Wellington Says AI Agents Kill Process Alpha, Shift Edge to Human EQ
Ai4
Autonomous AI agents are turning modeling speed and data retrieval into mere table stakes, forcing asset managers to derive alpha from qualitative judgment and executive assessment rather than technical execution.
## Background
Published on YouTube by Ai4 on August 20, 2026, this session features a conversation between John Divine of U.S. News & World Report and Garence Staraci of Wellington Management. The discussion examines how investment software is evolving from simple chatbots to autonomous agents, analyzing how these workflows impact stock selection, alpha commoditization, hiring shifts, and the integration of machine reasoning into institutional investing.
## TL;DR
Garence Staraci outlines the progression of AI in asset management across three eras, culminating in today's agentic workflows that directly interface with financial models and research pipelines. As automated agents commoditize traditional 'process alpha' like rapid filing retrieval and standard modeling, investment edge is shifting toward long-term qualitative judgment, management assessment, and emotional intelligence. Consequently, hiring practices are deprioritizing pure coding skills in favor of high EQ, financial intuition, and hybrid AI tech leadership. Ultimately, agents function not as autonomous decision-makers—due to vital fiduciary and accountability constraints—but as cognitive multipliers that stress-test human hypotheses and filter high-velocity market noise.
## Atmosphere & style
The dialogue maintains an intellectually rigorous, pragmatic, and measured tone, eschewing speculative tech hype in favor of institutional reality. John Divine facilitates a structured, inquiring narrative, while Staraci speaks with reflective, experienced authority, drawing from his background across both quantitative research and fundamental investing. The dynamic feels like an insightful executive fireside chat, balancing clear enthusiasm for agentic productivity gains with a grounded, candid appraisal of fiduciary duty and the enduring necessity of human intuition.
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=4s) Wellington Management combines fundamental investors, quants, data scientists, and engineers within its Science Equity Research team to ground AI development in alpha generation.
- [03:03](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=183s) The evolution of AI in investing spans three distinct phases: the Quant/NLP era, the conversational LLM era ('a great intern, but a terrible colleague'), and the current Agentic era of non-technical v-coding.
- [06:43](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=403s) Fiduciary duty, regulatory requirements, and human accountability prevent autonomous AI from making unilateral stock-selection decisions without human portfolio manager oversight.
- [09:11](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=551s) Process alpha—the historical edge gained from parsing filings and updating financial models faster than peers—is being commoditized as agentic AI standardizes information processing speed.
- [09:11](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=551s) Investment outperformance is shifting toward irreplaceable human qualities, such as qualitative management evaluation, narrative inflection detection, and proprietary internal research.
- [12:20](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=740s) Data science and analyst hiring criteria are evolving away from pure hard-coding ability toward high emotional intelligence (EQ), intellectual curiosity, and deep investing acumen.
- [12:20](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=740s) A new hybrid 'AI/Tech Lead' role is emerging within investment teams to manage localized agent infrastructure, curate team prompt libraries, and tailor customized workflows.
- [16:25](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=985s) Agentic systems rapidly compress high-velocity market noise into actionable signals, helping managers efficiently filter expansive opportunity sets into high-conviction targets.
- [16:25](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=985s) To prevent alpha erosion from off-the-shelf third-party tools, asset managers must codify and stress-test their proprietary investment philosophies into distinct, custom AI architectures.
- [18:31](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=1111s) Machine reasoning serves as a cognitive multiplier to stress-test human assumptions, uncover macro blind spots, and brainstorm alternative scenarios during unprecedented market regimes.
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=4s) Introduction and the Evolution of AI in Fundamental Investing
John Devine, assistant managing editor of investing at *U.S. News & World Report*, opens the session by introducing himself and welcoming Garence Staraci, Managing Director at Wellington Management. Devine outlines the panel's focus: exploring how artificial intelligence has evolved within fundamental investing across distinct technological eras, how human capital and hiring strategies are shifting, whether AI is commoditizing investment alpha, and what the future holds for fundamental stock selection workflows.
Garence Staraci introduces his role at Wellington Management, where he leads the Science Equity Research team within the firm's Global Investment Research group—the core engine of Wellington's investment platform. He describes the team as a unique hybrid composed of traditional fundamental investors, quants, data scientists, and technology engineers. This interdisciplinary approach supports a highly data-intensive investment process that leverages traditional fundamental data, alternative dataset streams, and cutting-edge agentic AI solutions. Reflecting on his own career background as a former quant equity researcher who drifted toward fundamental investing, Staraci notes how fascinating it has been to stand at the intersection of fundamental finance, data science, and technology. He observes that the industry's sentiment toward AI over recent years has swung dynamically through a full psychological spectrum—from deep skepticism, to unbridled excitement, to sudden doubts, and ultimately toward a pragmatic focus on alpha generation.
