Anthropic’s robot exposure index shows why capability is not ROI
Anthropic estimates physical automation can technically cover far more work than it can economically justify today. The gap is a useful framework for automation investment decisions.
Anthropic’s new robot exposure index makes a distinction that matters for almost every automation decision: a task can be technically feasible long before it is economically sensible to automate. The research estimates a large technical exposure surface for physical work, while the share that is cost-competitive today remains tiny.
The capability-cost gap
Why technical exposure is not a deployment decision
Exposure asks whether a system can perform a task under defined conditions. A business decision asks a harder question: can the complete system perform it reliably in the real environment at a lower total cost and acceptable risk? Hardware, integration, supervision, downtime, maintenance, exception handling and regulation can all change the answer.
The reusable lesson is the gap between capability and value. Teams often see a demo or benchmark and jump directly to an automation roadmap. A better sequence is feasibility first, then environment fit, full cost, failure handling and measurable business value.
A five-gate automation investment test
- Technical feasibility: define the exact task and the conditions under which the system can complete it.
- Environment fit: test whether those conditions match the real workplace, data and edge cases.
- Full cost: include integration, infrastructure, supervision, maintenance and exception handling—not only model or hardware price.
- Failure economics: estimate the cost of errors, downtime, escalation and human fallback.
- Business value: require a measurable improvement in cost, speed, capacity, quality or risk before scaling.
Capability signal vs investment signal
| Capability signal | Investment signal |
|---|---|
| A benchmark or controlled demo succeeds | The real workflow succeeds under production constraints |
| The system can complete the task | The system can complete it reliably enough |
| Unit technology cost looks low | Total operating cost beats the current process |
| Automation is technically possible | Automation produces measurable business value |
What the research does not prove
The estimates are a research model of task exposure and cost, not observed automation rates across US employers. The study’s task scoring, environment assumptions and cost model shape the results. The numbers also should not be transferred directly to software agents, where economics and deployment constraints are different.
Do not treat technical capability as an ROI forecast. Use it as the first gate. The investment case starts only after environment fit, total cost, failure handling and measurable business value are tested.
Sources & useful resources
- Anthropic: What work can robots do?— Primary research, Sep. 30, 2026