AI for Hotels: a Guest Journey Agent

How can artificial intelligence support a hotel guest before they arrive?

An event-driven agent starts when the booking is confirmed, not when the guest asks. It sends the welcome, answers trip-preparation questions by email, points to a pre-check-in portal when structured data is needed, offers services in the context of the conversation and, after checkout, asks for and organises the feedback.

A booking is not the end of the sales process. It is the start of a relationship.

Many hotels concentrate the digital process into two moments — booking and check-in. Between the two there can be days, weeks or months, and that time is almost always silent.

In that gap the guest has concrete questions: what time they can arrive, whether there is parking, whether an airport transfer can be arranged, whether there are good restaurants nearby, whether early check-in is possible, whether they can fill in their check-in details already. Almost all of them repeat — and almost all of them are, at the same time, a commercial opportunity and a chance to improve the stay.

What makes this agent different is that it opens the conversation

A chatbot waits for someone to open a window and ask. Here it is the confirmed booking that triggers the interaction: the event reaches the application, AWS Lambda loads the context of that booking and that property, and the agent prepares a concrete welcome email — name, dates, property, number of guests — that tells the guest they can simply reply to it whenever they need something.

From then on, email stops being a notification channel and becomes the interface of the conversation. «We land at the airport around 7pm, what is the best way to reach the hotel?» «We are travelling with two small children, what would you recommend nearby?» The guest installs no app, does not return to the website and does not restart a conversation from scratch: they reply.

The phase of the journey is itself context

This is what separates a guest journey agent from an assistant that answers questions. The same question asked three weeks before arrival and two days before arrival does not deserve the same answer.

Three weeks out, the subject is preparing the trip. Two days out, it is pre-check-in, the expected arrival time, the transfer and the last instructions. During the stay, it is services and the destination. After checkout, it is feedback. The agent uses the phase as part of the context, so what is worth saying changes on its own over time — without separate campaigns or mailing lists deciding for it.

And no message is treated as if it were the first. When the guest writes only «yes, we'd like the transfer», that only means anything inside the previous conversation — the booking, the dates, what has already been asked and what has already been requested are part of what the agent considers before replying.

Contextual upselling is the opposite of a campaign

A promotional email sent to every guest knows nothing about any individual's intent. A conversation does, because the customer has just explained it.

Someone who writes «our flight lands very early, can we leave our bags?» is telling you an early check-in might interest them, if the hotel offers one. Someone mentioning a wedding anniversary is opening a different door. The agent understands the context before putting anything forward — and the hotel's own business rules decide whether there is anything to put forward at all. The aim is more relevance, not more commercial messages.

The same holds for cross-selling: «how do we get in from the airport?» may make a transfer relevant; «we'd like dinner near the hotel» may make the hotel's own restaurant relevant. The service question and the commercial opportunity become the same conversation.

Recommending is not listing the twenty best things to do

«We only have one free afternoon, what would you recommend?» «We have small children.» «It is going to rain, what can we do?» These have no answer on a static attractions page, because they depend entirely on context — where the hotel is, how long the stay is, who is travelling, what they have already said they like.

For many guests the hotel is also the starting point for discovering the destination, and that layer — restaurants, museums, beaches, activities, transport, local suggestions — can be tailored to each property's actual area rather than being generic.

Not everything should be a conversation: pre-check-in is a portal

Each interface is good at something, and forcing everything through a chat is a common mistake. Email suits conversation, questions, suggestions, follow-up and anything that stretches over weeks. A portal suits structured forms, documents, option selection and actions that need confirmation.

So the agent's email can carry a personalised link to a pre-check-in portal, where the guest supplies their details before arrival: email → portal → structured data → Lambda → validation → arrival preparation. The two channels complement each other instead of competing.

The model writes. Lambda decides when, and with what.

The agent is not a language model sending emails. AWS Lambda receives or identifies the booking, works out which phase of the stay the guest is in, loads the right context, processes the replies arriving by email, identifies intent, calls the model, structures the conversation data, creates the pre-check-in links, runs the workflow rules, prepares the next communications and asks for feedback at the end.

The division is the same one we apply in other projects: the model understands and produces language; the software coordinates the process and holds the rules. A general model knows a great deal about Lisbon or the Algarve, but it does not know that hotel's opening hours, room types, policies or services — and it is the property's authorised knowledge that answers for those, not what the model believes about hotels.

After checkout, feedback stops being a number from 1 to 5

A traditional questionnaire makes the guest translate an experience into scales: room 1 to 5, cleanliness 1 to 5, breakfast 1 to 5. Exactly the part that matters is lost — the why.

A comment such as «we loved the location and the staff, but the room was a little noisy at night» can be structured into overall sentiment positive, positive aspects (location, staff), negative aspect (noise) and area involved (room, comfort). Over time that surfaces recurring themes — breakfast, cleanliness, noise, check-in, parking — at volumes nobody would read by hand. What to do about it remains the hotel team's decision.

Hospitality is human, and stays that way

The agent covers what is repetitive and what happens out of hours — the guest can reply at two in the morning, be in another time zone, plan the trip on a Sunday. In a global industry, much of the relationship happens thousands of kilometres away and weeks before anyone walks into reception.

But hospitality is not a task you automate. Complex situations, exceptions, problem-solving, face-to-face service and commercial decisions stay with people. What changes is how much time they have for them, once the repetition is out of the way.