Test your AI Agent

Testing is more than one hello message. Try real questions from last month tickets, edge cases from your policy, and angry-tone messages to see escalation behavior.

Keep a shared doc of test prompts the team reruns after major Knowledge updates.

Suggested test set

1. Order status for a fulfilled and an unfulfilled order.
2. Return outside the window with a polite refusal and option to escalate.
3. Product recommendation for a bestseller and a out-of-stock variant.
4. Request that should never auto-refund without a human.

When tests fail

Fix Knowledge first, then tone, then permissions. Most wrong answers trace back to missing or conflicting policy text.