AI Agents: Set up guardrails
What is it?
Guardrails restrict certain customer inputs before they reach an AI Agent. They keep conversations on topic, safe, aligned with your business policies, and within an agent's responsibility — so your bot doesn't engage with requests it shouldn't.
Both Routers and Specialists have a Guardrails tab, and they're configured the same way. What differs is which guardrails actually run — and that depends on whether the agent is used standalone or through a Router. That rule is the most important thing on this page, so it's covered first.
Who can use it
Guardrails are configured in Flow Builder, so they're available to users with Flow Builder access (organization admins).
The rule that matters most: routed vs. standalone
Where a guardrail runs depends on how the agent is reached:
- Standalone Specialist: the Specialist's own guardrails apply to incoming messages.
- Specialist reached through a Router: only the Router's guardrails apply. The Specialist's own guardrails do not run after the Router transfers the conversation.
Important: Do not rely on Specialist guardrails for conversations handled through a Router. If a restriction must always apply, put it on the Router (Router → Guardrails). Use Specialist guardrails only when the Specialist is used standalone.
This is the single most common source of confusion, so a good habit is: decide first whether the agent will ever be reached through a Router. If it will, its restrictions belong on the Router.
Add a guardrail
The steps are the same on a Router and on a Specialist:
- Open the Guardrails tab.
- Click Add guardrail.
- Give the guardrail a short name.
- Describe the type of input it should restrict.
- Enable or disable the guardrail as needed.
Guardrail names should use lowercase letters, numbers, and hyphens.
What guardrails can restrict
Guardrails can help block off-topic requests, inappropriate content, personal or restricted topics, and requests outside your support scope.
Examples:
- Name: off-topic — Description: Block questions unrelated to our products, services, or customer support.
- Name: unsupported-topic — Description: Block questions about legal, medical, or financial advice.
Where to put each guardrail
- Restriction must apply to every conversation, including routed ones → Router → Guardrails.
- Restriction only applies when a Specialist is used standalone → Specialist → Guardrails.
- Not sure whether the Specialist will be routed → put it on the Router to be safe.
Best practices
- Write specific descriptions — say exactly what to block, not just "block bad input."
- Put shared, always-on restrictions on the Router when using routed Specialists.
- Use clear, hyphenated names so guardrails are easy to scan.
- Test a request that should be blocked as part of validating any agent.
- Revisit guardrails after changing whether a Specialist is standalone or routed — the rule about which ones run may now apply differently.
Troubleshooting
- A guardrail doesn't behave as expected: if the Specialist is used through a Router, move the guardrail to the Router's Guardrails tab — Specialist guardrails don't run on routed conversations.
- A guardrail name is rejected: names must use lowercase letters, numbers, and hyphens only.