The Automation of Judgment: How AI Agents Are Reshaping High-Skilled Finance

The announcement landed quietly enough for a Tuesday morning in May. Ken Griffin, the founder and chief executive of Citadel—one of the most consequential market-making and hedge fund operations in the world—had told audiences that artificial intelligence agents were now automating tasks that his firm had long considered the exclusive domain of its most elite professionals. The word he reached for was "extraordinarily," a qualifier that in any other context might read as hyperbole but from a man who has spent four decades building one of the most technically sophisticated financial enterprises on earth carries different weight. This was not a venture capitalist hyping a demo. This was the owner of a firm that competes for, and wins, the very best minds from every top university on earth acknowledging that the cognitive work those minds once did is increasingly being done by machines.

That acknowledgment represents something of a structural break in how the financial industry discusses automation. For years, the dominant narrative held that artificial intelligence would handle the routine—the data processing, the pattern recognition, the execution of defined rulesets—while humans retained the judgment, the relationships, the capacity to navigate ambiguity. Griffin's framing suggests that boundary is no longer holding. When the chief executive of Citadel says AI agents are automating extraordinarily high-skilled work, he is not speaking about back-office functions or trade settlement. He is speaking about the analytical and decision-making tasks his firm has built its reputation on.

The hiring calculus at elite financial firms has long operated on a simple premise: acquire the most talented people, give them enormous resources, and let compound intelligence do the rest. Griffin's own stated hiring philosophy—that he wants people with high aspirations, tremendous perseverance, and grit—describes a profile built for environments where cognitive stamina matters more than raw analytical horsepower. Those qualities made sense when the bottleneck was human comprehension: the analyst who could process more information, hold more variables in mind, and sustain focus longer would outperform. The question now is whether those qualities remain the primary differentiator in a world where AI systems can process orders of magnitude more information without fatigue, without distraction, and at costs that decline rather than increase over time.

The financial industry has always had an uncomfortable relationship with automation. The first wave—the digitisation of trading floors in the 1970s and 1980s—eliminated the telegraph operator and the pit hand but created new categories of employment in systems management and quantitative analysis. The second wave, accelerated by algorithmic trading in the 1990s and 2000s, reshuffled the workforce again, elevating the role of the quant while reducing the discretionary power of the traditional portfolio manager. Each wave produced anxiety about job displacement and each wave ultimately produced more wealth, more trading volume, and more complexity than the previous one. The industry survived by absorbing the technology and reconstituting the work.

What distinguishes the current moment is the nature of the tasks now in play. Early automation targeted repetitive, rules-based processes—order matching, risk calculation, reporting. The analytical tasks that remained required judgment: the assessment of a management team, the reading of macroeconomic signals, the interpretation of incomplete data in conditions of genuine uncertainty. These were, in the industry's own taxonomy, the high-skilled functions. They were the reasons firms paid seven-figure salaries to professionals straight out of graduate programmes. They were the functions that justified the extraordinary resources poured into human capital development.

If Griffin is correct—and his firm processes more equity volume than any other market-maker in the world, giving him unusually good visibility into the dynamics of market participation—then those functions are now themselves subject to automation. The implications cascade in several directions at once. For the firms themselves, the economic logic is compelling: an AI agent does not require a salary, a bonus, a corner office, or a retention package. It does not jump to a competitor after three years. It does not have ambitions that need to be managed. The cost curve points in one direction, and it is not up.

For the professionals who currently occupy those roles, the calculus is more complex. The argument in favour of human judgment in finance has historically rested on several claims: that markets are not fully rational, that relationships matter for deal flow, that regulatory environments are too unpredictable for pure algorithmic response. Each of those claims retains some validity. Markets do exhibit anomalies that resist mechanical explanation. Institutional relationships do generate business that pure price competition cannot replicate. Regulatory change does require interpretive judgment that a rules-based system cannot easily anticipate. But none of those arguments is an absolute barrier to AI adoption—they are friction, not prohibition. And friction, in a competitive environment, tends to erode.

The more honest question may not be whether AI will replace high-skilled finance roles but which specific functions within those roles are most exposed. A financial analyst who spends sixty percent of their time gathering and normalising data is in a different position than one who spends sixty percent of their time building and maintaining client relationships. The former function is highly automatable; the latter is not, at least not in the near term. The firms that will navigate this transition most successfully are likely those that redesign workflows around that distinction—automating the data work ruthlessly while concentrating human effort on the functions where judgment, trust, and relationship are genuinely irreplaceable.

There is also a distributional question that the industry's public statements rarely address directly. The professional services economy of finance—law, accounting, consulting, investment banking—has historically served as a primary path to wealth for graduates from middle-income backgrounds who could compete on analytical merit. The economics of elite finance have always been starkly unequal, but the ladder existed. If the analytical core of those professions is automated, the question of who gets access to the remaining human-only roles becomes acute. The risk is not mass unemployment in finance—the sector has always been small in absolute employment terms—but the closing of a particular pathway to wealth accumulation for those without existing capital or connections.

Griffin's own language about hiring provides an instructive window into how the industry's self-understanding is evolving. The qualities he cites—aspiration, perseverance, grit—are precisely the qualities that are hardest to automate, because they describe disposition rather than information processing. They are also the qualities that have always characterised the industry's mythology about itself: the relentless meritocracy, the proof-of-work culture, the idea that the right kind of person can compete regardless of background. If those qualities are indeed what remains valuable in a world of AI agents, then the industry's culture may survive even as its economics transform. But the number of humans whose qualities matter to those economics may be substantially smaller than the number who currently benefit from the industry's extraordinary compensation structures.

The timeline for this transformation remains contested. Griffin's acknowledgment of current capability does not specify pace—his firm has long been among the most aggressive early adopters of new technology, and what is deployable at Citadel may be years from reach for smaller operators. Regulatory constraints, particularly in markets like the United States where financial services are heavily supervised, may impose friction on AI deployment that pure technological capability would not. And the risk management functions that sit at the centre of any market-making operation may prove more resistant to automation than the analytical tasks surrounding them, simply because the consequences of model failure in those domains are catastrophic in ways that are difficult to fully specify in advance.

What seems clear is that the era of comfortable assumptions about the limits of automation in finance is ending. The industry spent decades operating on the premise that the most sophisticated cognitive work was inherently human. That premise is now being tested against evidence rather than assumption—and the evidence, from a man who has built his career on processing evidence better than his competitors, suggests the premise is weaker than the industry believed. The professionals who thrive in the next decade will likely be those who understand which side of that distinction their current work falls on, and act accordingly.

This article draws on public statements by Citadel chief executive Ken Griffin and the firm's publicly documented market operations. Monexus has no corporate relationship with Citadel or any affiliated entity.

Wire provenance

This editorial synthesis draws on the following public wire/social posts:

• https://x.com/polymarket/status/1921874342010200453
• https://x.com/unusual_whales/status/1921497566678880512