Blog · AI for small and midsize businesses
How to control AI application cost per completed task
A low model charge can coexist with an expensive workflow. Cost per completed task includes unsuccessful attempts, retrieval, external tools and the people who verify results. Define completion first, then measure the operating unit that matters to the business rather than a single request.

In this article
An application may make several model calls, searches and retries before producing one acceptable result. Comparing the price of a single call misses that workload. Control AI expenditure by measuring a completed business task with its associated review and failure costs.
Define the denominator
For a draft-answer workflow, completion might mean a reviewer accepts the response as ready to send. A generated response that staff discard is an attempt, not a completion. Keep attempted, accepted, referred and abandoned tasks visible so cost reporting does not improve by excluding failures.
Use two measures: direct operating expenditure divided by accepted tasks, and fully loaded operating cost including allocated review and support divided by accepted tasks. State whether implementation cost is included or reported separately. Consistent definitions matter more than a low unit price.
A worked hypothetical calculation
Assume 1,000 attempted monthly tasks produce 800 accepted completions. Illustrative direct costs are MXN 900 for model use, MXN 300 for retrieval and storage and MXN 300 for external tools. Direct cost per completion is MXN 1,500 ÷ 800 = MXN 1.875.
Assume review requires forty hours valued internally at MXN 200 per hour, with MXN 2,000 of allocated support. Fully loaded cost becomes (1,500 + 8,000 + 2,000) ÷ 800 = MXN 14.375. All amounts are invented planning inputs, not vendor prices.
Track the causes of expensive cases
- Number of model and tool calls per task, including failures.
- Retrieval work and source size needed for the answer.
- Retries by cause, with a maximum attempt limit.
- Reviewer time and corrections by request type.
- Support and maintenance effort allocated consistently.
A technical owner should investigate unusual spending without exposing unnecessary customer information in logs. The process owner checks whether a cost reduction worsens answer quality or transfers effort to staff.
Set operational controls
Agree usage budgets, alerts and a rule for pausing unusually expensive task paths. Bound repeated attempts and refer unresolved cases instead of allowing indefinite retries. Keep a manual route for important work when a budget or service limit is reached.
Segment unit cost by task type. A rare complex inquiry may legitimately cost more than a routine lookup. First inspect avoidable repeated searches, overly broad context and unclear instructions. Replacing a model without testing the completed task may simply increase correction work.
Deloitte's framework includes implementation and maintenance economics. The NIST profile supports monitoring the operating context. Neither provides a universal acceptable cost per task.
Review the completion definition with the receiving team. If staff accept a draft but another department later corrects it, the task may not be complete at that first checkpoint. Include a reasonable downstream quality check and report its limits. Use the same definition when comparing different configurations so apparent savings do not come from moving work outside the measurement.
Bring one month of attempts, completions, review time and charges to an AI operating-cost review to identify the next useful control.