The Jobs Safest From AI, According to Microsoft and Anthropic Research

Lists of jobs that artificial intelligence (AI) supposedly can never replace are everywhere, but many rest on opinion. Two studies from Microsoft and Anthropic offer a more measurable picture because both draw on how people actually use AI.

Their results point the same way: work that depends on physical skill, being on site and face-to-face contact is the least touched by AI. Neither research team claims any job is safe forever, however, and those limits matter before anyone treats the list as career advice.

The Jobs With the Lowest AI Scores in the Microsoft Study

The paper, Working with AI, was written by Kiran Tomlinson, Sonia Jaffe, Will Wang, Scott Counts and Siddharth Suri of Microsoft. It first appeared in July 2025, and its latest revision is dated October 23, 2026.

The team analyzed 200,000 anonymized conversations from US Copilot users between January and September 2024. Conversations were matched to work activities in O*NET, the US government's occupational database, and summarized into an AI applicability score for each occupation.

The score captures whether AI is used often enough, and successfully enough, on activities relevant to a job. The lower the score, the fewer of that job's tasks show up in AI use. The table below shows a selection from the 40 lowest-scoring occupations.

Source: Tomlinson et al., Working with AI, table of the 40 lowest-scoring occupations. For comparison, the highest score belongs to interpreters and translators at 0.492.

Three Patterns Behind the Jobs AI Barely Touches

The first pattern is work whose core task is moving things in the physical world. Roofers, concrete finishers and heavy equipment operators fall into this group, along with highway maintenance workers and oil field roustabouts.

The second is work that involves direct contact with the human body. Phlebotomists, massage therapists, surgical assistants and oral surgeons work with their hands and deal directly with patients or clients.

The third is work that requires someone on site to keep a system running. Water plant operators, bridge tenders and ship engineers belong here. A chatbot can help look up information, but it cannot show up on location and do the job.

Anthropic's Finding: 30 Percent of Workers Have Zero Coverage

Anthropic reached a similar conclusion by a different route. A paper by economists Maxim Massenkoff and Peter McCrory, published on March 5, 2026, introduced a measure called observed exposure. It asks how many of the tasks that language models could theoretically speed up are actually being automated in professional settings.

The result: 30% of workers have zero coverage, because their tasks appeared too rarely in Claude usage data. The examples named are cooks, motorcycle mechanics, lifeguards, bartenders, dishwashers and dressing room attendants.

At the other end, computer programmers have the highest coverage at 75%, and data entry keyers stand at 67%. Dishwashers appear in both studies as a job with very low AI exposure.

Why a Low Score Does Not Mean Immunity

Microsoft's team says plainly that it would be a mistake to read a high score as a sign of job or wage loss. The authors write that AI's effects on employment and wages will depend on hard-to-predict business decisions and the emergence of new types of work. The same logic applies in reverse, so a low score is no guarantee of safety either.

There are other limits. Microsoft's data comes from 2024, and the authors acknowledge that models have improved dramatically since then. Its scores reflect one product, and comparisons between jobs mean more than the absolute numbers. Anthropic's data likewise comes only from Claude usage.

Both studies mainly look at generative AI such as chatbots, not robots or automated machinery. A forklift or packaging machine operator with a low score here could still be affected by warehouse and factory automation, which neither study measures. Readers should also keep in mind that Microsoft and Anthropic both sell AI products and used data from their own services.

All of these figures apply to the United States. Neither study provides scores for other labor markets such as Indonesia, where the mix of jobs and the pace of technology adoption differ.

What Is Already Happening in the Job Market

So far, Anthropic has found no systematic rise in unemployment among highly exposed workers since late 2022. The one signal is among young workers: the job finding rate for 22- to 25-year-olds in the most exposed occupations fell about 14% compared with 2022. The authors call this suggestive, and the figure only just reaches statistical significance.

Higher pay does not automatically mean more protection. Anthropic found the most exposed group earns 47% more on average than the unexposed group, and is more likely to hold a graduate degree. Microsoft found only a weak link between its score and wages, with a correlation of 0.13.

The World Economic Forum's Future of Jobs Report 2025 points in a similar direction. The survey of more than 1,000 employers projects 170 million new jobs and 92 million displaced jobs by 2030. The largest growth in absolute numbers is in frontline roles such as farmworkers, delivery drivers and construction workers.

Conclusion

Both studies show the same thing: generative AI today mostly touches work done at a screen, not work done in the field. That picture may change as AI and robotics develop, so the list is better read as a snapshot of today than a guarantee for the future.