---
[03:03](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=183s) The Three Eras of AI: From Pre-ChatGPT to Agentic Workflows
When Devine asks which era the industry currently inhabits and how AI agents are actively deployed in daily workflows, Staraci steps back to frame the evolution into three distinct historical periods. He emphasizes that the shift isn't merely about what AI technology can do, but **who** is empowered to use it. A continuous feedback loop exists between the user persona, the tools constructed, and the ultimate stock-selection process.
```
PRE-CHATGPT ERA CHATGPT / LLM ERA AGENTIC ERA
┌──────────────────────┐ ┌─────────────────────┐ ┌─────────────────────┐
│ • Quants & Data Sci │ ───► │ • Fundamental PMs │ ───► │ • Non-Coders / All │
│ • Complex NLP models │ │ • Summaries & Memos │ │ • V-Coding / Agent │
│ • Invisible to PMs │ │ • "Great intern" │ │ • Core Models/Data │
└──────────────────────┘ └─────────────────────┘ └─────────────────────┘
```
- **Pre-ChatGPT Era:** Heavy machine learning and natural language processing (NLP) were strictly the domain of quants and data scientists. They built sentiment models and early Q&A tools for public filings. These tools remained largely "invisible" behind the scenes to traditional fundamental analysts, who simply consumed the output without understanding how the "soup was made."
- **ChatGPT Era:** Generative AI democratized access across the broader fundamental landscape. Suddenly, every analyst felt like they had a personal research assistant for drafting memos and summarizing filings. However, severe bottlenecks remained—notably model hallucinations, lack of persistent memory, and lack of integration with proprietary internal knowledge. Staraci jokes that during this phase, ChatGPT felt like "a great intern, but a terrible colleague."
- **Agentic Era (Today):** The current wave allows non-technical investors to execute "v-coding" (visual/voice coding using AI) inside environments like VS Code with GitHub Copilot to build custom dashboards and screeners rapidly. Autonomous agents now interface directly with core investment tasks—pulling filings, updating financial models in Excel, and creating real-time feedback loops.
---
[06:43](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=403s) The Future Vision of Investment Workflows and Human Accountability
Devine inquires about the ultimate end-state vision Wellington is building toward as agents become deeply embedded in fundamental analysis. Staraci clarifies that while research and stock-picking draw the most attention, agentic AI is expanding across the entire institutional investment lifecycle—including trade execution, portfolio construction, risk management, and client RFP responses. In trading specifically, where structured data is abundant, AI integration is often further along than in fundamental equity research.
Within research workflows, agentic AI operates as a full-blown co-pilot that absorbs recurring grunt work and time-consuming manual tasks. This dramatically compresses the timeline between a human analyst forming a hypothesis and retrieving empirical evidence. The analyst's role fundamentally shifts from manual data retrieval to higher-level human evaluation and judgment. However, Staraci strongly emphasizes that AI will not be making autonomous stock-selection decisions on its own anytime soon:
> "It's not a technical problem... It's just the human sort of accountability, the judgment aspect into it, the fiduciary aspect to it. How are you going to fire the AI if it makes a bad pick?"
- **Operational Impact:** Agentic tools automate grunt work, drastically reducing time-to-insight for human researchers.
- **Scope of AI:** Extends beyond research into trading, risk management, portfolio construction, and client relations.
- **Fiduciary Responsibility:** Human oversight remains indispensable due to accountability, regulatory duties, and client reporting requirements.
---
[09:11](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=551s) Is AI Commoditizing Alpha? Process Alpha vs. Human Intuition
Addressing Devine's question regarding whether near-identical AI models risk commoditizing investment strategies and eroding market alpha, Staraci reflects on historical technological shifts—such as Microsoft Excel, Bloomberg terminals, and corporate email. Each major innovation compressed specific competitive advantages, shrinking traditional alpha pools while causing new ones to form elsewhere.
