Cost per successful task is the AI margin metric
August 27, 2026
Finance should not approve AI investment on request volume alone. A request can fail, trigger human repair, or require three expensive retries. The decision metric is cost per successful task: attributable AI spend divided by work that meets the business definition of done. Pair it with revenue or avoided cost to see contribution margin.
Use an outcome denominator
Define success for each workflow before opening the dashboard. A support answer may need user acceptance; invoice extraction may need a confidence threshold and no manual correction; a research workflow may need a delivered artifact. The denominator must be observable and owned by the business, not invented after the invoice arrives.
The CFO bridge
Begin with the approved baseline and bridge the change through volume, rate, model mix, and efficiency. Then show cost per successful task by product, customer segment, and workflow. A higher bill can be a good investment if successful outcomes and gross profit grew faster. A lower bill can be bad news if completion quality fell.
| Metric | Question it answers |
|---|---|
| Total AI spend | How much cash left the business? |
| Cost per request | What does an attempt cost? |
| Cost per successful task | What does usable work cost? |
| Contribution per task | Does the workflow create economic value? |
What to request at month-end
Ask for provider-reconciled spend, successful-task volume, p50 and p95 cost, failure and repair rates, and the three largest movements. Require a named owner for every material variance. This turns AI reporting from a technology invoice into a business review.
The metric does not exist to force every workload to be cheap. It exists to make expensive capability visible, comparable, and accountable.
Related
- AI spend variance analysis
- Proving AI ROI
- LLM cost monitoring