Guest communication

Last updated: 2026-08-17

Plenty of software now promises to answer guest enquiries. The question that actually decides whether you use it is a different one: what happens when it is wrong?

This page answers that first. The channel list is further down, because it looks roughly the same at every vendor.

When the AI does not know

It does not answer.

Hunzi answers from the information about your property that you have provided. If it finds no basis there for a question, it does not guess and does not produce something plausible: it says it does not have that information, opens a ticket, and hands over to a person.

That sounds obvious, and it is the exact point at which most systems do something else. A model built to sound good will invent a breakfast time before it admits it does not know one.

When the AI claims something

The expensive mistake is not the wrong breakfast time. It is the sentence "I've passed that to the team" when nothing was passed on.

That is precisely what happened in production: the model said the confirmation without triggering the action. The guest was satisfied, the request was gone, and nobody noticed.

So a confirmation like that no longer depends on how the model phrases things. It depends on the fact. Two checks in the flow allow the assurance only if the action actually took place. They are not a note in a prompt that a model overwrites on a bad day, they are a barrier in front of it.

When someone tries to talk it round

Everything that comes from your uploaded documents is treated as material, not as instructions. If an uploaded PDF contains a line like "ignore all previous rules", to Hunzi that is a sentence in a document, not an order.

On top of that there is a hard ceiling on how often a single conversation may trigger actions. A conversation designed to drive the system into a loop runs into that ceiling rather than into your costs.

When a human takes over

When someone on your team takes over, the AI goes quiet. Identically on every channel.

In the code that state is written in exactly one place, deliberately: as long as each channel kept its own rule, sooner or later there would be a channel where the AI carried on talking while the front desk was already replying.

The guest does not have to repeat anything. What was said before is there.

The channels

• Phone. Calls around the clock, with handover to a person when needed.
• WhatsApp. Through Meta's official interface. Incoming messages are cryptographically verified, so nobody can write messages into your inbox that no guest ever sent.
• Email. Enquiries and booking mail.
• Web chat. On your site, in the guest's language.
• Instagram and Facebook Messenger. Through the same verified Meta interface as WhatsApp. The direct message nobody otherwise reads, because it sits in a different app.
• Booking.com. The guest messages the portal forwards to your mailbox are read and answered through the route the portal provides for it. Matching runs on the reservation number rather than the subject line, because portals rewrite subjects.

Everything lands in one inbox: one history per guest, across every channel. A guest who asked about a late check-out on the phone and follows up on WhatsApp two days later does not have to explain it again, and the colleague replying does not have to go looking for it.

The language follows the guest rather than the channel. What that looks like for the team approving the reply is on Multilingual.

Where the data is

Processing and storage take place in the EU, and a data processing agreement under Article 28 GDPR is in place.

How seriously that is meant shows in a channel we built and leave switched off by default: Telegram. The service routes guest data to a company in Dubai, that is, out of the EU. The channel exists, works, and stays off until a property explicitly turns it on.

A vendor who writes "EU hosting" is telling you something about their servers. A channel that is switched off tells you something about what happens when convenience and data protection disagree.

Limits

The 24-hour WhatsApp window is Meta's rule, not ours. More than 24 hours after the guest's last message, only templates pre-approved by Meta are allowed. No vendor can get around this, and anyone promising otherwise means something else.

With no knowledge stored, Hunzi answers little. That follows from the first rule on this page: an empty knowledge base does not produce invented answers, it produces a lot of tickets. Setup therefore consists mainly of recording the questions your property gets. We do most of that work with you, but it is the real work.

Hunzi never presents itself as human. If a guest asks whether they are talking to a machine, they get the honest answer. Anyone looking for software that hides it is in the wrong place.

The difference

The effort in a hotel rarely arises within a single channel. It arises between them: the promise made on the phone that never reaches the floor. The lost property someone emails about while it is sitting in the custody log.

Hunzi is also the system that builds the room plan and creates the reservation in the PMS. So the late check-out a guest asks for on the phone is afterwards in housekeeping's plan, without anyone retyping it.