AI in travel and hospitality: what actually works

1. The data layer decides everything

Travel is one of the hardest data environments in commerce. A single trip touches a GDS or NDC connection, a PNR that gets rewritten many times, fare rules written partly as free text, hotel content from several suppliers that describe the same room in different ways, and a payment flow with its own exceptions.

When a company says "our AI assistant gives wrong answers," the cause is usually one of these:

• Room and rate mapping. The same "Deluxe King" appears under five names across suppliers. The model cannot tell which one the guest booked.
• Fare rules as prose. Change and refund conditions live in text blocks the model has to interpret, and interpretation is where errors start.
• Booking history spread across systems. Mid-office, CRM and email each hold part of the story. The agent sees one part.
• No owner for content quality. Descriptions, amenities and policies drift out of date and nobody is accountable.

The fix is unglamorous: data contracts between systems, a single mapped catalogue for property and room content, structured fare and policy data, and monitoring on all of it. It is also the work that turns a demo into a product.

2. Where AI agents earn their keep

Agents work best on tasks that are bounded, reversible and fed with structured inputs. In travel and hospitality that includes:

• Servicing triage. Reading incoming change, cancellation and complaint requests, classifying them and drafting the response for a human to send.
• Booking summaries. Giving a call-centre agent the full trip, past changes and preferences in one paragraph before they pick up.
• Guest Q&A from verified property data. Parking, check-in times, pet policy. Only from a source the hotel maintains.
• Content mapping and QA. Matching supplier room types, flagging missing amenities, spotting stale photos and descriptions.
• Disruption options. Assembling rebooking alternatives during irregular operations so a human approves in seconds instead of searching for minutes.
• Review and feedback analysis. Turning thousands of reviews into a short list of operational fixes.

3. Where they fail, and why it costs real money

The risk grows as soon as an agent can commit the company. Issuing tickets, applying refunds, granting exceptions, or explaining a fare rule it inferred from free text.

The well-known example is Air Canada. In 2024 a Canadian civil resolution tribunal held the airline responsible for incorrect bereavement-fare information its website chatbot gave a customer. The company argued, in effect, that the chatbot was responsible for itself. The tribunal did not accept that. Whatever your AI tells a traveller, you own it.

Practical guardrails:

• Let agents propose, and let rules or people dispose, for anything involving money or policy.
• Answer only from retrieved, verified sources, and say "I'll check with the team" when the source is missing.
• Log every agent decision with the inputs it used, so you can audit and improve.
• Test with the ugly cases: split tickets, partial refunds, group bookings, multi-property stays.

4. The opening for independent hotels

For years, independent and mid-size hotels could not afford the technology chains used: personalization, smart guest messaging, pricing support, strong direct booking. AI changes the cost of building and running that software. A small engineering team using AI tools can now ship in weeks what used to take a year.

That makes a different question worth asking: not "which vendor do we buy" but "which two or three tools would most change our margin, and can we have them built around our own data." This is the case I made at the Hospitality Investment Forum, and it is the one hotel owners ask me about most.

5. When the guest is an AI agent

A growing share of trip planning starts in an AI assistant instead of a search box. Next, assistants will compare, hold and book. When that happens, the assistant is the one reading your website.

It will favour what it can verify. That means:

• Structured data for rooms, amenities, policies, location and prices, using schema.org markup.
• Consistent content across your site, OTAs and metasearch, so the assistant finds one truth.
• Clear answers to the questions guests ask, written as plain text on your own pages.
• Bookable connections through your booking engine, channel manager or APIs, so an agent can complete the transaction.

This is SEO again, with a new reader. The hotels and travel brands that make themselves easy for machines to read will get recommended. The rest will be summarized by someone else's data.

6. A first 90 days

1. Weeks 1–3: map the data behind your top five customer questions. Find where it lives and who owns it.
2. Weeks 4–6: fix the worst gap, usually content mapping or booking history, and add monitoring.
3. Weeks 7–10: launch one bounded agent use case with a human approving every action.
4. Weeks 11–13: measure handle time, accuracy and escalations. Then decide what the agent may do on its own.

Frequently asked questions