An AI concierge is a guest-facing system that handles inquiries, bookings, modifications, and in-stay requests through natural language — typically on WhatsApp, sometimes web chat or voice. It's not a chatbot with scripted intents; it's an LLM-grounded agent that understands what guests actually mean and acts on it.
Done right, an AI concierge handles 70-85% of guest messages without staff involvement. Done wrong, it produces hallucinations, books non-existent rooms, and erodes guest trust. Here's how to deploy one without the typical six-month consulting engagement.
What an AI concierge should do
Three categories of work:
- Pre-arrival. Booking inquiries, room availability, pricing, payment, confirmation, pre-arrival upsells, special requests, ID collection.
- In-stay. Room service orders, housekeeping requests, restaurant reservations, local recommendations, complaints, late checkout.
- Post-stay. Folio explanations, payment links, reviews, repeat-stay offers.
It should NOT do: anything legally regulated (medical advice, legal questions), anything financially irreversible without human confirmation (refunds beyond a small threshold), or anything outside your hotel's actual offering.
Architecture decisions that matter
1. Grounding, not hallucination
The biggest failure mode of LLM concierges is making things up — a non-existent gym, a fictional pool, a room category you don't sell. The fix is retrieval-augmented generation: every answer is grounded in your actual hotel data (room inventory, F&B menu, amenities, policies) rather than the LLM's general world knowledge.
Hotelary's WhatsApp AI uses this pattern: every guest-facing answer is constrained to data the hotel has actually configured. The LLM cannot promise a sauna if you haven't entered a sauna in your amenity list.
2. LLM-based intent classification, not regex
The 2018-era chatbot pattern was: regex match the message, route to a pre-written script. This breaks on every variation: "101 cleaned" works, "room 101 is done" doesn't, "saaf ho gaya 101 ka" certainly doesn't.
Modern AI concierges use the LLM itself to classify intent and extract entities. "Cancel my booking for next Saturday" yields {intent: cancel_booking, date_reference: "next_saturday"}, regardless of phrasing or language.
3. Voice messages first-class
In India in particular, 30-50% of WhatsApp messages from guests are voice notes. A concierge that ignores voice messages effectively ignores half the inbound. Speech-to-text in the guest's language, then the same intent pipeline, is non-negotiable.
4. Confidence-based escalation
For low-confidence interpretations or sensitive scenarios (refunds, complaints with emotional language), the agent should hand off to a human staff member rather than guess. The escalation should be framed naturally — "Let me get you to my colleague who can help with this" — not as an error message.
The deployment playbook
Week 1: Knowledge base
- Document every room category, with descriptions, amenities, max occupancy, photos, and pricing.
- Document every F&B outlet, hours, menu, dietary options.
- Document every amenity (pool, gym, spa, parking) with hours and policies.
- Document local recommendations (top 10 restaurants, transport, landmarks).
- Document hotel policies (check-in time, cancellation, kids, pets).
This is the grounding data. The AI cannot know what you don't write down.
Week 2: Connect to operational data
- Real-time room availability from your PMS.
- Booking creation and modification.
- Folio status (charges, balance).
- Housekeeping room status.
- F&B order placement.
The AI's value comes from doing things, not just answering. A concierge that can't actually book a room is a glorified FAQ.
Week 3: Voice + multilingual
Wire speech-to-text. Test with native Hindi, Hindi-English mix, regional language samples. Calibrate the LLM prompt for code-switching ("kal ka rate kya hai?").
Week 4: Soft launch
- Enable on one channel (WhatsApp) for one segment (returning guests).
- Have a staff member shadow every AI response for the first 100 conversations.
- Collect every "wrong answer" and feed it back as either a knowledge-base correction or a prompt refinement.
Month 2-3: Expand and refine
- Open to all guests.
- Track resolution rate (% of conversations closed without staff).
- Track CSAT or simple thumbs-up/thumbs-down on the AI response.
- Re-train weekly based on new data.
Metrics that matter
- Resolution rate. Target: 70-85%.
- Median first-response time. Target: under 5 seconds.
- Booking conversion from inquiry. Target: 60-85% (compared to 15-30% for slow human-only response).
- Hallucination rate. Target: under 1% of responses contain factual errors. Audit weekly via random sample.
- Escalation rate. 15-30% is healthy. Higher means the AI is over-cautious; lower means it's over-confident.
What it costs
The actual LLM inference cost for hotel-scale conversation volume is small — typically ₹0.20-0.80 per conversation on modern models. The bigger costs are integration, knowledge-base preparation, and ongoing tuning. Budget ₹1-3 lakh for a basic deployment, ₹5-15 lakh for a production-grade one with full PMS/POS integration.
Or use a turnkey product like Hotelary's WhatsApp AI, which ships with the integration, knowledge-base structure, and tuning loop already built. Independent hotels typically reach 70%+ resolution within two weeks.
External references
- Skift's AI in hospitality coverage for industry-level trends.
- Hospitality Net for case studies and vendor analysis.
- For the broader context on AI adoption in hospitality, see our AI guest experience piece.
The hotels that deploy AI concierges well in 2026 will compound an enormous operational advantage over the next three years. The ones that don't will spend the same staff hours answering "what time is breakfast" for the rest of the decade.
