74% Capable, 0.3% Affordable: Anthropic's Robot Exposure Index Sizes Up the Physical Economy

Ask when robots will take physical jobs, and you usually get vibes. Anthropic’s latest research contribution, published September 30 under the plain-spoken title “What work can robots do?”, replaces the vibes with a number — or rather, with two numbers that pull in opposite directions. Using its own Claude models to grade roughly 7,600 physical job tasks from the O*NET occupational database, the team built a robot exposure index that reaches a startling headline: robots available today can already perform about three-quarters of physical tasks in the US economy, accounting for 34% of all working hours. The second number is the cold shower: those robots are cost-competitive with human labor for just 0.3% of job tasks.

How the index works

The methodology is what makes the study interesting rather than another automation thought experiment. The researchers defined robots as autonomous physical machines that sense and act — which includes a car wash that scans vehicles to adjust its sprayers, but excludes teleoperated surgical systems that remain fully under a surgeon’s control.

Physical tasks were identified by having Claude score O*NET task descriptions on a rubric of physical, cognitive, and interpersonal requirements. “Dig trenches” is physical; “Teach dance students” is physical but also cognitive and interpersonal; a typist operating an adding machine may use their hands but doesn’t count. This filtering produced 7,594 physical tasks drawn from a database of around 900 occupations and 19,000 task statements.

Each physical task was then rated on a four-tier exposure scale determined by how controlled an environment a robot needs:

• E0 — robots cannot perform the task at all
• E1 — robots can do it only in a purpose-built robotic environment, like a factory assembly line
• E2 — robots can do it in a structured human workplace, like a logistics warehouse
• E3 — robots can do it in fully unstructured environments, like a city road

Crucially, only demonstrated capabilities count. Claude was instructed to search for specific commercially available robots relevant to each task, assess their real-world deployments, and quote sources directly — a task is only rated exposed if a robot can do it roughly as well as a human, factoring in reliability, error rates, and speed. The task “Dig trenches,” for example, splits into sub-activities: autonomous retrofit systems for hydraulic excavators can cut linear trenches in open ground (E3), but no robot can carefully dig around buried pipes and cables, and since that careful digging dominates the work time, the task overall rates E0.

What robots can actually do

Weighting tasks by employment and time spent, the distribution is lopsided in a revealing way. Physical tasks split into roughly a quarter that robots cannot do at all (E0), about half that robots can do only in purpose-built environments (E1), 22% in structured human workplaces (E2), and just 2% in unstructured environments (E3). That final tier is dominated by driving — autonomous cars, tractors, trucks, and pavers.

Accordingly, nine of the ten most exposed occupations are vehicle operators. Taxi drivers top the index at 2.2 out of 3, with their median task by working time rated E3 thanks to autonomous vehicles. Shuttle drivers and chauffeurs score 2.0, with Claude citing Waymo’s robotaxis as performing some of their tasks. Warehouse and logistics work follows closely: autonomous mobile robots now drive to loading docks, enter trailers, and use suction grippers to load packages onto mobile conveyors.

The flip side is equally concrete. Nurses and general repair workers see little exposure because present-day robots can do little of their work even in highly controlled environments. Fine motor dexterity remains a moat: hair coloring, untangling wires, pressing paper tape into wet drywall compound. And the strongest machine capability — navigation — is precisely the one that doesn’t transfer to hands-on care work.

Validated against 50 years of labor data

The study’s most rigorous section backtests the index against history. The team rated robot exposure for job tasks as described in several years since 1977, then linked exposure to subsequent wage and employment changes. Jobs that were more exposed to then-existing robots experienced greater wage and employment declines in later decades, even after controlling for industry trends and confounders. The implication: the index is not just a description of capability but a leading indicator. If the pattern holds, taxi drivers and warehouse packers will see changes sooner than nurses and mechanics.

The backtest also quantifies the pace of progress: since 1977, robots have become able to perform about 2% of the physical tasks they previously couldn’t, each year. Steady, not explosive.

The 0.3% problem

The cost analysis is where the paper earns its pessimism about near-term disruption. For every exposed task, Claude estimated the annual cost of deploying the cited robots — annualized hardware costs over roughly 10-year service lives at an 8% cost of capital, plus maintenance, part-time human supervision, and energy — and compared that to what human workers cost.

The results show why “can do” rarely means “will do.” Packing and packaging is the largest occupation exposed to cost-competitive robots: a suite of machines costing over $2 million to purchase and install replaces the yearly work of about 14 workers, working out to roughly $45,000 per replaced worker against human compensation of about $49,000. That slim $2,500 annual margin matters — employment for packers and packagers has already fallen 22% since 2015, and the BLS projects the occupation will shed the 11th-most jobs of any by 2035. Robotaxis, meanwhile, are estimated to cost only around $7,000 more per year than taxi drivers, with regulation as the binding constraint.

But away from those edge cases the gap is wide. Welding robots with cameras and AI can weld autonomously, yet automating the full job — positioning heavy parts, climbing ladders, grinding finishes — costs about five times more than human welders. Dishwashers and janitors earn $25,000–30,000 less than welders, and their robot analogs remain several times more expensive. Postal Service mail carriers would need robot costs to fall 54% to reach parity. Overall, robots would need a roughly 70% cost decline to be cost-competitive for 10% of human work; at the historical ~3% annual price decline observed since the 1990s, that takes around 40 years. Half of physical work doesn’t cross over until 2085. Even in a fast-adoption scenario with costs falling four times faster and capabilities doubling in pace, the halfway point only moves up to 2050.

Who is exposed

The demographics invert the familiar LLM story. Workers in the top exposure quintile are 20 percentage points less likely to be female, 16 points more likely to be Hispanic, 55 points less likely to hold a bachelor’s degree, earn around $30 less per hour, and face more than double the unemployment rate of unexposed workers. Where LLM exposure concentrated in white-collar, college-educated work, robot exposure concentrates among male, less-educated, lower-paid workers.

Combining both technologies sharpens the total picture: about half of work is exposed to LLMs alone, but adding robots lifts the figure to roughly 81% of job tasks — all but one-fifth of employment. Transportation and moving jobs jump from under 15% LLM exposure to about 90% combined; office and administrative support approaches 100%. What remains unexposed is highly interpersonal, hands-on work: administering medications and monitoring patients, dressing children, greeting and seating guests.

Barriers beyond cost and capability include regulation (blocking about 14% of physical tasks, concentrated in healthcare, protective services, and education) and human preferences (a quarter of tasks, where people wouldn’t trust a robot or value the human interaction). Manipulation dexterity is the single biggest capability gap — half of physical tasks stay unautomated at scale until robots get better at simply touching and handling objects.

The bottom line

The paper’s dual finding is a useful corrective to both hype and dismissal. The capability frontier is much wider than deployment suggests — three-quarters of physical work is already within technical reach, mostly in structured settings. But hardware economics do not obey software economics: robots are physical capital that can’t be copied for near-zero marginal cost, and at current trajectories the affordable slice of that frontier expands by fractions of a percent per year. The plausible future the data sketches is not a sudden humanoid takeover but a slow, uneven crawl — packers and taxi drivers first, nurses and electricians much later — unless humanoid manufacturing drives a genuine break in the cost curve. Anthropic has released the underlying data, including Claude’s reasoning and cited sources for all rated tasks, so the crawl can now be measured.