What Does an AI Agent Actually Cost per Case?

This week's OpenAI–ServiceNow announcement illustrates how AI may move from answering questions into multi-step enterprise workflows. Anthropic's enterprise controls describe usage analytics and spend controls, but a usage dashboard alone cannot show whether a case was actually resolved or whether a person had to repair the result.

A Canadian insurer's administrative service team might pilot an assistant to classify incoming correspondence and prepare a response for review. The business should measure the entire handling process; the example is administrative and does not authorise automated coverage decisions.

For a defined period, add model usage, orchestration and tool costs, allocated infrastructure, human review, corrections and relevant support. Divide by cases resolved to an agreed quality standard. Report abandoned cases and cases escalated to humans alongside that figure so the denominator does not conceal failures.

For example, if total pilot cost is $1,200 and 300 cases meet the agreed resolution standard, the observed cost is $4 per resolved case. This is an illustration, not a benchmark. A fair comparison requires similar case complexity, quality standards and accounting for the existing workflow.

Break the results down by case type. A low-cost model may be economical for straightforward classification but expensive when it repeatedly calls tools or sends ambiguous cases back for correction. Routing difficult cases to people or a different model can be sensible, provided the routing itself is evaluated.

Set per-workflow budgets and alerts, cap runaway retries, and inspect unusually expensive cases. Review quality, turnaround and staff workload with finance. Optimising cost without checking missed errors or customer outcomes can simply move expense elsewhere.