AI and the future of work
AI is coming for desk work first. Here is what stays human.
The clearest signal about which jobs are exposed to AI is not coming from pundits. It is coming from the people building the technology and from the largest labor-market studies, and they agree on the shape of it: the work most exposed is predictable, screen-based knowledge work, and the work least exposed is hands-on, in-person, and accountable.
That is the entire premise of this site. While the headlines focus on what AI might take, we map the other side: the careers that still need a human, with real U.S. labor data and a transparent AI-resistance score for each one.
What the AI builders and researchers are saying
The labs: entry-level white-collar work is most at risk
Anthropic CEO Dario Amodei warned in 2025 that AI could eliminate up to half of entry-level white-collar jobs in fields like tech, law, consulting, and finance within one to five years, and push unemployment toward 10 to 20 percent. Early signs are visible: new-graduate hiring at large tech firms has fallen sharply. Axios, May 2025 ↗
Microsoft Research: exposure tracks knowledge work, not hands-on work
A 2025 Microsoft Research study measured how applicable generative AI is to each occupation using real usage data. The most exposed roles are knowledge and office work: interpreters and translators, writers, customer-service and sales roles, and administrative jobs. The least exposed involve manual labor, physical presence, and operating equipment, which is exactly the profile of the careers listed here. Working with AI, Microsoft Research 2025 ↗
World Economic Forum: clerical roles shrink, people-and-hands roles grow
The WEF Future of Jobs Report 2025 projects clerical and administrative roles, cashiers, accountants, and (newly) graphic designers among the fastest-declining jobs as generative AI spreads, while care, skilled-trade, and frontline roles keep growing. WEF Future of Jobs Report 2025 ↗
Why some work stays human
Across these forecasts the same four traits keep a job hard to automate. They are the traits our WontReplace Index measures:
- Physical, hands-on work. Care delivered to a body, a system fixed on site, a structure built in the real world. There is no screen for a model to reach through.
- Real-time human judgment and trust. High-stakes calls in messy, unpredictable situations, and relationships people want a qualified, accountable human for.
- Unpredictable environments. No two patients, job sites, or crime scenes are alike, and constant improvisation is the opposite of what automation does well.
- Licensing and accountability. When the law requires a credentialed person to sign off and carry the liability, a human stays in the loop by design.
An honest caveat
No one can predict this precisely, and the major forecasts disagree on timing and scale. That is why our score is a dated, transparent model we re-publish as AI advances, not a guarantee. We would rather show our work and update it than pretend a number is certain. See the full methodology.