Meta Platforms CEO Mark Zuckerberg has told employees that the company’s AI agent technology isn’t developing as fast as expected, a candid admission that one of Meta’s biggest bets in the AI race hasn’t delivered the momentum executives were counting on.
Zuckerberg made the comments during an internal town hall, according to Reuters, saying that progress in agentic AI over the past four months had not “accelerated in the way we expected.” For a company spending at the scale Meta is, that’s a notable thing to say out loud.
The stakes are high because AI agents represent something more ambitious than the chatbots and assistants most people are familiar with. Rather than just answering questions, agents are supposed to take actions, complete multi-step tasks, and help users and businesses manage workflows with minimal manual input. For Meta specifically, that could mean AI systems that handle advertiser tools, customer service automation, creator support, and small business services across Facebook, Instagram, and WhatsApp.
Zuckerberg didn’t sugarcoat where things stand. He said Meta’s internal restructuring hadn’t been as “clean” as it could have been, that bets placed on the new structure “haven’t come to fruition yet,” and that executives had misjudged the timing. Coming from a CEO who typically projects confidence about AI, that’s a fairly direct acknowledgment that the plan hasn’t gone to schedule.
Meta has been pushing hard to keep pace with OpenAI, Google, Anthropic, and Microsoft in a race where falling behind feels costly. The company has poured money into models, data centers, chips, and talent, and is projected to spend as much as $145 billion on AI infrastructure this year alone. That level of investment makes the agent slowdown matter more, not less, when you’re spending that much; “not yet” is an expensive answer.
The technical challenges are real and not unique to Meta. AI agents are genuinely hard to make reliable. They have to correctly interpret user intent, take the right actions, avoid compounding mistakes across multiple steps, and work across different software systems, often in real time. Even the most capable models available today can fail badly when tasks get complicated, long-running, or dependent on constantly changing information.
The organizational side has created its own friction. Meta carried out a significant restructuring earlier this year, cutting jobs and moving thousands of employees into AI-focused roles. Zuckerberg had previously acknowledged that the company made mistakes during that transformation while promising more stability ahead. The latest comments from the town hall suggest the disruption isn’t fully behind them yet.
A separate issue surfaced alongside all of this. Meta paused a controversial internal program that tracked employee mouse movements and digital activity for AI training purposes, following a data-security review. Chief Technology Officer Andrew Bosworth told employees no employee data had actually been used in AI training, and said any future version of the program would run on an opt-in basis. The episode is a reminder of a tension that every AI company is navigating: the need for more data to improve models, which runs directly into growing employee and user concerns about privacy and consent.
There’s also a compute problem. According to a Financial Times report cited by Reuters, Google reportedly limited Meta’s access to Gemini models after Meta requested more AI capacity than Google could actually provide. That shortfall apparently disrupted and delayed some internal AI projects, which is striking for a company of Meta’s size and illustrates just how tight AI computing capacity remains even at the top of the industry.
For more on how Meta’s AI infrastructure strategy fits into the bigger picture, see our coverage of Meta’s AI cloud business plan.
Zuckerberg told employees he still expects Meta to start seeing meaningful benefits from its AI investments within the next three to six months. That’s a reasonably specific timeline, which means it’s also a window that employees, investors, and developers will be watching closely.
The overall picture is complicated. Meta is still one of the biggest spenders in the AI race by any measure, and Zuckerberg hasn’t shown any sign of backing away from his ambitions. But progress on the agent side has been slower than planned, and the gap between enormous investment and practical products that actually work has proven wider than expected.
The real test is coming. Meta needs to demonstrate that all this spending ultimately produces tools that users and businesses genuinely rely on. Right now, it’s a high-investment project with clear ambitions and uneven results. The next few months will say a lot about whether the gap is closing.
0 Comments