Your AI bill is a judgment problem, not a token problem

Every finance team that calls us starts the same way: "we need to get our AI spend under control." And every time, the first thing we look at isn't the model mix. It's whether anyone decided what AI should be doing in the first place.

Token prices are a red herring. A frontier model costs a few cents more per call than a smaller one. That's real money at scale, but it's a rounding error next to the actual cost driver: a high-effort, multi-step task running on the expensive model because nobody set an effort level, for a question that a cheap model answers just as well.

The bill is a record of ten thousand small decisions

Nobody approved that spend line by line. It's the sum of individual people, in individual tools, making individual judgment calls about which model to reach for and how hard to push it, because that's the only lever most AI tools give them. Multiply one slightly-too-expensive judgment call by thousands of requests a month, and you get a bill that doesn't map to any decision anyone actually made.

This is why cutting a model from the approved list rarely moves the needle. The spend wasn't really about the model. It was about effort with no accountability attached: nobody had to justify "high effort" versus "low effort" because nobody was ever shown the choice.

What actually brings a bill down

Three things, in order: define the flows people actually run, set the effort level per flow instead of per request, and put a budget in front of each team so spend has an owner before it happens, not after.

• Flows turn "whatever model felt right" into a decision your AI team makes once, for everyone.
• Effort levels (low, medium, high) separate "answer a question" from "do deep, multi-step work." Most tasks are the former.
• Team budgets with alerts mean Finance sees a number trending toward a problem weeks before the invoice does.

None of this is about restricting what people can ask AI to do. It's about making sure the cost of doing it well is a decision, not an accident. When Support's "draft customer reply" flow runs on a smaller model at low effort because 220 sampled replies showed no quality difference, that's not a budget cut. That's someone finally checking whether the expensive option was earning its keep.

The bill was never really about tokens. It was about judgment at scale, and judgment, unlike tokens, you can actually govern.