Labor Dept AI jobs data: measure your own adoption

The federal government wants to know how much work AI is actually doing, and it has decided the fastest way to find out is to ask the companies selling it. That is a reasonable engineering decision and a strange measurement decision, and the Labor Department AI jobs data effort is worth watching for both reasons.

What actually happened

Axios reported on August 26 that the Department of Labor has signed memorandums of understanding with a set of technology companies — OpenAI, Google, Meta and Amazon among them — to share data on how businesses are deploying AI and where they expect adoption to go next. The stated goal is to supplement traditional labor statistics, which move on a quarterly cadence, with something closer to real time.

The rationale from acting Labor Secretary Keith Sonderling is blunt: the government does not have the data, and the large companies most affected by AI do. Bloomberg Law covered the same push from the department's side.

The terms are the part nobody has published. We don't know what fields these agreements cover, how the data is aggregated, or what gets released publicly versus stays internal to the department — so we're not going to characterize it. What's established is that the agreements exist and who signed them.

Why it matters for your business

Here's the structural problem. The companies supplying the data sell seats and tokens. The metrics they can produce cleanly are the ones they already bill on: accounts provisioned, API calls, monthly active users. Those are adoption-shaped numbers. They are not outcome-shaped numbers, and the gap between the two is where every disappointing AI rollout lives.

Which means the national picture of "AI adoption" will likely run ahead of the national picture of AI doing useful work — and you'll read headlines built on the first one while making payroll against the second.

Do not outsource this measurement. Before your next hiring or non-hiring decision, get three numbers for your own shop: how many hours a week a specific task consumed before you automated it and after; your cost per finished task, tokens plus review time, not cost per token; and the rework rate — how often a human has to redo the output. That's a spreadsheet and two weeks of honest logging, and it will beat any benchmark published about your industry.

The other reason to have your own numbers: when a vendor raises prices or a policy shifts, you want an argument grounded in your workload, not in a statistic somebody else's telemetry generated.

Sources: Axios, Bloomberg Law.