What an AI receptionist can actually do — and what it can't

Most of what is written about AI receptionists is written by people selling them. This is too, so start from the part that is usually left out: the limits.

An AI assistant on WhatsApp is not a replacement for your front desk. It is a filter. It removes the twenty identical questions a day that nobody enjoys answering, so the people at the desk can do the part that actually requires a person. If a vendor tells you it replaces staff, they are either overselling or building something you should not want.

What it genuinely does well

The same questions, endlessly. What time is check-in, what is the wifi password, where do I park, is there a lift, can I leave my luggage, are dogs allowed. In an apartment or a small hotel these are most of the inbound volume, and every one of them has a fixed, correct answer that already exists somewhere in writing.

The hours nobody covers. A guest landing at midnight who cannot find the entrance is the single most expensive moment in hospitality: it produces a phone call to a private number, a bad night, and often a review that mentions it. An assistant that answers at 00:40 with the right instructions costs almost nothing and prevents all three.

Languages, without effort. A property in a tourist city receives messages in a dozen languages. The assistant reads the guest's language and replies in it, drawing from information written once in one language. This is where the gap between an AI assistant and a set of canned replies becomes obvious.

Consistency. Three people at the desk give three slightly different answers about the ZTL. The assistant gives the same one every time, and when the rule changes you update it in one place.

What it must never do

This list matters more than the previous one, and it is the part to interrogate a vendor about.

It must not invent. If the answer is not in the property's own information, the correct behaviour is to say so and hand over to a person. A system that guesses plausibly is worse than no system, because a confident wrong answer travels further than silence.

It must not hold access codes. Keybox and door codes should not be in the assistant at all. Not filtered, not restricted — absent. If the code is not in the system, no amount of insistence, social engineering or clever prompting can extract it. Any vendor who tells you their assistant "won't share" the code is describing a rule, not a guarantee.

It must not confirm bookings, change dates or touch payments. These are commitments with financial consequences. The assistant can collect the request; a person decides.

It must not receive identity documents in chat. This is the one that gets overlooked most often. When a guest photographs their passport and sends it, that image lands in the chat history, in the system logs, and potentially in the model's context. The right design intercepts attachments before they reach the model and redirects the guest to a secure upload link.

It must not pretend to be human. Beyond the ethics, this is now law in Europe. Under Article 50 of the EU AI Act, applicable since 2 August 2026, a system interacting with people must be designed so they are informed they are dealing with an AI, no later than the first interaction. Note the wording: designed so. An instruction in a prompt is a probability, not a design. The disclosure has to be produced by the system itself.

The design questions to ask a vendor

Four questions separate a serious product from a demo that looks good.

Where does the assistant get its facts, and what happens when the answer isn't there? If the honest answer is "the model figures it out", walk away.

Are access codes stored in the system? The right answer is that they never enter it.

What happens when a guest sends a photo? The right answer is that it is stopped before the model sees it.

How is the AI disclosure generated? If it lives only inside the prompt, it is not reliable.

What guests actually ask

There is a gap between what hotels expect the assistant to handle and what it really receives. Very few guests want a conversation. They want one fact, quickly.

Parking is almost always first, especially in cities with restricted traffic zones, where the honest answer is complicated and expensive to get wrong. Then arrival instructions, then wifi, then luggage storage before check-in and after check-out. Restaurant recommendations come later, during the stay.

The practical consequence is that the quality of an AI receptionist has very little to do with the model and almost everything to do with how well the property's information is written down. An assistant with a thorough, specific description of the parking situation on that street outperforms a more sophisticated one with a vague paragraph. This is unglamorous, and it is the whole job.

What it will not fix

It will not improve a property that has a real problem. If the apartment is noisy, the assistant will explain the noise politely and the review will still mention it.

It will not increase direct bookings on its own. It answers people who have already booked, or who are already talking to you. It improves the experience and frees up time, which is worth paying for, but it is not a distribution strategy.

And it will not remove the need for a human. It reduces how often one is needed, which is a different and more honest claim.

The reasonable expectation

A well-built assistant handles most routine questions, at any hour, in the guest's language, and passes everything else to a person with the context attached. The reception stops answering the same question for the fortieth time and starts answering the ones that need judgement.

That is a smaller promise than the one usually made, and it is the one worth buying.

---

Published July 2026. The AI Act references reflect the rules in force at the time of writing; we review this article every six months.