Watch an AI handle a real hotel inquiry, end-to-end.
Three captured conversations with a 100-key boutique resort in Krabi. Each one runs the full pipeline — intent classification, cultural-context retrieval, property data tools, drafted reply, human operator approval — and you can replay every step.
Hero scenario
Direct booking inquiry, Thai.
A guest asks in Thai about sea-view availability for three nights in July. The AI checks rates, retrieves Thai-language honorific conventions and VAT-disclosure rules, drafts a reply, and the receptionist approves.
Two costs are quietly eating Thai hotel margins.
OTAs take 15 to 20% of every room they send you. The only real defence is a direct-booking channel fast and credible enough that guests choose it over the OTA.
Guest messaging is the other half. At a Thai boutique property of 80 to 120 keys, that means 30 to 80 inquiries a day on WhatsApp and LINE, in Thai, English, and Mandarin. It quietly turns into a full-time job, or the job that doesn’t get done in time to keep the booking.
The Thai-market AI tools that should solve this don’t. Most have no LINE. None have Thai cultural context. We built the one that does.
Cameo scenarios
Pre-arrival airport transfer, Mandarin.
A guest arriving tomorrow asks in Simplified Chinese to arrange airport pickup. The AI looks up the transfer SOP, retrieves notes about Krabi airport routes and ferry-timing, drafts a Chinese-language confirmation, and the operator approves with no edits.
MICE corporate retreat, English.
A corporate buyer asks about a 30-room retreat with meeting space. The AI knows this is out of its auto-quote scope, retrieves the MICE-handoff template, and drafts a structured proposal-request reply for human follow-up.
Real cultural retrieval
k=4 RAG over an 11-document corpus on Thai honorifics, VAT phrasing, mainland-vs-traditional script signals, MICE register.
Of 11 surveyed vendors, none of HiJiffy, Whistle, Asksuite, Canary, Akia, Bookboost, Myma.ai, Duve, or Quicktext carry this.
Real tool use
getAvailability, getRate, getPolicy, getSop — every call, argument, and result is inspectable inline.
Runs on your data contract, not a vendor’s PMS lock-in.
Real human approval
Captured operator approve / edit / reject decisions. The diff between the AI draft and the sent message is shown.
Every send is human-approved. Your brand voice, your judgement — not an autonomous chatbot.
A reference build, not a deployed product.
AI Front Desk runs end-to-end on real Claude, real Voyage embeddings, and a real receptionist console. Three captured conversations with a synthetic 100-key resort prove the methodology. No hotel is in production on it yet. The first one to deploy isn’t buying a proven SaaS. They’re partnering with Tres on the first real-world rollout of a reference architecture we’ve spent 50 hours building and stand behind.
Pilot partner gets: a fully configured pipeline for their property, a six-week AI Operations Diagnostic to define the data contract and tune the cultural corpus to their brand voice, mandatory shadow mode during rollout (no AI message goes live until the team is confident), and ongoing operating support at pilot-partner rates.
In return: honest feedback during the rollout, and permission to publish the deployment case study. Anonymised or named, the partner’s choice.
How the pipeline works →
Cultural retrieval, tool use, and human approval — the three things that make this different.
11-document cultural corpus →
Honorifics, holidays, MICE register, channel conventions. Markdown source on GitHub.