Mark Zuckerberg is building an AI agent to help him run Meta. According to the Wall Street Journal, the system retrieves information that would normally require coordination across multiple teams and management layers, giving him direct access to data that would otherwise take days to surface through traditional channels.
This is not a productivity tool. It is an attempt to solve a structural problem that every organization shares: the distance between the person who has the information and the person who needs it to make a decision.
What Meta Already Built
Meta’s internal AI ecosystem is further along than most companies realize. MyClaw gives employees direct access to internal files and chat logs without navigating bureaucratic handoff points. Second Brain, built on Anthropic’s Claude, functions as a personal chief of staff, organizing tasks, surfacing insights, and retrieving institutional knowledge. The choice of which AI model powers these executive tools is itself a decision framework challenge as the enterprise model market reshuffles every year.
The results are measurable. Susan Li, Meta’s CFO, reported during the Q4 2025 earnings call that output per engineer has risen 30% since the start of 2025. Employees who fully adopted the AI tools saw 80% output gains year over year. Meta backed the direction with $2 billion, acquiring Manus, a general-purpose AI agent developer, in December 2025, and established Meta Compute as a new top-level organization led by Santosh Janardhan and Daniel Gross, the latter recruited from Safe Superintelligence.
The company has forecast capital expenditure of $115 billion to $135 billion for 2026, nearly double the $72 billion it spent in 2025. The infrastructure bets across the industry are substantial, but Meta’s internal tooling shows where the company expects the returns to come from.
Why the CEO Tool Is Different
Employee AI tools solve for speed: faster coding, faster research, faster task management. The CEO tool solves for something else entirely. It addresses the gap between when a leader needs an answer and when the organizational structure delivers one.
In a company of 70,000 people, information passes through multiple layers before reaching the executive suite. Each layer adds delay. Each layer also adds interpretation, context loss, and the natural tendency to present information in the way the presenter thinks the CEO wants to hear it. The result is that the person making the decision often works with a filtered, delayed, reinterpreted version of the raw data.
That dynamic is not unique to Meta. In a 200-person company, the same filtering happens. The layers are fewer, but the information still gets summarized, packaged, and reshaped at every handoff. The information gap Zuckerberg is solving is the same one that boards face when overseeing AI initiatives. The question is not whether this happens in your organization. It is how much decision quality you are losing because of it.
Zuckerberg told investors that projects once requiring entire teams can now be handled by a single talented person. Applied to the executive function, that means the coordination overhead between a question and an answer, the meetings, the email chains, the intermediate reports, may itself be the problem worth solving.
What This Means for Your Organization
The McKinsey approach to internal AI was to map every role against two questions: does it require human judgment, and does it require human presence? Zuckerberg’s tool asks a third question that most organizations have not considered: does the path information takes to reach the decision-maker actually need to exist?
Not every layer is unnecessary. Middle management adds judgment, context, and prioritization that raw data does not provide. And not every AI deployment improves decisions. There is a strong case for using AI more selectively, especially when cognitive dependency replaces the critical thinking that made the human layer valuable in the first place. But some layers exist purely because of organizational structure, because the information lives in one system and the decision-maker sits in another, with three people in between whose job is to move it across.
A practical starting point is to map the journey of one critical decision you made recently. Trace the information from its origin to your desk. Count the handoffs, estimate the delay at each, and ask a simple question at every layer: did this step add judgment, or did it just add distance?
The line between a personal AI assistant and a secondary CEO is thinner than most leaders assume. The companies that recognize this early will redesign their information flow deliberately. The ones that do not will keep making decisions on data that was already old by the time it arrived.
Questions this article gets
What AI tools does Meta use internally?
Meta employees use MyClaw for direct access to internal files and chat logs, and Second Brain, built on Anthropic's Claude, which works as a personal chief of staff for organizing tasks, surfacing insights, and pulling institutional knowledge. Output per engineer is up 30% since the start of 2025, with power users seeing 80% year-over-year gains.
What is Manus and why did Meta acquire it?
Manus is a general-purpose AI agent developer that Meta acquired for 2 billion dollars in December 2025. The acquisition supports Meta's broader push to embed AI agents across the organization, from individual contributor tools to executive-level information retrieval.
What is the difference between employee AI tools and executive AI tools?
Employee tools like MyClaw and Second Brain optimize productivity by handling tasks, organizing information, and accelerating coding. An executive AI tool solves a different problem: it bypasses the organizational layers that filter, delay, and reinterpret information before it reaches the CEO. The bottleneck it addresses is structural, not operational.
How can CEOs assess their own information access bottleneck?
Start by mapping the path a critical piece of information takes from the person who has it to the person who needs it to decide. Count the layers it crosses, estimate the delay at each, and identify where context gets lost. Then ask which of those layers exists because of organizational structure rather than because it adds judgment or value.