Rob Seamans

A few high-level notes on the paper: 1. We ask where the marginal value of work AI cannot yet do may rise as AI speeds up the rest of the job. - Such work can become a bottleneck, especially where more of the surrounding tasks in a job can be done by AI. We call this "bottleneck pressure." 2. Using a task-based framework, we show that for an AI-infeasible task, this pressure is simply the exposure of the occupation posting it. - So you get a new, theory-derived measure that potentially shows where AI could raise the marginal value of work it cannot yet perform. 3. The magic happens at the task level. The same task has different bottleneck pressure in different jobs. 4. To take this measure to the data, we map 288 million U.S. job postings from 2018 to mid-2026 to O*NET tasks and use their GPT-4 feasibility ratings from Eloundou, @sj_manning, Mishkin, and @danielrock's "GPTs are GPTs" (2024). 5. O*NET can be a very coarse proxy for advertised work. - About 90% of the task mentions matched from postings fall outside the posting occupation's own O*NET list, although those statements themselves come from O*NET. 6. Indeed, about 53% of mention-weighted variation in pressure on AI-infeasible work lies within tasks, across occupations. 7. In every year, AI-infeasible tasks within occupations faced more pressure, on average, in senior than junior postings, about 6.5 percent more in the first half of 2026. 8. Over the full sample, mean occupation pressure rose from 17.6 to 18.7, or by 6 percent, about 75% from task mentions shifting toward higher-pressure occupations. (Thank you to @robseamans for pushing me to do 7 and 8!) Overall, reporting bottleneck pressure alongside exposure shows where improving the work AI cannot do may pay off more as capabilities expand.