Lisa, AI concierge

The problem

A small villa or hotel operator spends much of the day relaying messages. Guests write at any hour, often in languages the team does not speak. Each request has to be passed to a driver, chef, therapist or housekeeper, usually in a group chat in Thai, and then the answer has to be passed back to the guest. Details get lost along the way: a pickup time, a flight number, a room.

What we built

Lisa is the AI concierge of Kove Samui Beachfront Villas, a related Kove business with two beachfront villas on Koh Samui. She takes over that relay work. She answers guests immediately, books services with the complete details, tracks each booking until the provider confirms it, and reminds providers before the job. The host keeps oversight and steps in only when judgement is needed.

Lisa is an AI agent rather than a scripted chatbot. She reads the whole conversation, the guest’s booking and everything already arranged for them, decides what to do, and acts through about 20 tools, such as checking availability, creating a housekeeping ticket, sending a booking to a provider group or looking up a destination.

For guests, on WhatsApp

• She answers questions about the villas, amenities, house rules, check-in and check-out, and local recommendations, from a knowledge base the host maintains.
• She checks live availability and rates from the property management system, including minimum-stay rules.
• She books airport and local transport, private chef meals (one meal or several across a stay), and in-villa massage and nail treatments.
• She takes housekeeping and maintenance requests and tells the guest when staff have acknowledged or completed them.
• She replies in the guest’s own language. Guest messages so far include English, German, Thai and French.
• She understands photos and shared locations: a screenshot of a restaurant, a photo of a problem in the villa or a dropped map pin is handled like a text message.
• She answers travel questions using Google Maps: drive or walking time, when to leave, and a place’s opening hours. If a guest books a ride to a place that will be closed on arrival, she checks with them before the booking is sent.
• She only promises what the villa actually offers, checks with the host otherwise, and honours the complimentary airport transfers included with longer stays.

For drivers, chefs and housekeeping, in LINE group chats

• Every job arrives as a card with an English section followed by a Thai section, each with the booking number, and one-tap Accept and Unavailable buttons.
• An accepted job is confirmed to the guest. A declined job gets an automatic follow-up asking for other times. When several jobs are open, replies are matched by booking number.
• Changes and cancellations arrive as cards in the same two-language format, and if a provider declines a change, the original booking is restored.
• Short reminders go out before each job, with the lead time set per team, such as one hour before a transfer. Chef bookings with several meals get a reminder for each meal.
• Unconfirmed jobs are chased until a provider responds, and the host is alerted if nobody does.
• Providers can type /jobs to list their upcoming confirmed jobs, and the housekeeping team can list open requests.
• Messages in the group chats are translated between English and Thai as a quoted reply.

For the host

• Escalations and alerts go to a management group chat, and the host can answer a guest by quote-replying to an alert.
• The host can send Lisa a booking they arranged themselves. She forwards it to the provider and reports back when it is confirmed.
• Lisa can be paused for a conversation while the host handles it personally.
• Passport photos are read automatically and queued for Thailand’s mandatory TM30 guest registration, and filed with the TM30.io service when the host starts it from the dashboard.
• An admin dashboard shows searchable conversations, bookings, tickets, the knowledge base, provider group settings and diagnostics.

Voice

Lisa also answers phone and web calls about availability and bookings, in a natural voice and with live availability. Booking requests from calls are recorded and passed to the host.

How she is built

Context instead of scripts. Every message goes to a large language model together with the context a villa manager would have: the whole conversation with the time of each message, the guest’s booking and profile, open requests and service bookings, the knowledge base, and the current date and time. The model decides what to do; it is given context rather than hard-coded rules.

Safeguards written in code. Where a mistake would be costly, code checks the facts instead of trusting the model:

• Dates: weekdays are calculated from the date, never guessed.
• Flights: flight numbers are checked against the airlines that actually fly to Samui, and pickup times are compared with the scheduled arrival.
• Transport: van capacity, duplicate bookings on the same day and complimentary transfer allowances are checked, and a destination that will be closed on arrival stops the booking.
• Places: a drop-off must be a place name a driver can find, not a map code.
• Prices: all price arithmetic is done in code.
• Housekeeping: a room is only put on a ticket if the guest actually named it.

Channels. Guests use WhatsApp and voice calls. Service providers and villa staff use LINE group chats. The host uses a management group chat and the admin dashboard.

Reliability and privacy

• What is stored: the conversation, including a private copy of any photo a guest sends; a guest profile with preferences such as language, dietary needs and special occasions; bookings and requests; and passport details for guests who need TM30 registration.
• How long: a conversation stays active until 30 days after check-out, then moves to an archive, and a short stay summary lets Lisa recognise a returning guest.
• Passports: images are kept in private storage with no public links. Before a registration is filed, the filing service’s own reading of the passport must match the stored details, and any mismatch goes to the host instead. The host’s passport uploads are read only by the passport reader, never by the chat model.
• No lost messages: a message that cannot be saved is queued and saved again automatically, and if a guest’s history cannot be loaded, the guest gets a holding reply and the host is alerted, rather than the AI answering without context.
• Alerts: the host is alerted if a background process stops, saving or archiving fails, a TM30 filing fails, or a guest’s flight is cancelled or delayed by an hour or more.
• Testing: an automated suite of more than 400 tests runs before each release.

Result

Lisa has been in live daily use with real guests and service providers since March 2025.

Coming next

• A single inbox for messages from online travel agency bookings, answered with Lisa’s help. Built, not yet live.
• Asking guests for their passports ahead of arrival for TM30. Running in a test mode that sends nothing to guests.
• Keeping booking records in step with notes the host adds.
• Richer knowledge retrieval, an analytics dashboard and a feedback loop.