How to put an AI agent on your WhatsApp, and what has to happen first

The request usually arrives like this: “I want an AI agent to answer my WhatsApp.” And the conversation almost always stalls at the same point, which is not the one people expect. It is not the model, not the prompt, not the price. It is the line.

This article is the order worth doing things in, awkward part included.

First: the line has to be the official one

WhatsApp has two worlds that look alike and never touch.

One is the app on your phone. The other is the official WhatsApp Business API, which is the only route by which a system can read and answer messages with Meta’s permission.

A number lives in one or the other, never both at once. When you register a number on the official API, it stops opening in the app. That sentence is usually the surprise of the first meeting, so it is better to have it clear up front: if your team answers from a phone today, that phone stops being the tool.

And one warning worth writing even though it is uncomfortable: there are services that connect WhatsApp by reading the app’s QR code, bypassing Meta entirely. They work until Meta notices, and then the number gets suspended. Losing the line your customers write to is not a risk worth the savings.

Second: the 24-hour window drives the design

This is the rule that reshuffles the most projects, and the one almost nobody mentions before selling.

• Within 24 hours of the customer’s last message, your Met replies with free text. It can talk, ask, quote, book. No format restrictions.
• Outside those 24 hours, the only thing that can go out is an approved template: text you registered beforehand in the WhatsApp Business manager and Meta reviewed.

This is not a limitation of the platform you hire. It belongs to the channel, and it applies to everyone equally.

The practical consequence is that your use cases split into two groups, and it pays to sort them before you start:

What you want to do	Who starts	What you need
Answer questions, quote, book	The customer	Nothing extra
Remind someone of tomorrow’s appointment	You	Approved template
Chase an overdue invoice	You	Approved template
Recover an abandoned order	You	Approved template

Everything in the bottom half waits as long as Meta takes to approve your template. Which is why you register them on day one, not on the last day.

Third: the Met answers on its own, automation is something else

There is a confusion here that costs weeks: believing you have to draw a flow for the agent to reply.

You do not. The Met assigned to the channel receives the conversation and resolves it inside the conversation. There is no diagram to build for it to answer a question, quote or book. That is simply its job.

Automations are for the other half: whatever has to happen without anyone writing. The day-before reminder, Tuesday’s receivables sweep, the reaction to an order that changed status. Those are scheduled jobs that fire on their own.

Mixing the two leads to projects that try to model every possible branch of a conversation in a diagram, which is exactly the work a Met came to eliminate.

Fourth: with no systems connected, you have an expensive chatbot

A Met with no tools answers with whatever you wrote into it and nothing more. If you want it to say “yes, we have it in a medium” or “Thursday at 3 is open,” it has to be able to look that up.

The connection order we recommend follows the questions you get most:

1. The CRM, which is native and asks for no credentials. Everyone who writes becomes a contact with their history, and without that there is no follow-up to speak of.
2. Whatever answers your most frequent question. If you sell product, inventory. If you sell time, the calendar. If you collect, receivables.
3. Everything else, once the first one is genuinely running.

The temptation to connect everything in month one is strong and usually backfires: it multiplies what has to be tested and pushes back the day the Met answers its first real message.

What to review in the first week

Once it is running, these are the four things worth watching. They are not dashboard metrics: they are the ones that catch a problem while it is still cheap to fix.

• Conversations that ended without a useful answer. Read them in full, not the summary. What the Met is missing is written there.
• Handoffs to a person. If there are many, that is not a failure: it is the list of what comes next.
• Templates Meta rejected. It is almost always the wording, not the intent.
• Answers that sounded right and were wrong. These are the dangerous ones and you only find them by reading. A wrong fact stated confidently is indistinguishable from a correct one.

Where to start if you have nothing

If you are starting from zero, the shortest path we have seen is this: begin Meta’s business verification on day one, because it is the only thing you cannot accelerate. While that moves, pick one concrete job and get it working end to end before adding the second.

Booking is usually the best first job, because the result shows up the same day and it does not depend on integrating your inventory.