September 19, 2026 • 4 min read
AI & Automation Specialist
I design AI-powered communication systems. My work focuses on voice agents, WhatsApp chatbots, AI assistants, and workflow automation built primarily on Twilio, n8n, and modern LLMs like OpenAI and Claude. Over the past 7 years, I've shipped 30+ automation projects handling 250k+ monthly interactions.
If you enjoy the content that I make, you can subscribe and receive insightful information through email. No spam is going to be sent, just updates about interesting posts or specialized content that I talk about.
Heads-up: the PDF itself is in Spanish for now. This page is in English so you can decide whether it is worth running through a translator — the document is one table and three short lists, so it survives one fine.
Meta shipped an AI agent inside the WhatsApp Business app: it learns from your chats, your catalogue and your documents, answers on its own, and right now it is free. If you already have a bot running, or you were about to build one, the question is a fair one: why build what the platform is giving away?
This is the table I use to decide, with an agent of my own in production at 300 to 1,000 conversations a day. It is not a case for building: in half the situations the right answer is to wait for Meta’s button and use it.
Meta compares its agent against bots put together over a weekend, not against an API integration. The table compares the six criteria that hurt later: cost, traceability, handoff to a human, integrations, channels and eligibility. That is where the real difference shows up, and it is not answer quality: it is what you can see, what you can change, and who decides whether your bot keeps existing.
The comparison table across those six criteria, including which Meta agent is actually being compared and who the per-message charge starting 1 October really applies to.
Where each option breaks. On Meta’s side: you cannot read or version what it knows, the handoff has no queue and no state, and if they pull your eligibility you are left without a bot. On yours: token cost in long conversations, the strict validation that stops it inventing prices, and observability, which you build yourself.
The three cut-off criteria for each column: when to go with Meta, and when you need your own infrastructure.
Three questions to answer before choosing: what happens the day it answers wrong, how many chats you will have in six months (not today), and what the bot has to touch besides replying.
Business owners and operations teams handling WhatsApp who are deciding between switching on Meta’s agent and building something. There is no code in the document, there are criteria. If you do write code, the section on limits is the useful one.
Is Meta’s agent free?
Today it is, and paid plans have already been announced. The per-message charge starting 1 October applies to service messages sent through the Business Platform, that is, to an API integration, not to the agent living inside the app.
How many chats a day can Meta’s agent handle?
Volume of replies is not the problem; the handoff is. Below 200 chats a day you will not notice that the handover to a human has no queue and no alert. Above 500 it is a customer waiting that nobody knows exists.
Can it read my stock or my CRM?
Meta’s agent works with your catalogue and your chats. If the sale depends on stock, CRM, payments or scheduling, you build that yourself or someone builds it for you.
What happens if it quotes the wrong price?
With Meta you open a ticket and wait: there is no prompt to read and no change to revert. With your own stack you have the message log, the tool that failed and what it cost. If nobody in your business is going to read those logs, that control is worth nothing and Meta is enough for you.
The PDF is in Spanish. Grab it with the button above, run the three questions against your own operation, and decide before you build. There is no free tier that saves you from architecture: pick the failure you can afford, not the option that is free this quarter.