Blog / WhatsApp

A WhatsApp Chatbot for Travel Agencies: Automate the Itinerary, Never the Refund

1 November 2026 · 8 min read · By Runix B

A WhatsApp Chatbot for Travel Agencies: Automate the Itinerary, Never the Refund

A customer is standing in an airport at 11pm, typing "is my transfer still booked?" to a travel agency that closed at six. That one message is the whole case for a WhatsApp chatbot for travel agencies. It is also the case against getting it wrong.

Travel is a strange fit for automation. Most questions are dull and answerable: pickup times, what's included, whether the hotel has a lift. A few are expensive: a cancelled flight, a refund dispute, a visa question where a wrong answer strands someone. The job is telling those two piles apart before the bot does.

This is the last post in our series on WhatsApp automation. It picks up from the setup overview, docs-grounded auto replies, lead capture and webhook setup for self-hosting. Here we apply all of it to one industry, and we're blunt about where the bot should stay quiet.

Written by the ChatterMate team. We build an open-source AI support agent, so weigh our product claims accordingly. We've limited them to what we could verify, and said where we couldn't.

Why agencies end up on WhatsApp anyway

Nobody planned this. Customers started messaging the agent's personal number, then the shop's number, then a "business" number someone set up on a spare phone. By the time you notice, half your bookings are being discussed in a thread that only one employee can see, and that employee is on holiday.

Travel is unusually message-heavy for a few reasons. Trips have long gaps between booking and departure, so there's a steady drip of small questions. People are abroad, often on patchy data, and a text thread survives a bad connection better than a phone call. And the stakes are time-sensitive in a way a normal shop order isn't.

We won't hand you a "X% of travellers prefer messaging" statistic. We went looking, and the figures floating around travel-tech blogs mostly trace back to vendors selling messaging software, or to articles with no checkable source. If you find a number you like, follow it to the original before you put it in a deck.

What we can say from the plumbing side: WhatsApp's customer service window gives you 24 hours from a customer's last message to reply freely, and Meta's documentation is explicit that outside it you're limited to approved templates. For a travel agency that rule is the whole game. It shapes what the bot can do at 2am, and what it can't.

What a travel bot should answer on its own

Start with what's boring and stable. If the answer lives in a document you already have, the bot can read it.

Itinerary logistics. Pickup times, meeting points, what time check-in opens, which terminal. These come from your own trip documents, so they're ideal grounded answers. The catch is that a general knowledge base is the wrong source. A customer asking "what time is my transfer" wants their transfer, not the generic one. Unless the bot can see booking data, it should give the general rule and point to the booking confirmation. Don't let it guess a time.

What's included. Breakfast, airport transfers, excursions, baggage. Package descriptions are written in marketing prose and answered in a hurry, which is how disputes start. Rewrite them as plain statements, including the negatives: "Airport transfers are not included in the Budget package." A bot with that sentence will say it. A bot without it will improvise.

Destination basics. Currency, plugs, tipping, dress codes at temples. Low stakes, high volume, well documented. But keep these as sources you wrote or checked, not whatever the model remembers about a country's customs.

Opening hours and contact routes. Who to ring out of hours, and the emergency line. This one is simple and it is the most valuable thing in the knowledge base. Put it first.

We've covered how to write these pages in help-center content an AI can actually use. The same rules apply: one question per page, the answer in the first two sentences, exceptions right after.

What it must hand to a person

Here we have an opinion, and it's firmer than most vendors will give you. Some things shouldn't be automated even if a model could produce a fluent answer.

  • Visa, passport validity and entry requirements. These change, vary by nationality, and the cost of being wrong lands on the traveller. Point to the official government source and hand off.
  • Anything medical or insurance-related.
  • Refunds, cancellations and compensation. Not because the bot can't read the policy. Because a customer in a dispute wants to hear from someone who can decide.
  • Disruption. A cancelled flight at 11pm is not an FAQ. It's the moment a customer decides whether they'll book with you again.
  • Whether a booking is protected under the consumer-protection rules in your country. That is a legal answer, and it's yours to give.
  • Payments. Never ask for card numbers in a chat thread, and tell the bot never to repeat or request them. Send a payment link from your booking system instead.

Write these as handoff triggers, not as "please be careful" in the prompt. "Be careful" is a hope. A trigger is a rule. Our human-handoff post goes through how to set them up so the customer isn't asked to repeat themselves.

And design the handoff message for a tired person on a phone. "Thanks for your patience, I've passed this to the team and someone will reply here by 9am" beats a silent wait. Name a time you can keep. If it's 11pm and nobody's staffed until morning, say so. Customers forgive a closed office. They don't forgive silence.

