Uncapped Usage Was a Design Choice

Uber didn't blow its AI budget. Uber designed a system that guaranteed the blowout, and then was surprised when the system worked.

The internal guidance encouraged employees to use AI tools "as much as possible." Four months into a twelve-month budget cycle, finance pulled the plug. TechCrunch reported the cap this week. The story has been framed as a cost-control problem. It isn't. It's the predictable endpoint of a governance pattern that's been building across enterprise IT for two years, and you can see the corrective model forming on the other side of the same week.

Here's the trajectory. In 2024, enterprise IT signed AI tool contracts that priced like utilities — per token, per call, per query — but procured them on the SaaS template that a decade of seat-license buying had made automatic. Legal reviewed an MSA built for fixed-fee software. Finance budgeted a line item that assumed predictable per-user costs. Adoption teams were told to push usage, because the seat-license playbook says wasted seats are the failure mode. None of that is wrong for SaaS. All of it is wrong for a meter.

By mid-2025, the cracks were visible. Sam Altman publicly conceded what every enterprise AI buyer already knew — token costs at scale don't behave like SaaS economics. A few public companies disclosed AI cost overruns in earnings notes. The agentic-system blowouts that surfaced through May were the early tremors. Uber is the headline version: a name big enough, a number large enough, that the seat-license fallacy stops being a procurement footnote and starts being a board agenda item.

The Morgan Stanley story arrived the same week and looks like a different topic. It isn't. CNBC reported the bank is opening its wealth management platform to external AI agents. The coverage framed it as first-mover boldness. Read it the other way. The bank is moving slowly and constrained-by-design because its fiduciary and regulatory posture won't let it do otherwise. It has to know who owns the agent, who controls the prompt boundary, who carries liability when an agent gives client-touching advice, and what the per-interaction cost is before the interaction happens. The pioneer framing flatters the bank. The constraint-driven framing explains why its architecture is the one CFOs should be studying.

(If you're an unregulated tech company looking at this comparison and noticing that a regulated bank has better AI cost governance than you do — yes, that's the point.)

The seat-license fallacy is what connects both decisions. Uber priced unlimited consumption as a benefit. Morgan Stanley priced bounded consumption as a control. One treated the meter as a perk; the other treated it as a liability surface. The budget blowout and the controlled rollout are two outputs of the same upstream choice: how much of your cost base do you let your employees, or your agents, run up before you find out.

I don't know whether the next twelve months produce a wave of public restatements or just quiet internal caps applied without disclosure. The practical answer is that most large enterprises don't know their actual AI spend run-rate well enough to disclose it cleanly, and that opacity is doing a lot of work right now. The gap matters.

Here's the prediction. By the end of Q4 2026, at least three Fortune 500 companies will disclose AI tool overruns large enough to require an earnings-call mention, and the corrective pattern they describe will look structurally identical to what Morgan Stanley is shipping now: per-agent cost ceilings, prompt-boundary controls, and explicit ownership of the liability layer. If the seat-license fallacy persists as the default procurement template through 2027 — if finance teams don't learn from each other's budget lines that fast — I'm wrong.

The CFOs who internalize that distinction this quarter will save themselves a Uber memo next year.