AI Is Good at the Words in Your Travel Business. It's Bad at the Facts.

Search "AI for travel agents" right now and much of what ranks is written for airlines, platforms and software buyers: chatbots, AI agents, customer-service tools. Search "ChatGPT for travel agents" and you mostly get lists of prompts, with little about what can go wrong.

Neither answers the question a working advisor actually has. Which parts of my week can I hand to a chat assistant, and which parts will it quietly get wrong?

The short answer: AI is good at words and order. It is bad at facts that change.

The one test to run on any task

Before you hand a task to AI, ask what the task is made of.

If it is made of wording, tone, structure or sorting, AI is fast and often good, as long as you supply the facts. A first reply, a recap email, a messy page of call notes turned into a clean brief. You supply the facts, it supplies the sentences.

If it is made of facts that live somewhere else and change without telling anyone, AI is the wrong source. Today's fare. This week's cabin availability. A supplier's cancellation terms. The entry rules for a country your client is transiting through.

The AI Workflow System for Travel Agents puts the risk plainly: a chat assistant doesn't reliably know those things, and "It can still write as if it does." The draft reads just as confidently when it is wrong. And even facts you did supply can bend on the way through, which is why the check never goes away.

What AI is actually good for

The first reply. When an inquiry lands, a quick reply that sounds like a person helps. AI can turn "thanks, here's what happens next, here's how to book a call" into a warm short reply quickly. The guide's inquiry workflow tells it to "not quote prices or suggest specific dates or suppliers," which keeps the fast reply from making a promise you haven't checked.

Messy notes into a clean brief. After a discovery call you have half-sentences, a budget that may be per person or total, and a "maybe March?" AI is very good at sorting that into headings: travelers, occasion, dates, budget, must-haves, deal-breakers. The useful instruction is that anything unclear or missing gets labeled, not guessed. A brief that says "Budget: UNKNOWN whether per person or total" is more valuable than one that picks.

Summarizing a long thread. Forty emails with a supplier and a client, three changes of plan. AI can pull out what the client wants, what's confirmed and what's still open. You check the status words against the source, because a summary can put something in the wrong column as easily as it can leave it out.

Rewriting supplier copy in your voice. Ship and resort descriptions are written to sell. AI can turn them into something that sounds like you. The rule is that it keeps every fact and adds none, because AI is also very good at adding things. A balcony becomes a "private oceanfront retreat."

Laying options side by side. Two quotes on different bases, one per person, one total, one with gratuities and one without. AI can lay them out on matching terms quickly. You still do the math yourself, and anything a quote doesn't say gets written as "not stated," not filled in.

The hard messages. A schedule change, a canceled tour, a client who is upset. These are the emails advisors put off, because the wording matters and the facts are stressful. AI is good at calm, clear, specific wording. It must be kept on a short leash with facts: only the options the supplier actually confirmed, no refunds or credits nobody offered, no promises about the outcome.

Admin and follow-up. Payment reminders, document checklists, a welcome-home note, a weekly list of what's due. Repetitive writing is the easiest place to start.

Ideas, clearly labeled as ideas. Ask for things to do on a free afternoon and AI will happily give you a list, complete with opening hours it made up. The guide's version asks for idea types only, labeled "Ideas to verify, not confirmed," with no hours, prices or booking details. You check each one before it reaches the client.

One task, start to finish

Here is what that looks like on a job almost every advisor does: turning call notes into a brief. This is an illustration built to show the kind of slips to look for, not output from any particular tool.

What the advisor pastes, already anonymized:

What a first draft of the brief might say:

What the advisor catches:

• "Maybe March or April" became "March." Flexibility is information, and it just disappeared.
• "Around 8k" became "$8,000 per person." The notes never said per person or total. That one slip changes the whole trip.
• "No driving" became "private driver." Maybe they would love one. Maybe they meant trains and ferries. Either way, the client never asked for it.
• "Italy, first time" vanished. That shapes the pace, the hotels and how much hand-holding the trip needs.
• "Not sure about tours" vanished. It belongs under open questions.

