Build evaluation cases from approved examples: ambiguous requests, spelling variations, mixed-language messages, unsupported questions and requests that need escalation. Have business owners review the answers and action outcomes. Include unavailable APIs, duplicate webhooks and an unanswered human queue in the test set.
Estimate cost from the number of intents, languages, knowledge sources, integrations and review responsibilities. For AI-assisted flows, include model usage, retrieval maintenance and monitoring. Launch with bounded use cases whose owners can correct the source material and inspect failed conversations after release.