Robot jobs are often discussed as though the only question is whether a machine can perform a task. A new Anthropic analysis makes the picture much more awkward. The researchers estimate that existing robots can perform 74% of physical tasks in the United States in at least some setting, representing about 34% of all working hours, yet they calculate that robots are currently cheaper than human labour for only 0.3% of job tasks.
That gap between capability and economics is the most useful part of the study. A robot can technically do something without being a sensible replacement for a worker. Hardware costs, installation, maintenance, structured workspaces, safety rules and the awkwardness of real environments all sit between a laboratory demonstration and widespread automation.
Robot jobs depend on more than whether a machine can do the task
Anthropic’s exposure measure asks how much of an occupation’s work present-day robots can perform and how controlled the environment must be. A machine that succeeds only in a specially designed cell is treated differently from one that can operate in an ordinary workplace. That distinction prevents a familiar error in automation debates: confusing a demonstration with a deployable job replacement.
The researchers also separate physical tasks from cognitive and interpersonal ones. That matters because a large language model and a robot expose different parts of the economy. Software can already affect writing, analysis and administration, while physical machines are relevant to driving, handling goods, cleaning, production and other work that requires action in the real world.
The 74% figure sounds more dramatic than the economics
The headline capability estimate is large because it counts work a robot can perform under some circumstances. Many industrial machines are excellent when the environment is predictable. They can repeat a movement precisely, transport goods along known routes or carry out narrow manufacturing operations for long periods without tiring.
Humans, however, are unusually good at cheap general-purpose adaptation. A worker can step around a box, recognise that an object has been moved, use a slightly different tool, answer a question and continue after a small surprise. Reproducing that flexibility with sensors, software, grippers, safety systems and maintenance can be expensive even when the individual task looks simple.
Cost is the biggest brake on robot jobs today
Anthropic estimates that robots are cost-competitive for only 0.3% of job tasks at present. Its calculation attempts to compare the annual cost of robotic deployment with the labour cost of producing the same output. The authors stress that these are estimates, not a forecast that every task below a price threshold will immediately be automated.
The study’s historical scenario is also sobering for anyone expecting instant mass replacement. If robot costs fell at roughly 3% a year, the authors estimate it would take about 40 years for robots to become cost-competitive for 10% of human work. New designs, cheaper manufacturing or faster capability gains could change that path, but the exercise shows how much the economics matter.
Warehouses and driving look more exposed than nursing or repair
The occupations closest to change are not simply the most physical. They are jobs where machines already have a foothold and the environment can be structured. Driving is highly exposed because autonomous vehicles already perform the core activity in some conditions. Warehouse packing is another obvious case because goods can be standardised, workstations redesigned and repetitive movement measured.
The public labour data help put that example in context. O*NET reports 555,500 hand packers and packagers employed in the United States in 2025, with a median annual wage of $36,280 and projected employment decline over the following decade. That does not prove robots will cause the decline, but it shows why employers have a strong incentive to investigate automation where labour is repetitive and margins are tight.
The hardest physical work is often the least structured
Some jobs look manual yet remain difficult to automate because the world refuses to stay still. General repair work requires diagnosing unusual problems, reaching awkward places and improvising with tools. Nursing combines physical movement with judgement, communication, reassurance and responsibility. Those are not merely separate tasks that can be automated one by one without changing the job.
The study also identifies tasks held back by regulation and human preference. People may accept a machine moving boxes while resisting one that dresses a child or handles an intimate care task. Safety rules can likewise prevent a technically capable robot from taking over work where mistakes carry severe consequences. Automation therefore depends on social permission as well as engineering.
Better robot learning could change the cost equation
Capabilities are moving. LiveAIWire has covered robots learning tasks from ordinary videos and orchard robots designed to pollinate and pick fruit, both examples of researchers trying to reduce the amount of specialised programming and environmental preparation required. If a robot can learn a broader range of behaviours, the same hardware becomes useful for more work.
That is important because a machine used for one narrow task has to justify its cost from that task alone. A more general robot can spread the purchase price across many jobs. The economics of a warehouse robot change if it can unload, sort, pack and clean rather than performing only one motion at one station.
Exposure is not the same as unemployment
A task-level study cannot tell us how many people will lose jobs. Occupations are bundles of activities, and technology often changes that bundle before it removes the role. Workers may supervise equipment, handle exceptions, communicate with customers or take on tasks that were previously uneconomic. Companies can also use automation to increase output rather than simply reduce headcount.
LiveAIWire has also examined evidence that the mere presence of robots can change human performance, a reminder that work design matters alongside replacement. A robot introduced into a workplace changes pacing, responsibilities, training and collaboration even when the number of employees stays the same.
The study itself has an important limitation
Anthropic used Claude as part of the process for rating tasks and estimating robotic costs, drawing on cited examples and public information. That gives the analysis breadth, but it also means the numbers should not be treated as an audited census of every robot installation in America. Cost estimates can vary sharply by workplace, volume, maintenance needs and local wages.
The authors themselves present the measure as a way to understand exposure and identify where change may arrive first. It is most useful as a map of pressure points, not a timetable. A figure such as 74% tells us that physical automation has a surprisingly broad technical footprint. The 0.3% figure tells us why most workers do not yet see that footprint turning into an immediate economic substitute.
The real robot-jobs question is when capability becomes cheap enough
That is the point worth carrying into future discussions of humanoids and intelligent machines. Impressive movement is only one part of adoption. A robot must be reliable, safe, maintainable, flexible and cheap enough to beat the human alternative after all the hidden costs are counted. In many workplaces, that remains a high bar.
The study therefore gives both automation enthusiasts and sceptics something uncomfortable. Robots can already do far more kinds of physical work than many people realise, but the business case for replacing human labour is still narrow. The next phase of robot jobs will be decided where those two facts meet: the places where capability is already real and cost begins to catch up.
Businesses can redesign a workplace around a robot
The economics also change when a company is willing to redesign the environment. Warehouses can use standard containers, marked routes and fixed picking stations. Factories can present parts in predictable orientations. Those changes make robotic perception and manipulation easier, but they are investments in their own right and may only make sense at sufficient scale.
This helps explain why the same task can be automatable in one workplace and uneconomic in another. A high-volume site can spread integration costs across millions of repetitions. A small business handling constantly changing objects may find that a human worker remains cheaper and more flexible even when a robot has demonstrated the underlying movement elsewhere.
There is also a financing effect. Robots often require capital up front while labour is paid over time. Interest rates, expected utilisation, tax treatment and the risk that hardware becomes obsolete can therefore alter the decision even when a simple annual cost comparison looks favourable. Two companies doing identical work may reach different automation decisions because their scale and cost of capital are different.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.