"Is AI taking jobs?" is one of those questions everyone has an opinion on and almost nobody has actually joined the data on. So I tried to settle it with numbers instead of takes — and built the whole thing in Bonacci Studio.
The honest answer surprised me.
The Data Problem
Four public datasets, none of which speak the same language: BLS/OEWS employment, O*NET's occupation spine, the Anthropic Economic Index (what people actually use AI for), two competing academic "AI exposure" scores, and Indeed's hiring data. Three different occupation coding schemes between them.
The real work — the part that usually kills projects like this — is the crosswalk that resolves all of it to one key. That's the node in the middle of the canvas, and it's where Bonacci Studio earned its keep: ingest → crosswalk → one clean spine, then I just asked the agent in the dock to test the thesis. It wrote the PySpark, ran it, and rendered the charts inline. No notebook, no export.
What the Data Says
- In aggregate, AI has not cut jobs. US employment is up ~3.4% since 2019.
- Sort occupations by "AI exposure" and the most-exposed look basically like the least-exposed — and which way the tiny gap leans flips depending on which exposure measure you trust. The two standard academic measures agree only 58% of the time. That disagreement isn't noise to hide; it's the finding.
- The real split isn't "exposed vs not." It's how AI is used. Occupations where AI augments people grew (+1.3%); occupations where it automates the task shrank (−1.4%).
- The clearest signal is in hiring, not headcount: job postings in high-AI-exposure sectors are running below their pre-pandemic baseline while low-exposure sectors sit above it. Demand has bent before employment has.
The Verdict
AI hasn't taken the jobs. It's quietly bending the hiring curve — and it's doing it along the automation line, not the blunt "exposure" line everyone argues about.
The two mainstream academic AI-exposure measures agree only 58% of the time. Any analysis that picks one and runs with it is arguing with half the foundation missing.
How It Was Built
The messy multi-source join, the crosswalk, the analysis, the charts, the writeup — that used to be a two-week project. In the studio it was an afternoon, and the agent did the parts I'd normally dread.