Meta’s Agent Bet Starts With 200 Million Users

WhatsApp Business had 200 million monthly users in 2023, before Meta AI connected to Gmail or Google Calendar. Meta had the audience before it had the agent.

The graph was built to predict attention; agents make it an execution context

Meta’s social graph answered an advertising question: given what a person watches, follows, shares, discusses, and buys, what should appear next? Feeds converted those signals into ranked content, and the advertising system sold access to the resulting attention. The graph was valuable because it improved prediction.

An agent asks what the system may do next, given who a person is, whom they know, which business they are dealing with, and what they have authorized. The agent uses those relationships as context for action rather than inputs to a forecast.

Meta has introduced that transition in pieces. Meta AI can connect to Gmail and Google Calendar to perform tasks such as creating daily updates. Its AI Mode can search public Facebook posts across Groups and Reels to generate answers. One feature imports task context; the other searches social context.

Meta presented those capabilities separately rather than as one continuous personal agent. But an assistant inside a Meta surface could gather context from social activity, connect to a work tool, produce an answer, and return to a conversation where another person or business is already present.

To turn those pieces into delegated work, Meta must govern which context is available, where judgment enters, who can approve an action, how the result is reviewed, and what happens when it fails. Meta already owns many of the interfaces in which users would request and observe those decisions.

Distribution can outrank exclusive model ownership

At the research frontier, the company with the strongest model may win. At the product layer, companies must still place a model in an interface, connect it to tools, supply context, assign an identity, and secure permission to act. If capable models become available from several suppliers, the scarce layer moves upward to the system that can deploy them safely and habitually.

Meta is reportedly willing to consider Google or OpenAI models for Meta AI and features in its social apps. That challenges the assumption that its own model leadership is assured, but it also separates model procurement from distribution. A rival model inside WhatsApp would still enter through Meta’s account system, operate under Meta’s product rules, and meet customers inside a Meta-owned conversation.

Google could supply both models and the Gmail and Calendar services to which Meta AI connects. OpenAI could supply models and a powerful standalone assistant. Meta owns Facebook’s public and private social contexts, Instagram’s creator and business relationships, and the messaging threads of WhatsApp and Messenger. Any outside supplier would still depend on Meta to reach users through those surfaces.

Meta can make each layer increase demand for the next. When a user connects another tool, the assistant can handle more tasks. When businesses move more conversations through agents, Meta needs more inference capacity. Existing social and messaging surfaces reduce the cost of placing assistance in front of users, though Meta still has to make them choose it.

Business messaging is the shortest route from assistance to action

Meta’s clearest commercial strategy puts an agent inside a customer conversation, attaches it to a verified business identity, and lets it move the exchange toward support, a booking, or a sale.

WhatsApp Business grew from 50 million monthly active users in 2020 to 200 million in 2023. That growth preceded the current agent push, giving Meta an installed base of businesses and customers already communicating in the same threads.

Meta Business Agent extends that design across WhatsApp, Instagram, and Messenger, where it can answer customer questions and close sales. The same product push opened Meta’s advertising ecosystem to third-party AI tools for creating, managing, and analyzing campaigns. A business can move from audience acquisition to conversation, assistance, transaction, and continuing customer contact without leaving Meta’s commercial system.

When an agent answers customer questions, businesses still decide refund limits, escalation rules, approved claims, payment authority, and when a person must take over. Meta must sell businesses the decision system around the model, not just the model itself.

A business can compare the cost of handling a conversation, the speed of a response, and whether the interaction produced support or a sale. A consumer assistant must establish both a habit and a monetization model; a business thread begins with a job and an economic measure.

Openness supplies intelligence; platform rules allocate demand

Open-weight models let Meta recruit developers, spread experimentation, and remain relevant beyond its own applications. Meta said Llama models had been downloaded about 350 million times, while usage through cloud providers more than doubled from May to July. It later introduced a Llama API for fine-tuning and evaluation, adding a managed route into the same model family.

A developer may download Llama, yet WhatsApp’s Business API terms barred general-purpose chatbots beginning in January 2026, affecting assistants from companies including OpenAI and Perplexity. Under those terms, outside developers can use Meta’s models without gaining equal access to Meta’s messaging gateway.

