Meta Business Agent on WhatsApp vs a Custom WhatsApp AI Agent

What is Meta Business Agent on WhatsApp?

On June 3, 2026, at its Conversations event in London, Meta introduced Meta Business Agent and said it was rolling out to businesses worldwide. TechCrunch reported that Meta had spent nearly two years testing these support agents inside WhatsApp Business in markets including India and Mexico. Meta says more than one million businesses already use a Business Agent on WhatsApp and Messenger.

According to Meta's announcement, the agent answers questions about your business, recommends products from your catalog, books appointments, qualifies leads and can close sales. It replies in the customer's local language, in a tone you set, and you decide when a team member takes over. A morning briefing on overnight chats is still in testing.

It comes in two forms. The version inside the WhatsApp Business app is set up from the Tools menu. Meta's product page says it learns from your Facebook Page, past chats and website, and you can upload a catalog or price list. Your controls cover what it learns from, its personality, who it talks to (everyone, new customers only, or people arriving from ads) and which topics it hands back to you. The same page says it is available to eligible businesses in select markets, so check your own app before planning around it.

For larger companies, Meta is building the Meta Business Agent Platform. Meta describes it as a way to build and deploy agents at scale with enterprise controls, guardrails and measurement, connected to hundreds of outside systems. Shopify and Zendesk are the examples Meta names, and it says the platform works alongside the WhatsApp Business Platform.

What is a custom WhatsApp AI agent?

A custom WhatsApp AI agent is an application your team or a partner builds on the WhatsApp Cloud API. Incoming messages arrive at your server as webhooks. Your code decides what to do with each one: look up the customer, search your documents, call a language model, run an action in another system and send the reply back through the API.

The Business Platform rules still apply. Free-form replies are allowed only inside the 24-hour customer service window that opens when a customer messages you; outside it, you send pre-approved templates to customers who have opted in. According to Meta's Cloud API overview, each business number can send up to 80 messages per second by default.

Take a distributor in Riyadh. A customer writes in Arabic asking whether an order has shipped. The agent matches the phone number to a customer record, reads the order status from the ERP and replies in Arabic with the tracking details. If the customer then asks to change the delivery address, the agent drafts the change and places it in a queue for a staff member to approve before anything is written back.

Who controls the data and the actions?

With Meta Business Agent, Meta runs the model and the guardrails. You shape the behavior through knowledge sources, tone, audience and handoff topics, which is enough when the answers sit on your website or in your catalog.

A custom agent gives you control at the level of each action. You decide which records it may read, which fields it may change, what needs a human sign-off and what gets logged. Transcripts and decisions can be stored in your own database, in the region you choose. That matters when you are planning for rules such as Saudi Arabia's Personal Data Protection Law, or when your auditors ask who changed a record and why.

Two points apply to both options. Messages still travel through WhatsApp, which Meta operates. And Meta's platform terms, last modified on September 23, 2026, prohibit using WhatsApp Business Platform data to build or improve AI models, with one exception: you may fine-tune a model for your exclusive use. If you hire a vendor, ask in writing how it handles your conversation data.

Which option connects to your CRM and ERP?

The app version of Meta's agent works from your catalog, website and chat history. Deeper connections belong to the Meta Business Agent Platform, which Meta says links to hundreds of systems. TechCrunch's report adds Shopee to Shopify and Zendesk as named examples. If you already run those tools, that route is worth a look.

A custom agent connects to anything with an API or a database, including systems no marketplace connector covers. Common examples:

• Reading live stock and price lists from Odoo, SAP Business One or a custom inventory database
• Creating or updating leads and deals in HubSpot, Salesforce or Zoho
• Booking slots in a clinic or service scheduling system with its own rules
• Checking credit limits or payment status before confirming an order

Approvals, handoff and languages

Meta's agent hands a chat to a person based on the topics you flag, and you can step in at any time. Meta says it replies in the customer's local language but does not publish a list of supported languages, so test it with your own customers' messages.

A custom agent can apply business rules to approvals. Refunds above a set amount go to a manager. Order changes wait for the warehouse. Anything touching a contract goes to sales. In Pakistan and the GCC, language is a practical concern too: customers mix Arabic and English, or write Urdu in Latin script, often in the same chat. A custom build lets you test against real transcripts, decide how the agent should answer mixed-language messages and tune it until the replies read naturally.

Safock's own textile product, Mavick, follows the same pattern. Its agents read ERP data live and pause for a person to approve before they change anything.

How do the cost models compare?

Meta's announcement says getting started with Meta Business Agent is free and that paid subscriptions for different business sizes are coming. TechCrunch reported that the agent will be included in some tiers of the WhatsApp Business Premium subscription, and that large businesses will pay by the tokens they use. On the developer side, Meta added a Meta Business Agent message category on July 1, 2026. As of October 2026, Meta's pricing page lists one global rate of 2 US dollars per million tokens and says a typical message uses 20,000 to 25,000 tokens, or roughly 4 to 5 US cents.

A custom agent has more cost lines. You pay once for the build, then for hosting, for language model usage with your chosen provider and for Meta's messaging fees. Meta has charged per delivered message since July 1, 2025, with rates that depend on the message category and the recipient's country. Its pricing page also says that from October 1, 2026, service messages beyond a free tier of 1,000 per business number per month are charged again.

The build cost depends on the systems the agent touches, the languages it handles and the approval paths it needs. A fixed quote should follow a short discovery phase.

Lock-in and the policy line Meta has drawn

Meta's agent lives inside Meta's products, along with its settings, the knowledge you gave it and your corrections. Moving to another approach means rebuilding that work. A custom agent keeps the code, prompts, knowledge base and logs in your hands, and the same logic can serve your website chat. The WhatsApp channel stays on Meta's terms either way.

Meta's terms also bar providers of AI technology from using the Business Platform when the AI is the main product being offered. According to respond.io, the rule took effect for all existing users on January 15, 2026. It was aimed at general-purpose assistants. A business using an AI agent for its own support, bookings or sales is a different case, and Meta's terms allow you to retain an AI provider as your solution provider.

Which one should you choose?

Use Meta Business Agent when most questions can be answered from your website and catalog, your team is small and you want something live this week.

Choose a custom WhatsApp AI agent when replies depend on live data in your own systems, when actions need approvals and an audit trail, or when you need to know where every transcript is stored. Safock builds custom WhatsApp AI agents on the WhatsApp Business Platform, and the client owns the code and documentation. A 30-minute call is usually enough to tell which side of the line you are on.