications become much more immediate.
Primary sources
- Legate-Yang & Massenkoff — What work can robots do? (Anthropic, Sep. 30, 2026)
- Anthropic — Claude plays robotics (Jul. 9, 2026)