The 24-hour window, applied to a trip

Here's where travel gets awkward in a way restaurants and plumbers don't.

A customer messages in March to ask about a September trip. You answer. The window closes the next day. In July you want to send a "your documents are ready" note. That needs an approved template, because the window is long shut. In September the customer messages from the airport, and the window reopens for another 24 hours.

So plan your templates before you plan your bot. The useful ones for an agency are usually dull: booking confirmation, documents ready, a pre-departure reminder, a post-trip review request. Meta sorts template messages into categories (marketing, utility, authentication) and bills them differently, and since 1 July 2025 it charges per delivered message rather than per conversation. Rates depend on the recipient's country, and Meta's pricing page also describes a free window when a customer arrives from a Click to WhatsApp ad, plus a monthly allowance of free service messages per number. Those numbers have changed more than once this year. Check the live page for your markets before you model costs, because we don't trust any blog post's table, including ours, to still be current.

One practical point: don't use the marketing category for something that's really a utility message. Meta reviews template content, and miscategorised templates get rejected or reclassified. Write a reminder as a reminder.

Is a travel bot allowed on WhatsApp?

Yes, with a caveat worth knowing. In late 2025 Meta updated its business terms to bar general-purpose AI assistants from the platform, effective 15 January 2026. TechCrunch's report says Meta described the change as aimed at chatbots distributed as products in their own right, not at businesses using AI to serve their own customers, and gave a travel company running a customer-service bot as an example of what remains fine.

Take the shape of that seriously. A bot that answers questions about your trips, bookings and policies is on the right side of it. A bot that will happily chat about anything is not. Keep the agent narrow, and keep a human reachable. Read Meta's current terms yourself before you put a business number at risk.

How we'd set this up in ChatterMate

ChatterMate connects to WhatsApp through the Business Cloud API with your own Meta app, and its README lists template and media support and 24-hour session handling. That's the verified part. Here's the order we'd work in, and none of it is exotic.

1. Build the knowledge base from your trip documents

Load package descriptions, terms, FAQs, and the out-of-hours contact page. Rewrite the vague ones first. The walkthrough is in how to train an AI chatbot on your docs.

2. Pull thirty real messages from your real inbox

Typos included. "does the tour include lunch", "my driver isnt here", "can i change names on the booking". Notice how many of those are about a specific booking. That tells you where the bot will hit its limit, because a grounded agent answers from documents, and a booking is a record.

We haven't verified a built-in connection from ChatterMate to a booking system or CRM that would let it look up a customer's reservation over WhatsApp. Our lead-capture page is about the web widget, and the README's CRM list doesn't match it exactly. If you need per-booking lookups, ask us or check the repository before you plan around it. Until then, treat booking-specific questions as handoffs.

3. Run all thirty before the number goes live

Mark each as right, wrong or "didn't know". Wrong is the number to drive to zero. A "didn't know" is a gap in your content, which is cheap to fix. Our testing checklist has the longer method.

4. Write the handoff triggers

Visa, refund, disruption, payment. Keywords plus a fallback: if the bot isn't confident, it hands off. We wrote about confidence scores if you want the reasoning.

5. Read the first fifty conversations

All of them. You'll find the question nobody told you about, and it'll be about luggage.

Where this goes wrong

Three patterns, all avoidable.

The bot answers a booking question from general text. "Your transfer is at 9" when the transfer is at 7. Fix: general rule plus a pointer to the confirmation, never a specific time it can't see.

The handoff goes nowhere. The bot says "a team member will help" and the team's inbox is empty at night. Fix: honest hours in the message, and a real out-of-hours number for emergencies.

The template list is an afterthought. You can't message anyone outside the window, and the first time you need to, you discover your templates are still in review. Fix: submit them early.

Should you bother

If your agency gets more than a handful of repeat questions a week and the same two people answer them, yes. If your business is mostly bespoke, high-touch itineraries where every question is about one client's situation, a bot will mostly hand off, and you're better off spending the effort on faster human replies. That's a fair outcome, not a failure.

For the small agency in between, the realistic win is narrow: the plain questions answered at 11pm, the hard ones routed to a person with the context attached. That's worth having. It isn't a replacement for anyone.

ChatterMate is open source and free to start, with 50 free AI replies a month, and you can self-host it. You can try it at chattermate.chat, or read the code on GitHub and judge for yourself.

Where this fits: WhatsApp

See how ChatterMate handles this in the product.

Read about WhatsApp

Plug it in

Add your website and start teaching your AI agent.

Start free