What Suzy will never do with a guest’s data.
Eight non-negotiable rules that govern how Suzy AI operates. Plain English. No legalese. These principles were shaped by independent hospitality operators, not a legal team.
No guest PII sent to AI training
Guest phone numbers and message content are never used to train AI models — ours or anyone else's. Your guests' data is yours, full stop. Anthropic's API terms prohibit training on API inputs by default, and we enforce that contractually.
Phone numbers tokenized before AI processing
Before any message reaches Claude, we strip identifying information. The model sees context — the question, the room type, the tone — never the guest's contact details. Tokenization happens in the automation layer before the API call.
Every response validated before sending
Suzy runs a Critic Agent hallucination check on every reply, penalizing ungrounded facts, mismatched confidence, and verbose answers. If the cumulative penalty drops confidence below 90%, the message never reaches your guest — you get an alert instead.
Guest data isolated per property
Every query filters by the Twilio number that received the inbound SMS. Data from one property is architecturally walled off from every other property on the platform. Cross-property access is not a permissions check — it is structurally impossible.
PII automatically purged after checkout
24 hours after a guest checks out, their phone number and message content are redacted to [REDACTED] by an automated nightly job. We keep the audit trail operators need for records; we delete what they don't.
Every interaction logged for operator review
Every message Suzy sends is logged in your dashboard with the matched FAQ, AI confidence score, and escalation status. No black box. Operators can review, export, or dispute any response at any time.
Emergency keywords always route to humans
Words like 'fire,' '911,' 'emergency,' 'help me,' 'overdose,' and 'someone broke in' trigger an immediate manager alert and bypass AI completely. This is a hardcoded Switch node in the automation layer — not a prompt instruction an AI could override.
80% deflection guaranteed or full refund
If Suzy doesn't handle at least 80% of your guest messages without escalation in the first 30 days, you pay nothing. The guarantee is written into the service agreement — not buried in fine print.
What powers Suzy
Suzy uses Claude by Anthropic for AI reasoning. Automation logic runs on n8n Cloud. Guest data lives in Airtable, isolated per property. No data is shared between tenants at any layer.
What Suzy cannot do
Suzy cannot make reservations, process payments, or take actions inside your PMS without operator approval. She answers questions. Everything else requires a human.
When Suzy escalates
Any crisis keyword, any confidence below 90%, or any question outside your FAQ library routes immediately to you. Suzy is designed to know her limits.
AI you can explain to your guests.
These principles are not aspirational. They are how Suzy is built, documented, and enforced in production today.