The Argument
Most enterprises fund AI through a central "innovation" pool. Finance sees one line item; product teams see unlimited tokens for demos. That arrangement creates moral hazard: engineers optimize for impressive outputs, not marginal cost per decision. Finance absorbs run-rate creep until the board asks why digital COGS jumped 18% with nothing shipped to revenue.
Chargeback is not accounting theater. It is the feedback loop that makes AI operable at scale. When spend attaches to workflow owners and decision volume, product leaders face the same trade-offs they already manage for headcount and cloud—scope, model tier, kill switches. Until $/decision has a name on the P&L, no organization will optimize routing, cap tool use, or retire workflows that fail a unit-economics test.
This memo is about allocation, not architecture. The Rebuild Threshold answers when to replace a system; AI Engineering defines the discipline that keeps systems alive. Here the question is simpler and more urgent: who pays for each inference every month?