Skip to main content
Guardrails function as persistent safety mechanisms that establish boundaries for AI responses. They define what the artificial intelligence must never say or do, regardless of user requests, Knowledge, Skills, or tool results. Before transmission, all messages traverse through guardrails. When activated, guardrails either block or escalate communications—without exception.

​
What Goes in Guardrails?

  • Privacy regulations (example: withhold sensitive account details absent confirmed identity)
  • Legal constraints (example: refrain from price comparisons)
  • Escalation thresholds (example: avoid promising refunds exceeding 30% without authorization)
  • Safety procedures (example: never share customer identity with third parties)

​
Where to Find Guardrails

Open Build > Agents, select an agent, then open Persona. Put hard response restrictions in the guardrails area.

​
Good Guardrail Examples

  • “Never reveal sensitive account details to anyone without verified identity”
  • “Never promise compensation above 30% without human approval”
Effective guardrails employ absolute language—applying universally rather than conditionally.

​
What Not to Include

  • Product or service information - use Knowledge
  • Conditional directives - use Skills
  • Communication style - use Style

​
Key Difference: Skill vs. Guardrail

Skill: Activates when a scenario matches, such as “if the customer asks for account details, explain that identity verification is required.” Guardrail: Evaluates across replies and prevents sensitive data sharing unconditionally.