AI and infrastructure
Failed AI Calls: Decide What Counts Before the First Dispute
By the Orvlin team · · 2 min read
Billing failed ai requests: what to decide
A model call uses resources but produces no usable answer. The provider may still report usage, while your buyer expects to pay only for completed work. Neither expectation should be settled ad hoc after an incident.
Define success for the commercial unit and retain status evidence. Separate provider-cost metering from billable customer actions. If retries are included in an outcome price, their costs still matter to your margin.
A concrete scenario
A document task makes three model calls and completes once. An offer priced per completed document may charge one unit while recording all three calls as costs. An offer priced per token would need a different explanation.
A detail teams often miss
Do not infer business success solely from HTTP status. A technically successful response can fail validation or require human correction. The billable definition should reflect what the customer bought.
Before you put this into production
- Define failure and partial completion states.
- Separate cost events from billable actions.
- Preserve retry relationships.
- Publish and test the agreed charging policy.
Frequently asked questions
- Should failed calls always be free?
- Not universally. It depends on the agreed unit and terms. Explain the policy clearly, especially when customers pay for attempts or underlying usage.