AI and your job

Most lists of “jobs AI will replace” mix a few anecdotes with a lot of anxiety. This page does something narrower and more useful: it shows how much the tasks of each occupation overlap with what current AI systems can already do, using research anyone can check. The book this site is built around, Living in Exponential Times, insists on one distinction that the headlines skip: being exposed to a technology is not the same as being replaced by it.

Spreadsheets exposed accountants in the 1980s, and the profession grew. Exposure tells you where the pressure will be felt first. What that pressure turns into — fewer jobs, different jobs, or better-paid jobs — depends on decisions made by people, firms and institutions. That is the part you can still influence.

The exposure index is the AI Occupational Exposure (AIOE) score developed by Felten, Raj and Seamans (2021), expressed here as a percentile from 0 to 100: it links the abilities each occupation requires to the abilities where AI has made measurable progress. A 99 means the occupation sits at the very top of that overlap. The second column comes from “GPTs are GPTs” by Eloundou and colleagues (2024): the share of an occupation's typical tasks that a large language model, or software built on one, could complete at least twice as fast at the same quality.

The two measures do not always agree, and the disagreement is informative. Translators score 79 on the index but have 84 % of their tasks exposed, the highest share in the table: the work is almost entirely text. Teachers score 82 on the index yet only 34 % of their tasks are exposed, because so much of teaching happens in a room, with people. Accountants and psychologists both reach 99 on the index; drivers sit at 14, waiting staff at 22. The figures are occupation-level averages from US data, not a reading of your exact job, and they describe what is technically possible, not what employers will choose to do.

Where exposure is high, the first thing to go is rarely the job. It is the routine middle of the work: drafting the first version, summarising the file, reconciling the accounts, answering the question that has been answered a thousand times. Customer-service agents (70 % of tasks exposed) and sales representatives (71 %) feel this early, because their days contain many repeatable exchanges. Programmers and data scientists see it as a change of pace: code and analyses that took a morning now take an hour, and the bar for what counts as a good day's output moves.

The economic effect follows the task, not the title. When a task becomes cheap, the people who did it either move up to the decision the task was feeding, or compete with a machine on price. The difference between those two outcomes is mostly skill and timing.

Across the table, the tasks that resist are consistent: judgment when the case does not fit the rule, responsibility that someone has to sign, trust built over time, physical presence, and care. A radiologist's software can flag a shadow; it does not explain the diagnosis to a frightened patient or carry the liability. A lawyer can have the case law in seconds; the strategy in front of the judge is still theirs. Nurses (57 on the index, 33 % of tasks) and cooks (30 and 12 %) sit lower precisely because their work is made of hands, timing and attention to people.

The book proposes a simple method. Automate the routine: hand the repeatable, predictable part of your work to AI and reclaim the time. Augment your judgment: use AI as a copilot for what you already do — analysing, drafting, exploring options — while you stay in charge of the decision. Anticipate what comes next: invest in what the machines do not close, such as judgment, relationships, creativity and the habit of learning continuously.

Start small and concrete. List the ten tasks you do most often in a week, mark the ones a model could draft for you, and try it on one of them this week. The professions at the top of this table are not doomed; they are the ones where people who adapt early will gain the most.