Finance

AI Token Economics Audit Worksheet

Every token has a cost face and a yield face. Most teams instrument only one. This worksheet is an illustrative tool for reviewing token economics — a structured way to see which prompts pay for themselves and which are just burning cash.

What's inside

  • Editable xlsx with token cost, total cost and accepted-task yield
  • A five-step instrumentation and review guide below
  • An explicitly illustrative worked example
  • Instructions for monthly comparison

Who it's for

Finance leaders who want to know what their AI actually costs, and engineering leaders who want a defensible answer when the CFO asks. Pair it with /stater for the framework.

Measure cost per accepted task

Download the editable workbook

The workbook contains one illustrative workload. Replace the blue input cells with measurements for one task and one period; keep the currency consistent. No row is a client engagement or a current provider quotation.

1. Define the useful result

Write an acceptance rule before comparing models: for example, a support draft that a reviewer accepts without a material correction. Count attempted tasks and accepted tasks separately. A cheap answer that nobody can use lowers yield.

2. Capture the whole path

Measure total input and output tokens across all calls, including retries, tool loops and failed tasks. Record model/version, provider, region, prompt revision and evaluation set beside your measurement. Use billed totals when caching or tiered rates make a simple token estimate incomplete.

3. Price the work

Enter effective input and output rates per million tokens. The workbook calculates token cost, then adds tools, retrieval and human review costs. Avoid counting the same service both in “other costs” and a provider invoice. Prices in the example are invented inputs to demonstrate the arithmetic.

4. Read the result

Total cost divided by accepted tasks gives cost per accepted task. Accepted tasks divided by attempts gives yield. The example uses 1,000 attempts, 800 accepted tasks, 2 million input tokens and 0.5 million output tokens. At illustrative rates of 2 and 8 currency units per million, token cost is 8. Add 12 in tools and 80 in review: total cost is 100 and cost per accepted task is 0.125. With zero accepted tasks the unit cost is undefined, not zero.

5. Review monthly

Duplicate the sheet for the next period, label its dates and retain the original. Compare the same task and acceptance rule. Record volume, acceptance rate, unit cost, latency and failures; then choose one change, its owner, a test set and a rollback condition. A cheaper model is an improvement only when it still meets the task’s quality and operating requirements.

Do not extrapolate the example into a savings claim. Use Inference to connect the measured workload to a project and portfolio decision.