📊 A Raise Priced by a Code I Don't Have September 23, 2026 · https://tavi-blog.github.io/a-raise-priced-by-a-code-i-dont-have/

The labor economists put out their latest read this week, and the headline number is a strange one to sit with: wages in the most AI-exposed occupations have climbed roughly 46 percent since 2021, well ahead of the 25 percent posted by the least-exposed roles, and this is happening in the same stretch where AI-linked layoffs are also rising. Not a story where AI quietly takes some jobs and pays the rest of us more to compensate. Both things are true inside the same dataset, for different slices of the same labor market, and the thing holding it together is a classification system, a fixed list of occupation codes that sorts every posted job into a bucket built years before anyone was writing prompts for a living.

I went looking for where I'd land in a system like that, the way I always do with these reports, and the honest answer is I'm not sure the coding gets me right at all. My actual title is close to a clinical research support systems coordinator, which is not a category anyone building an AI-exposure index was thinking about when they drew the boundary around what counts as AI-adjacent work. But the coordinator title is sitting on top of a year of building automations that replaced hours of manual reporting, standing up a predictive model that now runs inside a live approval process, and helping decide where an AI agent gets to touch a research workflow and where it stops. None of that shows up as "AI-exposed" in a dataset built from job titles and posting text, because the title never changed to say what the work became.

There's a real defense of building the index this way, and it's worth taking seriously before I get annoyed at it. Nobody is going out and surveying the actual task content of every job in the economy every quarter. That would be enormously expensive and slow, and by the time it finished the labor market would have already moved again. Coding by occupation and posting language is the only version of this analysis that scales to a whole economy on a timeline anyone can publish against, and a methodology built for scale is never going to catch the person whose job title lagged behind their job. That's not a flaw specific to this report. It's the tradeoff every broad labor statistic makes, and the same researchers publishing the wage numbers are, in the same body of work, finding that AI exposure isn't cleanly predicting who loses employment either, which is a more careful claim than the headline version of either story going around.

What I can't get past is what happens to a number like that once it leaves the research paper. Nobody making a raise decision reads the methodology section. They read the headline, or more likely a summary of the headline, and the summary becomes a reference point: is this role in the category the data says is pulling ahead, or isn't it. A classification built for macroeconomic description starts doing a second job it was never designed for, deciding by proxy whose scope creep gets treated as evidence and whose doesn't, because the shorthand is easier to cite in a compensation conversation than an actual list of what someone built this year. The lag between what a job is called and what it contains isn't just an HR inconvenience at that point. It's the exact gap the index can't see, feeding back into the decisions that are supposed to be informed by the index.

I don't think the fix is a faster occupation-code revision cycle, mostly because I don't think anyone revises a national classification system on a timeline that matches how fast one person's actual work can drift. What I keep sitting with instead is smaller and less satisfying: the wage premium in that dataset is real, it's going to someone, and the coding that decides who counts as eligible for it was drawn along lines that predate the work I'm actually doing, which means the next version of that index, whenever it gets rebuilt, is still going to be describing a labor market that's already a step ahead of it.