Blog Industries

AI in travel and hospitality: guest messaging and use cases.

What a hotel AI agent actually does, how to run guest messaging safely, which systems it connects to, and where a person has to stay in charge.

Fig. 0A guest’s question, answered in their language

For hotels, tour operators and restaurants, the first AI system worth building is a guest-messaging agent: it answers the routine questions in the guest’s language at any hour, reads the live reservation, and makes bookings and changes in your property system within your rules. A person takes over anything emotional, unusual or costly. This guide covers what that agent is, how to run it safely and how to roll it out.

What is a hotel AI agent, and how is it different from a chatbot?

A chatbot answers from a script or a fixed set of articles. A hotel AI agent reads the live reservation, checks real availability and rates, and acts in the property management system (PMS) on the guest’s behalf, inside limits you set. The difference decides what the guest can get done in one conversation.

QuestionScripted chatbotHotel AI agent
“Can I check in early tomorrow?”Quotes the standard check-in timeReads the booking, checks room readiness and arrival, and answers for this guest
“Move my booking to Friday.”Sends a link or the front-desk numberChecks availability and the rate rules, makes the change in the PMS and confirms
A message in Arabic, then EnglishOften fails or restartsReplies in the language each message arrives in, with the earlier conversation in view
An angry guestRepeats the menuRecognises the situation, stops, and hands over with a summary

Our own AI agent runs 24 hours a day in whatever language the customer writes in, and carries a 4.8 out of 5 customer rating. It is a WhatsApp sales agent for a solar company, handling 70% of that company’s monthly sales with 99% uptime, and the same discipline applies to a hotel agent: live data, limits enforced in code, a fast route to a person. Our guide to WhatsApp AI agents explains how that works.

Where does AI help in travel and hospitality?

Guest messaging before, during and after the stay

Pre-arrival questions, directions and check-in details, requests during the stay (towels, late checkout, a taxi), and follow-up afterwards. Research from Boston University’s hospitality school, based on surveys and interviews with US and Chinese hotel guests, found that guests prefer messaging apps for non-urgent needs, and before and after the stay. It also advised keeping other channels open for emergencies. That is a good brief for an agent: take the routine volume, and never become the only way to reach a person.

Bookings, changes and cancellations

Checking availability, quoting from your live rates, taking bookings, and making changes or cancellations within your policies, then writing them back to your booking system. For tour operators and agencies, the same applies to itineraries, transfers and excursions.

Restaurants

Taking reservations by message, answering menu questions from your approved menu data, managing waitlists and confirming bookings. Allergen and dietary answers must come from your recorded menu information, never from the model’s guesses, and should point guests to staff for anything serious.

Reviews and reputation

Drafting personal replies to reviews in the reviewer’s language for a manager to approve, and summarising recurring issues across hundreds of reviews so you know what to fix.

Staff knowledge

An assistant that answers staff questions from operating procedures and property information, useful for new and seasonal staff.

How should AI guest messaging work: draft or auto-send?

This is the design decision that matters most, and it is made per type of request, not once for the whole agent. The agent either sends a reply itself (auto-send), or writes a draft that a team member approves in one click (draft mode). A sensible starting map:

RequestModeWhy
Facilities, opening hours, directions, parkingAuto-send from the property fileThe answer is a recorded fact
Availability and rate quoteAuto-send from live PMS dataNumbers come from the system, never from the model
New booking, or a change under a flexible rateAuto-send after the guest confirms the detailsRules in code decide what is allowed
Refund, compensation, upgrade or any exceptionDraft for staffIt commits money or inventory
Complaint, accessibility need, medical or safety issue, special occasionHand over to a personThese are emotional or high-stakes
Allergen and dietary questionsRecorded menu data only, plus a pointer to staffA wrong answer can hurt someone

What is a confidence threshold, and where should it sit?

A threshold is the line below which the agent stops auto-sending and drafts or asks instead. The model’s own claim that it is “90% sure” is not a reliable signal, so we build thresholds from checks that can be measured: whether the answer was found in a recorded source, whether the request type was recognised clearly, and whether the details validate (the dates parse, the room type exists, the quoted rate matches the system). Set the thresholds per request type from test results, start strict, and loosen them only when the scored tests and staff corrections justify it. When nothing in your sources answers the question, the right reply is “I’ll check with the team”, not a plausible guess. Our guide to evals before features covers how to score this before each change.

When should the agent hand over to a person?

