Multilingual AI Concierge
A three-star hotel in the greater Paris area welcomes international guests from six language groups. Reception was understaffed in the evenings and on weekends. We built a multilingual concierge agent that responds via website, email, and WhatsApp, escalates complex requests to the team via Slack, and handles translation in both directions.
Starting situation
The hotel sits in a business district outside Paris, twenty minutes from Orly Airport. Roughly 60% of guests are international — French, English, German, Spanish, Italian, Dutch. Reception is well staffed during the day, but in the evening and at night usually only one person is on. Inquiries via the website, the booking portal, and WhatsApp added up to several dozen per day, many of them repeat questions about arrival, breakfast, late checkout, and local restaurants. The team spent a substantial part of its time answering the same standard questions — time that was missing for personal service at the front desk.
Solution
We built a concierge agent that operates under the hotel's name and responds in six languages. The agent knows the full operational knowledge base of the hotel — location, directions, room amenities, breakfast hours, nearby restaurants with rating and distance, public transit, emergency contacts, on-call pharmacy, laundry service, pet policy. The technical architecture is based on Langflow for agent logic, n8n for escalation, and Slack as the team channel. Complex or uncertain inquiries — complaints, refunds, technical issues in the room, security topics — are escalated to the team automatically, with full conversation history and translation. A set of trigger phrases ("I want to speak with a person", "connect me with reception", and equivalents in all six languages) leads to immediate handoff.
Result
The agent handles a large share of routine inquiries fully autonomously. Response times for standard inquiries are under 30 seconds, around the clock. Reception reports noticeably fewer evening and night calls for trivial questions, which lets the team focus on the actual front-desk service. Exact percentages vary seasonally, which is why we do not publish concrete figures here — in the initial conversation we are happy to discuss the detailed metrics.
What we learned
Two things mattered more than expected. First: the escalation rules. A concierge agent that wants to solve everything itself frustrates guests in complaint situations — we had to expand the trigger phrases multiple times to reliably recognize indirect requests for human contact. Second: the cultural tone per language. French requires more polite greetings than English, Dutch tolerates more direct answers than Italian. A "neutral translation" is not enough — tone has to be calibrated per language group. Both extended the build time but stabilized the result.
Similar problem?
Send us a short note and we reply within one business day.
Request a project