Staraci distinguishes between **process alpha** (the historical advantage gained by quickly processing filings, updating financial models, and coming to a conclusion faster than peers) and fundamental human insight. As agentic systems automate information processing across the entire market, pure speed and information processing efficiency will be heavily compressed. Consequently, edge shifts toward long-term horizons and qualitative human elements that algorithms cannot replicate—evaluating corporate management teams, detecting narrative inflections, tapping proprietary internal research, and leveraging emotional intelligence to "gauge the room."
- **Process Alpha Decay:** Rapid information retrieval and financial modeling speed are rapidly becoming table stakes rather than sources of outperformance.
- **New Edge Drivers:** Differentiated viewpoints, understanding qualitative nuances, assessing management quality, and recognizing macro narrative shifts.
- **Proprietary Data:** Internal corporate research and direct access to management gain renewed strategic importance.
---
[12:20](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=740s) The Evolution of Hiring: From Pure IQ to EQ and New AI Roles
Turning to human capital and organizational design, Devine asks how hiring criteria and desired skill sets have shifted over time. Staraci notes that the traditional profile of both technical data scientists and fundamental analysts is undergoing a profound transformation as soft skills rise in prominence over pure technical execution.
```
HISTORICAL HIRING MODERN HIRING
┌───────────────────────┐ ┌───────────────────────┐
│ • Heavy Hard Coding │ ───► │ • High EQ & Curiosity │
│ • Elite Quant / CS │ │ • Investing Acumen │
│ • High Pure IQ Focus │ │ • AI Stack Leadership │
└───────────────────────┘ └───────────────────────┘
```
For data science hires, the industry historically prioritized top-tier coding ability and computer science expertise to build custom tech stacks fast. Today, with generative AI and agents streamlining coding, firms favor high emotional intelligence (EQ), intellectual curiosity, passion for investing, and clear communication skills over raw programming chops.
Staraci also points out an emerging hybrid role inside investment teams: the **AI/Tech Lead**. These professionals possess both strong financial acumen and deep technical understanding, allowing them to construct team-specific prompt libraries, customize workflow agents, and manage the team's localized AI infrastructure. Meanwhile, the core responsibilities of traditional junior analysts—historically centered around financial modeling and monitoring news wires—are being fundamentally re-evaluated as autonomous systems prove capable of executing those baseline tasks smoothly.
- **Data Science Shift:** Prioritizing communication, curiosity, and investment passion over hard coding capabilities.
- **Emergence of AI Leads:** Dedicated team members who curate prompt engineering libraries, direct AI agents, and manage local tech stacks.
- **Analyst Role Transformation:** Traditional analyst responsibilities in modeling and data entry face structural shifts as technology automates entry-level tasks.
[16:25](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=985s) Filtering Signal from Noise and Codifying Differentiated Alpha
A primary motivation behind implementing agentic AI solutions in investment management is automating complex workflows across massive, volatile market universes. With an overwhelming volume of market headlines and data breaking every day, the fundamental challenge centers on distinguishing genuine signal from market noise. AI drastically compresses the timeframe required to ingest, process, and evaluate breaking news, giving investment teams the agility to decide whether to react immediately or ignore the distraction. By filtering out irrelevant noise, these systems help managers efficiently narrow down expansive opportunity sets into high-conviction targets.
Beyond initial screening, Wellington leverages AI to customize, codify, and rigorously stress-test its proprietary investment workflows at scale. Staraci emphasizes that relying solely on off-the-shelf, third-party AI workflow platforms presents significant competitive and operational risks, as widespread adoption inevitably commoditizes standard investment strategies. To preserve their alpha generation edge, fundamental managers must build unique, differentiated supporting architectures that embody their own distinct decision-making frameworks.
- **Signal Extraction:** Rapidly parsing high-velocity news flow to separate actionable fundamentals from transient market volatility.
- **Workflow Codification:** Translating unique in-house investment philosophies into scalable, automated AI-supported pipelines.
- **Edge Protection:** Avoiding commoditized third-party tools so outperformance is not diluted when market participants access identical models.
---
[18:31](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=1111s) Reconciling Human Intuition with Machine Reasoning
Addressing how to "square that circle" when human portfolio managers disagree with machine forecasts—such as conflicting outlooks on interest rate trajectories or asset allocation—Staraci explains that AI currently functions as a collaborative research assistant rather than an autonomous authority. Human investors contribute irreplaceable historical intuition, macro cycle awareness, and nuanced pattern recognition honed over decades of market shifts. However, in unprecedented market regimes where "history doesn't repeat," relying entirely on past playbooks can leave blind spots.
