From ad targeting to personal agents
Meta’s third-quarter 2026 results show how AI can add value to a mature business. Revenue rose 26% year over year to $64.60 billion, including $63.06 billion from advertising. For the fourth quarter, Meta forecast revenue of $73 billion to $76 billion. AI is helping the company target ads more precisely, improve their relevance and recommend content to its large user base.
That model does not require selling consumers a separate AI subscription. AI can strengthen products and increase the returns from an existing business. Muse represents a more ambitious next step: making AI a consumer product in its own right.
The initiative began with the launch of Muse Spark in April 2026. Meta’s longer-term vision is personal superintelligence: AI that helps people handle difficult tasks in ways tailored to their needs. That points beyond chatbots, which mostly respond to prompts, toward agents that can gather information, plan, use tools and coordinate tasks within the permissions they have been given.
A shopping assistant, for example, might compare products, check compatibility, track prices and prepare a purchase for the user to approve. Meta can distribute assistants through Facebook, Instagram, WhatsApp, Messenger and Meta AI, placing them in services people already use rather than asking them to adopt a new platform.
Reuters reported that Muse attracted more than 10 million users in the US and Canada in the six weeks after launch. That is an early sign of uptake, not evidence of durable use or profitability.
The infrastructure bill
Meta’s advertising business gives it a financial base for AI development, but building advanced models and running agents at scale requires substantial spending on networks, energy, specialized processors, computing infrastructure and engineering teams.
In its third-quarter report, Meta listed capital expenditures of $34.91 billion, including principal payments on finance leases. It raised its full-year capital expenditure forecast to $135 billion–$145 billion. Those figures capture the central tension in AI economics: the technology may create significant value, but it takes large investments to build, and the returns are uncertain.
Agents could be particularly expensive. They process large amounts of contextual data, call external tools and carry out multiple stages of reasoning. For the economics to work, the value an agent creates must exceed the cost of operating it. Meta’s advertising business shows that AI can produce measurable commercial returns; it does not show that every new AI product will be profitable.
Personal agents could also create opportunities in enterprise applications, developer ecosystems, paid AI services and commerce. Those are possible revenue sources, not guaranteed ones. Their value will depend on how often people use agents and how much useful work those agents do.
More autonomy, more exposure
An agent that can access files, workplace applications, calendars, email or financial processes presents different security risks from a system that only generates text. The broader its permissions, the greater the potential consequences of stolen credentials, malicious instructions or mistaken actions.
Meta has described safety measures for Muse, including safeguards for agents handling sensitive resources. But no single technical measure can remove every risk that comes with greater autonomy. In my view, controls need to grow alongside agents’ ability to act.
Personalization brings a related concern. An agent may need access to conversations, schedules, preferences, documents and other details about a person’s life to be useful. Taken together, that context can reveal a great deal.
That makes transparency central to the product, not an afterthought. Users need to know what an agent can access, why it needs that access, how information is stored and whether it is used for commercial or advertising purposes. They also need a way to revoke permissions and understand how an agent reached an important decision.
My guess is that distribution will be the easier part of Meta’s plan. The harder question is whether people will trust an agent with consequential tasks when its operator also runs a major advertising business. Familiar apps can put an assistant in front of users; they cannot, by themselves, make it reliable or make its use feel appropriate.
The interface is the business
Muse points to a broader change in how people may use computing. People have long navigated digital services through apps, websites and search. Generative AI introduced conversational interfaces; agentic AI could make software responsible for coordinating steps toward a user’s goal.
That shift could affect commerce, customer service, education, health administration, software development and business operations. Competing in this market will depend not just on model performance, but also on computing infrastructure, useful applications, operating costs, security and public trust.
Meta has advantages in both monetization and distribution. The unresolved question is whether it can turn those advantages into sustained agent use while controlling infrastructure costs and protecting users. A large potential audience is not the same as a large population willing to delegate important tasks.
For companies, the shift means AI may become more than a productivity tool or a feature inside existing software. It could become an operating layer for digital services. For Meta, that makes the agent’s everyday usefulness and the boundaries around its access at least as important as the model behind it.
Daily AI news
Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.
Only what matters — every day
Follow on X