  • The guest asks for a human. Always, immediately, with no argument.
  • Sensitive topics. Complaints, accessibility, health, safety, bereavement and anything with money or blame attached.
  • Two failed attempts. If the agent has not resolved the request after two tries, it hands over rather than looping.
  • Out of hours. The guest is told honestly what happens next and when, and the case is queued for the morning team with the full conversation.

The handover carries the conversation, the reservation, the language and a one-line summary, so the guest never repeats themselves. WhatsApp’s own Business Messaging Policy allows automation inside the 24-hour customer service window as long as there are prompt, clear and direct escalation paths, and limits messages outside that window to approved templates (the policy page showed an update date of 23 September 2026). Our guide to fallback chains covers what the agent does when a system it depends on is down. For the wider picture of support automation, see AI customer service.

What are the most common messaging mistakes?

  • Looping replies. The same answer repeated to a guest who clearly did not get what they needed.
  • Tone mismatch. A cheerful reply to a complaint, or a stiff one to a guest who wrote casually.
  • No memory of the stay. Asking for a booking reference the system already holds, or forgetting what was asked ten minutes ago.
  • Promising what policy does not allow. A late checkout or a refund offered by the model rather than approved by a rule.
  • Stale information. Policies copied into a prompt months ago, so the agent quotes last season’s breakfast hours.
  • Ignoring the channel’s rules. Sending free-form messages on WhatsApp after the 24-hour window has closed.

Which systems does a hotel AI agent need to connect to?

Integration with the property systems is usually the biggest piece of work, and it decides what the agent can do. The same pattern holds whichever vendor you run.

System typeWhat the agent does with itWhat to watch
Property management system (PMS)Reads reservations and guest profiles, checks availability, creates and changes bookingsWhich operations the API allows, write access limits, and rate limits
Channel manager and central reservationsKeeps availability and rates consistent across channelsDelay between systems, so a quote is checked again at booking time
Booking engine and paymentsHands the guest to a secure payment stepCard numbers must never pass through the chat
Restaurant reservation and point-of-sale systemsTakes tables, reads the approved menu and allergen dataMenu data being current and owned by one person
CRM and review platformsReads guest history, drafts review repliesConsent for marketing messages, and who approves a reply

Property platforms such as Mews, Oracle OPERA Cloud and Cloudbeds are the kind of system an agent integrates with, and the depth of what their APIs allow varies. As one example, Mews documents connector operations for adding, updating and cancelling reservations. Which of those an agent may call, and under what limits, is a decision for your team, not a default. Where a system has no usable API, design around a slower route such as file exports or a staff-approved step. Our services page describes how we build these integrations.

Is revenue and pricing AI the same thing?

No. Dynamic pricing and demand forecasting are machine-learning and optimisation problems that work on booking history, events and competitor rates, not on conversations. They are a different tool from guest messaging, usually bought as part of a revenue-management system, and a language model is the wrong core for them. Where they meet is at the edges: a language model can explain a rate change to a manager in plain words, and the guest agent should quote only the rates that revenue management has published in the PMS.

What do guests think, and how do you keep the human touch?

A 2026 study led by the University of South Florida, which surveyed 44 hospitality practitioners and 101 recent hotel guests, found that guests value AI for quick, round-the-clock help with simple requests but prefer a person for requests with emotional weight, such as choosing a restaurant for an anniversary dinner. Staff were more enthusiastic than guests. The authors recommend a hybrid model in which AI handles the conversation until a person should step in. The design rules that follow:

  • Say it’s AI. Guests should know when they’re talking to an assistant. In the EU, Article 50 of the AI Act requires people to be told they are interacting with an AI system, unless that is obvious from the context (see the rules below).
  • Hand over fast. Complaints, special occasions, accessibility needs and anything emotional go to a person straight away, with the full conversation.
  • Never promise what you can’t deliver. Room upgrades, refunds and exceptions stay within limits set in your systems, not left to the model.

Let the AI answer the hundred routine questions so your team can make the one moment that guests remember.

How do you protect guest data and payment details?

