Multilingual AI Concierge
An Austrian 4-star city hotel wanted to give guests a faster way to reach the front desk without the language barrier. We built an AI concierge that handles guest requests in eight languages via WhatsApp and room QR codes, integrates with the Mews PMS and the room-service POS, and escalates anything sensitive to a duty manager on shift.
Background
The hotel has roughly 110 rooms, two-thirds international guests, peak occupancy above ninety percent during trade-fair seasons. The front desk regularly received the same requests in broken English at 23:30 — extra towels, late check-out, advice on a restaurant nearby, a taxi to the airport at 5:00 a.m. — and the receptionist on the night shift either knew the answer or had to wake up someone who did. Guests using a non-native language often phrased requests three or four times before being understood, which slowed everyone down and produced avoidable friction in review channels. The hotel director had already trialled a chatbot from one of the larger PMS vendors but found it too rigid and English-only. He wanted something that felt like a competent front-desk colleague in the guest's own language, that could see what was actually in the PMS, and that knew when to step back and hand a call over to a human.
Solution
We built a concierge agent that reaches guests through two channels: a WhatsApp Business number shared on check-in, and a QR code in each room that opens the same conversation in the guest's preferred language. The agent supports German, English, French, Italian, Spanish, Czech, Polish and Arabic. It is built on Claude, with conversational orchestration in Langflow and integration glue in n8n. PMS integration is bidirectional with Mews: the agent can read reservation context (room, length of stay, package booked) and write back simple actions such as late check-out requests for staff to confirm. Room-service orders flow into the kitchen POS via a small middleware service. All conversation data sits in PostgreSQL on Hetzner Frankfurt, encrypted at rest, with a documented Article 28 processing agreement. The agent is explicitly conservative: anything touching billing, complaints, medical issues or visibly upset guests is handed straight to a staff member with a one-paragraph context summary.
Outcome
Roughly sixty percent of inbound guest messages are now resolved end-to-end by the agent without staff involvement. Night-shift volume on the reception phone has measurably dropped: typical late-evening interruptions for towel requests or breakfast-time questions are gone. Average response time on guest WhatsApp messages is well under one minute, including in Arabic and Polish, which the previous chatbot did not handle at all. Booking review scores on language-related comments have improved noticeably in the first three months. Staff have reframed the concierge in their daily standup as "the polite colleague who covers the obvious stuff" — that adoption framing turned out to matter as much as any single feature.
Lessons Learned
Two things stand out. First, multilingual quality is not just a matter of model choice; it is a matter of grounding. Our first prototype produced linguistically clean German answers that quietly invented a swimming pool the hotel does not have. We rebuilt the system prompt and tool-calling layer so that anything factual about the hotel — opening hours, amenities, restaurants, transfer prices — has to come from a vetted knowledge document, not the model's general knowledge. That single change made the agent trustworthy. Second, escalation paths matter more than reply quality. The hotel director was clear from day one that he would rather see the agent step back too often than overstep once. We built a small set of escalation triggers (sentiment, billing keywords, medical hints, anyone asking for the manager) that route the conversation to a staff WhatsApp group with one-tap take-over. Trust with the staff was earned in the first two weeks largely through that escalation behaviour, not through any clever language feature.
A similar problem?
Drop us a short note and we will reply within one business day.
Request a project