What AI Actually Does in Hotel Representation

The short version

Most of what is sold to hotels as "AI distribution" is three different promises wearing one label: a system that answers questions, a system that matches properties to travellers, and a system that transacts — checks availability, quotes rates, takes bookings.

The first one works. The second one works in a narrow form and is routinely oversold. The third one is a booking engine with a new name, and if a vendor is promising it alongside the other two, that is the claim to press hardest on.

This is a guide to telling them apart, written by people building the first one.

The job AI is genuinely good at

A luxury travel advisor's day is full of questions with answers that already exist somewhere.

Do the sea-view suites have a bath or a shower? Does the spa do couples' treatments? What is the transfer from the airport actually like at nine at night? Is the kids' club open in November? Can rooms 12 and 14 connect?

Each of those has a correct answer that your team knows. Each currently costs an email, a time-zone delay, and — this is the part that hurts — sometimes a lost booking, because the advisor had a client on the phone and moved on to a property that answered.

This is a retrieval problem, and it is the thing language models are actually reliable at: take a body of material somebody wrote, find the relevant part, and put it into a sentence. Not judgement, not prediction. Lookup and phrasing.

Two properties of a system that does this honestly:

It answers only from your material. Not from a general model of what luxury hotels are like. If the answer is not in what you published, it is not in the answer.

It says when it does not know. This is the whole ballgame. A system that invents a plausible spa treatment because your competitors have one has not saved your team an email — it has put a false statement in front of an advisor who will repeat it to a client, and you will find out at check-in. An assistant that says "the property has not published anything about that" is more useful than one that is right 90% of the time, because the advisor knows which sentences to trust.

The claims worth refusing

"AI matching finds the right travellers for your property"

There is a real version of this and an oversold one.

The real version: an advisor describes what they need in ordinary language and the system searches property material by meaning rather than by keyword, so "somewhere quiet for a honeymoon, walkable to a village" surfaces properties whose descriptions are about that. Useful, and unglamorous.

The oversold version claims the system knows which advisors will book you, from their booking history and their clients' demographics. Be direct with any vendor selling this: where did the booking history come from? Advisor production data is exactly what the industry does not have — it is the thing this whole problem is about. A matching engine trained on data nobody possesses is a ranking with a story attached.

Our own version is narrower than we would like: what a hotel searches is a directory built record by record, filtered by destinations placed, clientele and location. Real, and not a machine that reads a client brief and returns a shortlist. That one is being built and is not here.

"Automated availability, rate quotes and booking"

This is the claim to check against a live screen before you sign anything. Answering "what is included in the honeymoon package" from a document and answering "is a sea-view suite free on the 14th" are unrelated engineering problems: the second needs a live, correct connection to your inventory, and getting it wrong quotes a room you do not have.

A system doing document retrieval can be built in months. A system safely quoting availability is a booking engine and needs PMS integration, and no amount of language model changes that. We do not do the second one. Skipper answers questions and stops there; the booking happens where it always happened.

"Verified attribution, automatically"

Attribution is not an AI problem at all, and vendors bundling it in are hoping you will not notice the seam.

Knowing which advisor drove which booking requires a record created at booking time and checked against your system afterwards. Somebody records it, somebody verifies it. The value is real and the mechanism is bookkeeping, not intelligence. Any vendor describing attribution as something the AI does should be asked, slowly, who checks it, and against what.

An illustrative scenario

The following is an illustrative scenario, not a customer. The property and the situation are a worked example; the screens and the steps are the real product, and the figures belong to the scenario rather than to a measured result.

Say a 60-room resort in Sri Lanka. Two people handle trade sales. Their standing complaint is not that advisors do not know about them — it is that answering advisors properly takes about a day, and by the time the reply goes out the client has been sent three other options.

What they load. Room types and how they genuinely differ, not the brochure paragraph. What each rate includes. The three experiences they are actually known for. Their cancellation terms. And then, because someone finally wrote them down, the fourteen questions the trade desk answers every single week — the transfer, the monsoon months, whether the villas suit multi-generational groups, what "half board" covers there.

That last file takes an afternoon and is the one that does the work.

What changes for the advisor. An advisor in Delhi with a client on a video call asks whether the two-bedroom villas have a private pool and how far the beach really is. She gets an answer in the property's own words with the source attached, and keeps the client in the conversation. She does not wait a day. Nobody at the resort was involved.

What changes for the trade desk. The questions that reach them are the ones worth their time: a group of eighteen, a wedding, an advisor asking whether the GM will hold four villas over Diwali. The transfer question stops arriving.

What they can measure, and what they cannot. They can see which advisors asked about them, which offers travelled, and — once bookings are recorded and confirmed — which advisors actually placed guests. Say four advisors ask regularly and two of them place clients. That is a real finding they did not have before.

They cannot see how many bookings the faster answers caused. Nobody can. The claim worth making is the narrow one: an advisor who gets an answer in a minute stays in the conversation, and an advisor who waits a day sometimes does not.

⚠ Two things the scenario deliberately does not include. No availability was quoted and no rate was confirmed by a machine — every booking went through the channel it always went through. And no percentage improvement is claimed, because the property has not run a controlled comparison and neither have we.

Questions to put to any vendor

Where do the answers come from? If it is anything other than material the hotel published, ask what happens when the model is confident and wrong.
Show me it saying "I don't know." Ask on a live screen, about something the property has not published. If it always produces an answer, that is the defect, not the feature.
Does it quote availability or rates? If yes, ask to see the PMS integration, not the chat window.
Who verifies an attributed booking? If the answer is "the system", ask against what.
What is your matching trained on? If the answer includes advisor booking history, ask where they got it.
What does it refuse to do? A vendor with no answer has not thought about the failure modes, which is the entire risk with this technology.

Where we actually are

Skipper answers advisor and hotel questions from hotels' own published material, with sources, and says so when a hotel has published nothing. It does not check availability, quote rates, or book.

Penguins runs in one market, India: over 800 advisor agency records we researched and built ourselves — collected records rather than sign-ups — and over 400 luxury properties in the catalogue an advisor can research. Hotels also get a market chapter covering entry rules, trade movement and demand, with every signal dated and checked by a person before it appears; where a category has nothing verified, it says so instead of printing something plausible. Entry for advisors is by invitation and a directory claim, not an open sign-up.

Want to test the "I don't know" behaviour yourself? Ask us for a walkthrough and try to make it invent something. That is the demo worth having.

What AI Actually Does in Hotel Representation (And What It Does Not)