Anthropic's robot jobs study says the bottleneck is cost, not capability: the window for skills is now

On 30 September 2026 Anthropic published research titled "Can we predict the jobs robots will do?", by Russell Legate-Yang and Maxim Massenkoff. The authors used Claude to classify around 19,000 job tasks across roughly 900 US occupations from the O*NET database, building a "robot exposure index" that asks whether a present-day robot can already perform each physical task, and in what kind of environment. Their headline finding is that robots can already perform 74% of physical tasks in the US, accounting for 34% of working hours. But robots are cost-competitive with people for just 0.3% of job tasks, and if robot prices fall at their historical rate of around 3% a year, it would take 40 years for that share to reach 10%.

The study separates two questions that are usually run together: what machines can do, and what it makes economic sense for them to do. Only 2% of physical tasks in unstructured settings such as city streets are within reach of today's robots. Taken together with large language models, around 80% of job tasks by working time are exposed to one technology or the other. The most exposed occupations are dominated by vehicle operators, with taxi drivers at the top of the index. Workers in highly exposed jobs are 55 percentage points less likely to hold a bachelor's degree and earn around $30 an hour less than unexposed workers. The authors caution that AI-powered robots could leapfrog today's constraints, and suggest watching occupations such as drivers and warehouse packers for early signs of disruption.

Dan Fitzpatrick, The AI Educator, says the figure that matters is not the 74% but the 55 percentage points. The people whose work is most exposed are the least likely to have been through higher education, and in practice they are also the least likely to be offered any AI education at all. Most AI literacy effort today is aimed at graduates, office workers and the professions. In his view the greater risk for schools, colleges and employers is not that they adopt AI too quickly, but that they read a 40-year horizon as permission to wait.

Fitzpatrick also points to the combined figure. If around 80% of working time is exposed to robots or language models, then the question for every leader is which parts of a role are the doing and which parts are the thinking. His standing advice is to outsource the doing, not the thinking: let the tools take the repeatable task, and invest in people's judgement, oversight and ability to check what a machine has produced. That applies as much to a warehouse as to an office.

For school and college leaders, the practical step is to treat vocational and technical pathways as the front line of AI literacy rather than an afterthought. Learners heading into logistics, driving, construction and care need to understand how automated systems are directed, supervised and corrected, because those are the parts of the job that stay human. For employers, Fitzpatrick suggests mapping tasks rather than job titles, starting with the roles this study names, and extending AI training to frontline staff on the same terms as head office. For policymakers, the authors' own suggestion stands: watch the exposed occupations now, before the pressure arrives.

Fitzpatrick works with schools, colleges, employers and government bodies on adopting AI and building AI literacy, and writes regularly on what AI means for education and the future of work.

Dan Fitzpatrick is available for interview and comment on this story.