How does the AI stay grounded and safe?

Guests trust the AI because it never makes things up. Three automatic guards keep every reply grounded in your facts, safe for unconfirmed guests, and in the right language. You don't configure these — but understanding them helps you trust the AI enough to let it work alone.

Grounding guard

Every factual claim the AI makes must exist in the facts pack — the structured knowledge built from your listing and rules. No claim is allowed unless it can be traced back to something you provided.

This means the AI can't infer facts from other facts. From "blackout curtains," it cannot claim "soundproof windows." From "has a washer," it cannot claim "has a dryer." If a fact is missing, the AI does one of two things:

• Asks the guest a clarifying question to gather the missing detail, or
• Hands the conversation to your team.

It never guesses, and it never blends two known facts into a new, unverifiable claim. This is what keeps the AI accurate across hundreds of conversations with no human review.

Disclosure guard

Sensitive operational details are only shared with guests whose booking is confirmed — and even then, only when the details exist in the facts pack.

• Unconfirmed bookings (inquiries and cancelled bookings): the AI withholds the full address, WiFi credentials, access codes, and directions.
• Confirmed bookings: the AI may share operational details like the address and access instructions, but only if they're actually in the facts pack.

The disclosure guard prevents the most damaging failure mode of a guest-messaging AI: leaking entry details to someone who isn't a real guest.

Language guard

The AI always replies in the guest's language, auto-detected from the incoming message. A guest writing in Vietnamese gets Vietnamese replies; a guest writing in French gets French replies.

The guard also prevents the AI from copying language from your rules or host messages. If you write a rule in English, the AI will follow the rule's content but deliver it in the guest's own language — never pasting your English text verbatim to a non-English speaker.