Anthropic: "Physical AI Could Replace 74% of Human Labor... The Problem Is Cost" [Global AI Briefing]

[Financial News] A study has found that artificial intelligence (AI), which is rapidly replacing knowledge work in office jobs such as coding, document writing and information searches, could perform 74% of physical tasks—including driving and packing goods—when combined with robots. Physical AI, which uses generative AI as its brain and robots as its body, could replace human labor in driving, logistics and manufacturing.
Russell Leggate and Maksim Massenkoff of Anthropic's Economic Research team analyzed the potential of Physical AI to replace 74% of physical labor tasks in the United States in their report, *What Work Can Robots Do?*, released on September 30 local time. That figure is equivalent to 34% of total work hours in the United States.


The report examined 7,594 physical labor tasks that must be performed directly by people—including driving, moving and sorting goods, and operating machinery—among 19,000 tasks across approximately 900 occupations in the United States. It assessed the potential for those tasks to be performed by Physical AI.
The study found that Physical AI could perform 23% of physical tasks in environments designed for robots, such as modernized factories. In structured environments such as warehouses where people work, 10% of physical tasks could be performed by Physical AI. In unpredictable, unstructured environments such as roads, however, only 1% could be performed by the technology.
Driving occupations, including taxi drivers, had high exposure to Physical AI because of autonomous vehicles, while moving and sorting robots are already being used in warehouses. By contrast, Physical AI currently has limited capabilities in tasks requiring human interaction or delicate hand movements, such as nursing and general repairs.
However, Physical AI is unlikely to immediately replace human jobs because of cost concerns.
The report analyzed that although Physical AI can technically perform a substantial share of physical tasks, only 0.3% of all tasks could be performed more cheaply than by people.
Even if robot prices continue to fall by approximately 3% annually—the same pace as in the past—the report projected that it would take until 2085, about 40 years later, for the share of tasks in which Physical AI is cost-competitive to reach 10%.
The report's authors explained, "AI robots can perform most physical tasks, but they are far more expensive than people." They added, "Unlike software, robots are hardware and cannot be easily replicated or deployed, so it is essential to distinguish between the ability to deploy AI robots and their cost."
[email protected] Lee Gu-soon Reporter