How to Run an AI Spend Month Close

Freeze the reporting window and catalog

Record the time zone, invoice periods, catalog versions, credits, commitments, and late-arriving usage. Do not silently reprice historical requests with today's model rates.

Export provider totals and the gateway ledger to a controlled close workspace. Preserve source files and calculation versions.

Reconcile totals before distributing reports

Compare provider spend with gateway-attributed spend by provider and account. Investigate gaps from direct keys, consoles, asynchronous jobs, rounding, failed calls, cache billing, and missing integrations.

Keep an unattributed line rather than spreading unknown cost across teams. Assign an owner and due date to each material variance.

Validate savings claims

Routing savings require a documented baseline model and price at request time. Cache savings require proof that an eligible provider call was avoided or discounted. Separate shadow-mode opportunity from realized savings.

Report the cost of routing classifiers, evaluators, embeddings, and retries so the net figure is credible.

• Provider invoice total
• Gateway ledger total
• Allocated and unattributed coverage
• Budget variance by team
• Realized routing and cache savings
• Quality or rollback exceptions
• Direct-key and console leakage
• Actions for the next period

Close with decisions

A close is complete when material variances have explanations and owners, chargeback or showback statements are distributed, and policy changes are approved. Carry recurring problems into automation: scoped keys, price verification, budget alerts, and route evaluation.