Skip to main content
Use this page as a checklist when an Agent is not answering the way you expect, or when you want to prepare a polished launch.

​
What you will learn

  • How to make Knowledge Sources easier for the Agent to use
  • How to write Behaviour settings that create consistent answers
  • How to test before launch
  • How to use Conversations and Analytics after launch

​
Make Knowledge Sources specific

Agents work best when your source content is explicit. Avoid vague statements that require guessing.
Well structured text source placeholder
Strong source content usually includes:
  • Clear product names, plan names, and feature names
  • Exact steps customers should follow
  • Known limitations and eligibility rules
  • Links written as plain visible URLs when the Agent should share them
  • Separate sections for pricing, support, setup, and troubleshooting

​
Use Q&A for critical answers

Use Q&A when a question needs a precise, repeatable answer. This is useful for refunds, pricing rules, handoff rules, business hours, and compliance-sensitive responses.
Q&A source best practice placeholder
Keep each Q&A focused. If one answer covers several topics, split it into multiple entries.

​
Write Behaviour like an operating guide

The System Prompt should explain the Agent’s job, audience, and boundaries. The Guardrails field should cover things the Agent must avoid or escalate.
Behaviour best practices placeholder
A strong Behaviour setup includes:
  • Who the Agent helps
  • What the Agent can answer
  • When the Agent should say it does not know
  • How it should ask for missing context
  • When it should collect a lead or direct the visitor to support

​
Choose tone and verbosity intentionally

Use Friendly for conversational support, Professional for sales or B2B workflows, and Neutral for factual help centers. Use Short for fast FAQ answers, Balanced for general support, and Detailed for technical or process-heavy guidance.

​
Test with real customer questions

Before launch, test the Agent with questions from actual conversations, support tickets, sales calls, or website search terms.
Agent preview testing placeholder
Test for:
  • Accuracy: the answer matches your source content
  • Completeness: the visitor gets the next step
  • Boundaries: the Agent avoids unsupported claims
  • Tone: the response fits your brand
  • Recovery: the fallback message is helpful

​
Monitor after launch

Use Conversations to inspect individual threads and Analytics to find patterns across many interactions.
Conversation review placeholder
Analytics dashboard placeholder
Turn repeated weak answers into better Text sources, Q&A entries, or Behaviour instructions.

​
Next steps

Knowledge Sources

Learn how to add, edit, delete, and train sources.

Conversations

Review live conversations and identify content gaps.