Validation Checklist
Use this checklist before enabling a new AI Task, changing a prompt, or rolling out a new custom insight.
A. Data prerequisites
- Conversations have transcripts/threads available
- Conversation volume is sufficient to validate (recommended: 10–30 samples)
- Conversations represent your typical and edge cases
B. Task configuration checks
- Task is mapped to the intended Custom Field(s)
- Field types match output types (number/date/dropdown/text)
- Filters are appropriate (not overly strict, not overly broad)
C. Output format checks (JSON recommended)
- Output is valid JSON (no extra text outside JSON)
- Keys match mapping attributes exactly
- Values are within allowed ranges / labels
- “Unknown” or “not mentioned” behavior is defined and consistent
D. Explanation quality checks
- Explanations are concise (1–3 sentences)
- Explanations cite evidence from transcript/thread
- Explanations avoid speculation (“maybe”, “probably”) unless explicitly allowed
E. UI verification
After enabling/adjusting the task:
- Field values appear in Conversation Details → Analytics
- Dashboards update correctly (if applicable)
- Search filters work (e.g., numeric comparisons like CSAT < 3)
- Drilldowns work (bucket labels navigate to matching conversations)
F. Governance checks
- Stakeholders agree with definitions (QA/CX/Sales)
- Prompt version/change notes are documented
- Rollout plan defined (pilot group first, then broader)