Agentic AI for hotels: how an AI concierge handles 85% of guest messages

An AI concierge handles most hotel guest messages because most guest messages are the same twenty questions asked a thousand ways. Check-in time, parking, late checkout, the nearest pharmacy, the Wi-Fi password. askHermis, which we built, answers 85% of guest queries instantly and routes the other 15% to a person. That split is the product. A concierge that knows what it does not handle is the one a hotel can rely on.

What makes a concierge "agentic" rather than a chatbot?

A chatbot follows a script. An agent carries out a task. The difference shows up the moment a guest writes "we land at 11pm, is that a problem, and can we leave the bags somewhere Friday" in one message. A decision tree has no branch for that sentence.

An agent reads the whole message, in the guest's language, and works out that there are two requests inside it. It pulls the booking. It checks the property's late arrival policy and its luggage rule. It drafts one reply that answers both, and it sends it on the channel the guest wrote on. askHermis does this across eight channels in one inbox, WhatsApp, Instagram, Messenger, email, Viber, web and more, so a guest who starts on Instagram and follows up on WhatsApp is one conversation to the hotel rather than two tickets.

The mechanism is grounding. Every answer is drawn from the property's own knowledge, its policies, its amenities, its directions, its house rules, and the agent is told what is outside that set. The model supplies the reading and the writing. The hotel supplies the facts.

What happens when it is not sure is the part to look at before anything else. A message that matches nothing in the grounded set, or that mentions money, safety or a complaint, is not answered. It is handed to a person. It is plain work, and it is the work that separates a concierge a hotel keeps from one it switches off after the first bad review.

Where does the 85% come from?

From scope. The questions a property receives most often are the ones the agent is given grounded answers for. At most hotels a short list covers the majority of traffic: arrival and departure times, amenities and their hours, directions and transport, booking changes, house rules.

askHermis publishes the figures for what happens after that: 85% of guest queries answered instantly, a 1.2 second average response time, 50x faster replies, and 38+ languages. Those are askHermis's numbers, and they are the reason the product exists.

What the numbers do not say is as useful as what they do. The 15% that goes to a person is not a failure rate. It is the complaint about a noisy neighbour, the guest asking whether the pool is safe for a child with a cast, the request that mentions a refund. An agent that tried to answer those would be cheaper for a week and expensive for a year.

The agent takes the questions that repeat. The person keeps the ones that matter.

What does the front desk actually save?

Estimated, not measured: 3 to 5 minutes per guest message, once you count reading it, checking the PMS, switching between four messaging apps and typing the reply. A 60-room hotel handling 40 messages a day is spending 2 to 3.3 hours a day on this, most of it between 9pm and 9am when the desk is thinnest. Your numbers will differ, and a look at your own inbox replaces these with yours.

The saving lands on the 85%, so say it out loud: the agent removes something like 1.7 to 2.8 of those hours, and the rest still needs a human. A hotel expecting the whole 3.3 hours back will be disappointed and will be right to be.

Who this suits: a property or a group where the same questions arrive in five languages and the night shift is one person. Who it does not suit yet: a boutique hotel whose entire proposition is that the owner answers every message personally. The tool does not make that hotel better. It makes it ordinary.

How do you start without betting the whole front desk on it?

Narrow. One channel and the ten most common questions, with a human reading everything else. Widen the scope as the resolution rate proves out on your own traffic, not on someone else's case study. askHermis reports most properties fully operational within 3 to 5 days.

Two things to check before signing with any vendor, ours included. Ask what the agent does when it is not sure: the honest answer names a handoff, a timeout and a person. Ask for the resolution rate on your message history, not on the average customer's. A demo on the vendor's sample questions tells you nothing about your inbox.

We have not run a concierge for a property outside hospitality's usual channels, so if your guests write to you somewhere askHermis does not connect, treat the 85% as their figure rather than a promise about yours.

How many of yesterday's guest messages at your property were one of the same twenty questions, and who answered them at 2am?