AI's emerging reasoning capabilities serve as a powerful cognitive multiplier, allowing managers to rapidly stress-test underlying assumptions, brainstorm alternative scenarios, and uncover unconventional perspectives they might have otherwise missed. While teams remain vigilant regarding the inherent training biases of machine learning models, combining human judgment with machine-driven reasoning fosters deeper, more rigorous investment analysis before the session concludes with closing thanks.
- **Cognitive Multiplier:** Using machine reasoning models as brainstorming partners to challenge preconceived human biases.
- **Stress-Testing Assumptions:** Rapidly simulating unconventional scenarios when macroeconomic dynamics deviate from historical precedents.
- **Symbiotic Decision-Making:** Pairing human cycle expertise with machine-driven breadth to achieve a resilient, stress-tested market view.
## Notable quotes
> [03:03](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=183s) great intern, but a terrible colleague — Garence Staraci
> [03:03](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=183s) soup was made — Garence Staraci
> [06:43](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=403s) It's not a technical problem... It's just the human sort of accountability, the judgment aspect into it, the fiduciary aspect to it. How are you going to fire the AI if it makes a bad pick? — Garence Staraci
> [06:43](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=403s) How are you going to fire the AI if it makes a bad pick? — Garence Staraci
> [09:11](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=551s) gauge the room — Garence Staraci
> [18:31](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=1111s) square that circle — John Devine
> [18:31](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=1111s) history doesn't repeat — Garence Staraci
## Mentioned
Garence Staraci, John Devine, Wellington Management, U.S. News & World Report, Ai4, ChatGPT, GitHub Copilot, VS Code, Microsoft Excel, Bloomberg terminal
## Video description
Garence Staraci - Managing Director, Wellington Management
John Divine - Assistant Managing Editor, Investing, U.S. News & World Report
As investment software evolves from chatbots to autonomous agents, how do fundamental managers protect their competitive edge? We discuss how new workflows impact stock selection and whether widespread tech commoditizes outperformance. Learn how hiring needs are shifting to favor irreplaceable human skills, and how to navigate trust within the black box. Join us to uncover near-term predictions for investment workflows.
According to Garence Staraci of Wellington Management, the rise of agentic artificial intelligence is shifting fundamental investing from speed-driven information retrieval toward qualitative human judgment and proprietary workflow design.
The Evolution from Quant NLP to Agentic Workflows (3:03)
Staraci outlines three distinct phases in the financial industry's adoption of artificial intelligence. In the pre-ChatGPT period, complex natural language processing models were confined to quants and data scientists, operating invisibly to traditional analysts. The subsequent generative AI wave broadened access by providing personal research assistants capable of drafting summaries and memos; however, these systems suffered from hallucinations, lacked persistent memory, and operated outside proprietary internal knowledge pools.
Automation of Process Alpha and Fiduciary Constraints (6:43)
In the current agentic era, non-technical investors leverage AI coding tools to construct custom screeners, pull filings, and update financial models directly in Excel. Staraci notes that this automation diminishes traditional 'process alpha'—the historical edge gained simply by retrieving data and updating spreadsheets faster than peers. Because these routine processing tasks are increasingly commoditized, fundamental advantage is shifting toward qualitative factors, including assessing management teams, identifying narrative inflections, and interpreting proprietary internal research.
While agentic tools are expanding across trading, portfolio construction, risk management, and client RFP generation, Staraci stresses that AI will not autonomously execute final investment picks. In trading, where structured data is plentiful, adoption has moved faster than in equity research. However, regulatory accountability, fiduciary obligations, and client reporting require human portfolio managers to retain ultimate ownership and responsibility for capital allocation.
Organizational Redesign and Changing Talent Profiles (12:20)
The widespread availability of AI coding and automation is restructuring investment personnel requirements. Historically, firms prioritized elite computer science backgrounds and raw programming power for data science positions. With AI handling baseline programming tasks, Staraci explains that hiring has shifted toward candidates exhibiting high emotional intelligence, intellectual curiosity, clear communication, and financial acumen.
Investment organizations are also introducing dedicated hybrid roles, such as AI or Tech Leads, who combine market knowledge with technical expertise to manage local prompt libraries and specialized agents. Concurrently, the career development of junior analysts faces structural adjustments, as the data collection and financial modeling routines that previously formed their core training are automated.
Synthesizing Machine Reasoning with Human Intuition (16:25)
To prevent investment strategies from being diluted by widely accessible third-party models, Staraci emphasizes that managers must embed their own distinct investment frameworks into proprietary architectures. These automated systems help analysts filter noise from high-velocity market news flows and isolate high-conviction ideas.
