On June 4, 2026, Apple approved Poke as the first standalone third-party AI agent allowed to run inside iMessage (TechCrunch). It's a marker for where this is all going: agents are moving into the apps people already text in, instead of asking anyone to learn a new one.
We've been building in that space for a while. This is a walkthrough of one of our agents — a WhatsApp AI agent that reads what lands in your inbox, sorts the signal from the noise, and files the important stuff in Notion. The part that makes it usable in a real business: it never logs or acts on anything until you approve it.
What it does
The agent sits on a WhatsApp number and watches what comes in. For every message, it does three things:
- Reads and understands it. Not keyword matching — it works out what the message is actually about.
- Classifies it. A new lead, a customer question, a supplier update, or noise.
- Routes what matters. Anything worth keeping becomes a structured entry in Notion, tagged and ready to act on.
The result: the messages that used to scroll off the top of your phone before you could deal with them now show up as organized tasks, without you sorting a thing.
The clever part: it asks before it acts
Most "AI inbox" tools fail the trust test the moment they do something you didn't expect. We designed around that from the start.
Before the agent writes anything to Notion or takes an action, it sends you a summary on Telegram with two buttons: approve or skip. One tap. Nothing irreversible happens without your yes. You're approving from your phone in the time it takes to read a text.
That single design choice is what makes the agent something you'll actually keep running. It's a tireless assistant that still checks with you, rather than a black box you're quietly afraid of.
How it works
Here's the shape of the system, without the wiring diagram.
- WhatsApp in. A connector watches the WhatsApp number and forwards each new message to our orchestrator.
- The orchestrator. A Cloudflare Worker is the brain. It runs at the edge, holds the logic, and calls the model that reads and classifies the message.
- The decision. The model decides what the message is and whether it's worth filing. Low-value noise stops here.
- Human approval. For anything that matters, the Worker sends you a Telegram message with approve / skip buttons and waits.
- Notion out. On approval, the agent writes a clean, tagged entry to your Notion database. On skip, it drops it.
The whole thing runs on infrastructure you own, so there's no third-party tool sitting between you and your own messages, and nothing acts without a human in the loop.
Why this matters for a small business
Phone and chat are where small businesses actually lose leads. A message comes in while you're with a customer, you mean to deal with it later, and later never comes. Multiply that by every channel and the leak is real money.
A triage agent plugs that leak without adding another inbox to check. It watches the channel, surfaces what needs you, and keeps a clean record in the tool you already use. You stay in control, and you stop losing the thread.
The Poke news is the same idea pointed at consumers. The version that pays off for an operator is the one wired into your business: your channels, your database, your approval.
For more on how we approach this, see what an AI automation agency actually does.