When Ken Griffin — founder of Citadel, one of the world’s most quantitatively sophisticated hedge funds — told Stanford students he went home “fairly depressed” after watching AI work inside his own firm, the investment world paid attention. The work that previously required master’s and PhD-level quantitative analysts — weeks or months of structured research — is now being completed in hours or days. Griffin didn’t say this as a prediction. He said it as a report from inside the building.

The Verifiability Principle

In May 2026, Andrej Karpathy published a summary of his Sequoia AI Ascent fireside chat that contains the clearest available framework for understanding which expert jobs fall to AI first: LLMs automate what you can verify. Traditional computers automate what you can specify in code. LLMs automate what you can check after the fact.

This distinction is decisive for finance. Quantitative finance — the work Citadel’s PhDs have done for decades — runs on verifiable outputs: model predictions can be backtested, signals can be checked against market data, and errors surface directly in profit-and-loss statements. When the reward signal is unambiguous and continuous, AI systems can train themselves to perform. Citadel’s quantitative workflow is structurally the ideal environment for AI automation.

The pattern extends beyond hedge funds. NYU Stern finance professor Aswath Damodaran — known as the “Dean of Valuation” on Wall Street — created AI systems specifically trained on thousands of his published valuations. His public framing: he is “trying to stay ahead of my bots.” The world’s most prominent valuation practitioner describes his professional work as a race against AI systems that already know his methodology. That’s not a warning about the future. That’s a description of the present.

Where the Pattern Holds — and Where It Breaks

The verifiability principle predicts domain dominance with precision. Finance automation accelerates because the outputs are checkable:

  • Quantitative research: model outputs backtested against historical data
  • Financial modelling: formulas either reconcile or they don’t
  • Regulatory document review: answers cross-checkable against statute text
  • Earnings analysis: claims verifiable against filed numbers

But the same principle also predicts where AI stalls. Market timing, founder assessment, LP relationship management, and strategic judgment all share a common structural problem: the reward signal is ambiguous. Was the investment decision right? You find out in 5-7 years, not 5 seconds. You cannot run reinforcement learning on a signal that resolves in a decade.

A November 2025 McKinsey survey of 1,993 organisations found 62% were experimenting with AI agents, but only 23% were scaling them at enterprise level. The 39-point gap between experimentation and scale is where verification complexity lives. The use cases that scaled fastest — code review, document summarisation, data extraction, financial modelling — are predominantly verifiable. The ones that stalled are judgment calls without clean feedback loops.

Why This Matters for Investors and Founders

The verifiability principle is the most reliable map of which expert roles compress first and which remain defensible. For investors: any firm whose workflows are built on verifiable analytical steps — quant research, document review, financial modelling, sector scanning — faces headcount compression within 24 months. Not because AI becomes “intelligent” in a general sense, but because the work is checkable, and checkable work is trainable.

For founders: the opportunity is not to build AI that replaces judgment. It’s to build AI that handles all the verifiable steps upstream of judgment, so human decision-makers can deploy their finite attention exclusively on the 20% of the work that actually requires them.

Griffin’s depression is structurally understandable. He built a firm on the theory that if you hire enough PhDs with enough computing power, you can find signals others miss. That theory still holds. But the PhDs are now agents.

The Charaka View

Our Analytical Council pipeline embodies this exact distinction at operating scale. The verifiable half of investment analysis — financial mechanics, regulatory flag scanning, traction metric extraction, market size triangulation — runs on AI agents, in parallel, in hours. The one task that stays human: the conviction call that integrates all of it into a funding decision.

Manthan’s data on 13,900+ companies and 333 postmortems confirms what Griffin’s admission implies: the verifiable half of expert finance work is already automating. The judgment half isn’t. Yet. The question every investment professional should be asking isn’t whether this is happening. Griffin answered that on a Friday night when he went home depressed. The question is: what exactly is left for you to do that a model can’t check?

This analysis draws on Benzinga’s report on Ken Griffin’s Stanford talk, Andrej Karpathy’s Sequoia AI Ascent 2026 summary, The Data Exchange’s profile of the Damodaran AI valuation system, and McKinsey’s State of AI November 2025 report. Human editorial oversight applied.

This analysis is informational and does not constitute investment advice, a research report, or a recommendation to buy, sell, or hold any security.

Charaka Notes by Manthan Intelligence. Subscribe