A WhatsApp AI agent, because that is where your customers already are

Most businesses discover WhatsApp by accident. Someone puts a number on the site, customers start messaging it, and within a year there is a phone on a desk that one person has to watch. It works right up until it does not: messages arrive after hours, the person watching it goes on holiday, and nobody can tell you how many enquiries were lost because nothing was ever written down.

A WhatsApp agent fixes the specific version of that problem you have. It answers the questions you answer twenty times a week, collects what you actually need before a human gets involved, books the slot, and passes the whole conversation to a person the moment it should. On the official Business API, so it does not get your number banned.

On a phone call, silence is failure. Every response has to land inside about a second or the caller starts talking over the agent, which means the whole stack gets designed around latency: streaming speech-to-text, a model chosen partly for speed, text-to-speech, and telephony routing, all running concurrently on every second of every call. A large part of building a voice agent is defending that budget.

On WhatsApp, three seconds is normal and nobody notices. That single difference removes the telephony stack, removes the latency engineering, and lets us use the model that answers best rather than the one that answers fastest. The conversation design is the same work. Everything underneath it is not, which is why the price is roughly half.

There is a well-trodden shortcut here: an unofficial library driving WhatsApp Web, no Meta approval, live the same afternoon. It works, until Meta notices the traffic pattern and bans the number, which is usually the number printed on your business cards and every listing you have.

We build on the WhatsApp Business Cloud API. That means business verification, a template approval step for anything you send first, and a real rate limit. 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.

Good candidates: the questions you answer every day, qualifying an enquiry before a person spends time on it, taking a booking, chasing a confirmation, telling someone where their order is. All of these are bounded, checkable, and boring, which is exactly what makes them safe to automate.

Bad candidates: anything where being wrong is expensive, anything that needs judgement about a specific person, and anything you would not be comfortable with a new hire doing unsupervised in week one. We will say so during the call rather than take the money and find out together.