The fixed brief reads:

The read-through caught five problems before the client saw any of them. The guide's version of this workflow tells the AI to "Use only what's in my notes" and to label anything missing "UNKNOWN or NOT PROVIDED." Instructions like that point the draft in the right direction. They do not make the read-through optional.

What it is bad at, and why it matters to you

Prices, availability and status. A chat assistant doesn't see your supplier portal unless you've connected it, and even then the draft needs checking. If a draft mentions a price you didn't give it, treat it as invented until you trace it. The same goes for status. A cabin that is "quoted" can become "your cabin" in a draft, and "held" can become "confirmed." Those are different promises.

Supplier policies. Cancellation terms, deposit rules, what's included. If it didn't come from the supplier, in writing, it isn't a fact yet.

Entry, visa, passport and health requirements. These vary by nationality and transit country, and they change. The guide treats this one as non-negotiable: the AI writes "[OFFICIAL SOURCE]" and you check the official source yourself.

Travel insurance. What an unlicensed seller may say about coverage varies by state, and your host agency may add its own rules. The guide's approach is to have AI record only that protection was offered and what the client decided. Anything beyond that follows your licensing and host agency rules.

If you think "the AI said it" is a defense, there is a case worth knowing. In February 2024, a British Columbia tribunal ruled on Moffatt v. Air Canada. The airline's website chatbot told a customer, booking a flight after his grandmother died, that he could claim a bereavement discount after buying his ticket, within 90 days. The airline's actual policy did not allow that once travel was completed. According to the American Bar Association's summary of the case, Air Canada argued the chatbot was a separate legal entity responsible for its own actions. The summary says the tribunal found Air Canada "still bore responsibility for all the information on its website, whether it came from a static page or a chatbot." The award was about $650 CAD, plus interest and fees. Small money, big lesson.

That was an airline's public chatbot, not an advisor's reviewed email, and the legal details for your business are a question for your host agency or a lawyer. This is not legal advice. The practical point is simpler: whatever reaches the client carries your name, not the tool's.

What goes in the chat, and what never does

The other thing AI is bad at is being your privacy filter. By the time it could flag a passport number, you've already sent it.

The guide's rule is to strip identifiers before anything is pasted: names paired with trip details, passport numbers, birthdates, card numbers, confirmation numbers, health notes tied to a named person. Placeholders go in instead. "Couple, mid-50s, anniversary, 7 nights, prefers quiet" gets you a better draft than a name and a birthday would. Watch the combination, too: a rare detail plus a date plus a hometown can point to one person even with no name attached. Real details go back in afterward, in your own email or CRM.

Check your tool's training settings too. In ChatGPT, for example, OpenAI's help page says that when "Improve the model for everyone" is off, "your new conversations won't be used to train OpenAI models." It also notes that ChatGPT Business and Enterprise workspaces are not used for training by default. Other assistants have their own settings, and menus move, so look in your tool's data or privacy settings. A training setting is one piece of the picture. Your agency's privacy policy still decides what may be pasted at all.

How to set it up so it helps

Starting cold every time makes weak output more likely. A few habits help:

• Tell it who you are once. Your name, your business, your fees, your tone, the phrases you never use, and the rules: use only the facts I give you, never invent a price or a policy, flag what's missing. The guide calls this a Context Block. Save it as a Project or custom instructions so you aren't pasting it daily.
• One client and one trip per chat. Mixing clients is an easy way for the wrong detail to land in the wrong draft.
• Ask for two sections. The draft itself, clean, and a separate list of what you must verify. Copy only the draft.
• Read every factual line against its source. Every time. That review is the part your clients are paying you for.

The guide sums up the process in six words: "Choose. Anonymize. Draft. Verify. Approve. Send."

Will AI replace travel agents?

This is opinion, not a forecast. Look at the lists above. The jobs AI does well are mostly behind the scenes: the first-draft wording, the sorting, the reminders. The jobs it does badly are the ones clients notice most: knowing what's true today, knowing which supplier will actually help at 11 p.m., and being the person accountable when a flight is canceled.

An advisor who uses AI for the words can put more of the week into the judgment, as long as the review stays in.

The line to keep

Use AI for the words and the order. Keep the facts, the status and the promises for yourself.