4 minBusiness
AI Exposure Depends on Tasks, Not Job Titles, Analysis Finds
A new analysis argues that workers' exposure to artificial intelligence is determined by the specific tasks they perform each week, not by their job titles or seniority. It offers a five-step method for measuring individual risk.
Workers' exposure to artificial intelligence is determined less by their job titles than by the specific tasks they perform each week, according to a new analysis that offers a practical method for measuring individual risk. The argument draws on research showing that tasks vary substantially within the same occupation, meaning two people with identical titles can face very different levels of automation risk.
The analysis cites a 2023 paper by researchers from OpenAI and the University of Pennsylvania that scored occupations task by task. That study estimated that 80% of U.S. workers could see at least a tenth of their tasks affected by AI, while about 19% could see at least half. But those averages obscure what matters for any one person: exposure depends on how much of a workweek can be specified briefly and checked quickly. Seniority offers protection only when it changes that mix of tasks.
To measure exposure, the analysis proposes a five-step process. The first step is to list actual tasks from a real week rather than relying on a job description. That means recording everything that consumed more than half an hour — a usability readout, a stakeholder review, a status update — along with hours, frequency, and who reviews the result. The list should also include the unrecorded work between tasks, such as unblocking a teammate or offering a quick answer that saves someone a day. That work rarely appears in a job description or on a calendar, yet it is often where complexity is absorbed.
For a checklist, the analysis points to O*NET OnLine, the Labor Department's occupational database, which lists 31 tasks for web and digital interface designers. It also cites a 2009 paper by economists David Autor and Michael Handel, which found that tasks vary enough within occupations to predict wage gaps among people holding the same job.
The second step is to score each task's exposure using what the analysis calls a spec-and-check test. The first question asks whether a finished version of the task could be described in two sentences that someone else could follow. The second asks whether someone could confirm the output is right in less time than it took to produce. A task that is quick to specify and quick to check is exposed, regardless of job title. Generating twenty variations of an empty state passes both questions, while deciding which of three problems to fund this quarter fails both.
The analysis warns that difficulty is a poor guide to exposure. It cites a 2023 experiment led by Fabrizio Dell'Acqua in which 758 Boston Consulting Group consultants did far better on 18 tasks inside a model's abilities but were 19 percentage points less likely to get the right answer on one task outside them. Coauthor Ethan Mollick calls this the jagged frontier: tasks that look equally hard can sit on opposite sides of an invisible wall.
The remaining steps involve weighting the task list by hours, since a few tasks typically fill most of the week, and then using that weighted picture to identify where human judgment adds value that AI cannot. The analysis illustrates the point with the history of automated teller machines. Economist James Bessen showed that ATMs took over dispensing cash and taking deposits, and the number of tellers needed to run an average urban branch fell from 20 to 13 between 1988 and 2004. But banks opened 43% more branches, so total teller jobs held. The title stayed the same while the job underneath it was rebuilt.
The broader lesson is that job titles are an unreliable proxy for automation risk. A calendar, a task list, and an honest assessment of which outputs can be quickly specified and checked offer a more accurate picture of where AI is likely to change the work — and where human judgment remains essential.
3
