A job title is a bundle, typically fifteen to thirty distinct tasks stitched together under one role. The task-based model, developed in the economics literature by David Autor and, more recently, Daron Acemoglu and Pascual Restrepo, insists on analysing automation exposure at the level of the individual task rather than the job as a whole, because a job rarely disappears wholesale. What changes is its composition: some of its tasks get absorbed by a machine, some stay firmly human, and the job that remains looks different from the one that existed before, without necessarily employing fewer people.
This distinction is why headline estimates of “jobs exposed to AI” are so easy to misread. Goldman Sachs’ widely cited 2023 estimate that generative AI could affect the equivalent of 300 million full-time jobs globally was a task-level estimate, roughly a quarter to half of the workload inside two-thirds of exposed jobs, not a prediction that 300 million people would be fired. Reading it as the latter overstates the disruption for any specific worker and understates how unevenly that exposure actually lands across a single role.
The practical use of the model is that it converts an abstract, anxiety-inducing question, “will AI take my job”, into a concrete, answerable one: which of the fifteen to thirty tasks I actually do this week does a model already perform competently, and which don’t. That list, not the job title, is the real unit of exposure, and it’s the only version of the question that produces an actual plan rather than a mood.