Realistic conversations, not scripted demos
Test with the messy, real phrasing an actual visitor would use — typos, vague first messages, mixed-language input if relevant — not the clean, cooperative test conversation that always makes a demo look good.
Edge cases and out-of-scope questions
Deliberately ask things outside the agent's knowledge boundary to confirm it says so honestly rather than guessing — see How to Prepare a Knowledge Base for an AI Agent for how that boundary gets defined in the first place.
Handoff triggers
Confirm the specific conditions that hand a conversation to a human actually fire correctly — see When Should AI Hand a Lead to a Human.
Failure paths
Test what happens if a downstream system (CRM, Telegram, email) is unreachable during a live conversation — the visitor should see an honest message, not a silent failure. See What Happens When an AI Automation Fails.
End-to-end, not component by component
The full path — website, agent response, CRM log, follow-up trigger — is tested as one connected flow before launch, matching how Revenue System describes launch and handover. See AI Sales Agent Build Timeline for where this testing phase sits inside the overall build schedule.
Questions
Do you test with real customer data?+
No — test conversations use synthetic scenarios, not real visitor data, during pre-launch testing.
What if an issue is found after launch?+
See What Happens When an AI Automation Fails for how post-launch issues are monitored and handled.
How long does testing take?+
It's part of the build timeline described in Revenue System — not a separate, open-ended phase.