At a glance
- Source
- Tomlinson, Jaffe, Wang, Counts and Suri (2025). Working with AI: Measuring the Applicability of Generative AI to Occupations. arXiv preprint 2507.07935, Microsoft Research (version 6, 22 December 2025). doi:10.48550/arXiv.2507.07935 · Not peer-reviewed: other researchers have not yet checked it for a journal.
- How it was done
- A language model labelled each conversation with the work activities the user wanted help with and the work activities the AI itself performed. The researchers added how often the AI completed the task and how much of each activity it covered, and turned this into an “AI applicability score” for each occupation. Data analysis. It builds a measure from existing data, such as job and task records.
- Who took part
- About 200,000 anonymized conversations with Bing Copilot from January to September 2024: 100,000 sampled at random, and 100,000 that users had rated with a thumbs up or down.
- Where
- United States. Conversations with Bing Copilot, Microsoft’s free AI chat assistant, matched to the work activities of US occupations in the O*NET database.
“Our measures of AI applicability indicate where AI is likely to change work, which is a part of the automation vs. augmentation question, but AI’s effects on employment and wages will depend on hard-to-predict business decisions.”
What it found
- People ask AI to find, write and explain information. The most common requests were gathering information, writing and communicating. The AI itself mostly provided information, explained, taught and advised.
- What the user wanted and what the AI did often differ. In 40% of conversations, the work activities the user wanted help with and the work activities the AI performed did not overlap at all.
- Knowledge work scores highest. Interpreters and translators scored 0.492, historians 0.462 and writers and authors 0.454. Sales representatives of services and customer service representatives were also in the top ten. Most occupations had some overlap.
- Physical work scores lowest. Dredge operators, bridge and lock tenders, and water and wastewater treatment plant operators scored 0.
- A weak link with pay. The score rose only slightly with wages (employment-weighted r = 0.13). Earlier predictions had found a strong link.
- Close to an earlier prediction. The score lined up with an earlier prediction, by Eloundou and others, of the share of each job’s tasks that language models could speed up (r = 0.73).
The paper lists the 40 occupations with the highest scores and the 40 with the lowest. Those lists were widely reported as the jobs AI will replace. The authors reject that reading. In the paper they write that concluding that high-scoring occupations “will be automated and thus experience job or wage loss” “would be a mistake”. They give the example of ATMs: the machines took over a core task of bank tellers, yet the number of teller jobs rose as banks opened more branches. See the claim check on the “40 jobs”.
When the result holds
The score ranks occupations against each other. The authors write that “relative comparisons are more meaningful than absolute score values”, because an activity counts only if it reaches a usage threshold they chose (0.05% of conversations).
The ranking holds for one product in one period: Bing Copilot, a free consumer assistant, in the United States, in January–September 2024. It describes overlap with what people asked for. It does not measure productivity, hiring or job losses.
The scores changed between versions. The first version (10 July 2025) was titled “Measuring the Occupational Implications of Generative AI”. The sixth version (22 December 2025) gives less weight to physical tasks. Eight of the first version’s top 40 occupations dropped out of it, and passenger attendants fell from 3rd to 12th. Cite the version you use.
Limits
- The researchers work for Microsoft, the data come from Microsoft’s own product, and Microsoft sells Copilot. A finding that AI chat is useful across most jobs supports that business.
- A preprint on arXiv. It has not been through peer review.
- The first version notes that it is “very difficult (or impossible)” to tell which conversations were for work and which for leisure.
- O*NET lists the activities of a typical US job. Your staff may do different tasks, and the work between tasks is not in the list.
- Other AI tools, used by other people, may show a different pattern.
Practices that use this study
This study supports:
Read next
- “Microsoft study reveals the 40 jobs most at risk from AI”The list exists: a July 2025 paper by five Microsoft researchers ranked 40 US occupations highest on an “AI applicability score”. The score measures how far an occupation’s work activities overlap with what people asked Bing Copilot to do, and how well Copilot did it. It is neither a count nor a forecast of jobs lost. The paper says reading it that way “would be a mistake”, and the authors wrote in August 2025 that their data “do not indicate” that jobs will be replaced. The list itself changed in the December 2025 version.
- Fewer young people hired into AI-exposed jobs: US payroll dataEmployment of 22–25-year-olds in AI-exposed occupations is 19% below where it would be had it kept pace with less exposed peers. Experienced workers show no such gap. The gap comes mainly from fewer young people being hired.
- AI exposure and employment in 23 countriesThere was no clear link between how exposed an occupation was to AI and how its employment changed. In occupations with heavy computer use, more exposure went with faster employment growth.
Claim checks that cite this study
- “Microsoft study reveals the 40 jobs most at risk from AI” Needs correcting