AI costs behave one way in development and a completely different way in production. In development you control the inputs, run a handful of calls, and the cost is trivial. In production the workload meets real traffic, real edge cases, and real scale, and the cost stops being something you can estimate from a spreadsheet. Tracking AI costs in production is a different discipline from budgeting for them.
Why production costs diverge from estimates
Three things change the moment a feature ships. Volume scales with users rather than with your test runs. Inputs get messier, which means longer contexts and more retries than your clean test cases ever produced. And failure modes appear that never showed up in development, from rate-limit retry storms to runaway agent loops. Each pushes production cost above the estimate, and none is visible without tracking built for live traffic.
Track calls per request, not just calls
In production, one user action often triggers several model calls behind it. Track the ratio of model calls to user requests, and watch it over time. A rising ratio with flat user traffic is the signature of retries or an expanding workflow, and it is the earliest warning that a production cost problem is forming.
Track cost per outcome
Raw production spend tells you what you paid, not whether it was worth it. Track cost against the outcome the feature produces in production: cost per completed task, per resolved ticket, per served customer. That ratio is what tells you whether a scaling feature is scaling profitably or just scaling.
Attribute production spend to features
A single production billing account usually serves several features at once. Tag each call with the feature that triggered it, so production cost breaks down by feature rather than arriving as one total. Attribution is what lets you find the one feature quietly driving the majority of the bill.
Watch it continuously
Production cost can spike in a day. Monthly reconciliation finds the spike after the money is spent. Set anomaly alerts on day-over-day cost and on the calls-per-request ratio, so a production cost event announces itself while it is still small enough to fix cheaply.
Production is where AI cost stops being theoretical. Track calls per request, cost per outcome, and feature attribution continuously, and the surprises that come with real traffic turn into signals you catch in time.