WhatsApp Business Automation With a Human Handover Path

Most of the guest questions a Belizean hotel receives never arrive by email. They arrive on WhatsApp, at nine at night, from someone who has already booked and wants to know whether the airport shuttle is included. WhatsApp business automation is a practical answer to that pattern: a helper that answers the easy, repeated questions from facts you control, logs everything, and steps aside the moment a person is needed.

I build and run that as a module inside Direct Belize, the multi-tenant hotel website and direct-booking platform I operate from my own servers. Multi-tenant means one system serves many properties, each with its own site, settings and data kept separate. The AI concierge is part of it, not a bolt-on from a third party you have never met.

What does an AI concierge on WhatsApp actually do?

It answers the questions your front desk answers twenty times a week. Check-in and check-out times. Whether there is an airport transfer and how it is arranged. What tours you can book. Parking. Whether the restaurant is open on Sunday.

It answers on two channels at once: the chat window on your website and your hotel's WhatsApp Business number. The guest does not have to learn anything new. They message the number that is already on your Facebook page and your booking confirmation.

Every conversation, from both channels, lands in one inbox in your admin area. Your staff can read threads, search them, and see what guests keep asking about. That inbox is often the most useful part. It tells you, in the guest's own words, where your website is unclear.

Where do the answers come from?

From two places only: your website text, and a facts list your staff maintain. The facts list is a plain set of statements about your property. Check-in is at 2pm. Late check-out is subject to availability and arranged at the desk. The shuttle from Philip Goldson International is arranged on request and quoted by the manager.

The assistant answers from that material and nothing else. It does not browse the internet, and it does not guess. If a guest asks something the facts list does not cover, it says so and offers to pass the message to a person, rather than inventing a plausible answer.

This matters more than any clever phrasing. An assistant that invents a price is worse than no assistant at all, because the guest will hold you to it. I wrote about the same constraint for ordinary website chat in a chatbot that never makes things up, and the rule is identical here. Prices and availability are never generated. They come from a person or from the booking system.

When should the AI hand over to a human?

The hand-over is the design, not an afterthought. Every thread in the inbox has a Stop AI control. One click and the assistant goes quiet on that conversation. Your staff member types, the guest sees a normal WhatsApp reply, and nothing in the guest's experience changes.

When the matter is settled, staff can let the assistant resume on that thread. The conversation history stays intact either way.

I set the hand-over triggers conservatively. In practice, a person should take over when:

• The guest is asking for a price, a discount, or a refund.
• The guest wants to change or cancel a booking.
• The question involves availability on specific dates.
• There is a complaint, or anything that reads as upset.
• The guest asks three times and is clearly not getting what they need.
• Anything medical, legal, or safety related.

The short version: keep a human path for anything involving money. Everything else is fair game for the assistant.

What do you need before it can go live?

WhatsApp is not open to anyone who wants to send automated messages. Meta, which owns WhatsApp, requires a business account and a verification step. The practical requirements are:

1. A Meta WhatsApp Business number, separate from a personal WhatsApp account.
2. Business verification with Meta, which means supplying registration documents for your organisation.
3. A facts list, written and agreed by whoever actually knows the answers.
4. A named person who will read the inbox each morning for the first month.
5. A decision on which model answers the messages.

Verification is usually the slow part, and it is out of my hands and yours. It can take a few days or most of a week depending on how quickly Meta reviews the documents. Everything else can be ready while you wait.

Claude API or a local model?

The module works with either, and the choice is yours.

The official Claude API is a paid service from Anthropic. You pay per message volume, the quality of the drafted replies is high, and the guest text leaves your server to be processed. For most hotels this is the straightforward option.

A local model runs on my servers through colibrì, the self-hosted AI service I operate. It presents an OpenAI-compatible interface, meaning it speaks the same language as the common commercial APIs, so the same module works against either without rewriting. There is no per-message cost and the text stays on infrastructure I control. The replies are a little plainer.

You can start on one and move to the other. The facts list, the inbox and the hand-over controls do not change.

What I have learned running it

Write the facts list first. Not after the number is verified, not while testing. The quality of the facts list is the quality of the assistant, and writing it forces a useful conversation internally about what your actual policies are. Most properties discover two or three things their own staff disagree about.

Review the inbox daily for the first month. Read what guests asked and what the assistant said. You will find gaps in the facts list in the first week and fix them in minutes. After a month the pattern settles and a weekly skim is enough.

Keep the assistant narrow. It is tempting to let it try to handle bookings end to end. I do not. It answers questions and routes people. The booking happens on the booking engine or with a person, where availability is real.

Expect it to improve your website. Half the questions the assistant receives are things your site should have said plainly. Fixing the site reduces the messages, which is a good outcome even though it makes the assistant look less busy.

Where to start

If you run a property, an agency or any organisation where the same questions arrive on WhatsApp all day, here is a sensible order:

1. For one week, write down every question that comes in on WhatsApp. No system, just a notebook or a shared document.
2. Group them. You will find a dozen questions account for most of the traffic.
3. Write the facts list: the answer to each of those, in plain sentences, agreed by the person with authority to set policy.
4. Mark which ones must always go to a human. Money, dates, complaints.
5. Start the Meta business verification, since it runs on its own schedule.

With those five done, the technical part is short. Everything I build is delivered with source code, documentation and in-app help, and you keep ownership of it.

If you would like to talk through whether this fits your operation, you can get in touch.