When machine outputs conflict with human views, Staraci characterizes AI as a collaborative sparring partner rather than an authoritative decision-maker. While experienced investors provide cycle awareness and qualitative pattern recognition, machine reasoning models help stress-test baseline assumptions, identify overlooked data, and simulate alternative macroeconomic scenarios where historical playbooks no longer apply.
As technical barriers fall and baseline analytical tasks are commoditized across the industry, investment firms must reconcile algorithmic reasoning with human accountability while defining how junior talent will develop practical judgment without traditional modeling apprenticeships.
## Background
该视频由频道“Ai4”于 2026 年 8 月 20 日发布在 YouTube 平台。视频记录了《美国新闻与世界报道》助理主编约翰·迪瓦恩与惠灵顿管理公司董事总经理加伦斯·斯塔拉奇的深度对谈，围绕投资工具从自动补全向自主 Agent 演化、基本面经理如何构筑护城河、超额收益是否面临商品化平庸化，以及投研招聘需求向不可替代的人类技能转移等议题展开。
## TL;DR
对话回顾了 AI 在基本面投资中经历的 NLP 黑盒、ChatGPT 助理及复合 Agent 编程三大纪元，指出由信息搜集与手速驱动的“流程 Alpha”正在被技术全面压缩。未来的超额收益将转移至机构专有数据、管理层深度定性沟通以及洞察非共识拐点等人类独特禀赋上。与此同时，硬核编程门槛的降低推动资管招聘转向注重商业好奇心与高情商，并催生了常驻团队的“AI 工程师”编制。最终，受限于人类信托责任与法律问责机制，AI 在可预见的未来仍将扮演高效降噪与投资假设“压力测试员”的副驾驶角色，无法取代人类的终极投资裁决。
## Atmosphere & style
整场对话洋溢着买方一线从业者特有的务实、严谨与敏锐思辨，完全摒弃了空洞的技术炒作。主持人迪瓦恩提问直击行业隐忧且妙语连珠，斯塔拉奇则凭借横跨量化与基本面的复合背景沉着拆解，生动的行业比喻与深度的信托审慎感交织。观感如同一场高密度的顶尖闭门研讨，既有对自动化浪潮的理性拥抱，又处处透露出对人类直觉与制度底线的清醒坚守。
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=4s) 投资界对 AI 的认知经历了从怀疑、兴奋、质疑到务实探索的完整周期，目前全行业聚焦于探寻 AI 能否切实赋能超额收益（Alpha）的创造。
- [03:03](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=183s) AI 在基本面投资经历了三大演进纪元：从量化专属的“NLP 早期黑盒时代”，到通用文本助手的“ChatGPT 时代”，再到人人皆可辅助编程并深度结合业务的“Agent 时代”。
- [04:00](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=240s) 大语言模型早期常被视为“优秀的实习生与糟糕的同事”，由于模型幻觉、缺乏长期记忆且无法调取机构私有知识，难以直接推动垂直投研跃升。
- [06:42](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=402s) Agent 在投研中最核心的价值在于大幅压缩“提出投资假设到获取事实证据”的时间跨度，将分析师从重复性杂务中解放，专注于人类主观判断。
- [08:00](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=480s) 全自动 AI 自主选股存在制度性底线：AI 无法承担法律与信托责任（亏损时无法解雇 AI），资产管理的核心决策必须由人类最终负责。
- [09:15](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=555s) 底层基础模型的趋同重演了 Excel 与彭博终端普及的历史，技术的标准化不会消灭竞争优势，而是会重塑工作流并转移竞争维度。
