Mark Zuckerberg is outsourcing his thinking to AI. It shows in his new ‘personal AI agent’ experiment

A recent Reuters investigation offers a revealing glimpse of where that idea can lead. Meta’s Project OT envisioned an “AI-native” company in which AI agents would take over much of the work performed by thousands of employees. Teams could become dramatically smaller. Some layers of management could disappear. Zuckerberg himself has been using what Meta calls a “CEO agent,” allowing him to retrieve answers that previously required going through several layers of staff.

The attraction is obvious. Every organization contains friction. Meetings take time. Information gets lost between layers. There’s push back.  Decisions move slowly. AI can compress all of this.

Yet some friction serves a different purpose. It comes from people who disagree, see problems differently, remember inconvenient facts or ask questions that others would rather avoid.

That distinction matters enormously as AI moves from helping people perform tasks to helping leaders think.

We call these two forms of friction coordination friction and cognitive friction. Coordination friction comes from the mechanics of collective work: scheduling, documentation, data reconciliation, approvals and communication. AI can reduce much of this friction with obvious benefits.

Cognitive friction comes from the resistance between ideas. Two people may look at the same evidence and reach different conclusions because they have different assumptions, experiences or ways of framing the problem. That friction can be uncomfortable. It also can be generative.

This creates a peculiar danger for people at the top of organizations.

Imagine a CEO who once received information through several layers of people. Each layer introduced interpretation. Each person brought a different mental model. Someone might challenge the premise of a question. Someone else might point out an anomaly. Another might say, “We tried this three years ago, and here is what happened.”

An AI agent can make that entire process dramatically faster. It can retrieve the information, synthesize it and present a coherent answer.

And coherence can feel like intelligence.

The danger arises when AI becomes a sophisticated mirror.  Instead of creating distance from an existing worldview, it can make that worldview more articulate, comprehensive and persuasive. The user experiences an apparent external intelligence while interacting with a system that has learned from the user’s preferences from their previous questions, prompts, assumptions, and accumulated information.

This is self-referentiality automated at scale.

There is a useful distinction here between two possible roles for AI: peacemaking and sensemaking.

A peacemaker resolves contradictions. It finds common ground, smooths disagreements and produces an answer that hangs together.

A sensemaker does something harder. It exposes contradictions. It identifies hidden assumptions. It searches for evidence that does not fit. It constructs the strongest argument against the user’s position. It keeps competing interpretations alive long enough for them to teach us something.

For an individual thinker, we argue that AI should function more like a sparring partner than a mirror.

This distinction becomes especially important for CEOs because their organizations already tend toward self-reference. The higher someone rises, the more information is filtered before reaching them. The very efficiency of an AI system can intensify that tendency. A CEO agent may give its user faster access to the organization while simultaneously reducing exposure to the people who would have challenged the organization’s assumptions. (maybe: AI is The New Organization Man – the ultimate corporate “yes-person” that ideally fits into the organization, respects hierarchy, and suppresses individual judgment and self-reflection.  Instead of Groupthink we have Algothink)  

This may help explain a larger paradox emerging at Meta.

Project OT was built around the premise that AI could allow smaller, more “talent-dense” groups to accomplish work previously requiring much larger teams. Meta ultimately abandoned its most aggressive workforce-reduction plans after internal data raised questions about whether dramatically increased AI-assisted coding was translating into comparable gains in user-facing products.

The lesson extends far beyond Meta.

AI makes it increasingly attractive to build end-to-end systems. In science, this means systems that can generate hypotheses, run experiments, analyze results and produce scientific papers. In organizations, the emerging equivalent runs from information gathering through synthesis, decision-making and execution.

The more complete the loop becomes, the more important the points of friction become.[BU1] 

A fully integrated system can become exceptionally good at optimizing what it already believes matters. Its greatest weakness may emerge at the boundary where someone asks whether the system is solving the right problem.

Science provides a useful warning. Scientific progress depends on variation. Different laboratories pursue similar questions using different methods. Researchers make different bets. Most fail. Occasionally, an unexpected result opens an entirely new direction. Our paper argues that end-to-end AI systems could compress this variation by steering exploration toward what existing data suggest is likely to succeed.

Organizations face a similar choice.

AI can make coordination extraordinarily cheap. That is a tremendous opportunity. It also means that disagreement can become easier to eliminate, alternative interpretations easier to compress and decisions easier to accelerate.

The real leadership challenge of the AI era may therefore be learning where friction is a cost and where friction is a source of intelligence.

The best AI for a leader may sometimes be the system that slows the leader down: the one that says, “Here is the strongest argument against your position,” “Here is what your model cannot explain,” or simply, “Here is another way to see the problem.”

AI can make leaders faster. The harder challenge is making sure it also makes them more capable of seeing beyond themselves.