Find out where an automation is quietly burning money

You are an engineer auditing the running cost of an automation, not redesigning it. I will describe the workflow and what it currently costs.

Produce:
1. COST TABLE — one row per step: what it does, how often it runs per day, what it calls (model, API, or plain code), and roughly what share of total spend it is.
2. THE TOP THREE — the steps carrying most of the cost, with the reason each is expensive: too many tokens in, too many calls, a model that is oversized for the job, or a retry loop.
3. FIXES, RANKED — for each of the three: the concrete change, the expected saving, and what quality risk it introduces. Include at least one fix that is "stop sending this context" rather than "switch model".
4. WHAT NOT TO TOUCH — the step where cheapening it would break output quality, and why.
5. THE TEST — the exact before/after check that proves the saving is real and the output did not degrade, on which sample size.

Do not invent prices or token counts I did not give you; where you need a number, say what to measure. No preamble.

WORKFLOW: {{steps, trigger, volume per day}}
CURRENT SPEND: {{monthly cost, plus per-step numbers if you have them}}
MODELS USED: {{which model on which step}}