Loading…
From the first agentic models to saturation
An average white-collar workweek
Most exposed today
Share of the job's working time AI can do now.
Least exposed today
Jobs held back by in-person work, people and relationships.
How to read this
Jobs
Each occupation is split into the tasks that fill its week. Open a group, then a job, to see which hours AI can take now and in the next two generations.
Task types
Every task in every job maps to one of these 29 types. Each is scored for how much of it the frontier model can do at professional quality, now and for the next two generations, up to a ceiling set by things intelligence can't change. Select a row for the evidence behind it.
Adjust the scoresTest your own assumptions. Changes stay in this browser.
Method and sources
Snapshots
One snapshot per Claude Opus release. Each scores what the best publicly available model could do on that date. Earlier snapshots were scored afterwards from benchmark results and are marked as backfilled. Abbreviations: GDPval, expert-graded comparison with professionals' deliverables; METR, length of task a model finishes on its own; RLI, Remote Labor Index, share of whole freelance projects done to professional standard; OSWorld, computer use.