Interactive FinOps worksheet
Calculate AI agent workflow cost across models, tools, runtime, retries, human review, exceptions, attempts, and business-accepted outcomes.
A completed model call is not necessarily a valid business result. Reconcile accepted outcomes with the workflow owner or system of record.
Evidence: Attempt count, accepted outcome count, rejection reasons, and reconciliation period
Retries, fallbacks, approvals, queue handling, and exception resolution can materially change unit economics even when provider cost looks low.
Evidence: Correlated retry events, active review time, loaded labor rate, and exception time
Use an approved, repeatable method to allocate runtime and shared tool costs. Keep the measurement window, rate source, exclusions, and workflow version beside the result.
Evidence: Reproducible input sheet and dated methodology
Measure the same workflow before and after a routing, prompt, tool, or approval-policy change. Track cost, acceptance, latency, and failure rate together.
Evidence: Comparable versions, sample sizes, quality gates, and confidence range
A provider invoice knows an API key, not a customer, a feature, or a workflow version. That mapping cannot be reconstructed later, so every day without it is a day of spend you will never be able to analyse.
Evidence: Tenant or customer identifier resolved from the authenticated caller and stored with token usage
Smaller tool results, a repaired cache prefix, and a deleted step carry no quality trade-off at all. Changing the model on a customer-facing response carries the most. The largest line item is rarely the safest place to start.
Evidence: A ranked change list with an evaluation requirement stated against each tier
Export model, tool, runtime, retry, review, exception, and accepted-outcome inputs.
Compute total workflow cost and divide it by accepted outcomes.
Review exclusions, allocation rules, sample size, and acceptance criteria.
Change one policy variable and compare the full outcome economics.
Cost per accepted outcome is the total of model, tool, allocated runtime, retry, human review, and exception-resolution cost, divided by the outcomes the business actually accepted.
Two rules decide whether the answer means anything. Accepted outcomes are results reconciled with the workflow owner or a system of record, because a completed model call is not an accepted outcome. And every attempt belongs in the numerator, including runs that failed, were retried, or were rewritten by a person - excluding them is the most common way a unit-cost figure flatters itself.
Record the measurement window, the rate source, the workflow version, and the acceptance definition beside the result, or it cannot be compared to a later one. Open this page in a browser to edit the interactive worksheet; nothing entered there is submitted to BlueBear.