Where generative AI improves the travel guest experience

Generative AI changed what a travel chatbot can do. The old rule-based bots answered from a script, so they were limited but predictable. Generative systems reason over your actual knowledge and respond in natural language, which makes them far more capable and introduces a new failure mode: they can be fluently, confidently wrong.

That trade is the whole story of generative AI in travel. The upside is real and worth pursuing. The downside has a name and a court case attached to it. This piece walks through where generative AI genuinely improves the guest experience, where it creates risk, and how to capture the first without inheriting the second. For the wider map of where these wins land, see our overview of AI in hospitality across the guest journey.

What changed when chatbots got generative

A traditional chatbot matches a guest's question to a pre-written answer. If the question is phrased oddly or falls outside the script, it fails politely and the guest gets a human. Limited, but safe.

A generative model works differently. It understands the question in natural language, reasons over the knowledge it was given, and composes a response on the fly. That is a genuine leap in capability, and it is why guest-facing travel AI got dramatically better in the last two years. The mechanics of that shift are covered in our explainer on generative AI versus traditional chatbots.

The new capability comes with a new responsibility. Because a generative system composes answers rather than retrieving fixed ones, it can compose an answer that sounds right and is wrong. In travel, where answers carry money and contracts, that is the risk to manage. The guest experience improves when generative AI is pointed at the right jobs and constrained on the wrong ones.

Where it genuinely helps

Three areas stand out, because they play to what generative models are good at.

Natural conversation over messy knowledge. A guest rarely asks a question the way your FAQ is written. They ask "is this place okay for a family with a toddler and a dog," which touches three policies at once. A generative assistant grounded in your real property and policy information can synthesize an answer across all of them, in plain language, the way a knowledgeable concierge would. Guests want this: in Oracle and Skift's Hospitality in 2025 study of 5,266 consumers and 633 hotel executives, 77% said they were interested in using automated messaging or chatbots for customer service requests.

Multilingual service without a multilingual team. Travel is global, and staffing native speakers for every guest language is impractical for almost any property. Generative models handle dozens of languages natively, so a guest can ask in Portuguese and get a fluent answer drawn from the same knowledge base that serves English speakers. This is one of the clearest experience wins, because the alternative is a translation app or a guest who simply does not get helped.

Personalization that reads context. A generative assistant can weave a guest's booking details, stated preferences, and history into a relevant response rather than a generic one. That capability matters because the personalization gap is wide. Medallia's 2024 study found 61% of consumers would spend more for a personalized experience while only 23% said their recent hotel stays felt highly personalized. Generative AI can close part of that gap at scale, as long as it stays grounded in real data.

Underneath all three is the same engine: generative AI is good at understanding intent and composing a relevant answer from a body of knowledge. The places it improves the guest experience are the places that job description fits.

Where it creates risk

The risk is hallucination, and travel has the canonical cautionary tale. Air Canada's website chatbot told a customer he could apply for a bereavement fare retroactively. He could not; the airline's actual policy required requesting the fare before travel. When he was refused, he took it to the British Columbia Civil Resolution Tribunal, which in February 2024 ordered Air Canada to pay $812 in damages and rejected the airline's argument that the chatbot was a separate entity responsible for its own answers.

The ruling matters beyond the dollar figure, which was trivial. It established that a company owns what its AI tells customers, the same way it owns a static webpage. A generative assistant that invents a cancellation policy, quotes a fare that does not exist, or promises a refund that is not owed is creating a liability the company will answer for. We catalog more of these in our AI hallucination examples, and the travel ones tend to be expensive.

So the risk is concentrated exactly where generative AI is most tempting to deploy: customer-facing answers about policies, fares, and money. The capability that makes it useful (composing fluent answers) is the same capability that makes a wrong answer dangerous. That is the tension to design around.

How to get the upside without the downside

The good news is that the failure mode is manageable. The teams getting generative AI right in travel tend to do four things.

Ground the model in real, current knowledge. A generative assistant should answer from your actual policies, fares, and property information, retrieved at the time of the question, instead of from whatever it absorbed in training. This is the difference between an assistant that reflects your real cancellation policy and one that invents a plausible-sounding version. Grounding is the single biggest lever on accuracy.

Make it conservative about uncertainty. The most important behavior in a travel assistant is knowing when it does not know. A system that hands off to a human when it is unsure, the way our Agent 5 model does, avoids the Air Canada failure by design, because it declines to invent the policy it cannot confirm. A confident guess is the enemy. A clean "let me get someone who can confirm that" is the safe behavior.

Keep the high-stakes actions gated. Answering "what time is checkout" is low-risk. Promising a refund or confirming a fare exception is high-risk. The generative layer can handle the conversation while sensitive commitments stay behind confirmation steps or human review. Capability and authority are different settings, and travel rewards keeping them separate.

Measure accuracy alongside coverage. A generative assistant that answers everything looks impressive until you check whether the answers are right. Track the error rate and the rate of clean escalations alongside the resolution rate, because in travel a wrong answer is worse than a deferred one. The metric that matters is correct resolutions, and a high coverage number means little without it.

None of this blunts the upside. A grounded, conservative, well-gated generative assistant still gives guests natural conversation, multilingual service, and real personalization. It just does so without composing a fictional policy on the way.

The honest framing is that generative AI moved travel guest service from scripted and safe to capable and accountable. The capability is what guests feel, and it is genuinely better. The accountability is what teams have to engineer, because the same fluency that delights a guest can, ungrounded, cost you a tribunal hearing. The properties that win with generative AI are the ones that treat its confidence as something to earn and verify, so the guest gets the smart concierge and never the confident liar.