AI Cost Management Playbook

Attribution and caps, in that order. Tag every workload with an owner, a product, and an environment, reconcile the next invoice against internal usage so you know where the money actually goes, and put a hard cap on every workload so nothing can run away while you work. Only then move to forecasting and unit-cost reduction. Skipping attribution to chase savings first means optimizing spend you cannot even assign, and the biggest leaks stay invisible.

As a range, never a point estimate. Start from the measured run rate, split into price per unit and units consumed, then state a low, expected, and high case, because an agent that loops, retrieves, and calls tools can make identical tasks differ in cost by more than an order of magnitude. Size the budget against the high case or explicitly accept the risk of not doing so, and keep falling prices and rising usage as separate inputs so a flat bill cannot hide a volume problem.

Three structural levers, applied in order of safety. Cache the stable prefix of repeated prompts so fixed content stops being billed at full price. Move work nobody is waiting on, like bulk classification or overnight enrichment, to the provider's discounted batch path. Then trim retrieved context to what the task actually needs, which compounds inside agent loops because every step re-reads it. After each change, confirm with the people who use the output that nothing degraded before declaring the saving.

Bring it into the inventory instead of punishing it. Search expense records, vendor invoices, and access logs for tools nobody registered, then give each one an owner and a cap, or retire it if it earns nothing. Punishing the discovery teaches people to hide the next tool on a personal card, which moves spend somewhere invisible. Longer term, a shared gateway with per-team entitlements removes the reason shadow tools appear, because the sanctioned path becomes the easy one.