Methodology

The source data

Every figure comes from Future of Work with AI Agents (Shao et al., 2025) and its accompanying WORKBank database. Researchers surveyed 1,500 U.S. workers and collected ratings from 52 AI experts, covering 844 real occupational tasks across 104occupations. The tasks themselves come from the U.S. Department of Labor’s O*NET database (licensed CC BY 4.0).

The two core measures

Worker desire to automate (1–5).Workers who actually do a task were asked: “If an AI system could do this task for you completely, how much would you want it to?”, from 1 (not at all) to 5 (entirely). We average across everyone who rated each task.

AI capability today (1–5). AI experts rated how capable current AI systems are of performing each task. We average across the experts who rated each task.

The Human Agency Scale (H1–H5)

Workers also placed each task on the Human Agency Scale, how much human involvement they think the task should keep, from H1 (AI handles it entirely) to H5 (humans essential throughout), with H3 meaning an equal human–AI partnership.

How we assign the four zones

We cross the two measures. A task’s desire or capability is “high” when it’s at or above the median across all 844 studied tasks, a desire median of 3/5 and a capability median of 3.5/5. That split is our own simplification of the framework, chosen for clarity; it is not an official cutoff from the study.

• Automation Green Light , Workers want it · AI can do it. Strong candidate to automate now.
• R&D Opportunity , Workers want it · AI can't yet. Wanted — but the tech isn't there yet.
• Automation Red Light , AI can do it · workers don't want it. Capable — but automate with care.
• Low Priority , Low desire · low capability. Leave it to people for now.

Limitations

• The survey covers U.S. workers and was collected in early 2025, a snapshot, not a forecast.
• Worker desire is self-reported; AI capability is expert opinion, not a benchmark of any specific product.
• Figures are aggregates for an occupation or task, not statements about any individual’s job.
• Our high/low median split is a deliberate simplification to make the data browsable.
• AI capability moves fast; an R&D Opportunity today may become Green Light tomorrow.

Attribution & citation

Automatable is an independent project and is not affiliated with, or endorsed by, Stanford University or the U.S. Department of Labor. If you use this data, please cite the original study:

@misc{shao2025futureworkaiagents,
  title  = {Future of Work with AI Agents},
  author = {Shao, Yijia and Zope, Humishka and Jiang, Yucheng and
            Pei, Jiaxin and Nguyen, David and Brynjolfsson, Erik and Yang, Diyi},
  year   = {2025},
  eprint = {2506.06576},
  archivePrefix = {arXiv}
}