Validate and refine before launch
Pressure-test the agent in the Launch section and tune behavior in Train, so the first live week teaches you about edge cases, not avoidable setup mistakes.
Test with real customer language
In Launch, do not validate against idealized prompts alone. Use the ambiguous, messy phrasing customers actually send and watch whether the agent still grounds its answer.
Validation is not just about correctness. It is also about tone, clarity, and the ability to admit uncertainty without sounding broken.
Run a tight refine loop
- 1Validate common tasks first in Launch: order status, shipping, returns, and account questions.
- 2Capture the turns where the answer becomes vague, overconfident, or unsupported.
- 3Refine in the right place — adjust persona or escalation in Setup, fix coverage in Knowledge, or use Train to capture behavior memories — one change at a time.
Know when the first pass is good enough
- The agent answers common questions directly and without filler.
- Escalation happens for uncertainty instead of fabricated certainty.
- Operators can explain what changed between each validation round.
Where do I validate the agent before launch — Launch or Train?
Use the Launch section to pressure-test live prompt behavior with real customer phrasing and confirm answers stay grounded before traffic increases. Use Train to capture behavior memories and review real conversations once you start finding recurring gaps.
What is the safest way to decide when the agent should escalate?
Start narrow. Escalate on refund disputes, sentiment spikes, account risk, or any place where a wrong answer creates operational cost. Expand only after reviewing real transcripts.