TLDR: An AI agent now screens early-stage deals 537 times faster than a human analyst, yet venture management fees keep rising — because the fee was never only the price of screening labour, and judgment stays scarce.
An LLM agent screened 61,814 ventures 537 times faster than a human analyst
The economics of venture screening just changed. A peer-reviewed study led by Silvio Vismara and co-authors ran a large language model (LLM) agent across 61,814 early-stage ventures and found it operated 537 times faster than a human analyst with no loss of categorisation quality. The same agent delivered a roughly 70 per cent improvement on the Calinski–Harabasz Index, a standard measure of how cleanly clusters separate — the machine sorted the funnel more sharply, not merely more quickly.
Screening is the part of the venture workflow most exposed to this shift. Harvard Business Review, surveying how generative AI is reshaping venture capital, describes a workflow in which sourcing and first-pass diligence move from artisanal craft to industrial throughput. The labour that once justified a team of analysts now runs at the speed of inference.
Management fees climbed from $150M per top firm in 2005 to $1.6B by 2025
While the cost of screening collapsed, the price of access kept rising. Dan Gray of Odin, writing on what he calls the magical money tree of management fees, traces the inflation directly. A pre-ZIRP (zero interest-rate policy) successful company once justified around $10 million in total management fees across its private life; the modern held-private equivalent justifies closer to $250 million. In 2005, the five top-fundraising venture firms each generated roughly $150 million in subsequent fee income. By 2025, that figure had climbed to roughly $1.6 billion.
Gray frames the mechanism as Horowitz’s Law — capital flows to wherever the fees are highest. Companies now stay private far longer and raise far more: Rivian raised $10.5 billion in private capital before listing, against Tesla’s roughly $200 million. Each additional private dollar under management is a dollar that earns a fee, and the structural incentive favours keeping the meter running.
The principal–agent lens explains why fees pay for two different things
The agency-cost framework set out by Michael Jensen and William Meckling clarifies what a fee actually buys. A limited partner (LP) hires a general partner (GP) as an agent, and the management fee compensates that agent for two distinct things bundled together: labour and judgment. Labour is the screening, the spreadsheets, the first calls — the work AI now does in a fraction of the time. Judgment is conviction, access to oversubscribed rounds, and the willingness to back a contrarian thesis before it is consensus.
AI collapses the cost of the first while leaving the second untouched. That is the tension at the heart of the modern fee: the industry charges artisanal prices for screening that has become industrial, even as the genuinely scarce inputs — judgment, access, conviction — remain exactly as scarce as before.
Public-market managers already run closer to 0.5% than 2%
A reference point already exists for what fees look like once labour is commoditised. Public-market managers, who screen thousands of listed names with heavily automated tooling, run closer to 0.5 per cent than the venture standard of 2 per cent. The gap between those numbers is, increasingly, a gap in narrative rather than in cost structure. Pressure on that narrative is building through concrete instruments: in the United States, the Institutional Limited Partners Association (ILPA) reporting templates and management-fee offset standards give LPs a common language to ask where a fee goes, and European and Swiss LPs are importing the same templates into side letters.
What to do, by seat
General partners should separate the fee narrative into its two components and price them honestly. Let AI absorb screening throughput, then articulate the fee as payment for judgment, access and conviction — the inputs that do not scale with compute.
Limited partners should use the ILPA reporting templates and fee-offset provisions to ask what proportion of the management fee now funds work that AI performs, and benchmark the answer against the 0.5 per cent public-market managers charge for comparable scale.
Founders should read the fee structure as a signal of where an investor adds value. An investor whose value is screening offers little an AI agent cannot; one whose value is access, conviction and judgment offers the scarce thing.
References
- Vismara, S. et al. Generative AI-powered venture screening. ScienceDirect. https://www.sciencedirect.com/science/article/pii/S105752192500835X
- Harvard Business Review. How Generative AI Is Reshaping Venture Capital. https://hbr.org/2025/11/how-generative-ai-is-reshaping-venture-capital
- Gray, D. (Odin). The Magical Money Tree of Management Fees. https://blog.joinodin.com/p/the-magical-money-tree-of-management
