Anthropic estimates that demonstrated robot capabilities cover 74% of physical working time in the United States, equivalent to 34% of all working time. Its cost model finds only 0.3% of all working time currently economical to automate with robots. These are employment- and time-weighted estimates, not counts of jobs that will disappear.[1]
The study defines robots as autonomous physical machines that sense and act. It rates tasks by the environment required: a purpose-built robotic workspace, a structured human facility or an unstructured environment such as a public road. This matters because a robot that welds one repeated joint inside a designed cell has not demonstrated that it can do the full job of a welder across variable sites and preparatory work. Its exposure and cost estimates use Claude-assisted assessment of task descriptions, documented robot capabilities and working-time assumptions. They are not direct trials of robots doing every sampled job.
The capability–economics gap
| Measure | Value | Context |
|---|---|---|
| Physical working time exposed to robots | 74% | The estimate includes tasks possible only in purpose-built or structured environments.Source 1 |
| Share of all working hours represented | 34% | This is weighted by employment and estimated time spent on tasks.Source 1 |
| Work exposed to robots or LLMs | About 80% by working time | The technologies affect different portions of work and overlap in some areas.Source 1 |
| All working time currently cost-competitive | 0.3% | Cost estimates include equipment, installation, operating expenses and coordination assumptions.Source 1 |
| Illustrative slower scenario | Half of physical work cost-competitive by 2085 | The study’s scenario combines capability and cost trends; it is not a prediction.Source 1 |
A large purchase can leave a small annual margin
Occupations are bundles of activities. A warehouse packer may assemble boxes, inspect goods, apply labels, move material, resolve exceptions and coordinate with people. Automating several high-volume steps can change staffing even when no machine performs the complete occupation. Conversely, a demonstration of one task can overstate disruption if the remaining activities are expensive or difficult to integrate.
In the study’s modelled packing example, equipment exceeding $2 million serves the equivalent yearly workload of about 14 people. A roughly ten-year life, 8% capital cost, maintenance, supervision and energy leave an annual cost near $45,000 per worker-equivalent. Against $49,000 compensation for a human whose work is 97% exposed, the estimated saving is only about $2,500. This is a modelled case, not a verified factory saving.[1]
Structured environments are part of the investment
Many robots become capable only after the workplace is redesigned. Floors may need markers, items must arrive in consistent orientations, safety zones must be enforced and upstream systems must provide reliable data. Those changes are not evidence that automation is false; they are part of its capital cost. Firms with high, stable volume can justify that engineering sooner than smaller sites with variable work.
Why technical exposure does not become immediate adoption
| Barrier | What it changes | What to measure |
|---|---|---|
| Capital cost | Raises the volume needed to justify a deployment | Installed cost, financing, service life and utilisation.Source 1 |
| Environment design | Moves complexity from the robot into the workplace | Facility modifications and lost flexibility.Source 1 |
| Incomplete task coverage | Leaves handoffs and exceptions for people | Share of end-to-end workflow completed without rescue.Source 1 |
| Reliability and downtime | Makes a nominally cheaper task expensive when recovery is frequent | Successful cycles, repair time and human supervision.Source 1 |
| Regulation and preference | Limits adoption even when capability and price are favourable | Permits, liability and customer acceptance.Source 1 |
Falling hardware cost compounds slowly
A simple illustration shows why time matters. A 3% annual decline leaves a cost at about 29.6% of its starting level after 40 years, a reduction of roughly 70.4%. That is substantial, yet an expensive system can remain uncompetitive for decades if integration and supervision do not fall with hardware. Anthropic’s scenarios are richer than this arithmetic and also allow capability to improve; the example explains why a modest annual trend can still imply a long transition.
The labour consequences will also differ by occupation. The study finds greater exposure among jobs that are more physical, lower paid and more often held by men, while nursing and general repair remain less exposed because dexterity, interpersonal work and unstructured environments are difficult. Policy should therefore monitor specific tasks, regions and worker transitions rather than treating “robots” as one uniform shock.
Sources & further reading
- What work can robots do? Anthropic · 30 Sep 2026 · Accessed 14 Oct 2026