Robots ยท a separate score
What robots
can do today.
Our AI score is about software: chatbots and tools that read, write and reason. Robots are a different question, so they get their own score. A job can be safe from one and exposed to the other.
Where the robot score comes from
It comes from What work can robots do?, published by Anthropic in September 2026. The study went through 18,796 job tasks. For each physical task it looked for a real robot that has done that work, and recorded 56,933 reports as evidence: 20,877 of robots in real use, 25,743 of robots on sale and 10,313 of demonstrations.
Four levels
The question for each physical task is where a robot would have to work to do it.
In a purpose-built robot cell (3,967 tasks). A space built around the machine, with parts held in known positions and people kept out. A factory line or a paint booth.
In an ordinary workplace (1,103 tasks). A place built for people and kept predictable. A hospital pharmacy, a hotel corridor, a warehouse.
Anywhere (105 tasks). A setting taken as found. A private home, a building site, a pavement.
Hands only (2,419 tasks). Physical work that no robot in the evidence could finish, even in a space built for it.
Most robot-ready tasks sit at the first level. That means the work can be done by a robot only after someone builds the setting for it, which costs money and time. Very few tasks can be done by a robot anywhere.
What the number on a job page means
It is the share of the job's working time spent on physical tasks a robot has been shown to do, at each level. It says what is technically possible. It does not say an employer will buy the robot, or when.
What a robot costs, and when it could cost less than a person
Being able to do a task is only half the question. The other half is price. For each job we compare what an employer pays for an hour of a person's work with what a robot costs for the same hour.
The person. Average US pay for the job, divided by 2,080 working hours a year, plus benefits. Wages are 70% of what an employer pays, so we divide by 0.70.
Fixed robot, in a space built for it: about $18 for an hour of a person's work. The price is a $90,000 industrial robot, times 3.5 for the finished station around it (guards, tools, programming and fitting). Of the 397 jobs where at least a fifth of the time is this kind of work, the robot is already cheaper on paper in all of them. For this kind of work, price is not what holds robots back. Building the space around the robot is.
Mobile robot, in an ordinary workplace: about $40 for an hour of a person's work. The figure comes from a company filing by the maker of one warehouse humanoid, as reported by Fortune. Two robots per person is a bank's estimate. Of the 142 jobs where at least a fifth of the time is this kind of work, the robot is cheaper on paper today in 49, would be by 2030 in 70 more, by 2035 in 23 more, and later in 0.
Anywhere. No robot on sale can work in homes, streets or building sites as it finds them, so there is no price to compare.
The forecast. We assume robot prices fall 10% a year, which is what the published forecast below works out to. Each job page shows the full bill: the price to buy the robot, the cost to run it, and how long it takes to pay for itself. You can change the year and the number of shifts.
What this leaves out. The cost of changing the workplace, breakdowns, the people who supervise the robots, and whether the robot is good enough. We also use one price for each level, when real machines differ: a self-driving truck is not priced like a floor-cleaning robot. A robot that is cheaper on paper may still not be bought. Treat the dates as the earliest that price stops being the obstacle, not as a date your job changes.
Sources for the prices.
Fortune, 16 Oct 2026. Reports Agility Robotics' company filing: its next robot costs about $400,000 over five years, which is about $20 an hour on two shifts. JPMorgan estimates two humanoids are needed to match one worker. The $90,000 industrial robot is the article's own example. A Roland Berger study puts running costs at about $2 an hour.
NIST report GCR 16-005, August 2016. A US government study, from interviews with industry experts, found a finished robot system costs 3 to 4 times the price of the robot itself. We use 3.5. This figure is ten years old.
US Bureau of Labor Statistics, June 2026 figures. Wages and salaries were 70.0% of what private employers spent on workers ($32.82 of $46.89 an hour). Benefits were the other 30%.
Bain & Company forecast, Nov 2025 (as reported by Yicai Global). Humanoid robot production costs are forecast to fall about 70% by 2035. That works out at about 11% a year; we use 10%. We could not find this figure on Bain's own site, so it rests on the news report.
Bain & Company, April 2025. Humanoid robot unit costs fell by at least 40% between 2022 and 2024, which is faster than our forecast setting.
Robotics and Automation News, 31 Mar 2026. Shows how wide humanoid prices are: from about $16,000 for basic models to over $250,000 for industrial ones. We use the industrial end, because that is what is working in warehouses today.
Who is building these robots
The developers cited most often in the evidence: Boston Dynamics, Diligent Robotics, ANYbotics, Amazon Robotics, Vitestro, Chef Robotics, GrayMatter Robotics, University of Liverpool (Cooper group), Built Robotics, Simbe Robotics, Figure AI, FANUC America.
Jobs where robots can do the most in an ordinary workplace
Jobs with at least 20,000 US workers, by share of working time.
Jobs with the most hands-only work
Physical work no robot has been shown to do.
Limits
The ratings, and the estimates of how much time each task takes, were made by an AI model (Claude) reading published sources. They were not made by people watching the work. A demonstration is not the same as a robot you can buy, so each job page says whether the evidence is in real use, on sale or a demonstration only. The data is used under the CC BY 4.0 licence and includes information from O*NET.