The experience premium stopped being a slope this year.
In AI-exposed occupations it became an arch. Three years out, the market pays more than it used to. Twenty-five years out, it pays less. Nobody updated the comp bands.
The retention bump
A CHRO at a mid-market management consultancy shows me a comp-band proposal in September. Schooling premium held at prior-year levels. Experience premium indexed linearly to years. A Senior Director with 25 years in a role her analytics team flags as top-quartile AI exposure. The proposal: a 2.5 percent retention bump to match the external market.
I ask what the external market is doing. She says that is what the retention bump is for.
What the paper measured
The paper comes from Levy Yeyati and co-authors at Universidad Torcuato Di Tella. Eight and a half years of US Current Population Survey records. 253 occupation clusters, 655,919 person-month observations. The authors run a difference-in-differences on the top versus bottom quartile of the Felten-Raj-Seamans occupational AI exposure index, with ChatGPT's November 2022 release as the shock.
In exposed occupations, after that date, a year of schooling pays 0.55 log points less (s.e. 0.19). A year of experience at three years pays 0.26 log points more (s.e. 0.10). At fifteen years the effect is statistically indistinguishable from zero. At twenty-five years a year pays 0.12 log points less (s.e. 0.04). The slope flips sign at 18.1 years of potential experience (s.e. 2.1).
The authors state it plainly:
Codifiable knowledge becomes less scarce; the value of experience increasingly depends on whether it transfers to the new technology.
The pandemic placebo returns a schooling effect of -0.02 log points, p=0.94. Pre-trend tests on the main specification do not reject. The paper is working, US CPS only, not yet peer-reviewed.
The mechanism the authors propose is clean. Generative AI made codified knowledge cheap at the point of production, so the credential prices in less. Practice-based tacit knowledge still lives with the worker, so experience that complements the new tooling prices in more. Experience tied to the pre-AI stack prices in less because that stack is depreciating.
The retention bump I was shown was built on a slope that no longer exists in that role. Treatment in the paper is defined at the occupation level, so the number prices the seat, not the person. For every seat in an exposed role, stop asking how many years. Ask whether those years transfer.