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

An AI concierge can answer most of what hotel guests ask, in any language and at 3am, and it can take a booking when it reads live availability. It should not decide anything that needs judgement: complaints, refunds, accessibility, safety, and exceptions to policy go to a person, with the conversation attached.

The difference between those two lists decides whether an AI concierge saves your team work or quietly creates more of it. Most guides describe the first list at length and mention the second in a sentence. This one gives them equal weight, because the second list is where hotels get hurt.

For scale: across more than 190 properties running vGuest, 97% of guest conversations arrive on WhatsApp, and around 70% of the replies guests receive are sent by the AI. The other 30% are sent by people, and that is not a failure rate. It is the part of the job that should stay human.

What an AI concierge completes without staff

The work it does well is repetitive, factual and answerable from information the hotel already controls.

• Questions about the property. Check-in and check-out times, breakfast hours, parking, the pool, Wi-Fi, pets, luggage storage, directions. The answer comes from the hotel's own knowledge base, not from the model's general idea of what hotels usually do.
• Questions in languages the desk does not speak. A guest writes in Portuguese at midnight and gets a correct answer in Portuguese.
• Pre-arrival logistics. Sending the check-in link, confirming arrival time, explaining how to find the entrance.
• Simple service requests. Extra towels, a late check-out request, a taxi, logged as a task for the right department rather than answered with a promise.

Each of these has a right answer that does not change from guest to guest. That is the test: if two guests asking the same thing should get the same reply, the AI can send it.

What it can book, and what that depends on

An AI concierge can quote and take a booking only when it reads availability from a live source: the PMS, the booking engine, or a connected calendar. Without that connection it is guessing.

The failure is quieter than it sounds. A hotel or attraction types its availability into the assistant's settings once, and the list is accurate on the day it is written. From then on it drifts, fastest over holidays and long weekends, exactly when accuracy matters most. We have written up why manually entered availability is a trap and what connecting a calendar actually takes, because it is one of the most common questions we get from venues.

A wrong "yes, Saturday at 2pm is free" costs a complaint and a refund conversation. "I don't have that to hand, let me connect you" costs the guest thirty seconds. If the assistant cannot read live availability, the second answer is the right one.

What it must hand to a person

Some requests should never be closed by an AI concierge, however good the model is. Write them down as a rule, not as a setting someone switches off in a busy week:

• complaints, and any request for a refund, discount or compensation
• accessibility needs, medical conditions and allergies
• anything involving safety, a lost child or an incident in a room
• group bookings, events and contracts
• any promise that makes an exception to policy

Our list of the guest requests you must not automate goes through each one and how to write the rule so it survives staff changes. Deciding what is on the list is a management decision, not a technical one.

How a good handover works

The handover is the product. A guest who has explained a problem once and then has to explain it again to a person has had a worse experience than if there had been no assistant at all.

A good handover does three things:

1. It carries the whole conversation, plus the booking it belongs to, so staff start from context rather than a blank message. 2. It lands where the team actually looks: the shared inbox the front desk keeps open, or a ticket in the tool maintenance already uses. A notification nobody watches is not an escalation. 3. It tells the guest what happens next and roughly when, instead of going silent.

We covered what an escalation should carry and how to route it to the right person, including the cases where the right answer is simply the desk's phone number.

Where this breaks

• A stale knowledge base. The AI will confidently repeat last season's breakfast hours. Someone has to own the information, with a date to review it.
• Invented answers. An assistant that is not restricted to the hotel's own information will fill gaps with plausible fiction: a shuttle that does not run, a fee that does not exist. The safe behaviour is to say it does not know and hand over. Hotel Tech Insight's practical guide to AI concierges makes the same point about refusing unknown answers.
• Nobody reviews the handovers. If failed conversations are not read, the same failure repeats every day and never shows up in a dashboard.
• Measuring the wrong thing. "The AI answered 5,000 messages" says nothing about whether the answers were right.

How to test an AI concierge before guests see it

Test the unhappy path first. Before launch, send it the ten questions your desk answers most, then the ones that go wrong: a complaint, a wheelchair access question, a request for a sold-out date, a message in a language nobody on the team speaks. Check what it says and where each handover lands.

Then launch on one channel, usually WhatsApp, and read every handover and every correction for the first weeks. Our guide to what a 90-day guest messaging pilot should prove sets out the four numbers worth tracking and the decisions that settle the outcome.

When you compare vendors, ask each one to show the handover and the "I don't know" behaviour live, not just the happy path. Our comparison of AI guest communication platforms and our breakdown of [what drives hotel chatbot pricing](/blog/hotel-chatbot-pricing-what-actually-drives-the-number) cover the rest of the buying questions.

What to do first

Write two lists on one page: the questions your team answers every day that have one right answer, and the requests you would never let software close. The first list is what an AI concierge should take over. The second is your handover rule. Any vendor worth piloting should be able to show both working on your own property before you sign.