The evening and weekend messages
The ones that used to sit until Monday now get a real answer in under a minute, which is where most of the lost business was.
How-to — — by Mahmoud Zalt
What a WhatsApp chatbot really is, the Meta rules it has to live inside, what it should never handle, and what running one actually costs.
The reason WhatsApp is worth the trouble is simple: people read it. A message sent there gets opened in minutes by a customer who never opens your email, and that is the whole business case in one line.
The reason it is trouble is that Meta guards the channel carefully. The rules about who you may message, when, and with what wording are stricter than anywhere else you have automated, and every chatbot vendor lives inside them whether their marketing mentions it or not.
This guide covers what you are actually buying, what you need before you start, and the two or three decisions that separate a chatbot customers tolerate from one they hate. Our own take on it is an AI Employee at Sistava, which is the same idea framed differently: you hire a role and brief it, rather than drawing a conversation tree.
Software connected to your WhatsApp business number that receives incoming messages and replies without you. That is the whole definition, and everything else is a choice about how it decides what to say. There are two families, and the difference matters more than any feature list.
A rule-based bot is a decision tree. The customer taps a numbered option or types a keyword, and the bot sends the matching answer. It is predictable, cheap and never says anything embarrassing. It is also frustrating the second a customer asks something you did not anticipate, which is most of the time.
A bot built on a language model reads the message as written and produces an answer from what you told it about your business. It handles the awkward phrasing, the two questions in one sentence, and the customer who writes in another language. In exchange, you have to be deliberate about the boundaries, because it will attempt anything you did not forbid.
| Rule-based menu bot | Language model bot | |
|---|---|---|
| How it decides | Keyword or tapped button | Reads the actual message |
| Unexpected question | Dead end or a loop back to the menu | Answers, or says it will fetch a person |
| Other languages | Only ones you built | Usually handled without extra work |
| Setup effort | Drawing every path yourself | Writing a brief about your business |
| Main risk | Customers give up and call | It answers something it should have escalated |
A phone number, a Meta business account, and access to the WhatsApp Business Platform. The free WhatsApp Business app cannot host a chatbot. It gives you greeting messages, away messages and quick replies, all of which are fixed text, and none of which read the customer's question.
The number is the decision people rush. Once a number is registered to the Platform it behaves differently from the one on your phone, so most businesses either move their public number across deliberately or start a new one and redirect. Sort that out before you build anything.
Step six is where most of the value is. Real customers ask things nobody on your team would think to test, and the first fifty conversations will show you three gaps in your brief that no amount of planning would have surfaced. Fixing those is twenty minutes of editing and it changes the experience completely.
The 24 hour customer service window, and message templates. Meta's documentation states that when a user messages or calls you a 24 hour timer starts, that it resets if they message again, and that once it closes you can only send pre-approved template messages. Your chatbot is bound by that regardless of who built it.
In practice this makes chatbots strongly reactive. Answering somebody who messaged you is unrestricted and, for non-template messages inside the window, free. Starting a conversation means an approved template and a per message charge, which is why the good WhatsApp chatbots concentrate on inbound and the annoying ones do not.
Templates come in three categories that Meta defines: utility for following up on something the customer did, marketing for anything promotional, and authentication for one-time passcodes. Approval is not permanent either. Meta says templates can be paused for recurring negative feedback or low read rates, and disabled after recurring negative feedback.
It should handle the repetitive front half of every conversation, and it should never make a promise, quote a price it is unsure of, or apologise on your behalf. Those three cover almost all the damage a customer channel can do, and all three read as official because they came from your business number.
The ones that used to sit until Monday now get a real answer in under a minute, which is where most of the lost business was.
Hours, location, price range, availability, what you need from them, what happens next. Consistent every time.
Name, need, timing and budget gathered up front, so your callback is short and informed instead of exploratory.
A message in another language gets an answer in that language, without hiring for it.
The never-touch list should be written down and short enough to remember. Refunds and complaints, anything medical, legal or financial, anything about a specific person's account or data, and anything where the customer is already angry. Those go to a person with a warm holding reply, and the handover should feel like a feature rather than a failure.
One more decision worth making early is whether the same assistant works in your other channels. A customer who asked something on WhatsApp last month and emails you today should not be treated as a stranger, which is why the Sistava version keeps one memory across channels rather than running a separate bot per surface.
Two bills, not one. Meta charges per template message delivered, across the marketing, utility and authentication categories, at rates that vary by the recipient's country calling code. On top of that sits whatever your chatbot platform charges, which is usually a monthly fee or per conversation pricing.
The good news is that the inbound case is cheap. Meta documents non-template messages sent inside an open customer service window as free of charge, and utility templates sent while that window is open as free too. So a chatbot that answers people who messaged you costs far less to run than one built around campaigns.
Budget for the build as well as the running cost. If you are doing it directly with Meta, the effort is developer work by their own framing. If you are using a provider, the real cost is the week you spend writing and rewriting the brief, which nobody quotes for and everybody spends. Our own pricing sits on the Sistava plans page if you want a number to compare against, and Meta's charges are separate from any of it.
A chatbot does not create demand. If twelve people a week message you, it turns twelve slow replies into twelve fast ones, which is worth having and is not growth. Businesses that expect a messaging tool to fill a quiet week are disappointed for reasons unrelated to the tool.
It also cannot fix a business problem underneath the messages. If people message because your prices are not published anywhere, a chatbot that answers the price question forty times a week is a very efficient way to avoid updating your website. Look at what customers keep asking, then fix the cause.
And it does need an owner. Somebody has to read the escalations, notice when it is answering something badly, and keep the brief current when your prices or hours change. That is an hour a month, not a job, but the setups that go stale are the ones where nobody owns that hour.
No. The app gives you greeting messages, away messages and quick replies, which are fixed text rather than answers. A chatbot that reads and responds to what a customer wrote needs the WhatsApp Business Platform, which is Meta's API.
Not necessarily. Meta's own documentation frames direct Platform work as being for developers, but plenty of services connect on your behalf so you only write the instructions. If you want full control of the conversation logic, you are back to needing engineering.
Only with a pre-approved template, and only where you have consent. Meta's 24 hour customer service window means free-form messages are limited to conversations a customer started recently. Outside it, only approved templates send, and each one costs money.
A properly connected one on the WhatsApp Business Platform is exactly what Meta intends you to use, so the platform itself is not the risk. The risk is behaviour: WhatsApp's guidelines list repeated unwanted contact and bulk auto-messaging as prohibited, with bans among the consequences.
Being straightforward works better than pretending. Many will guess from the reply speed anyway. A short line saying an assistant handles first replies and a person picks up anything complex sets the expectation honestly and rarely bothers anyone.
The connection is usually the short part. The long parts are getting the number sorted, getting through Meta's account setup, and writing a brief good enough that the answers are worth sending. Allow for at least one round of template rejections if you plan any outbound messaging.