AI for travel agent admin: what it can and can't do

AI tools are strong on phrasing and weak on facts. That one split explains most of what they can and can't do in a travel office.

What's going on inside the tool

Most chat tools run on a large language model, a program fed a huge pile of written text so it can handle language jobs. Writing, summarising and translating are typical ones¹.

The key detail is how it writes. Models of this kind are first trained to predict the next word¹. Nothing is being looked up in a booking system. It's producing the words that most plausibly come next, given what you typed and what it has seen before.

That's why the first draft often reads well. It's also why it can be wrong. When an AI states something false as though it were true, people call that a hallucination². Chatbots can slip believable falsehoods into otherwise fluent text².

For a travel agent, a believable wrong detail is the worst kind. A flight time that looks right but isn't will get past a tired reader.

Where it fits in the office day

An agency mainly sells travel for suppliers³. Around each sale sits a lot of writing, and much of it suits a language tool, as long as you hand it the facts.

Good fits:

• Turning a few rough lines into a polite reply to an enquiry. See replying to travel enquiries.
• Laying out confirmed bookings as a readable day-by-day plan. See drafting itineraries.
• First drafts of requests and confirmations to suppliers. See supplier and booking emails.
• Reminder and thank-you messages that go around each booking. See reminders and follow-ups.
• Tidying tone, shortening a long email, or fixing awkward phrasing in your own draft.

Where a person still does the work

Some jobs aren't writing at all, so the tool shouldn't own them.

• Facts. Dates, names, times, places and reference numbers should come from your booking record, not from the tool's memory.
• Judgement. Whether a client's plan is sensible, or whether to offer a different option, is your call.
• Rules and advice. Entry rules, insurance, health and safety questions belong with the official source for that country or the supplier. A language tool shouldn't answer them for your client, and neither should an AI-drafted email.
• Sending. Nothing should go out unread. The routine in checking AI-written travel details works for any message.
• Client data. Passport and booking details need care before they go anywhere near a chat box. See keeping client details private.

A working rule of thumb

Give the tool your confirmed facts and ask it only to arrange and word them. Don't ask it to supply facts. Asking it "what time does that train leave?" turns a writing aid into a reference book. It isn't one.

Prompts matter a little, too. Shaping your plain-language request to get the output you want is called prompt engineering⁴. You don't need to study it. Say who the message is for, what tone you want and which facts are fixed.