This is an always-on messaging desk for a vacation-rental business. A Guest Message Webhook receives inbound guest messages — a channel manager or PMS like Hostaway, Lodgify, or Guesty forwards Airbnb and Booking.com threads to it, and email arrives through an IMAP-to-webhook bridge. Normalize Guest Message unifies every payload, the message is logged with status NEW, and AI Agent — Classify & Draft reads it against the property context (check-in and check-out times, WiFi password, house rules) and returns a structured verdict through Structured Output Parser — Guest Intent: intent, confidence from 0 to 1, an auto_reply_allowed flag, a ready-to-send draft reply, and an escalation reason. IF — Safe to Auto-Reply? passes only when auto-reply is allowed and confidence meets the threshold (default 0.85): the draft goes out via Send Guest Reply, the sheet flips to AUTO_REPLIED, and the host gets a Telegram summary. Anything else — refunds, complaints, access codes, safety issues, or low confidence — flips to ESCALATED and fires the Telegram — Escalate to Host alert with the guest's message, the AI's read, the reason, and a suggested draft. It exists because slow replies cost stars, and 2 AM WiFi-password questions cost sleep.
The 15-property Airbnb host losing evenings to check-in questions
A host with fifteen listings spends two hours every evening answering the same five questions: WiFi password, check-in time, parking, checkout, house rules. Those map exactly onto the default safe_intents list (wifi, checkin, checkout, houserules, amenities, parking, directions), and the answers live in CONFIG — Properties & Rules — checkin_time, checkout_time, wifi_password, house_rules_url — so the AI drafts from facts, not guesses. With the confidence threshold at 0.85, the routine questions get answered in seconds while anything ambiguous escalates. Practical tip: extend the system prompt with a per-property facts block (gate codes, parking spots, quiet hours) keyed off the property field from Normalize Guest Message, so a fifteen-property setup still answers precisely.
The co-host managing 30 listings who needs sleep
A co-host's response-time metric directly affects search ranking, but she cannot answer messages at 3 AM. The auto-reply branch is her night shift: Send Guest Reply POSTs the draft to the PMS reply endpoint, Update Sheet — Status AUTO_REPLIED records it, and Telegram — Notify Host (Auto-replied) gives her a morning summary of who got what answer. She wakes up to a log instead of a backlog. Practical tip: raise the confidence_threshold to 0.9 or higher in CONFIG — Properties & Rules — with thirty listings, she would rather escalate a borderline message than risk a wrong auto-reply at scale.
The boutique villa with strict house rules
A luxury villa owner keeps finding extra guests and late-night parties because the rules were buried in a PDF nobody read. The houserules intent plus the house_rules_url in config means every rules question gets an instant, consistent answer with the link — no more improvised replies from whoever happened to see the message. Early check-in and late checkout requests intentionally escalate, since those are revenue decisions, not FAQ answers. Practical tip: use the third-branch customization — extend the parser's intent list with upsell and route airport transfers and paid late checkouts to a human draft, turning escalations into revenue instead of chores.
The property manager who will not risk refunds going wrong
A manager with twenty doors refuses to let any software touch money conversations. The workflow agrees with her: refund and complaint intents never pass IF — Safe to Auto-Reply?, and the escalation alert includes the guest's message, the AI's interpretation, the escalation reason, and a suggested draft, so she replies in one minute instead of ten. The Draft Reply column in the sheet stores everything the AI ever sent, giving her a weekly audit trail. Practical tip: during the first two weeks, disconnect the true branch of the IF node into escalation entirely — calibrate on real messages, watch the confidence scores, and only reconnect auto-reply once the scores earn her trust.
The glamping operator with patchy connectivity
A rural glamping host is often offline for hours, which used to mean guests waited half a day for simple answers. Now the webhook, the AI agent, and the auto-reply branch run without her: routine questions resolve themselves, and the Telegram — Escalate to Host messages queue on her phone for the moment she has signal. The sheet's NEW to PROCESSING to AUTO_REPLIED or ESCALATED trail means she can reconstruct the whole day in five minutes. Practical tip: pre-filter obvious OTA auto-messages (booking confirmations) with an extra IF on message keywords before the agent runs, so the AI — and the API bill — only touches messages from real humans.