Robots can technically do three quarters of America’s physical work today. They can do it more cheaply than a human in 0.3% of cases. The distance between those two numbers is where the real story of automation lives.
Last week Anthropic published a study asking a deceptively simple question: what work can robots actually do? Within days, the headline had been carved into carousels and reels. “Robots can perform 74% of physical tasks.” “81% of all work is now exposed.” Taxi drivers, tractor operators and recycling workers, we were told, are next.
The headline is accurate. It is also the least interesting thing in the report. The finding that should change how leaders think is the one almost nobody shared.
I. What Anthropic actually measured
The study, by Russell Legate-Yang and Maxim Massenkoff, asks one question of every physical task in the US economy: can a robot do this today, and where? It draws on O*NET’s 19,000 task descriptions across roughly 900 occupations, and rates 7,594 tasks classed as physical.
The clever part is the scale. Exposure is graded by how much the environment has to be rebuilt around the machine:
| Tier | Where a robot can do the task | Share of all US work time |
|---|---|---|
| E0 | Nowhere, not yet | 12% |
| E1 | Purpose-built robotic settings, such as an assembly line | 23% |
| E2 | Structured human workplaces, such as a warehouse | 10% |
| E3 | Unstructured, open environments, such as a city road | 1% |
The other 54% of work time is cognitive and interpersonal, so it sits outside the robot rubric altogether. Add up E1 to E3 and you get the headline: 74% of physical work, or 34% of all work, can be done by a robot somewhere.
Two details deserve more attention than they received. First, the ratings were produced by Claude, searching for real deployed or demonstrated robots and citing them task by task. An AI is now auditing the capabilities of its own physical cousins. Second, the authors back-tested the method to 1977. Jobs that were more exposed to the robots of their day saw wages and employment fall in the decades that followed. This is not a speculative index. It has a track record.
II. The number nobody shared: 0.3%
Robots are cheaper than people for just 0.3% of US job tasks today. That is the finding the carousels left out, and it reframes everything above it.
Anthropic estimated the full annual cost of a robot doing each task: hardware, integration, maintenance, energy, insurance and the human supervision that still sits behind most machines. It then compared that cost with what a worker produces in a year. The results are sobering for anyone predicting imminent displacement:
| Occupation | Share of tasks robots can do | Robot vs human cost |
|---|---|---|
| Packers and packagers | 97% | Robot about $45,000 a year vs $49,000 for a worker, so roughly $2,500 cheaper |
| Taxi drivers | Most of the job (E3) | Robotaxi about $7,000 a year dearer, plus regulatory hurdles |
| USPS mail carriers | 94% | $166,000 in robots to match one $82,000 worker |
| Welders | High | Around five times the cost of a human welder |
| Dishwashers, janitors and cleaners | Up to 100% | Several times dearer; robot cleaners can be up to ten times slower |
Packers are the one large occupation where the maths already works, and the labour market agrees: their employment has fallen 22% since 2015. Everywhere else, the robot can do the job and the spreadsheet says no.
Then comes the timeline. Robot prices have fallen roughly 3% a year since the 1990s. At that pace it takes about 40 years before robots are cost-competitive for even 10% of work, and until 2085 for half of today’s physical work. In the authors’ fast scenario, with costs falling up to four times faster and capabilities doubling their pace, half of physical work still waits until 2050.
So the honest headline is this. 74% is possible. 0.3% is profitable. The ratio between them is roughly 246 to 1, and that ratio, not the 74%, is the variable every leader should be watching.
III. Moravec’s Paradox, now with a price tag
In 1988 Hans Moravec observed that what is hard for humans is easy for machines, and the reverse. Chess was solved before folding laundry. Anthropic’s data puts a price on that paradox. Atoms are expensive and bits are nearly free.
A language model, once trained, can be copied to a million desks at close to zero marginal cost. A robot cannot. Every packer replaced needs steel, sensors, integration and a technician. That is why the same report finds about half of all work exposed to LLMs alone, rising to 81% once robots are added, yet the cognitive half is moving far faster than the physical half.
