Economics What work can robots do? Sep 30, 2026 Download the PDF Key f

Economics What work can robots do? Sep 30, 2026 Download the PDF Key findings We present a robot exposure index based on how well robots can perform job tasks today. Robots, which we define as autonomous physical machines that sense and act, can perform three-quarters of physical tasks in the US, making up 34% of working hours, but mostly in limited settings. Workers exposed to robots are more likely to be male, less educated, and lower paid. For example, driving and warehouse jobs are highly exposed to currently available robots; nursing and general repair jobs are not, since present-day robots can do little of their work even in highly controlled environments. Overall, about 80% of job tasks by working time are exposed to either robots or LLMs. Robots do work where LLMs cannot. The remaining unexposed work is highly interpersonal or requires physical skills that robots today don’t have. While robots can do most physical work tasks today, they are much more expensive than human labor. Robots are cost-competitive for just 0.3% of job tasks. If robot price declines follow past trends, it will take 40 years for that share to reach 10%. Beyond price, factors including capabilities, preferences, and regulations pose further barriers to robot automation. Over the past 50 years, jobs more exposed to robots experienced greater declines in wages and employment than others. At the same time, job exposure has grown: each year, robots have become able to do about 2% of the physical work they previously couldn’t. Introduction Advances in large language models have raised the possibility of automating large swaths of work. But many jobs are physical. AI’s impact on the economy will in part depend on robotics.¹ Predicting the pace of robot advances is difficult, but enumerating capabilities today, we argue, can give insight into the coming years. We develop a measure of job exposure to robots, using Claude to assess how well present-day robots can perform work tasks. A job is more exposed when robots can do more of its tasks in less controlled environments. A robot is cost-competitive when it can do the same task for cheaper than a human worker. We find that robots can already perform 74% of physical tasks in the US, making up 34% of working hours. Robots and LLMs together expose all but one-fifth of employment. But we also find significant barriers to adoption: most robots require highly structured environments, and are cost-competitive with people for just 0.3% of work. If robot price declines follow past trends, it will take 40 years for that share to reach just 10%. After cost, the main barrier is capability, such as the dexterity needed to untangle wires. Human preferences and regulations further limit robot adoption for a significant share of tasks. Our core premise is that jobs are more likely to be impacted when robots can already do their work today. A backtest across 50 years validates this approach: from 1977 to today, jobs that were more exposed to existing robots experienced wage and employment declines in later decades. If the past is any guide, taxi drivers and warehouse packers will see changes sooner than nurses and mechanics. We expect that physical work will first be automated where robots have a foothold today. Robots today and tomorrow Most robots today operate in controlled environments like factories. Robots usually need to be programmed to interact with the physical world, whereas humans can adapt to their work environments. This has been a major hurdle for commercially viable robots.² But AI helps robots interpret and respond to their surroundings, allowing warehouse robots and autonomous vehicles to operate alongside humans.³ Many observers expect AI to improve robot capabilities quickly.⁴ Although spending on robots remains about 1% of total US equipment investment, business surveys suggest US robot adoption could nearly double within three years.⁵ And firms are investing billions to develop AI-powered robots that match human physical abilities.⁶ It’s difficult to predict exactly how these efforts will affect jobs and productivity. A ranking of occupations by current robot task coverage suggests where the impacts will appear first. Robots should affect jobs that they can already do before jobs they could do only with new technology and, in some cases, accommodating regulations. Current capabilities are also concrete and measurable, while forecasts of future capabilities must bet on which technologies will succeed.⁷ Robots still in development support our focus on current exposure: firms are testing humanoids in car factories and warehouses, structured environments where robots are already common.⁸ Measuring exposure Our analyses use data from O*NET, a database of around 900 occupations linked with descriptions of around 19,000 job tasks. We identify a set of physical job tasks that we think could not be automated without robots. To do so, we have Claude score task descriptions on a rubric measuring physical, cognitive, and interpersonal work requirements.⁹ The task “Dig trenches” is physical; so is “Teach dance students,” though this also requires cognitive and interpersonal skills. While typists use their hands to “Compute and verify totals on report forms, requisitions, or bills, using adding machine or calculator,” this task doesn’t count. Appendix A gives details, and Appendix F lists our prompts. Our task-level measure of robot exposure asks: can a robot today perform this task, and if so under what circumstances? We take robots to mean autonomous physical machines that sense and act, which includes car washes that scan cars to adjust sprayers and excludes teleoperated surgical machines fully controlled by a surgeon.¹⁰ We measure degrees of exposure by the kind of work environment a robot needs to perform a task.¹¹ Specifically, we sort tasks into four tiers of increasing exposure by where a robot can perform each task: E0: Robot cannot perform task. E1: Robot can perform task in a purpose-built robotic work environment, like a factory assembly line. E2: Robot can perform task in a structured human work facility, like a logistics warehouse. E3: Robot can perform task in an unstructured environment, like a city road. Figure 1 also displays this rubric. Figure 1: Task robot exposure rubric Job tasks identified as physical are rated on this robot exposure rubric. For example, the job task of unloading boxes from a truck at a warehouse is rated E2 if there’s a warehouse robot that can do that task. We think environmental control is a good measure of near-term automation risk. Because it’s hard for robots to adapt to unpredictable environments, most deployed robots work in engineered environments, like those that spray paint cars on assembly lines. While these difficulties are thought to have slowed physical automation in the past, AI-powered robots could better adapt to their environments.¹² To determine exposure, we instruct Claude to search for specific robots relevant to each task and assess their capabilities and operating environments, quoting sources directly. We ask whether a robot could perform a task in versions of that task’s typical work environment that are more or less structured. Getting robots to do seemingly simple tasks like loading a dishwasher requires many complex physical skills, so only demonstrated robot capabilities count.¹³ Robots must also do a task similarly well to humans, factoring in reliability, error rates, and speed. For instance, self-driving cars couldn’t “Drive taxicabs or privately owned vehicles to transport passengers” on real roads in the early 2010s. But they did drive around mock towns built for testing, a step toward today’s autonomous vehicles.¹⁴ Our rubric would rate driving passengers at exposure level E1 at that time (robot can perform task in a purpose-built robotic work environment) and E3 today (in an unstructured environment).¹⁵ This rubric requires many judgment calls. The O*NET task statements are often terse, and omit details that may be easy for humans but hard for robots.¹⁶ To describe work more concretely, we elicit detailed examples of how tasks are performed today, and how often these occur. Claude then scores exposure for these examples using web search to gauge robot capabilities. Cited sources must show robot deployments, commercial sales, or demonstrations, and results are similar if we omit ratings relying on demonstrations. Task exposure is set by majority rule: the least structured environment in which robots can do at least half of a task’s examples, weighted by time.¹⁷ Consider the task “Dig trenches.” We first ask Claude to provide examples describing how workers perform this task today and how often. For example, Claude estimates that 25% of the time, this task requires “cutting a linear trench in open ground.” Another 20% of the time, it requires “careful digging by hand around existing buried pipes, flowlines, cables and conduit.” These activit