- [10:03](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=603s) 依赖信息搜集速度与手动建模的“流程 Alpha”正在被 Agent 无情压缩，下一代 Alpha 护城河重新回归到机构私有数据、管理层定性评估情商与非共识拐点直觉。
- [12:20](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=740s) 资管团队招聘标准发生重大转变：纯硬核写代码的价值被 AI 稀释，商业好奇心、投资热情及跨领域沟通能力等软实力成为核心评估指标。
- [14:00](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=840s) 投资团队内部正诞生“AI 工程师”这一新型跨界编制，专职将团队特有的金融逻辑与前沿 AI 技术栈、提示词工程进行深度代码化绑定。
- [16:05](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=965s) 采购通用第三方 AI 系统可能导致投资策略趋同并引发系统性风险，机构必须依托自研高度定制的 AI 流程来固化并测试自有的投研框架。
- [18:31](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=1111s) AI 在高阶投研中应作为假设的“压力测试员”，利用推理能力迫使投资人跳出过往经验与认知盲区，检验投资逻辑的韧性。
- [19:47](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=1187s) 基本面投资的最佳范式是人机协同：在警惕模型固有训练偏见的同时，将 AI 驱动的深度思维倍增效应与人类资深市场经验深度融合。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=4s) 惠灵顿管理公司的跨界团队与 AI 演进背景
来自《美国新闻与世界报道》（U.S. News & World Report）的投资版面助理主编约翰·迪瓦恩（John Divine）拉开了对话序幕。本次讨论的核心主题是人工智能在基本面投资中的演化历程、人机协同（Human-in-the-loop）的定位、投研人才招募的变迁，以及 AI 工具普及是否会引发超额收益（Alpha）的平庸化与商品化。随后，惠灵顿管理公司（Wellington Management）董事总经理加伦斯·斯塔拉奇（Garence Staraci）介绍了自己的背景与团队定位。
加伦斯在惠灵顿领导着“科学股票研究团队”（Science Equity Research Team），该团队隶属于构成公司核心投研平台的全球投资研究部门（Global Investment Research）。这是一个高度跨界且背景多元的独特团队，既有传统基本面研究背景的投资人，也有专业的数据科学家与软件工程师。之所以组建这种跨界架构，是因为他们的投资流程高度依赖数据驱动与当下的 AI 技术——不仅涵盖传统财务基本面、价格量价数据，还广泛纳入了跨行业的另类数据（Alternative Data），并在当前流程中深度集成自主构建的 AI Agent 解决方案。加伦斯本人早年是一名量化股票研究员（Quant），后来逐渐转向基本面投资。这种横跨传统基本面、数据科学与工程技术的独特视角，使他得以近距离见证过去几年业内对 AI 态度的剧烈演变：从最初的怀疑、随后的兴奋，到中途的动摇，再到如今务实地探寻“它能否真正帮我创造 Alpha”。
- **团队属性**：跨行业、跨职能，汇集传统基本面分析师、数据科学家与软件工程师。
- **数据与技术范畴**：涵盖传统财务与行情数据、多行业另类数据，以及正在研发部署的 Agent 化投研方案。
- **核心心路历程**：行业态度经历了从“怀疑”到“兴奋”，再到“质疑”，最终落脚于“务实寻求 Alpha 赋能”的转变。
---
### [03:03](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=183s) AI 投资的三大纪元：从 NLP 黑盒到人人编程的 Agent 时代
针对约翰关于当前是否处于“半 Agent 时代”的提问，加伦斯梳理了 AI 在基本面投资领域经历的三大发展纪元。他特别强调，评判 AI 发展的维度不仅在于“AI 能做什么”，更在于“谁在使用 AI”。使用者群体的扩散直接形成了反馈回路，反向塑造着工具形态乃至选股范式：
```mermaid
flowchart TD
accTitle: AI in Fundamental Investing Three Eras
subgraph Era1["前 ChatGPT 时代 (Pre-ChatGPT)"]
A1["主要使用者: 量化人员 / 数据科学家"] --> B1["核心技术: 传统 ML / 早期 NLP 情感分析"]
B1 --> C1["基本面投资人角色: 纯下游黑盒消费者"]
end
subgraph Era2["ChatGPT 时代 (LLM 普及)"]
