Why this can't be left to the AI's judgment
Language models are genuinely good at conversation and genuinely bad at being trusted with zero-tolerance decisions, because they're probabilistic: even a well-built model will occasionally misjudge tone, miss context, or hedge on something it should have flagged immediately. That's an acceptable failure mode for "what time is checkout" and an unacceptable one for "there's smoke in the hallway."
So Suzy doesn't ask the AI to decide whether a message is an emergency. A separate, hardcoded check runs on every inbound message before it reaches the AI at all. If it matches crisis language, the message never goes through AI reasoning, it goes straight to an immediate SMS alert to the property manager, with an automatic SMS-carrier failsafe if the primary alert fails to send. This is checked first, every time, with no exceptions and no way for a model to reason its way around it.
The three rules the whole system is built around
Isolation
Every guest lookup is matched to the specific property the message came from. Data from different motels never mixes, so an escalation always reaches the right property manager, never a stranger's.
Deterministic safety
Crisis language, words and phrases around fire, medical emergencies, break-ins, and similar, is caught by hardcoded rules, not AI reasoning. This runs before the AI model ever sees the message.
Professional fallback
Outside of emergencies, if the AI isn't confident an answer is correct, it says so and hands off to a human instead of guessing. "I want to make sure I give you the right info" is the actual behavior, not a marketing line.
What this looks like in practice
A guest texts something indicating a real emergency. Before any AI reasoning happens, a hardcoded rule matches the message and fires an immediate SMS alert to the property manager's phone, with a backup carrier that sends the same alert if the primary one fails. The property manager is the one who calls 911 or responds directly, the AI's job is making sure that alert reaches a human instantly, every time, not handling the emergency itself.
What happens on a genuine wrong answer, not an emergency
Emergencies aside, mistakes are still possible, any system that answers thousands of guest questions will occasionally get one wrong (a stale WiFi password, a policy that changed last week and hasn't been updated yet). The response to that is also not left informal: there's a specific process, the property manager is contacted personally within an hour of a confirmed mistake, the underlying FAQ data is corrected within 24 hours so it can't recur, and there's a defined guarantee behind it.
The pattern across both cases, true emergencies and ordinary mistakes, is the same design principle: don't rely on the AI to self-police in the moment. Build the safety behavior as a rule that runs regardless of what the model decides, and build a real, specific process for when something still goes wrong.