Conversational AI for hotels: what an AI concierge can and cannot do

Conversational AI is a category, not a widget.

A hotel with a chat bubble on its website often assumes it has bought conversational AI. It has bought one surface. The category is wider and mostly invisible: which channel the guest chose to start on, whether the system understood what was actually being asked, what it is permitted to say, and what happens at the moment it should stop talking.

Those parts can be strong or weak independently, which is why two hotels running the same product can have entirely different experiences of it. A polished widget with a thin knowledge base produces confident nonsense. An excellent knowledge base behind a handover that lands in an unwatched inbox produces a guest who waited nine hours for nothing. The product is rarely the variable that matters most.

It is worth mapping the parts on paper before a demonstration, because a demonstration only ever shows the surface.

An AI concierge promises more than most products deliver.

A human concierge does three things. They know the guest, they know the city, and they act on the guest’s behalf: the table is booked, the car is arranged, the quiet side of the building is remembered without being asked for again.

Most software sold as an AI concierge does a version of the second. It answers questions about the property and the area from information the hotel supplied. Knowing the guest requires a reliable identity and a stay history, which usually sits in a different system. Acting on the guest’s behalf requires permission to change something, which most implementations sensibly do not grant.

This is not a complaint about the software. It is how to scope it. When a proposal uses the word concierge, ask which of the three jobs it performs, what it needs in order to perform them, and what it does when that information is missing. The answers turn a category name back into a specification you can hold a supplier to.

Answering in a language nobody on the team reads.

A model will answer a German enquiry in German whether or not the hotel’s knowledge base was written in German, and whether or not anybody on duty can check what was said. Two questions are being answered at once—is the fact correct, and is the German correct—and in practice only the first is ever reviewed.

The exposure sits in policy. A cancellation term, a deposit rule or a statement about accessibility is a commitment made in the hotel’s name. Translated on the fly at 23:00, it is a commitment nobody has read. The workable position is to translate the wrapping and not the rule: keep a short approved form of each material policy in every language you actually receive enquiries in, checked once by somebody who reads it, and let the model handle the conversation around it.

Then keep a readable record. If the team cannot read the transcripts, the non-English conversations become the ones nobody ever reviews—which is precisely the traffic operating with the least scrutiny. A translated view of the thread for the morning review is a small feature with a large effect.

• Approved wording for cancellation, deposit and payment terms.
• The fallback message used when there is no approved answer.
• The disclosure that the guest is speaking to an AI assistant.
• The out-of-hours promise, including the hour a person replies.
• Date, time and currency formats, which change meaning between markets.
• The sentence that hands the guest to a member of staff.

Voice is a different product with the same name.

Everything that makes text forgiving is absent on a telephone call. The caller cannot scroll back, so a wrong figure is simply wrong and gone. There is no interface to place a disclosure in, so it has to be spoken first, in a sentence short enough not to irritate. Silence carries meaning: a two-second pause that reads as thinking in chat reads as a dropped call on a phone.

People also interrupt. A voice system that cannot be spoken over mid-sentence will be hung up on, and the calls it loses are the impatient ones—which in a hotel are disproportionately the ones about tonight.

Voice earns its place on calls that are essentially routing, and on an overnight line at a property with no night porter. It struggles with anything the caller has to write down, with a name that needs spelling, with a booking reference mixing letters and digits, with an unfamiliar accent over background noise, and with payment of any kind.

• Read the spoken disclosure aloud and time it in seconds.
• Talk over the system mid-sentence and see whether it stops.
• Give a hard-to-spell surname at normal speaking speed.
• Read out a booking reference that mixes letters and digits.
• Call from a car, a bar and a station concourse.
• Transfer to a person, and listen to what that person hears first.

Where conversational AI genuinely fails.

Every category has a list of things it does badly, and a supplier who cannot produce one has not run enough live deployments. This is the list worth having in front of you during a demonstration.

Read the transcripts before widening the scope.

Containment rate—the share of conversations closed without a person—is the figure every supplier reports and the least useful one to optimise. It can be improved by making the assistant harder to escape, which is the opposite of what a hotel wants. Treat it as a description, never as a target.

A better routine takes half an hour a week and needs no reporting tool. Read ten conversations chosen deliberately rather than at random: three that were handed over, three the guest abandoned mid-thread, two the system answered with low confidence, and two in a language nobody on the team reads.

• Enquiries that arrived outside staffed hours, and what became of them.
• Questions that produced the fallback message, grouped by subject.
• Handovers not picked up inside the promised time.
• Conversations where the guest asked the same thing twice.
• Contact details captured that became a real enquiry.

Transcripts are personal data the moment they are useful.

The reading routine above only exists because conversations are stored, and a stored conversation containing a name, a booking reference or a phone number is personal data. The ICO’s guidance on AI and data protection sets out the questions to settle before that record exists: the purpose it serves, the lawful basis, who inside the hotel can see it, which supplier holds it and how long it is kept.

Retention is the item most often left undecided, and it is a five-minute conversation before launch and an awkward one afterwards. A hotel that keeps every transcript indefinitely because nobody chose a period has made a decision by accident. Ninety days is a defensible answer; so is a year, if there is a reason behind it. Having no answer is not.

The purpose of all this reading is not to score the supplier. Every fallback message is either a page that fails to explain something or a decision the hotel has not made, and both are cheaper to fix than the software is to replace. Widen the assistant’s scope only when the current scope has stopped producing that list.