A2["主要使用者: 全体基本面分析师"] --> B2["核心应用: 公告摘要 / 研报草拟"]
B2 --> C2["瓶颈挑战: 幻觉 / 无长期记忆 / 缺失内部知识"]
end
subgraph Era3["Agent 与协同编程时代 (当下)"]
A3["主要使用者: 无代码背景的普通投资人"] --> B3["核心工具: GitHub Copilot / VS Code / 复合 Agent"]
B3 --> C3["深度整合: 财务建模 / 财报抓取 / 自动化数据验证"]
end
Era1 --> Era2 --> Era3
```
1. **前 ChatGPT 时代**：这一阶段被机器学习和自然语言处理（NLP）技术主导。应用场景集中在量化情绪分析或针对公开监管文件的早期问答原型。这一时期的门槛被统计学、计算机工程与专业技能牢牢锁死，核心玩家仅限于量化分析师和数据科学家；对传统基本面投资人而言，AI 运作的底层逻辑就像一个隐形的黑盒，他们“根本不知道这碗汤是怎么熬出来的”，只能机械地消费最终输出的数据结果。
2. **ChatGPT 时代**：以大语言模型为代表的生成式 AI 迅速破圈，席卷了广泛的基本面投资圈。投资人第一次感觉自己身边多了一位随时待命的初级研究助理，能够极快地完成公开研报总结、备忘录初稿起草等机械任务。然而这一阶段的硬伤同样明显——模型存在幻觉、缺乏长程记忆、无法调取机构内部的私有投研知识库。加伦斯打趣道，当时的 ChatGPT 在基本面分析师眼里就像是“一个非常优秀的实习生，却是一个糟糕透顶的正式同事”，这阻碍了投研流程发生真正的垂直跃升。
3. **Agent 时代（当下）**：这一阶段的核心跃迁在于“原本不懂代码的人开始深度拥抱技术”。许多毫无编程基础的分析师借助 VS Code 和 GitHub Copilot 等工具开始自行构建监控看板与选股筛选器（Screener），业内甚至流行起“我正在进行 VS Code 式编程（I'm VS coding right now）”的说法。在此基础上，Agent 开始深度介入核心投研环节，全自动抓取财报、参与 Excel 财务建模，并与投资人的交互形成持续的反馈循环。
- **纪元界限**：前 ChatGPT 时代（量化专享的 NLP 孤岛） $\to$ ChatGPT 时代（泛用型文本助理与高幻觉并存） $\to$ 当前 Agent 时代（全民辅助编程与投研工作流深潜）。
- **经典生动比喻**：ChatGPT 时代的模型如同“出色的实习生、糟糕的同事”（"a great intern but a terrible colleague"）。
- **行为习惯转变**：零代码基础的基本面分析师开始借助 Copilot 自行编写脚本搭建仪表盘，彻底改变了工具构建的生态。
---
### [06:42](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=402s) Agent 的终极愿景：压缩证据链与不可替代的人类信托责任
在展望 Agent 技术的演进终局时，加伦斯指出其影响力绝不仅限于个股研究，而是正在渗透到整个投资全生命周期。在量化交易执行（Trading）端，由于高频数据密度极大，Agent 的落地速度甚至领先于基本面研究；同时，它也正在重塑风险管理、投资组合构建乃至客户端的 RFP（提案征询书）响应流程。
在基本面投研的核心场景中，Agent 的终极形态并非取代人类，而是作为一个全功能的资深副驾驶接管所有高耗时、高重复性的繁重杂务（Grunt work）。加伦斯强调，Agent 最核心的价值在于“大幅压缩人类提出投资假设到获取事实证据之间的时间跨度”。它把分析师从机械繁杂的数据搬运中彻底解放出来，将投研精力全面置换并聚焦于“人类的主观判断与评估”。
```mermaid
flowchart LR
accTitle: Agentic Workflow Compression
subgraph Traditional["传统投研工作流"]
direction TB
T1["投资假设 / 提问"] --> T2["耗时繁复的手动搜集 / 建模 (Grunt Work)"]
T2 --> T3["人类判断与最终决策"]
end
subgraph Agentic["Agent 赋能工作流"]
direction TB
A1["投资假设 / 提问"] -->|Agent 秒级自动化抓取验证| A2["证据链瞬间成型"]
A2 --> A3["深度人类信托判断 / 最终决策"]
end
```
针对未来 AI 是否会“全自动完全自主选股”这一终极假想，加伦斯给出了极其坚决的否定判断。他指出，这从根本上不是一个工程或算法技术问题，而是一个关乎**人类信托责任（Fiduciary duty）与问责机制**的制度性底线。约翰随即调侃一针见血地指出：“如果 AI 选错了股票，你根本没法解雇它”。加伦斯深表认同，并将其与税务规划或专业财务咨询做类比：无论算法自动化程度有多高，底层的最终责任人必须是人类。即便量化领域存在全自动模型实验，面向客户的资产管理行业在可见的未来绝不会允许“从 A 到 Z 全无人化”的黑盒投资决策。
