AI for travel and hospitality
A hotel assistant that cannot read the PMS can converse but cannot book. And one that answers well in Spanish and badly in English looks fine on average and is useless to half the guests. We measure both separately.
Key takeaways
The most valuable query is the one that arrives outside reception hours, and it is exactly the one that today gets lost or answered the next day, when the guest has already booked elsewhere.
Availability and rates are queried live against the PMS, never from an index. A stale rate in the chat is an argument at the front desk.
Quality gets measured per language, not on average: a system that answers well in Spanish and badly in English reports a good overall number and loses the foreign guest, who is usually the higher-rate one.
The best-returning case is not the virtual concierge but the direct booking: every reservation that skips an online travel agency saves its commission.
The five leaks in a hotel
All five happen away from the front desk, which is where all the staff are.
- A good share of bookings gets decided outside reception hours, and there is nobody there. The booking is not postponed: it goes elsewhere.
- Check-in time, pets, parking, how to get here from the airport. High volume, fixed answers.
- The concierge who knows what to suggest to each guest exists in a twenty-room hotel, not a two-hundred-room one.
- Demand triples and staffing does not, so service quality drops exactly when the most revenue is on the table.
- That is where the upsell lives and where the review gets asked for, and almost nobody does it.
Two decisions that determine whether it works
The first is the PMS. Availability, room type and rate change by date and by channel, so they get queried live at answer time and never come out of an index. The semi-stable — policies, amenities, directions, FAQs — does go into the index. An assistant quoting an old rate does not generate a booking: it generates an argument at the front desk on arrival day.
The second is language, and almost nobody measures it. A hotel in Colombia receives guests in Spanish, English and sometimes Portuguese, and the foreign guest is usually the higher-rate one. A system evaluated with a Spanish-only set reports a high number and may be answering badly in English without anyone noticing, because those conversations are few and the average dilutes them. So the evaluation set is built per language and the number is reported per language.
And one operating rule: in hospitality the handoff to a person has to be fast and frictionless. A guest with a problem in their room does not want to converse with a system, and the assistant has to recognise that in the first message.
What we ask for before wiring it to the booking channel
| What is measured | What it means | How it gets fixed |
|---|---|---|
| Availability and rate accuracy | On a sample: whether what was said matched the PMS at that moment, for that date and channel. | Live queries, never the index. It is the failure that turns a booking into a complaint. |
| Quality per language | The same evaluation set run in every language the hotel serves, with the number reported separately. | The weaker language gets more examples, and the translated documentation in the index gets reviewed. |
| Attributable direct bookings | Those that started in the conversation and ended in your own engine, measured with a control group. | Without a control group, the system gets credit for the booking that was going to happen anyway. |
| Handoff speed | How many turns it takes to reach a person when the guest has a real problem. | Intent classification before retrieval: a complaint does not enter the FAQ flow. |
For a small hotel whose operation fits in one person’s inbox, this is over-engineering: a good FAQ page and a WhatsApp number somebody answers cover nearly everything.
The conversation changes with several properties, several languages, a marked season and dependence on online travel agencies. There, every point of direct booking recovered pays for the system, and that can actually be measured.
Related services
AI chatbots and assistants
With the evaluation set run before it ever reaches production.
View serviceIntegrations and APIs
MCP servers against the systems already in operation.
View serviceAI agents
LangGraph orchestration, durable state and human approval on steps with consequences.
View serviceAI for retail and e-commerce
Catalogue, inventory and order status live, not a script.
View serviceMore from the blog
Nearshore AI development rates in Colombia
Colombia is the cheapest place in the Americas to buy senior engineering that shares your working day.
Keep readingTwelve questions to ask an AI vendor
Every proposal reads the same.
Keep readingThe 15 best AI development companies in Latin America (2026)
This list came out of reading the public pages of more than eighty providers across Colombia and the region.
Keep readingFrequently asked questions
Does it connect to our PMS?
Yes, over API or an MCP server against whichever system you run, and that is what separates an assistant that books from one that only informs. When the PMS does not expose availability and rates in a queryable way, we say so before quoting: we would rather ship a smaller, honest scope — amenities, policies, directions — than an assistant that improvises a rate.
What languages does it answer in?
Whichever the hotel needs; in Colombia that is usually Spanish, English and sometimes Portuguese. What sets us apart is not that it answers in several languages — any model does that — but that quality is measured in each one separately. A Spanish-only evaluation set reports a high number and may be failing in English without anyone seeing it, and the foreign guest is usually the higher-rate one.
Does it increase direct bookings against the OTAs?
It can, and the only honest way to know is to measure it with a control group. Answering the query at the moment it happens — at night, in the guest’s language, with real availability — removes one of the reasons people end up booking through an agency. Any vendor promising a percentage increase before knowing your operation is making it up.
Can it handle check-in or payments?
It can prepare check-in: collect details, send the form, confirm arrival time and leave everything ready in the PMS. Payment we recommend keeping on a gateway link rather than inside the conversation, because handling card data in a chat adds a compliance problem that does not pay for what it saves.
What happens when a guest has a real problem?
It escalates fast, and that is designed and measured. A complaint does not enter the FAQ flow: it is classified before retrieval and routed to a person with the context of what was already said. The metric is how many turns that handoff takes. An assistant that insists on solving what it cannot turns an annoyance into a one-star review.
How long does it take and how is it priced?
Four to eight weeks to production depending on integrations and the number of languages, with something running from week one. It is quoted with fixed scope, price and date in a proposal 48 hours after the first call. The price is driven by the number of properties, the languages, and how accessible the PMS is.
Book 15 minutes
Tell us what you are building, or what stopped working. You leave the call with a concrete answer: it can be fixed, it can be built, or it isn't worth it.