A robotic arm can sort boxes on a line designed around it. Ask the same system to handle every surprise in a warehouse, and the comparison with human work changes. That gap between performing a task somewhere and taking over a whole job is the central question in Anthropic's September 30 research.
Researchers Russell Legate-Yang and Maxim Massenkoff estimate that existing robots can perform 74% of physical tasks in the US labor market, weighted by time spent and employment. Those tasks account for 34% of all working time in their analysis. The headline number needs its operating conditions attached: much of this capability exists only in factories or other carefully arranged spaces.
Capability is only the first hurdle
The researchers rank tasks from those robots cannot perform to those they can handle in unstructured environments. In their assessment, about half of all physical tasks require a workplace built for the machine. Another 22% can be done in structured human facilities such as warehouses. Only 2% can be done in unstructured settings. Packing meals on a conveyor belt says little about how a robot would cope with a busy kitchen.
The economic gap is wider still. Comparing estimated machine and labor costs, the study finds robots cost-competitive for just 0.3% of all work. A feasible task may still require equipment purchases, maintenance, supervision, and changes to the workplace. For a company, the useful question is not simply whether a robot can do it. It is where, how reliably, and at what cost per completed outcome.
Exposure describes a technical possibility under particular conditions; adoption requires the whole operation to work economically.
What the index can and cannot tell us
The team started with descriptions of roughly 19,000 tasks in the O*NET occupational database. They used Claude to identify physical work, generate concrete examples, estimate time spent, and find evidence of existing robots. That gives the research breadth, but it also brings model judgments into the measurement. Short task descriptions may miss the awkward details of actual work. Demonstrations can count as capability evidence, although the authors say results are similar when demonstration-based ratings are removed.
The study also looks backward: occupations with greater robot exposure in earlier decades subsequently had larger wage and employment declines. That makes the index worth watching, but it does not establish that today's exposure will cause the same outcome in every occupation. Demand, regulation, capital costs, and redesigned jobs can all change the path.
Anthropic estimates that, if robot prices continue falling at the historical rate used in its model, cost competitiveness would reach 10% of work in about 40 years. This is a conditional scenario, not a scheduled forecast. The researchers acknowledge that rapid technical gains could change it and that preferences and regulation are outside parts of their projections.
For leaders considering automation, the practical starting point is repetitive work in controlled settings, measured across the entire process, including exceptions and human support. For workers and policymakers, the index can reveal areas of pressure without turning task exposure into a claim about jobs lost. That distinction is where the more useful debate begins.