AI Agents for Hotels and Tourism: Direct Bookings, Upselling and Reviews

AI agents for hotels attack the four fronts where an independent hotel or small chain loses the most money: bookings that drift to the OTAs (with their 15-25% commission), inquiries nobody answers in time, upsells that never get offered, and reviews left unanswered. A well-built AI agent replies in seconds in the guest's own language, turns questions into direct bookings, offers room upgrades at exactly the right moment and keeps your online reputation current without stealing hours from the front desk.

This is not about dropping a decorative chatbot on the website. It's about agents connected to the booking engine, the PMS and the hotel's real WhatsApp and email channels, executing tasks end to end. This article walks through the use cases that actually move a hotel's P&L, with cost ranges and a clear list of what not to automate.

Why is the direct booking the first battle?

Because every reservation that comes through Booking or Expedia leaves 15-25% of its value in commissions — and a share of those reservations started as a visit to your own website that didn't get a question answered. The average traveler asks before booking — parking, cots, check-in time, cancellation — and if your site can't answer at 10:30 p.m., they end up booking on the OTA that resolves everything in two clicks.

An agent on the hotel's website and WhatsApp changes that mechanic: it answers the question instantly, in the guest's language, checks live availability in the booking engine and closes the direct reservation inside the same conversation — applying a direct-channel perk if you want one (breakfast included, late checkout). It rarely needs aggressive discounts: most of the time, simply being available at the moment of doubt is enough.

The math is quick. A 40-room hotel with a €110 ADR at 65% occupancy bills around a million euros a year in rooms; if OTAs carry 60% of that, shifting just 5 points to the direct channel saves on the order of €9,000-13,000 a year in commissions. This is the kind of project that pays for itself in year one.

What can an AI agent do across guest communication?

It can run the full communication cycle — pre-stay, in-stay and post-stay — in multiple languages, across the channels guests already use: WhatsApp, email, website chat and the OTAs' own messaging. The key is that it doesn't just answer: it executes, because it's connected to the PMS and the booking engine.

• Pre-stay: confirmations, arrival instructions, online check-in, special requests logged straight into the PMS, and answers to the perennials — 70-80% of a hotel's inquiries are variants of the same 20 questions.
• In-stay: requests to reception or housekeeping over WhatsApp ("could I get an extra pillow?"), local recommendations, bookings for the hotel's own restaurant or spa.
• Post-stay: thank-you message, review request targeted at satisfied guests, and a direct-rate offer for the next visit.

Multilingual service deserves its own mention: for a Spanish hotel serving British, French, German and increasingly Asian guests, an agent that responds naturally in each guest's language removes a barrier no small hotel's front desk can cover around the clock. And where phone volume matters, voice agents now handle the basic availability and confirmation calls too.

Upselling: the sale the front desk never has time to make

Hotel upselling works when the right offer lands at the right moment — which is precisely what a busy human cannot do systematically. An agent can: a few days before arrival it offers the room upgrade to the guests it makes sense for (live availability, rate paid, history), proposes breakfast, parking or late checkout during online check-in, and sells the hotel's own services during the stay.

Two design rules separate useful upselling from spam: each guest gets at most one or two relevant offers — not a catalog — and prices and availability come from the PMS in real time, not from a template. Hotels that do this well add several points of revenue per available room without touching rates; for a mid-sized property that's tens of thousands of euros a year from a mechanic that, done manually, simply wasn't being done.

Reputation and reviews: answer everything, always, without burnout

Reviews are the second booking driver after price, and answering them — all of them, quickly and well — influences both future guests' conversion and your ranking on the platforms. The reality at most small hotels: bad reviews get answered when someone finds a gap, good ones almost never, and always weeks late.

A reputation agent turns this into a process: it collects reviews from Google, Booking and TripAdvisor, drafts responses tailored to what each review actually says (not "thank you for your visit" boilerplate), and publishes them or queues them for approval depending on the policy you set. Our honest recommendation: serious negative reviews — founded complaints, incidents — always get human sign-off; the agent prepares the draft and the context (what happened on that reservation according to the PMS), but the hotel has the final word. The aggregate analysis is a bonus that often outvalues the answering itself: "street noise appears in 15% of reviews for rooms 101-104" is an operational map most consultancies would charge dearly for.

What does hotel AI cost, and what return should you expect?

The ranges we work with for independent hotels and small chains:

Add 10-20% of the project per year in maintenance (tuning, OTA and PMS API changes, improvements) and inference costs, which for a mid-sized hotel usually land in the tens of euros per month. The return comes from three measurable sources: OTA commissions saved, upsell revenue, and front-desk hours freed — a reception position costs the hotel €22,000-30,000 a year, and a large share of that shift is currently spent answering the same 20 questions.

What NOT to automate in a hotel

Hospitality is the product, and some moments are irreplaceably human:

• Complaints in the heat of the moment. An angry guest at the desk or on the phone needs a person with authority to resolve. The agent detects, contextualizes and alerts; it doesn't appease.
• Serious negative reviews. Agent drafts, human signs. Always.
• VIP guests and groups. Weddings, events, recurring corporate: human relationship, with the agent supporting logistics.
• Pricing as strategy. An agent can feed revenue management with data and alerts, but a small hotel's pricing strategy belongs to the owner or revenue manager, not to an unsupervised system.

How to start without risking the operation

1. Measure the current picture: channel mix and commissions paid, inquiry volume per channel, response times, unanswered reviews.
2. Start with the inquiry-and-booking agent on web and WhatsApp: fastest return, least disruption to internal operations.
3. Connect the PMS and booking engine from day one; without live availability, the agent is a talking brochure.
4. Add upselling and reputation in phase two, with approval rules made explicit.
5. Review at 6-8 weeks against the baseline: direct bookings, upsell revenue, front-desk hours freed.

To see how these agents fit alongside the rest of the operation — maintenance, purchasing, admin — the full map is in our guide to AI agent use cases by department, and the budget detail is in how much a custom AI agent costs. And if you'd rather ground it in your own hotel's numbers, that's how we work at our AI agency in Madrid.

Frequently asked questions

Does it work with my current PMS and booking engine?

In the vast majority of cases, yes: the PMSs and engines common in the market expose APIs or connectors the agent can work with. When a legacy system has no API, the diagnosis phase weighs whether a custom connector is worth it or whether it's time to migrate that piece. What we don't recommend is deploying the agent without a PMS connection.

Do guests accept talking to an AI?

They accept fast, useful answers — that's what they actually judge. Our policy is to identify the assistant as such (the EU AI Act requires it anyway) and give it a clear escape hatch to a human at reception. The rejection never comes from talking to an AI; it comes from not being able to get out of it when needed.

What about guest data and GDPR?

Reservation and communication data are personal data, and processing must be designed for GDPR: legal basis, minimization, data-processing agreements with model providers and access logging. Hotels don't handle special-category data by default, but they do handle passports and payments, so the data-flow design is part of the project — not an add-on.

How long until the agent is live?

An inquiry-and-booking agent deploys in 4-6 weeks, including the engine connection and loading the hotel's knowledge. The full operation — upselling, reputation, reporting — typically completes in 2-3 months in phases, without ever stopping the operation.

Does this make sense for a guesthouse or vacation rentals?

Yes — often more than for a large hotel. For properties without a 24-hour front desk, the agent is the after-hours reception. For vacation rentals, the automated cycle of arrival instructions, incident handling and reviews is essentially the bulk of the guest operation.