MARK ZUCKERBERG takes home a $1 annual salary. His actual compensation package in 2024 — $27.2 million, nearly all of it security costs and private-jet usage — is a rounding error relative to the $700 million in annual dividend income his controlling stake in Meta generates. The man does not need the money. What he apparently does need is an AI agent to help him do his job. According to the Wall Street Journal, Zuckerberg is building a personal "CEO agent" that retrieves information he would otherwise have to extract through layers of subordinates. The tool is still in development, but its mere existence raises a question that $17 million-a-year executives across the S&P 500 would prefer not to confront: if an AI can absorb the information, synthesize the context, and surface the answer, what exactly is left for the chief executive to do?
The project sits inside a broader frenzy at Meta to make its 78,000-person workforce operate like an AI-native startup. Employees now use internal tools called My Claw and Second Brain — the latter built on Anthropic's Claude — to query documents, coordinate with colleagues, and dispatch their own personal agents to talk to other employees' agents. (There is, apparently, an internal message board where the agents chat amongst themselves, which is either the future of enterprise collaboration or the opening scene of a film nobody wants to watch.) Meta recently acquired Manus, a Singaporean startup specializing in autonomous personal agents, and established a new applied-AI engineering organization with an ultraflat structure: as many as 50 individual contributors reporting to a single manager. The company is, in Zuckerberg's words, "elevating individual contributors and flattening teams." AI adoption is now a factor in performance reviews. The atmosphere, some employees say, recalls the old "move fast and break things" era — albeit with considerably more anxiety about layoffs.
The $17 million question
But the CEO agent is not merely an internal productivity hack. It is a philosophical provocation wrapped in an enterprise software project. The median S&P 500 chief executive earned $17 million in total compensation in 2024, a figure that has risen nearly every year for a decade — justified, in theory, by the irreplaceable nature of executive judgment. CEOs are paid not to know things but to decide things: to weigh incomplete information under uncertainty, to make calls that cannot be delegated, to bear accountability for outcomes. The CEO-to-median-worker pay ratio now sits at 192:1, a number that only makes sense if the person at the top is doing something a system cannot.
Zuckerberg is not the only tech titan flirting with his own obsolescence. OpenAI's Sam Altman told an audience in New Delhi in February that AI superintelligence would soon be capable of running a major company better than any human executive — "certainly me." On a podcast last year, he went further, saying he would be "nothing but enthusiastic" when an AI could be a better CEO of OpenAI than himself. Google's Sundar Pichai has made similar concessions. These remarks tend to be received as charming self-deprecation. They are, in fact, something more interesting: the early-stage dissolution of the myth that C-suite judgment is a uniquely human asset.
The trouble is that corporate governance has no framework for a world where agents make consequential decisions on a CEO's behalf. If Zuckerberg's agent retrieves flawed data and he acts on it — approving a product launch, authorizing a capital expenditure, greenlighting a policy shift — the legal and fiduciary accountability still rests with him. But the causal chain has been altered. The CEO did not gather the information, did not weigh the sources, did not interrogate the assumptions. He reviewed an output. Boards of directors are not yet asking whether "the CEO used AI to inform this decision" belongs in risk disclosures, but they will. The SEC has barely begun to grapple with algorithmic trading liability; agent-assisted executive decision-making is several layers more complex. MIT Sloan researchers noted earlier this year that hallucinations and prompt-injection vulnerabilities have already slowed enterprise adoption of agentic systems. The C-suite is not exempt from those failure modes — and the stakes are considerably higher than a miscategorized expense report.
There is a more unsettling possibility still. If agents can retrieve, synthesize, and contextualize information faster than any human hierarchy, then the traditional justification for executive compensation — that the CEO's time and judgment are the scarcest resources in the organization — begins to erode. What remains is something closer to a figurehead role: the CEO as chief narrative officer, the human face required by regulators, investors, and the press. The job becomes less about making decisions and more about owning them, which is a very different skill set — and arguably a much cheaper one.
Zuckerberg, of course, is uniquely insulated from this reckoning. His dual-class share structure gives him 61% of Meta's voting power regardless of what any agent recommends. He can afford to experiment with his own replaceability because, in his case, the CEO and the controlling shareholder are the same person. Most executives do not enjoy that luxury. For them, the question posed by Zuckerberg's little side project is considerably less academic — and the answer may arrive on a board's agenda sooner than any corner office would like. ■
For more, join 75,000 subscribers getting tech's favorite brief here
