Anthropic's new research states that robots can perform about 74% of physical jobs in the U.S., which account for 34% of total working hours. However, only 0.3% of these jobs are cheaper than human labor, and a full replacement of blue-collar workers will take decades at the historical rate of price declines.

The paper "What work can robots do?" identifies cost as the primary barrier. Robot prices have decreased by about 3% annually since the 1990s. At this rate, a further reduction of about 70% is needed for robots to be cost-competitive in 10% of jobs, which would take approximately 40 years. Capability limitations are the second barrier, with examples such as the dexterity required to untangle wires. Human preferences and regulations will continue to block some scenarios.

The pocket of jobs where costs are equal is small but still covers about 300,000 workers, with robots able to perform 95% of their tasks. Packaging workers are the closest: the cost of purchasing and installing robots for boxing, inspection, labeling, and sealing exceeds $2 million, replacing the annual workload of about 14 people. With a lifespan of about 10 years, 8% capital costs, plus maintenance, part-time labor, and energy consumption, this equates to about $45,000 per person per year, while the total compensation for packaging workers is about $49,000, making robots approximately $2,500 cheaper annually. There are about 560,000 hand packaging workers in the U.S., a decrease of 22% since 2015.

The cost differences are much larger for other occupations. The cost of welding robots is about five times that of welders; cleaning robots cost about $280,000 per year, while janitors earn about $47,700; robot taxis are only about $7,000 more expensive per year than drivers but face regulatory hurdles. At the historical rate, it will take until 2085 for half of physical jobs to be cost-competitive with robots; even if cost reductions accelerate to four times faster and new task learning doubles, this half will still take until 2050.

About 80% of working hours are already exposed to robots or large models, with robots filling the physical tasks that large models cannot perform. The paper back-tested data since 1977, finding that robots could not perform 62% of physical tasks then, now only 24% remain unperformable, with about 2% of new tasks each year being covered that were previously unperformable; jobs with higher exposure see greater declines in wages and employment. Workers highly exposed earn about $23 per hour, while those not exposed earn close to $53.

This is a cost curve rather than a capability announcement in pricing. Buyers of robots are warehouse and fleet operators, while sellers are equipment manufacturers and model companies; the event itself does not constitute immediate replacement orders. The packaging segment is close to replacement, while welding and cleaning still find humans cheaper. Beneficiaries are already scaled warehouse automation suppliers, while those under pressure are equipment narratives that have factored in humanoid robot penetration into valuations; the labor side first impacted is low-education, low-wage male blue-collar workers, not nurses and mechanics.

Apollo's chief economist Torsten Sløk quoted the paper, stating that even under aggressive assumptions, widespread replacement of physical labor will take decades. The paper notes that robots must maintain record price declines and performance improvements for decades for rapid physical automation to occur.

Source: Public Information