Robots can perform 74% of U.S. physical work tasks in at least one environment, according to new Anthropic research. Yet robots are currently cost-competitive for only 0.3% of total work. Both numbers can be true because robot exposure measures technical reach, while robot adoption must pass environment, reliability, integration, and cost constraints.
The useful forecast is therefore a funnel, not a headline percentage. Each improvement moves the bottleneck to the next layer. A better model expands capability; deployment still waits for the physical setting, operating reliability, workflow integration, regulation, user preference, and total economics to align.
What Anthropic's robot exposure index measures
Anthropic's What work can robots do? starts with the U.S. Department of Labor's ONET database. ONET contains descriptions for roughly 19,000 job tasks. The research classifies 7,594 as physical, then asks whether present-day robots can perform each task and how structured the environment must be.
The exposure rubric has four levels:
| Level | Meaning | Example environment |
|---|---|---|
| E0 | Current robots cannot perform the task | Work still requires missing physical capability |
| E1 | A robot can perform it in a purpose-built robotic setting | Factory cell or dedicated line |
| E2 | A robot can perform it in a structured human workplace | Warehouse or hospital corridor |
| E3 | A robot can perform it in an unstructured setting | Public road or open field |
Tasks are weighted by employment and estimated time spent. On that basis, physical work represents 46% of total U.S. work time. Robots receive some exposure rating for 74% of physical work, equal to 34% of all work time. Only 2% of physical tasks by work time reach E3. Most exposed work still depends on a purpose-built or structured environment.
This makes the index more informative than a binary can-do score. It captures environmental control as a capability requirement. It still measures exposure rather than observed deployment.
The 74% headline describes a wide technical frontier
Exposure aggregates very different situations. A robot that completes one task inside a dedicated cell and a robot that completes it in a normal workplace both count as technically exposed, while their deployment costs and integration burdens differ sharply.
The paper's examples show why. Warehouse movement fits structured spaces, known package shapes, marked routes, and repeatable handoffs. Drywall finishing can be highly exposed for spraying and sanding while messy tape application remains unexposed. Nursing and repair combine mobility, dexterity, perception, judgment, and interpersonal work in changing environments.
An occupation therefore faces two composition problems:
- A robot may cover many minor tasks while missing one task that controls the whole workflow.
- Several robots may cover separate tasks while adding handoffs, supervision, and coordination cost.
This is a bottleneck problem. Improving a non-binding task produces more theoretical coverage while the workflow remains constrained by its weakest required step.
Cost moves the forecast from possibility to pressure
Anthropic estimates task-level robot cost from purchase, installation, service life, financing, maintenance, energy, and supervision. It compares that annualized cost with worker compensation allocated to the same task.
Under this model, robots are cost-competitive for 0.3% of all work time today. The paper estimates that robots would need a roughly 70% cost decline to become competitive for 10% of current work. A historical quality-adjusted price decline of about 3% per year would take around 40 years to reach that point.
These figures describe a scenario, not a timer. New hardware, manufacturing scale, higher utilization, or better capability could accelerate the curve. Regulation, financing, integration, preferences, and wages could slow it. The paper itself labels cost comparisons as approximate and notes that task-by-task aggregation may double-count shared hardware or miss coordination cost.
The key signal is the gap between 34% technical exposure and 0.3% cost-competitive work. That gap identifies where adoption research belongs.
A five-gate robot adoption model
An enterprise evaluating automation can translate the research into five sequential gates.
Gate 1: Task coverage
Define the actual task with inputs, outputs, quality tolerance, speed, and failure conditions. A vendor demonstration establishes one point in this space. Production eligibility requires the full task distribution.
Gate 2: Environment fit
Record the structure the robot needs: fixed geometry, controlled lighting, marked routes, standardized objects, human separation, or predictable weather. Include the cost of reshaping the workplace.
Gate 3: Reliability and recovery
Measure success rate across shifts, edge cases, and degraded conditions. Track mean time between intervention, safe-stop behavior, recovery time, and the skill level required from the human supervisor.
Gate 4: Workflow integration
Map upstream supply, downstream handoffs, software interfaces, safety procedures, maintenance, training, and exception ownership. A fast robot can reduce total throughput when it creates queues or brittle handoffs elsewhere.
Gate 5: Risk-adjusted economics
Calculate total annual cost per accepted unit of output:
hardware + integration + facilities + financing + maintenance + energy + supervision + downtime + failure loss
Compare the result with the complete human workflow, including flexibility and the option value of reallocating labor. Run sensitivity analysis for utilization, failure rate, service life, and cost of capital.
An adoption probability should only emerge after all five gates have evidence. Exposure is the input to Gate 1 and part of Gate 2.
Where the study is strong, and where uncertainty enters
The study publishes a detailed rubric, task-level citations, a historical backtest, and an accompanying data release. Its historical analysis finds that occupations with higher past robot exposure later experienced larger wage and employment declines, after controls for industry trends and other factors. That gives the index predictive relevance.
Several layers still depend on model judgment:
- Claude classifies physical tasks and generates concrete task examples.
- Claude estimates time spent on tasks when O*NET descriptions lack that field.
- Claude searches for robot evidence and rates the least structured workable environment.
- Claude helps estimate robot costs and aggregate overlapping equipment at occupation level.
The paper uses majority rules, citations, alternative specifications, and historical validation to reduce error. Residual uncertainty remains material. The correct interpretation is directional and comparative: which occupations face earlier pressure, which environmental constraints dominate, and how wide the capability-to-economics gap remains.
How to use the index without overstating it
Policy analysts can use exposure to identify occupations that deserve closer measurement. Employers can use it to select candidate workflows for pilots. Workers and educators can use it to locate tasks where complementary skills, exception handling, and interpersonal work remain valuable.
None of those decisions require turning 74% into a displacement forecast. A robust dashboard should display at least four separate measures:
- technical task exposure by environment level;
- observed deployment and utilization;
- intervention-adjusted reliability;
- cost competitiveness under stated assumptions.
The same separation improves forecasts for software agents. Anthropic's earlier labor-market research distinguishes theoretical AI capability from observed use. Robotics adds two hard constraints that software analysis can sometimes postpone: the physical environment and capital cost.
FAQ
Does the study say robots can replace 74% of physical workers?
No. It says current robots can perform 74% of physical tasks by work time in at least one environment. Many require purpose-built settings, and task exposure does not establish whole-job replacement.
Why is only 0.3% of work cost-competitive?
Robots require hardware, installation, financing, maintenance, energy, supervision, and workplace integration. Human workers also cover multiple tasks and adapt to changing conditions, which raises the threshold for an economical replacement.
Which work is closest to adoption?
The index points toward vehicle operation, warehousing, packaging, and other tasks already feasible in structured environments. Actual adoption still varies by regulation, facility scale, utilization, and local labor economics.
Is the 40-year estimate a prediction?
It is a scenario that applies a historical 3% annual robot-price decline to today's task and wage structure. Faster capability progress or manufacturing scale can shorten it; integration frictions, regulation, preferences, and changing wages can lengthen it.
What should a company measure during a robot pilot?
Measure accepted output, intervention rate, recovery time, downtime, failure cost, environment modifications, supervision, maintenance, and end-to-end throughput. A demonstration success rate alone cannot establish adoption economics.