This week Meta made it easier for developers to plug a coding assistant, Claude, Cursor, ChatGPT, straight into a WhatsApp Business account and have it handle the setup: register the number, write the message templates, check the fine print. That's a plumbing story, useful if you happen to have a developer on call. The bigger story happened three months earlier, and it's the one that matters if you run a shop, a studio or a restaurant: Meta's AI agent that actually talks to your customers on WhatsApp went live worldwide back in June, after nearly two years of testing in places like India and Mexico. It's already running on more than a million business accounts.
If you've wondered whether to turn this on for your own business, here's what it actually reads, what "on its own" really means, and the one thing worth checking before you say yes.
What the agent reads before it answers
Say you run a yoga studio. You upload your class timetable, your pricing page and a short FAQ: what to bring, whether drop-ins are allowed, your cancellation policy. From that point on, when someone messages asking "is there a class at 6pm tonight", the agent isn't reasoning about yoga. It's reading the timetable file you gave it and answering from that, in whatever tone you picked under "skills": friendly, formal, brief. Connect it to your booking system through its own API, and it stops just answering questions. It can reserve the spot and confirm the payment itself, the same way a person would type it into your system by hand.
That's the whole trick, and it's worth saying plainly because it isn't magic: the agent is only ever as good, and as current, as the documents you hand it.
What "on its own" actually means
Once it's live, nobody reads the reply before it goes to the customer. Meta calls the safety net a "handoff": passing the conversation to a person "for specific events". But those events aren't defined for you by default. Unless you set them yourself, the agent decides, alone, which questions are routine and which ones should stop and wait for you, including the ones you'd genuinely want to see, a refund request, a complaint, something ambiguous enough that a wrong confident answer would cost you a customer.
Where it doesn't even apply
Worth knowing before you get curious: Meta excludes whole sectors from this outright, including health, finance, government, alcohol and gambling. If you run a clinic, a pharmacy or anything finance adjacent, this feature simply isn't available to you, whatever a headline implies. For a shop, a restaurant, a studio, a driving school or a workshop, it applies in full, today.
The honest limit
An agent like this is only as current as the file you handed it. If your class prices change and nobody updates the document it reads from, it will tell a customer the old price, confidently, in your own brand voice, at two in the morning if that's when they happen to write. Meta's own dashboard lets you send test messages before going live, and that's genuinely useful. But it tells you the agent works today. It doesn't tell you it will still be right in three months, once the timetable has changed twice and nobody remembered to update the file it reads from.
What to check before you switch it on
Before you turn this on, or if it's already on and you haven't looked, do two things. Send it a test message using your single most outdated or ambiguous piece of information, a class that changed time, a price that went up, and read exactly what it says back. Then set at least one handoff rule yourself, for anything involving money or a complaint. Those are the two moments you actually want to see with your own eyes, and Meta will not decide that for you.
This is the same question I wrote about a few weeks ago when Zapier's own adoption numbers pointed away from full autonomy for most small businesses: Do You Actually Need a Fully Autonomous AI Agent? covers the general case. WhatsApp is simply the first place most of you will actually be offered the choice, and it's already live, whether you've noticed it yet or not.