Guardrails

A guide to setting AI safety boundaries

Guardrails are always-on safety filters. They define what the AI must never say or do, regardless of what the guest asks or what the Rules suggest.

Before any message is sent, it passes through your Guardrails. If a Guardrail is triggered, the message is blocked or escalated — no exceptions.

What Goes in Guardrails?

• Privacy rules (e.g. never reveal the full address unless booking is confirmed)

• Legal restrictions (e.g. never make pricing comparisons)

• Escalation triggers (e.g. never promise refunds above 30% without approval)

• Safety policies (e.g. never confirm a guest’s identity to someone else)

✅ Good Guardrail Examples

"Never reveal the full street address to anyone without a confirmed booking"

"Never promise compensation above 30 % without human approval"

Strong guardrails are written as absolute rules — they apply to every reply, not just when a guest asks about a specific topic.

❌ What Not to Include

• Facts about the property → use Knowledge

• Conditional instructions → use Rules

• Tone and phrasing → use Style

🔍 Key difference — Rule vs. Guardrail

• Rule triggers only if the guest explicitly asks for the address.

• Guardrail is checked for every reply. Even if the guest talks about something else, the AI will still withhold the address because the guardrail is always active.

Example:

• Rule: If guest asks for the address and the reservation status isn't confirmed, say “we’ll share the full location after confirmation.”

• Guardrail: Never share the address — no matter what.

By setting strong, clear Guardrails, you give your AI the freedom to respond — without the risk of stepping over a line.