Evaluate outcomes, not just agent activity.

Task count and token usage describe activity. They do not establish that customers received correct answers or that a business process improved. State the intended outcome and the observable evidence for it.

Record how the same bounded process performs before adding an agent. Use representative cases, consistent definitions and authorised data. Account for human review and correction effort rather than measuring generation speed alone.

Check quality, reliability, latency and total operating cost together. A faster draft that requires extensive correction may not be an improvement. Record missing evidence and failures instead of discarding difficult cases.

A changed metric after an intervention does not prove the agent caused it. Consider changes in demand, staff, source quality and process rules. Use an evaluation design appropriate to the claim before publishing impact numbers.

Test proposed changes against the same quality and policy requirements. Record approval, version and rollback criteria. This guide describes an evaluation approach, not a claim of achieved ExtremeAI results.