Everyone's Writing About the Jobs AI Is Taking. Here's What's Actually Holding Up.

A quarter of the 88 occupations we researched this year landed in our low-exposure tier, for specific, identifiable reasons that have nothing to do with prestige.

The jobs in that tier -- electricians, dentists, chefs, and similar roles -- share a trait that has nothing to do with education level and everything to do with structure. Their core work is physical, situational, and unpredictable in ways that don't compress into training data. An electrician isn't doing the same wiring job twice; a lot of the value is judgment applied in a specific physical space, in real time, with real consequences for getting it wrong. A model trained on text can't step into that room and do the job. That's a genuinely different kind of durability than "some jobs just get automated slower."

Here's the part worth paying attention to even if your title didn't land in that tier: inside the roles getting hit hardest, there's a consistent durable core, and it's the same short list every time -- framing the right question instead of just answering the one you were given, getting a skeptical person to trust and act on a conclusion, owning an outcome instead of a task, handling the exception the standard process didn't anticipate, holding a relationship where the relationship itself is the value. None of that shows up on a resume as a line item the way "proficient in Excel" does. All of it is what's still getting paid for, at every exposure tier, in every function we researched.

It would be easy to read a low-exposure result as a reason to relax. The broader 2026 numbers argue against that: enterprise AI adoption is reported at 91% of organizations using it in some capacity, up from 78% two years ago, and generative AI use specifically has doubled in ten months to 65%. Only a third of that has scaled past pilot mode today -- and today is the operative word. Meanwhile white-collar payrolls specifically have contracted for 31 straight months while physical-and-situational trades haven't seen anything close to that pattern, which is itself informative: the current wave is hitting codified, text-based, checkable work hardest, and structurally physical, relational, high-judgment work has so far been largely exempt. So far is doing real work in that sentence.

Job Resiliens' own AI Job Risk Index research across 88 occupations, alongside McKinsey's 2026 State of AI enterprise adoption data and Axios/Quartz reporting on the 31-month white-collar payroll contraction.

Also worth reading: AI vulnerability score by job title · how we score AI resilience