Future of Work with AI Agents: Auditing Automation and Augmentation Potential across the U.S. Workforce
Citation. Shao, Y., Zope, H., Jiang, Y., Pei, J., Nguyen, D., Brynjolfsson, E., & Yang, D. (2025). Future of work with AI agents: Auditing automation and augmentation potential across the U.S. workforce. arXiv. https://arxiv.org/abs/2506.06576
Summary. Shao et al. introduce a worker-centered audit of AI-agent automation and augmentation potential across computer-compatible U.S. occupational tasks. The paper builds WORKBank from O*NET tasks, 1,500 domain-worker preference interviews, and 52 AI-expert capability assessments covering 844 tasks across 104 occupations. Its main contribution is a desire-capability map plus a five-level Human Agency Scale that separates worker appetite for automation from current technical feasibility and preferred human involvement.
Key claims
- Workers expressed positive attitudes toward AI-agent automation for 46.1% of audited tasks after considering job-loss concerns and work enjoyment; the main stated motivation was freeing time for higher-value work, with substantial sector variation.
- Combining worker automation desire with expert capability ratings yields four task zones: Automation Green Light, Automation Red Light, R&D Opportunity, and Low Priority. The authors report that 41.0% of Y Combinator company-task mappings fell in the Low Priority or Automation Red Light zones, suggesting current startup activity is not fully aligned with worker-desired, technically feasible opportunities.
- The Human Agency Scale ranges from H1, where an AI agent handles the task entirely, to H5, where continuous human involvement is essential; 45.2% of occupations had H3 equal partnership as the dominant worker-desired level, indicating broad interest in collaboration rather than pure automation.
- Workers generally preferred higher levels of human agency than current AI capability assessments might imply, creating a likely friction point as agent capabilities advance into tasks where workers still want control, validation, domain judgment, or interpersonal presence.
- Mapping tasks to core skills suggests a shift in the value of human competencies: information-processing skills such as analyzing information become less emphasized in high-agency/high-wage task space, while interpersonal and organizational skills gain importance and broader skill bundles may be required.
Evidence & limitations
- Evidence combines structured worker mini-interviews, ONET task statements, expert AI-capability ratings, and external mapping of AI startup activity. The task universe is restricted to computer-compatible work and filtered from a broader ONET set, so the findings are strongest for digital or digitally mediated occupations rather than the whole labor market.
- Worker desire is self-reported under a structured scenario rather than observed adoption behavior, and expert capability ratings estimate current technical feasibility rather than realized organizational performance, quality, accountability, or labor-demand outcomes.
- The paper is an arXiv preprint, and version 3 was updated in 2026; treat its WORKBank framework and descriptive results as useful map-making evidence, not as settled causal evidence about substitution or long-run workforce change.