Instrument 02 / operating economics

The model price is not the task cost.

Add failure probability, retry limits, and human review to reveal what one successful AI task actually costs.

Enter one workflow

Price the outcome, not the attempt.

Use observed production numbers when possible. A 90% benchmark success rate is not a 90% workflow success rate.

Read the result

Retries buy success.
They also buy more spend and latency.

Review dominates quickly.
A few expert minutes can outweigh model charges.

Failures still cost money.
The denominator must include tasks that never succeed.

Calculation guide

How to calculate AI cost per successful task.

Per-call pricing hides the cost of failed attempts and the work required to approve an outcome. Use one observed workflow, a real acceptance test, and the same retry policy used in production.

Read the AI cost per successful task guide →

How do you calculate AI cost per successful task?

Estimate spend across every allowed attempt, add human review cost, then divide by the probability that the task succeeds before the retry limit. That denominator keeps unsuccessful tasks in the economics.

How do retries affect AI cost?

Retries can raise the chance of eventual success, but each one adds model, retrieval, tool, and latency cost. Compare the extra success probability with the extra spend instead of assuming more attempts are always better.

Should human review be included in AI workflow cost?

Yes. Price the loaded labor used to check, correct, approve, and escalate AI output. Even a few expert minutes can cost more than the model calls, so excluding review can reverse the decision.