74% and 0.3% describe two very different robot markets
Anthropic analyzed roughly 900 U.S. occupations and 19,000 job tasks and estimated that current robots can perform 74% of physical tasks in at least some environment. That equals about 34% of all working hours.
Once purchase, installation, maintenance, supervision and energy costs are included, however, robots are currently cost-competitive with people for only about 0.3% of work.
The 74% figure measures technical exposure. The 0.3% figure measures the portion that also clears today's economics. The gap between them is where much of the Physical AI market still has to be built.
There are multiple meanings of 'a robot can do this'
The study ranks tasks from E0 to E3. E0 means no current robot can perform the task. E1 requires a purpose-built robotic environment, E2 works in a structured human workplace, and E3 works in an unstructured environment.
Across all working time, roughly 12% is physical work robots cannot do today, 23% can be done in purpose-built environments, 10% in structured human workplaces and 1% in unstructured environments.
So 74% does not mean robots can perform three-quarters of physical labor anywhere a person can. Much of that capability depends on shaping the workplace around the machine.
A successful demo is not the same as repeatable field performance
Anthropic's exposure measure requires demonstrated robots to perform tasks similarly well to humans, accounting for reliability, error rates and speed.
Even so, capability is estimated to block large-scale automation for around 70% of physical tasks. Manipulation is especially important: about half of physical tasks would not automate at scale without better touching and handling of objects.
For Physical AI, the useful question is not whether a robot has succeeded once, but whether it can preserve performance when the environment varies and the task repeats.
Economics remains a much narrower gate
For exposed tasks, the researchers estimate annual robot costs including purchase and installation, useful life, cost of capital, maintenance, energy and human supervision, then compare those costs with total human compensation.
Although robots can technically perform a broad share of physical work, only about 0.3% of all work is currently exposed to robots that are cheaper than people.
Can Do is therefore not the same as Should Deploy. A capability demonstration does not by itself produce an economic case for adoption.
Packaging shows how the 0.3% becomes economically viable
For packers and packagers, the study models a system of multiple robots that move goods, inspect items, erect boxes, label packages and seal them. Purchase and installation exceed $2 million.
The system is estimated to replace the annual work of roughly 14 workers. Annualizing fixed costs over service life and adding maintenance, supervision and energy produces a cost of about $45,000 per worker-equivalent, versus around $49,000 in human labor cost.
The economics work not because of a universal humanoid, but because the environment is structured, volume is high, work repeats and a large fixed investment can be spread across many years and tasks.
Task automation and job automation are not the same thing
Welding robots are mature, but welders also position heavy parts, climb to difficult joints, inspect quality, grind and finish materials.
When Anthropic estimates the robot system required to cover this broader bundle of work, it costs roughly five times as much as a human welder.
Automating nine tasks out of ten does not automatically automate the occupation. The last human-dependent task can become the bottleneck for total job economics.
At historical price declines, reaching 10% takes about 40 years — but that is not a forecast
Industry and government evidence reviewed by the paper suggests robot prices have fallen by roughly 3% per year since the 1990s. Extending that trend mechanically implies about a 70% cost decline and roughly 40 years for cost-competitive robot exposure to reach 10% of work.
The paper does not present this as a fixed 2060s prediction. Humanoid mass production, cheaper components, manufacturing improvements and rapid AI capability gains could make quality-adjusted costs fall faster.
In a fast scenario, robots become cost-competitive for half of physical work by 2050. The useful question is therefore not the date itself, but what compresses the cost and capability curves.
BANSEOG VIEW | The Physical AI market is built between capability and deployment
Looking only at 74% can overstate near-term automation. Looking only at today's 0.3% can understate the speed at which new deployment economics may emerge.
Banseog reads the gap as Capability → Environment Fit → Reliability → Integration → Economics → Deployment.
A robotics market expands when more tasks can be performed repeatedly in real facilities, integrated with existing systems and delivered at a total cost below the human alternative.
BANSEOG VIEW
Banseog View — Capability → Environment Fit → Reliability → Integration → Economics → Deployment
74% technical exposure and 0.3% cost-competitive work measure different stages rather than contradictory realities.
Real Physical AI adoption depends on environment fit, reliability, system integration, total economics and repeatable deployment.
As the market moves downstream, production engineering, reliability, field deployment, integration and fleet operations can become increasingly valuable capabilities.
SOURCES
Primary sources and references
- Anthropic — What work can robots do?
September 30, 2026. Confirms the ~900 occupations / ~19,000 tasks framework, 74% physical-task exposure, 34% of all working hours, 0.3% cost-competitive work, E0–E3 environment levels, capability and cost barriers, packaging and welding examples, and long-run cost scenarios.
The 74% figure is the share of physical tasks current robots can perform in some setting; 34% maps that exposure to all working hours. The 0.3% figure is the share of all work estimated to be cost-competitive at current robot costs. The study combines O*NET and BLS data with Claude-based task classification and cost estimation, so results depend on modeling assumptions and measurement error. Capability → Environment Fit → Reliability → Integration → Economics → Deployment is Banseog's analytical translation, not an official Anthropic framework.