AI concierge for hotels: what it can and can't do

GuestJourney editorial team, edited by Alon Shkuri

Published 1 November 2026 · Updated 2 November 2026 · 12 min read

Sam lifts a silver cloche, and a small white star glows underneath

An AI concierge for a small hotel can answer the questions that keep coming back, at any hour, from your own information. It should not promise rooms, quote group prices or guess about allergies, because you may have to answer for what it says: an airline lost a tribunal case over its chatbot's wrong advice. So the question to put to any supplier is where it stops.

Three AI mails before breakfast

Anna opens her inbox in the breakfast room of Hotel Kadeloft while the coffee machine warms up. Twenty rooms, two people at the desk on busy days, and this week three companies want to give her an AI concierge. One she deletes on the guarantee alone. The other two she reads over the first cup.

She has one real question, and neither mail answers it. Last night at 23:40 someone emailed to ask about parking, and she replied at seven. What would one of these have said at 23:40, and who stands behind it?

Hotel Kadeloft, Anna and these three mails are made up. No real company sent them.

The name tells you little

AI concierge, digital concierge, AI front desk, hotel AI agent: the words in the mails differ, and they don't tell you what you would be buying. Hotel Dive reported in 2025 on one platform that bundles four agents, among them a virtual concierge and a booking agent, and that can "answer guest calls and act on what it's told", for example by cancelling a booking. A chat on a website that only answers from the site's own text can carry the same label and does far less. Three questions sort them out:

Where does it live? On your website before the booking, on WhatsApp during the stay, on the phone line, or on a tablet in the room.
May it act, or only answer? Telling a guest where to park is one thing. Cancelling their booking is another.
Where do its answers come from? Your own text, your booking system, or the open internet.

This guide is about the kind a hotel of twenty rooms would start with: something that answers guests from your own information and hands you the rest.

What it can take over, and what stays with you

For a hotel of about twenty rooms we would split it like this. It is our judgment, and it is also how we built our own chat, so weigh it with that in mind.

AI can take over

Parking, breakfast, late arrival, directions
The same question in another language
Sending a guest to your booking page
A first draft of a review reply or an update, for you to check

Stays with you

Complaints
Group prices and special requests
Anything that commits the hotel, including a room or a price it cannot see
Allergies and medical questions, unless it repeats your own words
Deciding what gets published in your name

One thing belongs on neither list: it should not pretend to be a person.

At Kadeloft the split works like this. A guest asks whether the dog can come. The house rules say €15 a night, not in the breakfast room, so it can answer. A guest asks what twelve people would pay in June. That commits Anna, so it goes to Anna. A guest says their daughter has a nut allergy. If Anna has written down what her kitchen can promise, it repeats that word for word. If she hasn't, it goes to her.

What a good stop sounds like

These three are examples for Kadeloft, made up, of a chat that knows where its text ends.

"What would twelve of us pay in June?"

I can't give a price for a group. I'm asking Anna, who runs the hotel. Which address should her answer go to?

It does not guess, and it says who will answer and how.

"My daughter has a nut allergy, is breakfast OK?"

I don't want to guess about an allergy. I'm passing this to Anna so she can check with the kitchen. Which address should her answer go to?

Without a confirmed line from the owner, no reassurance at all.

"Do you have a room on Friday?"

I can't see which rooms are free. The booking page can, and it's one tap from here.

No promise, and the button goes to the hotel's own page.

For the detail on questions, handover and privacy, read what a website chat should answer. Run the six test questions there on any product before you buy.

Air Canada found out who answers for it

In November 2022 a customer asked the chatbot on Air Canada's website about bereavement fares. It said a ticket could be submitted for the reduced rate within 90 days of being issued. A link in that same answer went to a page saying the policy did not apply after travel. The customer relied on the chatbot, tried for months to get the difference back, and went to the Civil Resolution Tribunal in British Columbia.

Air Canada argued it could not be held liable for information from one of its agents or representatives, including a chatbot. In effect, the tribunal said, that made the chatbot a separate legal entity. It called that "a remarkable submission", and wrote:

It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot.

Civil Resolution Tribunal, Moffatt v. Air Canada, 2024 BCCRT 149, paragraph 27.

It added that Air Canada did not explain why customers should have to double-check one part of its website against another, and ordered the airline to pay C$812.02, of which C$650.88 was damages. It is one small-claims case in one Canadian province, so not a rule for your hotel. We would still run a chat on the assumption that its words are yours.

Three things follow for Anna:

It answers from text she wrote, and says it does not know when that text has no answer.
Her own pages agree with each other. The chatbot and the bereavement page said different things, and McCarthy's summary draws the lesson: accuracy across all customer interfaces. If the house rules say breakfast until ten and an old PDF says half past, a chat can find either. Clean that up before you switch anything on.
She decides now what happens when it gets something wrong. Our advice: honour it if you can, and fix the text the same day.

What the AI Act asks, and of whom

In the EU, Article 50 of the AI Act says providers must make sure that AI systems meant to talk to people are built so that those people are told they are dealing with an AI, unless that is obvious to someone reasonably well informed. It has to be clear, and it has to reach them at the latest at the first interaction. This has applied since 2 August 2026, and the AI Omnibus did not delay it.

