A low subscription price does not establish that an AI agent is inexpensive to operate. The useful unit is the cost of an acceptable result, including the work required to review and repair it.

Anthropic’s guide to evaluating agents distinguishes what an agent says it accomplished from the final state of the task. The same distinction belongs in a business case: a completed chat is not necessarily a completed job.

Define an accepted result

For a lead-research agent, an accepted result might be a correctly matched company, verified source links and enough information for a salesperson to decide what to do next. For an internal support agent, it might be a resolved request that stays resolved after review.

The definition should be written before the pilot. Otherwise, the easiest outputs to count may become the measure of success, even when employees cannot use them.

A worked example with hypothetical numbers

Assume a monthly pilot processes 500 research requests. Software and usage cost $100. Review takes five hours at an assumed $20 per hour, adding $100. Correcting failed outputs takes two more hours at the same rate, adding $40.

Total operating cost is therefore $240. If 400 results meet the acceptance criteria, the cost per accepted result is $0.60. These figures are illustrative assumptions, not observed performance or vendor pricing.

Notice that all attempted work contributes to cost, while only accepted results contribute to the denominator. Dividing by 500 instead would make the pilot look cheaper without improving its usefulness.

Compare equivalent work

The manual baseline must deliver the same quality and scope. If the agent produces short summaries while a person previously produced detailed reports, comparing the two preparation times would be misleading.

Also separate released time from cash savings. Freeing five hours can create capacity, but it does not automatically reduce payroll. The business needs a plan for how that capacity will be used.

Include costs that appear after the demonstration

  • Initial configuration and integration work.
  • Ongoing review and exception handling.
  • Monitoring, maintenance and changes to business rules.
  • Failed runs, repeated tool calls and duplicate work.
  • Consequences of incorrect external actions.

Keep setup expenditure separate from recurring operating costs, then show both over a realistic period. This makes it easier to see whether a promising pilot becomes worthwhile only at a higher volume.

A decision the numbers can support

Our recommended pilot report includes accepted-result rate, review time, correction time and operating cost. Compare those figures with the existing process before expanding access.

An agent can be useful even when it does not remove a job or eliminate review. The claim should match the evidence: faster preparation, more consistent coverage or additional capacity are specific benefits that can be tested.