Replace, augment, support: a nested task-level measure of generative-AI exposure for the full O*NET catalogue

This paper builds a task-level measure of potential exposure to generative AI in three nested layers. For every task, read in the context of its occupation, we first ask whether AI can replace its output at the quality of a typical worker (Re). If it cannot, we ask whether it can augment the worker enough to halve the time the task takes or to push its quality into the top quartile (Au), and if not, whether it can still support the task with some usable input (Su).

Three frontier language models rate all 24,087 task-activity pairs of O*NET 31.0 (923 occupations), working from a dated inventory of 33 capability benchmarks, and a Delphi round settles their disagreements. Weighted by time, the average occupation has 37.8% of its working time replaceable, 48.4% augmentable and 79.2% supportable. Raters agree most on replacement and least on the augmentation thresholds.

We also measure three bundling factors (relational nexus, organisational responsibility and authorship) that keep a job’s tasks tied to the person doing them, plus the share of working time in which AI helps an expert without replacing them. Everything is carried to European classifications through official crosswalks, with the Spanish CNO-11 as our application.