Meta’s advertising connectors create the opposite outcome for tools that strengthen its commercial workflow. An AI tool can help create, manage, or analyze a campaign, while a general assistant cannot use the same business interface to compete for the primary relationship with the user.

Developers also enlarge the security surface when they connect open models to credentials, tools, and execution systems, as the Hugging Face security arc demonstrates. Meta can distribute model weights broadly while keeping the path to consequential action bounded by platform rules.

The software rail terminates in concrete, power, and capital

An agent can appear weightless on a phone screen, but persistent assistance requires physical capacity. Every search across public posts, multimodal interaction, business response, and tool call returns to processors, memory, networking, cooling, power, and technicians inside a building.

Meta and BlackRock formed a venture for a one-gigawatt data-center campus in El Paso expected to cost about $14 billion, with capacity beginning in 2028. Separately, Reuters reported in October 2025 that Meta and Blue Owl Capital formed a joint venture to finance the $27 billion, two-gigawatt Hyperion data center in Louisiana. Meta retained roughly 20% of the venture’s equity.

Once Meta signs a power agreement, connects transmission, or builds a data-center shell, it cannot resize the commitment as easily as it can revise software. The company must forecast demand years before capacity arrives, making sustained use a strategic requirement.

The El Paso campus will not begin bringing capacity online until 2028. Meta and BlackRock must align construction, power delivery, equipment deployment, and product demand across different schedules. Decisions about agent features now help determine whether expensive capacity will be useful when it arrives.

The one- and two-gigawatt campuses also bind Meta’s software plans to land, construction, power, and financing. Every agent request inherits decisions made far beneath the message thread.

Permission to act is scarcer than access to a model

Meta can place an assistant in front of a user. It cannot make that user authorize the assistant to read a conversation, inspect a calendar, contact another person, spend money, or speak for a business. The more useful the agent becomes, the more consequential the permission becomes.

Privacy concerns already affect the category. Apple reportedly delayed its own AI-glasses launch partly because of concerns associated with Meta’s glasses and the category they helped define.

Meta must earn permission separately for each kind of observation and action. A camera on a face, an assistant inside a private message, and a bot answering for a business create different boundaries, even when the same account connects them.

Meta therefore needs consent architecture with permissions scoped to a task, visible before consequential actions, revocable after authorization, and reviewable when something goes wrong. A daily calendar summary may require one level of oversight. Sending messages, approving a purchase, changing an appointment, or making a claim on behalf of a business requires another.

Businesses still need a person to own the judgment embedded in refund limits, escalation paths, contact rules, and transaction authority. Meta must keep consequential actions visible without sending every routine step through an approval queue, concentrating human review where responsibility and risk are greatest.

Distribution is proven; delegation is not

Meta has stronger evidence for distribution than for mass agent behavior. Vendors still lack a shared definition of an AI agent, using the term for products ranging from conversational assistants to systems that execute long sequences of work. A user cannot grant informed permission to a category whose autonomy, tools, and failure modes remain unclear.

The 200 million WhatsApp Business users establish reach. They do not establish that consumers will routinely delegate personal tasks, that businesses will entrust agents with revenue-bearing decisions, or that either group will pay enough to support the infrastructure underneath them.

Giving a model more tools, relationships, and implied authority raises the cost of its errors. A poor answer from a standalone chatbot may waste time; the same error inside a customer account can change an appointment, mishandle a refund, or send an unauthorized message.

WhatsApp’s rules create another tension. By excluding general-purpose assistants, Meta protects its gateway, but users still have to prefer the allowed experience. The policy can concentrate usage while encouraging developers and customers to build elsewhere.

OpenAI’s move toward supervised agent work reflects the same operating constraint. Teams are using policy, checkpoints, logs, and named responsibility to turn autonomous steps into routine work.

The old signal is becoming a switch

If Meta joins its released pieces, a customer could ask a WhatsApp business to change an appointment while Meta AI consults a connected calendar. The business would still decide when software may alter the booking, and the user would still have to authorize calendar access. The workflow begins in a familiar thread but depends on two explicit grants of authority.

The old graph predicted which message appeared next. Meta now wants a message to trigger work across a calendar, a customer account, and a one-gigawatt data-center campus. For WhatsApp Business’s 200 million monthly users, that rail still ends at the same place: the permission screen.