The word in the text is providers, so as we read it the duty sits first with your supplier. It is still your website and your guest. Ask to see what the first message says. And tell them anyway: it sets the right expectation and costs you one line. We are not lawyers, and we did not check what applies outside the EU.

Whether it is worth it, in nights

Honest answer: we do not know of independent figures on bookings from an AI concierge at hotels of this size. The figures we have seen come from people selling them, and we are one of those people. Adrien Mathieu, co-founder of Kleio.ai, told The Future of Hospitality, the podcast GuestJourney sponsors, that with a chat built into the site they see "some massive jump in time of user engagement from a couple of percent of visitor engaging to a double digit percent of visitor engaging". That is engagement reported by a vendor, not bookings.

So do the sum on your own numbers. At Kadeloft, a night that moves from Booking.com to direct saves €16.50 in commission. Take off an assumed 3% of €110 for the booking engine and card fees, and €13.20 is left. Every €100 a product costs per month then needs about eight nights a month to move (100 / 13.20 = 7.6). Kadeloft sells 146 nights a month through Booking.com, so that is about one in eighteen. We cannot tell you whether that is easy or impossible, and guests who would have booked direct anyway don't count.

A more modest reason holds up better, in our view. In a hotel with one owner and no night desk, a question at midnight waits until morning, and the answer may be one tab away on an OTA's page. We do not know how often that loses a booking. You can find out what comes in, though: for two weeks, write down every question that reaches you by phone and email. If most of them are already answered on your site, the first job is the site. Later, the questions a chat gets will show what your website is missing.

Where it breaks: after the chat

Klaus Spiekermann, General Manager at Dolce by Wyndham, put it this way on The Future of Hospitality:

it's nice to have some chatbot that answers your frequent asked question online and says, okay. Check-in is this. But then you arrive at the hotel, and the team, beyond the reception has no clue what you actually want, has no idea where you're coming from, and this is then where AI breaks down.

Klaus Spiekermann, in Independents vs chains in 2026, Lead the Experience. GuestJourney sponsors this podcast.

He speaks from inside a global brand, and a hotel of twenty rooms has the same risk in a smaller form. Laura tells the chat at four in the afternoon that she will not arrive until after eleven, and asks whether someone will be there. If that sentence stays in the chat, she is standing on the quay at eleven in front of a locked door.

So make sure what the assistant hears reaches you, in the place you already look, and decide who tells whoever is at the desk that night. Decide as well what can wait until morning.

What guests type into it

Guests type what they like: a passport number, a medical condition, the reason for a last-minute trip. Somebody keeps that. In the EU, Article 28 of the GDPR says that processing by a processor is governed by a contract, and that you use only processors providing sufficient guarantees. If a supplier holds guest conversations for you, we would expect them to be a processor, and we would ask for that contract.

Cookies are the other question. A chat does not need them to answer, and on 8 of the 20 hotel sites we opened, the cookie banner already lay over at least a quarter of the phone screen.

The rest of AI: it drafts, you decide

For a small hotel the same test applies everywhere: it may draft and prepare, and you decide. Drafts for review replies in other languages, an update for your Google profile, a change to your site. Good uses save you typing and leave you in charge. That is also how our own helpers start: Marco's first three replies come to you as drafts, Gigi shows her posts by email first, and Wes shows a preview. For more from the trade, The Future of Hospitality has an episode on where AI helps a revenue team and where it quietly hurts.

Sam sits on a phone in the evening, next to a lit lamp

What to send back to the next sales mail

Anna kept two of the three mails. These are the questions she sends back. The last column is where she stops reading. All of it is our view.

Questions to ask a supplier of an AI concierge, and what the answers tell you. Our view
Ask	A good answer	A warning sign
Where do its answers come from, and can I see and change that text?	From your site and what you teach it. Here is the screen where you change it.	It "learns from the internet", or you can't see the text.
What does it do when it does not know?	It says so, and asks you.	It always has an answer.
What does it do with a group price, a complaint or a nut allergy?	It stops and passes it on. On the allergy it repeats your own words or nothing.	"It handles everything."
How does a question reach me, and how does the answer reach the guest?	A named channel, at set hours, and you can try it today.	A dashboard you have to remember to open.
Does it tell guests it is an AI?	Yes, in the first message. Here is the wording.	It signs off as a person.
May it book, cancel or promise anything?	Only what you switched on, and you can see when it did.	It can, and you find out from the guest.
Where does its Book button go?	To your own booking page. See direct bookings.	To a partner link or an OTA.
Does it set cookies, what does it keep about guests, for how long, and who signs the contract?	A list, a period in days, and a data processing contract.	"We are GDPR compliant", and nothing more.
Which of your figures are bookings, and who measured them?	Bookings, at hotels this size, and the name of whoever counted.	Engagement, time saved, or a percentage with no source.
How do I leave, and what do I take with me?	A notice period and an export of your questions and answers.	A long contract and no export.