The 20 most AI-exposed occupations
These are the 20 highest-scoring occupations of the 785 measured here, ranked by AI applicability: the overlap between the activities that make up the job and what an AI assistant was observed doing, and completing, in real use. Scores come from the 2026-07 data build. A high score describes overlap, not a probability that the job disappears.
Score = AI applicability (Microsoft Research, CC BY 4.0), data vintage 2026-07. Percentile = rank among all 785 occupations. Scores are shown to three decimals; the ordering uses full precision. Every occupation links to its task-level breakdown, and every field links to that field's own ranking.
What the score is, and what it is not
The number is a measure of overlap, not of risk. Microsoft Research classified 200,000 anonymized conversations with a consumer AI assistant against the O*NET work-activity taxonomy, then measured how often each activity appeared, how successfully the assistant completed it, and how much of the activity it covered. An occupation's score is built from the activities that fill its week. Nothing in the source data contains a time horizon, an employment model or a prediction, and the paper's own authors warn against reading applicability as a probability of job loss. The methodology page traces every step from their files to this table, and exposure, automation and replacement sets out how far the measure can be pushed.
What the ranking is good for is placement. All 20 occupations on this list sit in the High band, the top fifth of the distribution. The distance from first to 20th is 0.492 against 0.355, so the top of this distribution is a slope rather than a cliff. The next occupation down, Web Developers at 0.353, sits 0.0021 below the 20th. A place on this list and a place just off it are the same finding.
Which fields the top 20 come from
Arts, Design, Media & Entertainment supplies 7 of the 20. The other 13 are spread across 9 more fields, out of 22 in total.
| Field | In this top 20 | In the High band | Occupations in field | Field median |
|---|---|---|---|---|
| Arts, Design, Media & Entertainment | 7 | 9 | 36 | Elevated |
| Computer & Mathematical | 2 | 18 | 21 | High |
| Life, Physical & Social Science | 2 | 13 | 47 | Elevated |
| Office & Administrative Support | 2 | 22 | 50 | Elevated |
| Sales | 2 | 13 | 21 | High |
| Business & Financial Operations | 1 | 10 | 32 | Elevated |
| Education & Library | 1 | 39 | 60 | High |
| Personal Care & Service | 1 | 7 | 29 | Moderate |
| Production | 1 | 2 | 100 | Low |
| Transportation & Material Moving | 1 | 1 | 46 | Low |
The last column is the one worth reading twice: of the 10 fields listed, 7 have a median occupation below the High band. That is what it looks like when exposure follows the work rather than the industry, because a field can put an occupation at the very top of this ranking and still sit low overall. The High band holds 157 occupations, one fifth of the 785.
The High band contains nothing at all from 6 of the 22 fields: Legal; Healthcare Support; Cleaning & Grounds Maintenance; Farming, Fishing & Forestry; Construction & Extraction; and Installation, Maintenance & Repair.
What a place on this list does not mean
It does not schedule anything. The measure has no date in it. Whether overlap turns into redesign, wage pressure or displacement is decided by employers, regulators and customers, and none of those are in the dataset.
It is one assistant, over one window. The conversations were United States Bing Copilot sessions from January to September 2024, published in 2025. The study's own repository documentation names what that leaves out: programmers work inside AI-enhanced development environments, and legal, medical and financial work requires compliant AI tools. None of that traffic reaches a consumer chat log, so software and regulated professional work are likely to be further along than these numbers show. Hands-on and licensed jobs quotes that passage in full and cites it.
It says nothing about how many people hold the job. No wage or headcount weighting is applied anywhere on this site, so an occupation held by a few thousand people and one held by millions occupy the same kind of row.
The useful part sits under the row, not in it. Each occupation page breaks the score back into the individual work activities that produced it, with the measured value of each. That is where a plan comes from: which activities to hand over, which to keep. What to do if your exposure score is high covers the practical side.
If the answer is work of a different shape
Nothing on this page argues that a high score means leaving. Most of the useful response happens inside the job: hand the exposed activities to the tools, take ownership of the review, and make the shift visible to the people who decide. That is what the occupation pages and the guides are for, and it is where the evidence is strongest. Two other responses sit outside the current seat, and neither is covered anywhere else on this site. Both of them start with something that costs nothing.
Selling the work instead of the seat
An occupation ranks high here because the activities that fill its week overlap with what an assistant was measured doing. Those same activities are still worth money to somebody who does not employ you, and selling them directly is a response to a high score rather than a purchase. Start with the people who already know your work: former employers, colleagues, the clients and suppliers you dealt with. Direct work pays the full rate, because nothing sits in the middle taking a share. A marketplace is what you use when there is no such list to work from, or when you want the first few jobs to arrive without a pitch.
- Work out which of your activities are the sellable ones, from your own week Free
- List a service on Fiverr Partner link
Checked 30 August 2026: creating a seller profile and listing a service cost nothing, so this is somewhere to start rather than something to buy. Fiverr takes its share out of completed orders, and the current rate sits in its own payment terms, which is worth reading before you set a price. Quote the direct work first. Partner terms checked in August 2026. The link earns this site a commission at no extra cost to you, and it changes no score, ranking or plan here.
Looking for a different seat
The other response is a move, and the first question is where to look rather than what to pay. Three boards cover most of the market and cost nothing to search: LinkedIn and Indeed for the general market, We Work Remotely for remote-only roles. Work through those before considering a paid one. What a subscription board sells is filtering rather than access, which is worth something if bad listings are costing you hours, and worth nothing if they are not.
- LinkedIn Jobs Free
- Indeed Free
- We Work Remotely Free
- FlexJobs, a paid board that screens its listings Partner link
FlexJobs is a subscription rather than a free board: it bills on a recurring cycle after a short trial, and the current price is shown in your own currency on its own page. The part worth checking before paying is geography. On 30 August 2026 its remote jobs index listed 126 of 135 roles as United States only, against 9 open to countries outside it. Partner terms checked in August 2026. The three free boards above are listed first because for most readers they are the answer. The link earns this site a commission if you subscribe, at no extra cost to you, and it changes no score, ranking or plan here.
If your own job is not here
Most are not. This is 20 of 785. Search any occupation from the box in the header, open the full ranking to filter by field and band, or work from your actual week instead of a job title with the task self-assessment. The other end of the same distribution is on the 20 least exposed occupations.
- Occupation scores: the AI applicability values published by Microsoft Research in Working with AI (2025), released under CC BY 4.0 at github.com/microsoft/working-with-ai and used as published. Data vintage 2026-07.
- Occupation titles and field groupings: the O*NET 30.0 Database (2025), U.S. Department of Labor, Employment and Training Administration, used under CC BY 4.0. O*NET® is a trademark of USDOL/ETA.
- Rankings, band counts, field composition and medians on this page were computed from that score set, not quoted from anywhere. Band cut-offs, the percentile construction and every editorial layer this site adds are documented on the methodology page.
Figures on this page were computed on 2026-08-30 from the same 785-occupation score set the rest of the site uses. If the data is rebuilt, this page is rebuilt with it.