Robot exposure is a workflow map, not an automation forecast
Anthropic’s new index says robots can perform many physical tasks in some settings but are cost-competitive for very few. Workforce planning should separate capability, environment and economics.
What happened
Anthropic published a robot-exposure index on 30 September using O*NET tasks, Claude-assisted task classification, cited examples of demonstrated robots and historical back-testing. It reports that robots can perform 74% of physical tasks in some setting, representing 34% of working hours, but are cost-competitive for 0.3% of job tasks.
Why it matters
Exposure measures whether a machine can perform a task under specified conditions; adoption also depends on cost, reliability, workflow redesign, regulation and worker or customer preferences. Collapsing those layers would turn a useful task map into a false job-loss forecast.
Anthropic published a robot-exposure index on 30 September. The research starts from roughly 19,000 O*NET task descriptions across about 900 occupations. Claude helps identify physical tasks, search for demonstrated robots and classify the least controlled environment in which a machine can perform each task. The authors weight tasks by estimated work time and employment, release reasoning and citations, and back-test historical exposure against later wage and employment change.
The headline is large: the study estimates that robots can perform 74% of physical tasks in at least some setting, equal to 34% of working hours. The economic boundary is much smaller. It estimates current robots are cost-competitive with people for 0.3% of job tasks and projects, under historical price trends, that the share would take about 40 years to reach 10%. The paper also says regulation, preferences and capability gaps can block adoption.
Keep three columns
Use the dataset as a task inventory with three separate columns. Capability asks whether a demonstrated robot can perform the task and in what environment. Deployment asks whether the organisation can redesign the workplace, integrate the machine and meet reliability and safety requirements. Economics asks whether total cost—including supervision, downtime, insurance and transition work—beats the current process.
Do not convert an E1 task, possible only in a purpose-built robotic environment, into a claim that the ordinary workplace is ready. The study's E2 and E3 tiers distinguish structured human facilities from unstructured environments, but local variation remains material. A warehouse with standard packages and clean lanes is not the same workflow as a small mixed-goods site.
Validate the index where work happens
Select ten high-time tasks in one role. Observe the real environment, exceptions and handoffs. For each task, record the exposure tier, evidence date, robot cited, required workplace change, failure consequence, human recovery step and fully loaded cost. Run a time-limited pilot only where the task, environment and economics all pass.
The counterargument is that present-day cost can fall faster than historical trends and AI could improve dexterity or adaptation discontinuously. That is plausible. Conversely, Reuters reporting on Chinese humanoid factories documents current limits in dexterity, autonomy and commercial readiness despite strong hardware investment. Neither source can forecast a local adoption date.
The Skills Intelligence Role Dictionary can structure the task observation and expose where exception handling, coordination and safety work sit. It should not label a whole occupation automatable because some tasks score as exposed.
The immediate decision is to build a ten-task capability–deployment–economics table for one physical workflow. Use the index to choose what to inspect first, then let local evidence—not a headline percentage—decide the pilot.