It never takes someone else’s place. If a scenario is waiting for an answer, the scenario gets it. If someone on your team is working on the conversation in the Inbox, they are not overridden. And the thread is handed to it when the message arrives, not before: Meta can only be given a conversation if one is open, and it opens exactly when the person writes.
The agent only talks in the conversations Messaging Me hands to it. It only answers the contacts on its list, and Messaging Me keeps that list: you have no audience to choose and no numbers to add. When a customer writes and nobody is answering them, the conversation is handed to it and it answers right away; when a customer replies off a scenario, too. Before each template (scenario, campaign, Inbox), the contact leaves the list, so that their reply goes back to the scenario: during a scenario, the agent stays quiet. “Hand back” gives it the conversation, and it answers the customer’s next message; “Take back control” takes the contact off the list. A conversation the agent could not take (a customer who did not share their number, for example) goes to your team, in “To handle”.
Turned on, it is available; it only answers if you chose it. Who answers the customer is set on the Home screen, in “Who answers the customer”: Meta’s agent, an AI agent, a scenario, or your team. When it is not Meta’s agent, Meta’s agent, turned on, stays on standby: it only takes the contacts a scenario’s “Send to the MBA” block sends to it, and answers them right away. Turning it on when your team answers makes it the responder; turning it off when it is the one answering asks you to confirm, and messages then go to your team. The settings and the assistant tell you, before you turn it on, whether it will stay on standby.
Under the AI Agent menu, the MBA entry opens its settings, which set it up for real.
A header identifies the agent at the top of the settings: Meta’s logo, the words “Meta Business Agent” that say what the screen sets up, the number’s display name, the number itself, a badge that gives the status of the WhatsApp number, the required steps left to set (each one is a direct link to the tab where it is fixed, with the number of required settings already done out of the total), and the number of messages exchanged over the last 30 days in the conversations the agent held, once that number is known.
A category can be shown apart, in grey, under the steps: what we could not read at Meta, with its reason. These lines count neither in the steps you have left nor in the total of required settings, since nobody knows where they stand. Most of the time there are none, and that is the normal situation.
The number of messages counts the whole thread, incoming and outgoing, including campaign sends and what your team wrote after taking back control: it is therefore not a measure of what the agent itself produced, and it is not the same scope as the “messages exchanged” of the Home screen and the Performance Lab either, which leave out templates sent. Test conversations are not included, and as long as this number is not known, nothing shows. On a number that Meta has not opened yet, neither the steps, nor the grey lines, nor this number show: they would lead to a blocked screen.
The settings, in twelve tabs: Overview (the agent’s status), Assistant (set it up by talking to it), Activation (who talks to the customer), Business (your company’s information), FAQ (question by question, or in bulk from a file or an address, with a preview before writing and never duplicating an existing question), Guidelines (the tone, the procedures, what is forbidden, for example “never make up an opening time”; the tab used to be called “Skills”), Interactive messages (the buttons, lists, links, forms and carousels the agent sends itself when the situation you describe comes up, keeping the conversation), Tools (what the agent is allowed to do: set a tag, save a piece of information, send a block, start a scenario, call a system you connected), Files (up to 100 MB of knowledge documents), Websites (the pages it will read), History and Test (a sandbox where you talk to it without using up a billed conversation). This menu sits as a column to the left of the content; on a phone, it becomes the horizontal bar at the top again.
An interactive message is written with fixed content: the button labels, the rows of a list, the cards of a carousel are written in its record. Avoid having it filled by a tool’s answer: tried on a list filled with a customer’s tags, the send failed every time, and the agent answered “I can’t help you with that”, sometimes also to the following messages of the same conversation. For information specific to the customer, let the agent answer in text: without an interactive message, it quotes what the tool gave back very well.
Two situations block editing, and the screen tells them apart because they are not fixed in the same place: no number attached, which sends you to the Home screen, and Meta has not opened the agent on this number yet, whose banner gives the eligibility and carries a “Check again” link. Meta accepts every sector EXCEPT finance, the public sector, health, alcohol, gambling, over-the-counter drugs and matchmaking services; it also requires an authorized country, an account in good standing, and that no other conversational agent already runs on this number. The terms are signed in WhatsApp Manager: once signed, “Check again” reads the number’s status again.
The Activation tab carries the two settings that decide who talks to the customer. First, what the agent does when the customer asks for a human or when it refuses to handle the request: it hands over and the conversation moves up to “To handle”; or only during your opening hours, outside which it keeps the conversation rather than announce an absent advisor; or never. Then, how long a human keeps control after replying, the countdown starting again at every reply from the team and at every “Handled” (the delay runs from the more recent of the two): marking “Handled” does not hand back control. Once that delay is over, a customer who writes is handed to the agent right away, and it answers that message.
The message the customer reads when the agent hands over is chosen in the “Overview” tab: written by the agent, in the customer’s language; your own text; or Meta’s standard text, which is in English.
A cap for the agent, at the bottom of the Activation tab, over 1, 7, 14 or 30 rolling days. The two units do not cap the same thing. In tokens, it is a spending cap, shared by all the WhatsApp numbers of your Business Manager, including those of another workspace: the screen names that Business Manager. In agent replies, it is a cap per conversation: it does not cap the total spend. An estimate at Meta’s public price comes with the figure, and what shows is what Meta applies, read back from Meta. Meta gives the agent’s usage through no API: its invoice is what counts.
In Performance Lab, the Summary carries a “Meta’s agent” card over the last 30 days: the conversations it held, those waiting for your team right now, the messages it wrote, a cost estimated at the public price, and the tools it called, with their success and their average time.
Two things this screen does not promise, because they cannot be kept: preventing a human from taking back control (no lock exists, neither on our side nor at Meta), and preventing the agent from deciding a handover. “Never” only makes it keep the conversation instead of letting go. And when it does not know how to answer, it does not hand over: it points to the contact details in your knowledge base.
The Assistant tab lets you describe in plain words what you want to change (“add my Saturday hours”). It first reads what already exists, proposes one line per change, and nothing goes to Meta until you click Apply. A deletion is flagged apart, in red, and it only proposes one at a time: Meta has neither a trash nor an undo, so a bulk request gets a list and a question, never a purge. It will never take the agent out of service and will never create a connector, and everything it does can still be done by hand in the other tabs. It does not add documents: a document is uploaded through the Files tab. A guideline it proposes is complete (its name, when the agent applies it, what it does), but the list only shows its name, and a change replaces it in full: it does not see the current text, tell it what to keep.
The History tab exists for the same reason: Meta has no trash. A deleted FAQ, guideline, interactive message or document is lost at Meta, and this page keeps the only copy. “See the erased content” opens it again for you. Putting it back recreates it, it does not bring it back to life: you copy the content and get a new item. Nothing in it is purged.
Take back control. In the Inbox, on a conversation that Meta’s agent holds, the “Take back control” button really takes it from the agent, without any message going to the contact: the contact leaves the agent’s list, and the agent stays quiet for them until a conversation is handed to it again. Meta can refuse, and the screen then says so instead of letting you believe it is done. In any case, writing to the contact takes the thread for sure.
To spot these conversations in the Inbox list: a blue to purple gradient and a sparkle wand in front of the name. The list carries no written mention of who holds the conversation; it is there to spot the exception, and the exact holder is read in the header of the open conversation.