No industry has been shown to be safe from AI, but some kinds of work overlap with it far less than others. In a Microsoft Research analysis of about 200,000 Copilot chats, U.S. healthcare support, farming, construction, and building maintenance occupations had the lowest AI applicability scores, while computer and mathematical work and sales had the highest. The score measures how much of an occupation's tasks AI is already used for, so it describes task overlap and does not predict who keeps or loses a job.
Key takeaways
- The Microsoft study scored 22 U.S. occupation groups; healthcare support (0.05), farming (0.06), construction (0.07), and building and grounds maintenance (0.08) scored lowest.
- Computer and mathematical work and sales scored highest, both at 0.29.
- The Bureau of Labor Statistics projects healthcare support jobs to grow 13.3 percent from 2025 to 2035, and office and administrative support jobs to fall 4.0 percent.
- Low overlap and job growth do not line up cleanly: computer and mathematical work scores high and is also projected to grow 7.3 percent.
- Both sources describe tasks and demand, and neither says that people will or will not keep control of an industry.
What the Copilot study measured
Researchers at Microsoft (Kiran Tomlinson, Sonia Jaffe, Will Wang, Scott Counts, and Siddharth Suri) published "Working with AI: Measuring the Applicability of Generative AI to Occupations" on arXiv in July 2025, with the latest revision dated December 22, 2025. They analyzed anonymized U.S. conversations with Bing Copilot from January 1 to September 30, 2024, using two samples of roughly 100,000 each. Each conversation was matched to the work activities in the O*NET occupation database, and each occupation received a score combining how often its activities appear in chats, how often the AI completed the task, and how large its contribution was.
The authors list limits. The data cover one chatbot and no occupation-specific tools, the work activities in O*NET may differ from how a job is really done, and task completion does not capture productivity. They also note that effects on employment and wages depend on business decisions that are hard to predict.
Where the scores fall
The table lists the paper's applicability score for selected occupation groups next to the Bureau of Labor Statistics projection for the same group, published on August 27, 2026 for 2025 to 2035. The pairing is an illustrative comparison assembled for this article. Neither source computed it, and the two classify jobs by similar but separately maintained occupation groups.
| Occupation group | AI applicability score (2024 chats) | BLS projected change, 2025 to 2035 |
|---|---|---|
| Computer and mathematical | 0.29 | 7.3 percent |
| Sales and related | 0.29 | -1.4 percent |
| Office and administrative support | 0.26 | -4.0 percent |
| Healthcare practitioners and technical | 0.12 | 8.0 percent |
| Production | 0.11 | -0.4 percent |
| Healthcare support | 0.05 | 13.3 percent |
Within the lowest-scoring individual occupations, the paper's list includes phlebotomists, dishwashers, roofers, orderlies, and massage therapists. What they share is physical contact with people or materials, which a text assistant cannot supply.
What the Bureau of Labor Statistics adds
The BLS Employment Projections release for 2025 to 2035 expects total employment to rise 3.5 percent, from 170.3 million to 176.2 million jobs. Healthcare support is the fastest-growing group at 13.3 percent, followed by healthcare practitioners and technical occupations at 8.0 percent. Private healthcare and social assistance is projected to add over 2.2 million jobs, about 37 percent of all new jobs.
BLS attributes part of the decline in office and administrative support to automation tools, including AI-powered ones, and part of the sales decline to AI in sales processes. It also expects generative AI to limit some jobs in arts, design, entertainment, sports, and media. At the same time, it ties growth in computer occupations and in professional, scientific, and technical services to AI demand. Demand drivers other than AI also shape these projections, so the healthcare numbers cannot be credited to low AI overlap alone.
What these results do not show
The phrase "human domination" implies a contest, and the data do not measure one. A low score for roofers says that few chat requests resemble roofing tasks. It does not say roofers are insulated from other forces, such as swings in construction demand, and it does not say that high-scoring work is disappearing: BLS projects computer and mathematical employment to grow. The Microsoft authors themselves frame the score as a signal of where AI may change work, and say that new occupations may also emerge.
The figures are also American. Canadian occupation classifications differ, and this article does not apply the U.S. scores to Canadian jobs.
Reading claims about safe industries
- Does the claim name a measure, such as task overlap, or only assert safety?
- Is the data about a specific chatbot or about all AI systems?
- Does it separate task change from job loss?
- Are projections dated, and from an official agency?
- Does the argument account for other drivers of demand for the work?
FAQ
Which occupation group scored lowest for AI applicability?
Healthcare support, at 0.05, followed by farming, fishing, and forestry at 0.06 and construction and extraction at 0.07.
Does a low score mean jobs are protected?
No. The score counts how often AI is used for an occupation's activities in Copilot chats. It says nothing about wages, hiring, or other pressures on the occupation.
Is healthcare growing because AI is weak there?
The sources do not establish that. BLS projects healthcare growth, and the Microsoft study finds low overlap for healthcare support, but no source links the two as cause and effect.
Which occupation groups scored highest?
Computer and mathematical, and sales and related, both at 0.29, followed by office and administrative support at 0.26.
Related reading
- AI across U.S. industries
- AI on ships and in maritime work
- AI in the real estate industry
- AI tools used across industries
More in AI Tools and Industry.
Last verified: October 2026
Sources
- Tomlinson K, Jaffe S, Wang W, Counts S, Suri S. Working with AI: Measuring the Applicability of Generative AI to Occupations, arXiv 2507.07935, Microsoft Research, 2025 (version 6, December 22, 2025)
- Employment Projections: 2025-2035 Summary, U.S. Bureau of Labor Statistics, August 27, 2026