Most enterprise AI invoices that quietly grow into five and six figures share the same root cause — and it is not the one engineering teams reach for first. A field report on what actually drives runaway inference spend, and the discipline that brings it back to a number the CFO stops forwarding with a question mark.
AI bills that double every quarter are not pricing problems — they are architectural problems compounding. A smaller model and more caching usually capture less than a third of the available savings. The other two-thirds live in orchestration, prompt and context discipline, UX surface decisions, and governance.
Caching addresses one specific failure mode — repeated identical or near-identical calls. Most enterprise AI bills that grow quarter over quarter are not failing at caching; they are failing at orchestration, context discipline, and the absence of cost governance. Caching typically captures around a tenth of the available savings; the rest sits elsewhere.
Rarely. Cheaper-model substitutions look attractive in isolation but often cascade into more retries, more context, and more downstream corrections that wipe out the per-call savings. The first move is almost always visibility — proper per-request token tracing — followed by orchestration cleanup, before any model swap is evaluated.
Visibility wins (tracing, tagging, alerting) typically land in week one. Quick prompt and orchestration wins land in weeks two and three and usually account for a 20–40% reduction. Structural wins — the ones that produce the order-of-magnitude headline numbers — land over the following two months as design changes ship.