1. Count outcomes, not only calls
- Pick one customer-visible workflow.
- Count successful completed runs separately.
- Count failed, retried, abandoned, and human-rescued runs separately.
- Calculate cost per successful workflow.
Find the AI workflow that quietly breaks your unit economics
Use this before buying the audit. The goal is to stop looking only at provider bills and start measuring cost per successful workflow with a hard guardrail for ugly runs.
2. Tag the top 20 expensive runs
- Model used per step.
- Input and output token spend.
- Tool calls, context size, and memory injected.
- Whether the run reached a useful business outcome.
3. Check hard caps, not soft hopes
- Set a per-workflow cost ceiling.
- Set a per-user or per-team monthly ceiling.
- Kill or downgrade retry-heavy runs automatically.
- Separate staging, internal, and customer usage budgets.
4. Run the ugly-month scenario
- Normal month.
- 2x usage month.
- Retry-heavy month.
- One power user on a flat plan.
- Kill or reprice any plan that goes negative there.
What to send for the paid audit
- 7-30 days of redacted usage screenshots or CSV exports.
- Plan pricing and included limits.
- One expensive trace or example run.
- Any current retry, routing, or cap rules.
- No API keys.
- No passwords.
- No customer personal data.
- No unrestricted production access.