The two waves also strike different people. Robot-exposed workers are more likely to be male, less likely to hold a degree, and earn around $30 less an hour than unexposed workers. LLM exposure runs almost exactly the other way, towards the graduate, the analyst and the office. For the first time in industrial history, the white-collar worker is first in the queue.
My usual panel weighs in
The report itself convenes a useful panel, through the scholars it builds on.
- Rodney Brooks, the roboticist who has long argued that today’s humanoids will not learn human dexterity by watching video, would see his scepticism confirmed. Manipulation alone blocks large-scale automation for half of all physical tasks.
- Daron Acemoglu and Pascual Restrepo would point out that cost parity is not the finish line. Automation only lifts productivity when machines are much cheaper than people, not merely equal.
- Chad Jones and Christopher Tonetti would warn of weak links. The tasks robots cannot do become bottlenecks that cap the whole system’s gains, however brilliant the automated parts.
- David Autor would remind us that the question was never whether machines can do a task, but how much of the environment we are willing to rebuild for them. Anthropic’s E0 to E3 scale is essentially his idea of environmental control, made measurable.
The optimists have a counter. Each year, robots become able to do about 2% of the physical work they previously could not, and global humanoid shipments rose almost 300% year on year in the first half of 2026. Curves that look flat for decades can bend quickly. The panel’s caution is a statement about today’s cost structure, not a law of nature.
IV. Short-term turbulence, long-term abundance
The gap between can do and will do is not a reprieve. It is a runway. How leaders use it decides whether we arrive at abundance or at disruption without a plan.
For boards and executives. Stop asking which jobs AI can do and start asking where the cost curve crosses. Packers crossed this year. Taxi drivers are about $7,000 away, and regulation is the real gate. Map your own workforce the same way: capability, cost gap, regulatory gate, human preference. That is a far better planning tool than any exposure percentage.
For the white-collar workforce. The physical economy has a 40-year cushion. The cognitive economy has almost none, because software scales at zero marginal cost. The near-term turbulence will be felt in offices before factories. This is where the Long-AND, not Short-OR, mindset matters: the winning firms will redesign roles so that people and agents work together, rather than betting on one replacing the other.
For Asia Pacific. The study is about the US, and the region should read it carefully rather than copy it. Labour costs across our 14 markets vary by an order of magnitude, so the same robot can be cost-competitive in Singapore or Tokyo and uneconomic in Manila or Hanoi. The report itself notes that Japanese convenience stores already use robots to restock fridges, work that US stores still leave to people. Ageing, labour-scarce economies will cross the cost curve first, not because their robots are better, but because their people are scarcer.
For Singapore specifically. A small, high-wage, labour-constrained economy with world-class logistics is precisely where E2 automation, the structured warehouse and port, pays back earliest. The policy question is not whether to automate, but how to make sure the productivity dividend is shared rather than hoarded. That is a governance question as much as an economic one.
The least exposed work tells us where humans keep their edge: hands-on, face-to-face, and often regulated. Nurses, mechanics, carers and community workers. The report closes on an idea worth holding onto, that work itself may become more social and more interpersonal as machines take the rest. That is not a consolation prize. It may be the point.
V. The gap is the story
Viral headlines compress research into fear: 74%, 81%, your job is next. The research itself says something more useful. Capability is racing ahead; economics, regulation and human preference are setting the pace. The gap between them is where strategy, policy and governance actually happen.
We have been here before. The 1977 data shows robots reshaping work slowly, then decisively, for the jobs they could already touch. The difference this time is that the cognitive wave is arriving first and fast, while the physical wave gathers behind it. Leaders who read only the headline will either panic or dismiss it. Leaders who read the gap will plan.
Short-term turbulence. Long-term abundance. The distance between the two is a choice we are making now.
Sources
- Legate-Yang, R. and Massenkoff, M., What work can robots do?, Anthropic, 30 September 2026
- Massenkoff, M. and McCrory, P., Labor market impacts of AI: a new measure and early evidence, Anthropic, March 2026
Leave a comment