Meta's Business Agent is the cheapest way to put an AI on your WhatsApp number, and the least controllable one. A concierge built on top of a BSP costs more per conversation and gives you the tone, the data, and the system access that decide whether a conversation ends in a sale or a support ticket. The real choice is not "AI or no AI" — it is whether the answers your customers get are assembled by Meta from your public surfaces, or by you from your own systems.
This guide covers what Meta Business Agent is, how token billing changed the economics on 1 August 2026, and the four axes — cost, tone, data, integration — where the two options genuinely diverge. It ends with the case for using Meta's agent and leaving it at that.
What Meta Business Agent actually is
Meta Business Agent is Meta's own AI assistant, running inside the business messaging surfaces Meta controls: WhatsApp, Messenger, and Instagram. You do not host it, you do not choose the underlying model, and you do not write its system prompt the way you would for your own. You point it at business information — your profile, your catalog, the sources Meta lets you attach — and it answers customer questions in your inbox.
The appeal is obvious. There is no integration project, no vendor selection, no prompt maintenance. For a business whose inbound questions are genuinely about opening hours, stock, and delivery times, that may be the entire job.
What it is not is a system that knows your order history, your loyalty balances, or your CRM. Its knowledge boundary is whatever Meta ingests, and that boundary is the single most useful thing to understand before you choose.
The billing model changed the arithmetic
Meta ran a free build-and-test window from 1 to 31 July 2026, then moved Business Agent to commercial billing on 1 August 2026 at $2.00 per million tokens, a single global rate. Meta bundles the AI processing and the message delivery into that rate rather than invoicing them as two lines.
Reported real-world consumption sits around 20,000 to 25,000 tokens per message, which works out to roughly four to five cents per message. Treat that as an observation, not a planning input: token consumption scales with how much context Meta feeds the model per turn, so a conversation against a large catalog is not the same cost as a two-line hours question — and you do not control the context window that drives the meter.
Two separate changes land on 1 October 2026: service messages become billable, and utility templates sent inside the 24-hour window lose their free status. Those apply regardless of who writes the reply — the AI, a human agent, or a template. The practical consequence is that "replying free inside the window" stops being a cost strategy for anyone. For the full stack, including what a provider adds on top of Meta's own charges, see the WhatsApp Business API cost breakdown.
The comparison that matters is not $2 per million tokens against a provider's per-conversation fee. It is total cost per resolved conversation. An agent that answers 60% of questions cheaply and escalates the rest still leaves you paying human handling on the other 40% — and if it escalates the high-value ones because it cannot see an order, the cheap channel is quietly subsidising an expensive failure.
Tone is a product decision, not a setting
A concierge built on a BSP is a system you specify: the model, the prompt, the refusal behaviour, the escalation triggers, the phrasing of a price objection, the exact sentence used when a product is out of stock. You can version it, test it, and roll it back.
Meta's agent is a product you configure. You supply information and constraints within the controls Meta exposes; you do not own the response policy. For most support traffic that is fine. For anything where the wording is the product — regulated claims, medical or financial caveats, a distinctive brand voice, a negotiated B2B price — it is a real limitation, because "mostly right" becomes a compliance problem rather than a quality one.
This axis separates the two options most cleanly in regulated verticals. If your industry constrains what you may claim, an answer generator whose policy you do not control is not a shortcut available to you.
Data: what the agent can see, and what it gives back
Two distinct questions hide under "data", and they pull in opposite directions.
What the agent can see. Meta's agent reasons over what Meta has. Your own concierge reasons over what you connect to it: order status, shipping, loyalty balance, ticket history, stock in a specific store. The gap shows up precisely on the questions worth answering — where is my order, how much cashback do I have, can you reorder my usual — which are also the ones that convert.
What you get back. A conversation handled entirely inside Meta's agent produces a resolved customer and comparatively little structured signal in your own systems. A concierge you run writes intents, objections, and outcomes into your stack, where they can feed segmentation and the next campaign. If WhatsApp is a support channel for you, that loss is acceptable. If it is a revenue channel, you are giving away the exhaust.
Neither option changes your legal position: you remain the controller for the personal data in those conversations, and obtaining consent for business-initiated messaging is your responsibility either way.
Integration is the real dividing line
Strip away the positioning and the decision reduces to one question: does answering your customers' most common questions require a lookup in a system you own?
If the answer is no — hours, location, catalog browsing, generic pre-sales — Meta's agent is likely sufficient and cheaper than anything you would build, and choosing it is not a compromise.
If the answer is yes, an agent that cannot perform the lookup does not reduce your handling cost. It adds a deflection layer in front of the humans who still do the work, and it delays them. Measure this before choosing: take a week of real inbound messages and sort them into "answerable from public information" and "requires a lookup". The ratio decides it, and it costs an afternoon to find out.
When building your own is the wrong choice
Below a few hundred conversations a month, a custom concierge rarely pays for itself. The build cost is not the model — it is the integrations, the prompt maintenance, the escalation design, and the person who reads transcripts when quality drifts. That work is roughly fixed, so it amortises badly over small volume.
The same holds if your inbound is genuinely simple, if you have no system worth connecting, or if you are still validating whether customers will message you at all. In those cases run Meta's agent, keep cost per resolved conversation as the metric, and revisit when the escalation rate stops falling. Choosing not to build is a legitimate answer — and picking a provider you will fight with later is worse than using the native tool now, which is why the BSP selection criteria matter more than the AI layer on top.
A decision rule
Run Meta Business Agent when your questions are answerable from public information, your volume is low, and your brand voice is not a differentiator. Build on a BSP when the answer requires your data, when the wording carries regulatory or commercial weight, or when the conversation is a revenue channel whose signal you need back in your own systems.
Most brands end up with both: Meta's agent absorbing the trivial tier, a controlled concierge on the traffic that touches an order or a price. That is a defensible architecture, provided you decide deliberately which questions land where rather than drifting into it.
TikJoy runs the second kind — an AI concierge on the official WhatsApp Business API, connected to order and cashback data, with escalation and tone under the brand's control. See how the WhatsApp AI Concierge works.