Generative AI in Asset Management: Alpha is real but it’s not a “plug and play”
(A Curated CAIO’s read of the evidence and what leaders should do next)
Over the last two years since the popularity of ChatGpt, I’ve heard the same two claims repeated in finance and analytics circles:
- “GenAI will level the playing field because everyone can access the same tool.”
- “GenAI is hype it won’t generate durable differential performance.”
Last night I read a rigorous new academic paper, “Generative AI and Asset Management” (Aug 2025), which suggests both statements are incomplete. The data driven reality is more nuanced and actionable for anyone building AI strategy in a regulated, high stakes environment of active asset management. /
Below is my CAIO take: what the authors found, why it matters, and how I’d translate it into a practical curated roadmap.
Paper: https://www.business.rutgers.edu/sites/default/files/documents/generative-ai-and-asset-management.pdf
1. Adoption wasn’t gradual it was a step change
The paper documents a sharp rise in hedge funds’ GenAI usage after the release of GPT 3.5 and ChatGPT era capabilities, the fastest adoption was by larger hedge funds who were already using traditional predective data science.
After adjusting for a low false positive rate, they estimate:
- Approximately 21 percent of hedge funds adopted GenAI in 2022
- 40 percent plus in 2023
- approaching approximately a strong 60 percent by 2024 The paper itself was published in 2025 so we have to yet see 2026 where sophistication in LLM is a norm.
CAIO lens: That’s not “innovation theater.” That’s an operating shift. In enterprise terms: once adoption crosses a threshold, competitive advantage moves from whether you use GenAI to how well you operationalize the swiss knife at hand.
2. The performance effect is economically meaningful and statistically vaid for hedge funds
The headline result: hedge funds that adopt and rely more on GenAI show higher raw and risk adjusted performance in the post ChatGPT release and adoption period.
In their difference in differences (heenceforth called DiD) design, an interquartile increase in “GenAI Reliance” corresponds to approximately 2% to 4% higher annualized abnormal returns (varience depending on the factor model).
Important nuance: this performance benefit does not generalize cleanly to smaller non hedge fund asset managers in the same way.
CAIO lens: GenAI isn’t a universal performance booster. It’s an amplifier in essence and amplifiers reward firms with the right inputs: expertise of users, data, integration, and speed to decision. In plain english consider GenAI to be a new dataset purchased by the firm. Dataset itself won’t chnage anything but heavy analytics on it will!
3. The clever part: they measure GenAI usage via “decision influence”
One of my favorite in this paper is its view point on contributions, which is methodological.
Because direct telemetry on “who uses GenAI” is rarely observable, the authors build a proxy called GenAI Reliance using 13F holdings and ChatGPT generated signals from earnings call transcripts.
In simple plain English, the measure asks:
How much more of a fund’s portfolio changes can be explained when adding AI generated information, beyond standard fundamentals score?
CAIO lens: This is exactly the direction I encourage organizations to move:
Do not measure GenAI adoption by “usage.”
Measure it by decision impact and incremental explanatory power.
In business analytics terms: the question is not “Did people open the tool?” but “Did the tool change outcomes and how effectively you utilized the tool?”
4. Where GenAI helps most: firm specific understanding
The authors decompose AI signals into three groups:
- Macro.
- Firm policy.
- Firm performance.
They find GenAI’s performance link is the strongest for firm level policy and performance signal that is, where the information is unstructured, nuanced, and requires interpretation at scale. The expertise firm posses plays a significant role.
CAIO lens: This aligns with what I see across domains: GenAI tends to shine when the bottleneck is expert human attention on multi-interpretable text, not when the task is already structured and well modeled.
5. Talent is a force multiplier (and this is where inequality can widen
Even though tools like ChatGPT are broadly accessible to everyone and every firm equally, the paper finds not everyone benefits equally.
Larger, older, more expert, more active hedge funds extract more outperformance from GenAI, while smaller or less active firms don’t show the same gains.
They also show the effect is much stronger for funds that had hired AI skilled talent prior to ChatGPT performing predictive data science models (measured via job postings and AI related roles).
CAIO lens: This is the “people expertise + data + tools” triangle in action.
GenAI reduces friction, but it does not eliminate the need for:
- good data engineering
- model risk thinking
- domain expertise
- workflow design
If anything, GenAI may increase huge disparity because top firms can integrate it faster and more safely.
6. Micro evidence: trades aligned with GenAI signals do significantly better
The paper goes beyond fund level averages and looks at trade alignment.
They find that trades agreeing with GenAI recommendations deliver approximately 2.57% average quarterly return, versus approximately 0.35 percent for disagreeing trades (** a approximately 2.22% whooping gap**).
CAIO lens: This fact suggests GenAI isn’t just summarizing, it is influencing real portfolio choices, and “AI consistent” decisions are measurably different.
7. Market implications: efficiency improves, asymmetry spikes early
The authors also examine market structure effects:
- Price efficiency improves: stocks with higher GenAI influenced trade show stronger immediate earnings reactions and weaker post earnings announcement drift.
- Information asymmetry initially increases: bid-ask spreads are higher for stocks more exposed to GenAI driven trading early on, especially around earnings announcements, and then fade away as adoption expands.
CAIO lens: Every major technology wave has second order effect. If GenAI changes information processing speed and interpretation quality, it can reshape microstructure dynamics even if it is only temporarily.
A CAIO playbook: what I would do with these insights???
A) Build workflow integration before scaling!
The survey portion highlights that many hedge funds find workflow integration and in-house expertise very challenging.
That’s consistent with most enterprises: the hard part is not “the model” it is the operating system around it.
Action: pick 1 to 2 workflows where GenAI is naturally strong:
- earnings call summarization and extraction
- “what changed since last quarter” analysis
- policy intent signals
- expert interpretations on particular stock and company from announcement reports
B) Invest in AI literacy and evaluation, not only tools
If talent multiplies value, the strategic move is to develop:
- experts analyst training on prompting and verification
- workflow integration
- evaluation harnesses (accuracy, stability, bias)
- Human-in-loop feedback loops from decisions to model improvements
C) Measure adoption as decision influence
Take inspiration from “GenAI Reliance”:
- track how often GenAI outputs are cited in investment memos
- quantify time to insight improvements
- attribute incremental performance to GenAI assisted signals
- ask experts to interpret and use GenAI as a tool
D) Treat governance as accelerator, not a brake
The paper notes concerns like outages, reliability, and integration challenges.
In regulated environments, governance must be designed to enable safe security speed:
- approved use case catalog
- data boundaries
- audit trails
- human in the loop checkpoints
- pay attention to security
Closing thought
GenAI is not “free alpha. even if alpha is real.” But the evidence suggests it can be a real edge when combined with the right expert users, data, and workflow design, governance and that the advantage may concentrate among firms that operationalize it at its best.
My question for leaders (and builders):
Are you treating GenAI as a novelty… ??? or as a measurable capability embedded into how decisions get made????
If you are working on GenAI in finance or analytics, I’d love to hear:
Which part has been hardest for you cost, trust, evaluation, integration, governance, or expertise?