Guest messages contain names, travel dates, passport details, health and dietary information, and sometimes card numbers typed in by mistake. Design for that:

  • Keep sensitive fields out of prompts and logs unless strictly needed, and mask them where they must be stored.
  • Never take card details in the chat. The agent sends the guest to a hosted payment step, and detects and discards a card number if one is typed. Card data is governed by the PCI Security Standards Council’s PCI DSS, which your payment provider and acquirer will hold you to.
  • Apply the data protection law of each market you serve. In the UK and EU that means the UK GDPR and GDPR, including data minimisation and transparency. The UK regulator’s guidance on AI and data protection is the place to start (as of October 2026 it is under review following the Data (Use and Access) Act, so check the current version). A qualified professional should confirm what applies to you.
  • Treat guest messages as untrusted input. A guest, or someone pretending to be one, can write instructions meant to change the agent’s behaviour. Limits belong in code, not in the prompt.

Which rules apply?

  • EU. Article 50 of the EU AI Act requires AI systems that interact directly with people to be designed so those people are informed they are dealing with an AI system, unless that is obvious from the circumstances. As of October 2026 the obligation applies from 2 August 2026. Our EU AI Act guide for builders covers the wider picture.
  • You are responsible for what the agent says. In 2024 a British Columbia tribunal held Air Canada liable for wrong information that its website chatbot gave a customer about bereavement fares (Moffatt v. Air Canada, 2024 BCCRT 149). The tribunal rejected the argument that the chatbot was a separate entity responsible for its own actions. It is a single tribunal decision and not binding elsewhere, but it shows how a travel business should expect to be treated when its assistant gets a policy wrong.
  • Gulf. Hotels serving the UAE and Saudi Arabia have data-residency, Arabic-language and WhatsApp considerations. Our guide to building AI in the UAE covers them.

Where is AI the wrong tool, and where does a person stay in charge?

  • Service recovery. An upset guest needs an apology and a decision from someone with the authority to make one.
  • Safety, medical and accessibility issues. These go to staff at once. The agent’s only job is to hand over fast with the facts.
  • Pricing decisions and exceptions. Rules in the PMS set rates and policies. The agent quotes them and applies them, and never improvises them.

How should a hotel or travel company roll it out?

Start narrow and widen as the numbers allow:

  1. Pilot one property and one channel, usually WhatsApp or web chat, answering only from your recorded property information. No booking changes yet.
  2. Connect the PMS read-only. The agent can now see availability, rates and the guest’s own reservation, which turns generic answers into personal ones.
  3. Test before each step with real guest messages in every language you serve: price edge cases, unknown answers, hostile messages and a guest typing a card number.
  4. Add bookings and changes in draft mode. Staff approve each reply for a period, and their corrections become new test cases.
  5. Move the request types that pass to auto-send, one at a time, keeping handover and out-of-hours rules in place.
  6. Expand to more channels, more languages and more properties, reusing the same core and each property’s own information.

How do hotels measure it?

Agree the numbers before launch and review them weekly at first. Five that matter:

  • First response time, including out of hours.
  • Resolution without staff: the share of conversations finished without a person needing to step in or correct the reply.
  • Staff correction rate in draft mode: how often the team changes a draft, and why.
  • Handover quality: how often a handed-over guest had to repeat themselves, and how long it took a person to pick up.
  • Direct bookings and guest ratings from conversations the agent handled, compared with the same period before.

Keep a log of every message sent and every action taken in the PMS. It is what lets you answer a guest’s complaint, and what shows where to tighten a limit.

Questions hotels and travel companies ask

Can AI handle bookings and cancellations?

Yes, when it is connected to your PMS and bound by your rate and cancellation rules. The agent checks live availability, confirms the details with the guest, makes the change in the system and sends the confirmation. Anything outside the rules, such as a refund on a non-refundable rate, goes to a person.

Will it replace front-desk staff?

It takes the repetitive message volume, not the front desk. In the University of South Florida study, staff were more enthusiastic than guests, expecting relief from workload and staffing pressure. Your team keeps arrivals, problems and the personal touches.

Will guests mind talking to AI?

The research above suggests guests welcome quick, correct help with routine requests and want a person when a request matters to them. The risk is being stuck with an assistant that can’t help, so a fast route to a person matters as much as the AI itself.

Can it work across several properties?

Yes. One agent can serve a group, answering from each property’s own information and rules, and handing over to that property’s own team.

Does it work in Arabic and other languages?

Yes. Current models handle major languages well, including Arabic, and each language gets its own tests before launch.

In travel or hospitality?

Modulus Labs AI builds guest messaging and booking agents for hotels, travel companies and restaurants, on WhatsApp, web and email. Tell us how guests reach you today, and we’ll tell you what we’d build first.