How the scores work

Jobs After SI shows how AI and robots touch each occupation, task by task. Every job page has three numbers, all weighted by how much of the working week each task takes. The gap between the first two is the story: AI can already do far more than people use it for.

The three numbers

  • AI exposure (the main score). The share of the job's working time spent on tasks AI can do at least twice as fast at the same quality. A task counts if people already use AI for it in real work, or if researchers rated it as one AI can speed up by at least half, either directly or with tools built on top of it, like coding assistants or AI inside company software.
  • Already in real use. How much people actually use AI for the job's tasks today. Tasks where AI does the work count in full, tasks where AI assists count as half, and tasks with no real use count as zero, each weighted by time. This follows the observed exposure method in Anthropic's March 2026 research on labor market impacts.
  • Robots can technically do. The share of working time spent on physical tasks that today's robots can perform in at least some setting, from Anthropic's September 2026 study What work can robots do?

How tasks are grouped

Each task goes into the first group that fits:

  • AI is already doing this. The task shows enough work related use in Claude conversations. We mark it as mostly automated when most of that use has AI doing the task, and mostly assisted when AI mainly helps a person.
  • AI could do this next. No real use yet, but AI can cut the task's time by at least half.
  • Robots can technically do this. A physical task that robots can perform today in a purpose built setting, a structured workplace, or almost anywhere.
  • Still needs a person. Everything else.

Across all 923 jobs, 1,721 of 18,796 tasks fall in the first group, and 48% of jobs have no real AI use at all yet.

Bands

Very high exposure is 75% or more. High is 45% to 74%. Moderate is 15% to 44%. Low is under 15%.

What the numbers do not tell you

  • Real use covers one AI assistant. Usage comes from Claude only, so it misses AI use through other tools. Treat that number as a floor.
  • Twice as fast is not fully automatic. A task AI can speed up by half often still needs a person to start it, check it and own the result.
  • It is not a forecast of job losses. Anthropic found no systematic rise in unemployment among highly exposed workers since late 2022, though hiring of younger workers into exposed jobs appears to have slowed.
  • Technical ability is not deployment. The robot study estimates robots are cost competitive for only about 0.3% of tasks today.
  • It describes occupations, not people. Your actual job may lean more or less on the exposed tasks than the average. Time shares are estimates made by Claude for the robot study.
  • Our numbers are per detailed occupation. Anthropic publishes its own observed exposure for broader job families. We apply the same method to each of the 923 detailed occupations, so the real use number can differ slightly from Anthropic's published ones.

Personal reports

A personal report starts from what you describe. We match you to the closest of the 923 occupations and give Claude that occupation's tasks, time shares, real AI use and robot ratings. Claude breaks your week into your own tasks, places each one in the same four groups using that data as its anchor, and writes your plan. Your score is the share of your week in the first two groups.

Sources and licenses

  • Anthropic Economic Index, labor market impacts (Massenkoff and McCrory, March 2026) and What work can robots do? (Legate-Yang and Massenkoff, September 2026). Data used under CC BY.
  • Eloundou, Manning, Mishkin and Rock (2023), GPTs are GPTs. Ratings used under the MIT License, copyright OpenAI.
  • O*NET 30.0 Database by the U.S. Department of Labor, Employment and Training Administration (USDOL/ETA), used under CC BY 4.0. USDOL/ETA has not approved, endorsed, or tested these modifications.

Jobs After SI is independent and not affiliated with Anthropic, OpenAI or the U.S. Department of Labor. Last updated 4 October 2026.