New paper introduces "bottleneck pressure" to measure where work AI can't do gains value

soumitrashukla9 · x · 2026-10-24

Soumitra Shukla (Burning Glass Institute) released a new paper responding to Alex Olegimas's call for fast economic evidence on AI.

Instead of asking what AI can do, the paper asks where the marginal value of work AI cannot yet do may rise as AI speeds up the rest of the job.

Built "bottleneck pressure" from 288M US job postings mapped line-by-line to ONET tasks, combined with GPT-4 feasibility ratings from Eloundou et al. (2024)
About 90% of labeled task mentions fall outside the posting occupation's own ONET task list
53% of mention-weighted variation in bottleneck pressure comes from within tasks, across occupations
Pressure peaks where advertised work is evenly split between AI-feasible and infeasible tasks, ranking occupations differently than exposure metrics
As pressure rises, interpersonal and supervisory work makes up more of the AI-infeasible workload, physical work less

More from AGI Musings