AI WhatsApp automation: run a business number with agents
How to connect your own WhatsApp Business number to AI agents — bots that answer, drafts you approve from your phone, and dashboards for everything.
WhatsApp is where your customers already are, which is exactly why automating it is dangerous: a bot that says the wrong thing to the wrong person does it at conversation speed. This is the full walkthrough — connect a number, put agents on it, and keep a human hand on the wheel — the way Pocket does it.
First, what you are automating
Everything here runs on the WhatsApp Business Cloud API — Meta's official programmatic interface to a business number. If you just need a simple click-to-chat link for a website or flyer, our free WhatsApp link generator does that in the browser. For everything below, two things matter before anything else:
- It is a business number, full stop. These APIs automate numbers registered through a WhatsApp Business Account. Nobody's personal WhatsApp can be wired up, and any tool claiming otherwise is going to get that number banned.
- The number is yours. In Pocket each tenant connects their own WABA — their own phone-number id and their own billing with Meta. Conversations are billed to your account, not marked up by us, and you can leave with your number intact.
Connecting takes one of two paths: Embedded Signup (Meta's popup; we exchange the returned code for a token and subscribe your WABA to the app) or the manual route — pasting a phone-number id, WABA id, and System User token — which is what works before Tech Provider approval and for teams with an existing WhatsApp setup.
The architecture that survives contact with customers
The naive design — webhook hits a serverless function, an LLM writes a reply, it sends — falls over in a week. Not because the model is bad, but because real inboxes have context: order history, the conversation from Tuesday, the thing your team promised. And because sometimes the right answer is *not sending*.
The shape that works has three layers, and it is how Pocket is built end to end:
- A persistent bot, not a function. The WhatsApp bot is an agent with its own identity, memory, and toolset, running on an always-on machine — yours or ours — that can reach your data (orders, tickets, calendars) through the same tools any Pocket agent has.
- A confidence ladder. Messages the bot is equipped to answer, it answers. Anything outbound-sensitive or below your comfort line becomes a draft card on your phone — recipient, text, the whole thing — and one tap approves or edits it. Approving a message you have not read is not approval; the product enforces that.
- A dashboard, not vibes. Volume, response times, what the bot handled versus what it escalated — live, next to the conversation. Automation you cannot measure is automation you cannot trust.
Setting it up, concretely
- Connect the number under Channels → WhatsApp (embedded signup or manual credentials).
- Create a bot and give it the tools it needs — read-only to start: search the knowledge base, look up orders, check schedules.
- Point its automation at the inbox with a clear instruction: what it answers itself, what it drafts for approval, what it never touches (refunds, legal, anything with money).
- Run it for a week in draft-everything mode. Read the drafts. Where it is consistently right, loosen the leash; where it is consistently wrong, fix the instruction or the tool.
- Watch the dashboard until the escalation rate flattens. That plateau *is* your automation envelope — grow it deliberately, not by accident.
What this looks like from your pocket
The operator experience is the part stock videos skip: your phone gets one notification when a draft is waiting, not forty when a message arrives. Approve from the lock screen. The bot's whole timeline — every message, every decision, every tool call — is reviewable after the fact, so 'what did it say to whom' is a scroll, not a subpoena.
That is the honest version of WhatsApp automation: agents on your own number, a phone-shaped approval loop, and numbers on a dashboard instead of a feeling that it is *probably* fine. Set up a number in about ten minutes, and run it in draft mode free while you learn its edges.