Guardrails

Set boundaries, approved topics, escalation rules, and competitor handling for your AI demo agent.

Guardrails keep your agent on-topic, accurate, and aligned with your company's messaging. They prevent the agent from making claims it shouldn't, discussing topics it's not qualified for, or engaging with off-topic requests.

Topic boundaries

Define which topics the agent can and cannot discuss:

Approved topics

1. Go to Agent Demos > Guardrails > Topics.
2. Add approved topic areas (e.g., "Product features", "Pricing tiers", "Integration capabilities").
3. The agent engages confidently on these topics using trained knowledge.

Restricted topics

Add topics the agent should not discuss:

• Legal and compliance -- "I'd recommend discussing compliance requirements with our team directly."
• Custom pricing -- "For custom pricing, I can connect you with our sales team."
• Roadmap -- "I can't share specifics about upcoming features, but I can schedule a call with our product team."

For each restricted topic, configure the deflection response the agent uses.

Accuracy controls

• Confidence threshold -- Set a minimum confidence score (e.g., 0.8). If the agent's confidence falls below this, it acknowledges uncertainty: "I'm not sure about that -- let me connect you with someone who can help."
• Source citation -- Enable source references so the agent attributes answers to specific documentation.
• Hallucination prevention -- The agent only answers from trained knowledge. Questions outside its training trigger an escalation or deflection.

Escalation rules

Define when the agent should hand off to a human:

• Explicit request -- Prospect says "I'd like to talk to a person" or similar.
• Frustration detection -- The AI detects negative sentiment or repeated questions.
• Topic trigger -- Conversation enters a restricted topic area.
• Qualification threshold -- Prospect meets lead qualification criteria.

For each rule, configure the action:

• Transfer to live chat (if available).
• Offer to book a meeting.
• Collect contact information for follow-up.

Competitor handling

Configure how the agent discusses competitors:

• Acknowledge and redirect -- "That's a great product. Here's how we approach that differently..."
• Feature comparison -- Share approved comparison points from the training data.
• Decline to comment -- "I'm not the best source for comparing us to [Competitor]. I can share what makes us unique."

Set the strategy per competitor under Guardrails > Competitors.

Testing guardrails

Use the test console to probe your guardrails:

1. Ask questions about restricted topics.
2. Simulate frustrated or off-topic interactions.
3. Verify that escalation triggers fire correctly.
4. Review the agent's deflection responses for tone and accuracy.

Note: Guardrail violations are logged under Agent Demos > Analytics > Guardrail Events. Review these regularly to identify gaps in training or rules that need adjustment.