Account for the full task path
Plan for inference usage separately from the work required to orchestrate, verify and confirm an action. Model calls are one component. Reference lookups, deterministic checks, bounded recovery and destination reconciliation also contribute to delivering an inspectable result. A scoped proposal should make each component understandable.
Agree the billable units, rates, included allowances and treatment of unsuccessful work before an engagement. Rejected candidates, retries and interrupted runs can consume resources, so successful-outcome-only billing should not be assumed. The commercial terms need to explain how that usage is measured and which execution limits apply.
| Cost component | What a commercial discussion must define |
|---|---|
| Inference | Provider costs, usage units and treatment of repeated attempts |
| Orchestration and verification | Billable operations, included checks and execution limits |
| Confirmation and recovery | How reconciliation and interrupted work are accounted for |
| Integration and support | Any scoped implementation work and agreed support terms |
Estimate with representative tasks
A useful estimate begins with the operation and its document mix, not an assumed cost per token. Specify typical input sizes, expected volume, reference lookups, required checks and the permitted escalation path. Include difficult cases because they can change both the number of attempts and the amount of operational review.
Separate measured provider and execution usage from the team’s own review time. Compare like-for-like task definitions: preparing a candidate JSON object and confirming a persisted ERP draft have different boundaries. Use representative measurements for an estimate, including the frequency of exceptions and the work needed to resolve them.
Agree limits and service expectations explicitly
Define the attempt budget and task policy alongside the commercial scope. Specify what happens when a run cannot finish within its limits, including how already-issued writes are reconciled. A cost ceiling should not leave an uncertain business action without an owner or a recovery path.
For a pilot discussion, provide the destination system, approximate volume, representative inputs, required completion evidence and data-handling constraints. We can use that information to scope the engineering work and commercial questions. Availability, support response times, retention terms and any service commitments require a separate documented agreement.
Missing a detail or found a problem?
Send a documentation question →