WhatsApp automation that does the work, not just the replying

WhatsApp automation usually gets sold as a chatbot: a box on your number that answers questions. That is the smallest part of it, and on its own it leaves the expensive problem untouched. The enquiry nobody logged, the order update nobody sent, the renewal nobody chased, the person copying messages off a phone into a CRM by hand: none of that is a conversation problem, and none of it is fixed by a better answer.

What we build is the whole message layer. The chat is the surface: underneath it, the agent reads your calendar and your catalogue, writes to your CRM, fires off the events your systems already emit, collects what it needs in a form inside the thread, and hands the conversation to a person the moment it should. If all you actually need is the conversation, that is a smaller and cheaper build, and we will say so before you buy this one.

Since July 2025 Meta bills per message rather than per conversation, and one rule sets the price of everything you build. When a customer messages you, a 24-hour service window opens. Inside it you can reply freely, in any format, as many times as the conversation needs, and none of those messages are charged. Outside it you cannot send a free-form message at all: only a template Meta has approved, and that one is billed on delivery.

So the first design question on any flow is not "what should the bot say", it is "who spoke first". Automation that answers people is nearly free to run and gets cheaper as it gets better at ending conversations. Automation that starts them carries a per-message cost and a review process, so it has to earn its place. Most agencies never explain this, because most agencies are reselling a platform that hides it inside a bundled per-conversation price.

A chatbot answers. An agent answers and acts: it reads the diary, writes the record, and decides when to stop and fetch a human. Automation is the layer around both: the events from your systems that start conversations, the templates that carry them, the forms that collect what is missing, and the writes that mean nobody retypes anything afterwards.

You can buy the first without the rest, and plenty of businesses should. If what you need is a bot answering the same eight questions on your number, that is a smaller build, and we would rather point you at it than sell you this. What does not work is buying a chatbot and expecting the loop to close on its own: the reason the enquiry never reached the CRM is not that the bot was not clever enough.

There is a well-worn shortcut here: an unofficial library driving WhatsApp Web, no Meta approval, running the same afternoon. It works until Meta notices the traffic pattern and bans the number, and that number is the one on your cards, your listings, and every ad you have ever run.

We build on the WhatsApp Business Cloud API, which is now the only official option: the On-Premises API was deprecated in 2023, its final version expired on 23 October 2025, and numbers still registered to it stopped delivering messages altogether. Cloud API means business verification, template approval for anything you send first, and real rate limits. It is slower to start, and it is the only version worth putting a business on. Chatberry, which we built, runs on exactly this stack.

The request we most often turn down in its original form is "can we blast our list". You can: marketing templates exist, they are the priciest category, and they are the ones people report. Reports and blocks feed the quality rating on your number, and a number with a poor rating gets throttled, then restricted. It is also the number every customer you have already has saved.

What we build instead is smaller and works better: sends segmented off real signals in your data, an opt-out honoured on the first request rather than behind a confirmation step, and a cap on how often any one person can hear from you at all. One useful message a month keeps a number deliverable for years, which is worth more than any single campaign you could send this week.

Good candidates are bounded, checkable, and boring: the questions you answer twenty times a week, qualification before a person spends time on a lead, taking a booking, confirming an address, telling someone where their order is. All of them have a right answer that exists somewhere in your data.

Bad candidates are the ones where being wrong is expensive: changing a delivery address after dispatch, anything medical or legal that turns on judgement about a specific person, refunds above a limit you set, and anything you would not let a new hire do unsupervised in week one. Every build ships with a written list of things the agent must refuse and hand over, because deciding that list is most of the design work, and it is the part a platform subscription cannot do for you.