The Token Bill Is the Audit

Adoption was never the outcome. It was a proxy for one, and treating the proxy as the goal is exactly how you end up with a token bill that grows faster than anything it was supposed to produce.

That's the real story under this week's reporting that employers who pushed staff to use AI more are now watching it backfire as costs hit home. The tools didn't fail. The measurement did. When your north star is usage, people use. They just don't necessarily create value while they do it, and you built no way to tell the difference.

Seeing adoption isn't the business goal. It'd be like saying hammers are useful, so let's increase how often we swing them. Swing count is not a construction project.

The same week, DoorDash, Siemens and Airbnb turned up as enterprises quietly moving workloads to cheaper Chinese models to curb ballooning bills. The tidy read is that these buyers are now solving the problem. I don't buy that framing. Swapping a vendor to cut the per-token price doesn't answer whether the work got better. It just makes the same unmeasured spend cheaper. And routing enterprise workloads through Chinese model APIs carries compliance and data-sovereignty risk that the "just switch vendors" story leaves conveniently unpriced.

Swapping a vendor to cut the per-token price doesn't answer whether the work got better.

Neither the mandate nor the vendor switch was the mistake. The mistake was rolling out at scale with no framework to separate "we used it a lot" from "we do better work now." Governance and a measurement plan belong before the floodgates open, not after the invoice arrives.

What I don't know yet is how many of these programs get refined versus abandoned. The directional signal is clear though: the bill is now doing the audit nobody ran up front.