- **全流程覆盖**：除基本面选股外，Agent 正在渗透交易执行（进度超前）、风险控制、组合构建及客户服务。
- **效率本质**：压缩“提问 $\to$ 证据”的中间摩擦，把分析师的工时转变为“人类主观判断”。
- **问责悖论**：AI 无法承担法律与信托责任——“亏损时你无法解雇 AI”，因此人类主导地位不可撼动。
---
### [09:15](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=555s) 模型同质化与 Alpha 侵蚀：金融技术创新的历史镜像
约翰向加伦斯抛出了一个投资界高度关注的深层矛盾：随着底层基础大模型能力日益趋同，每家机构拥有的模型工具相差无几，如果所有人都用同样的 AI 工具在市场上挖掘超额收益，是否会导致大家的竞争优势相互抵消，最终造成 Alpha 的“平庸化与商品化（Commoditization）”甚至彻底消失？
加伦斯认为这是一个直击资产管理核心的深刻哲学问题，也是未来几年全行业必将共同面对的核心争鸣。尽管最终定论需要时间来检验，但他提出可以从金融科技的发展史中寻找参照标尺。回望过去几十年，金融界先后迎来了多项革命性的工具创新——从电子表格软件 Excel 的普及、彭博终端（Bloomberg Terminal）的标配化，再到电子邮件对沟通机制的重塑。每当一种大幅压缩时间成本、普及信息处理能力的新工具降临，市场都会产生相似的同质化担忧，但投研演进的本质规律始终在工具的更迭中被不断重写。
- **核心行业忧虑**：基础模型趋同可能引发 Alpha 竞争的同质化消耗。
- **历史参照系**：Excel、彭博终端与电子邮件的普及史，每一次工具普及都曾重塑工作流边界，并重新定义人类竞争优势的门槛。
### [10:03](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=603s) Alpha池的转移与“流程Alpha”的压缩
Garence Staraci 解释道，如同历史上的任何技术革新一样，部分现存的超额收益（Alpha）池正在萎缩甚至消失，但Alpha绝不会彻底归零，它只是转移到了其他地方。在传统的基本面投资中，以往的超额收益往往归属于那些能够最快速吞吐海量信息、提炼核心要点、推导投资含义，并迅速更新到财务模型中得出买卖结论的人。谁拥有更优质的专有信息获取渠道、更强大的技术栈，谁就能先人一步，这种曾被视为极其合法且坚固的护城河，被 Staraci 称为“流程Alpha”（Process Alpha）。
然而，在一个全自动化、由 Agent（智能体）驱动且处理流程干净利落的新世界里，“流程Alpha”正在被无情地压缩。针对市场上有人认为“信息处理周期的极度压缩会将Alpha推向更长投资周期”的观点，Staraci 指出，如果所有资金和玩家都转向长周期策略，那么长周期维度的Alpha同样会被迅速压缩。最终，Alpha的核心重新回归到那些人类独有的、难以被量化的能力上：机构自有的内部专有数据、与管理层对话的深度、识别市场叙事盲点与预期差的敏锐直觉，以及评估管理层质量与业务拐点的情商（EQ）和“读懂房间气氛（gauge the room）”的能力。
```mermaid
flowchart TD
accTitle: Alpha驱动力的演变路径
subgraph Legacy["传统 Alpha 来源（正被压缩）"]
A["信息检索与快速阅读"] --> B["手动更新财务模型与 Excel"]
B --> C["高响应速度的“流程 Alpha”"]
end
subgraph Future["下一代 Alpha 核心（人类不可替代）"]
D["机构私有数据与洞察"] --> G["超额收益新高地"]
E["管理层定性评估与情商 (EQ)"] --> G
F["识别错误市场叙事与拐点直觉"] --> G
end
Legacy -.->|Agent 全面自动化替代| Future
```
- **流程Alpha（Process Alpha）**：过去依赖个人手速、信息获取不对称与建模效率所带来的超额回报，正在被技术完全商品化与边缘化。
- **下一代Alpha支柱**：内部独有数据沉淀、深层次的管理层触达与判断、对市场非共识（Differentiated Viewpoints）的洞见，以及对拐点（Inflection）的直觉把控。
---
### [12:20](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=740s) 招聘标准的颠覆：从极客代码能力到投资好奇心与“AI工程师”
John Divine 顺势将话题引向人才招聘标准的变化：在“人机协同（Human-in-the-loop）”成为常态的背景下，资管机构如今寻找的是怎样的人才？过去与现在的核心技能价值发生了怎样的颠覆？Staraci 将其拆解为三类角色：数据科学家、新型跨界角色以及传统基本面分析师。
在数据科学领域，选拔标准发生了从“高IQ向高EQ”的剧烈转变。三到五年前，机构不惜重金招募技术最强、代码写得最快最深的人才，因为当时构建自研基础设施和定制技术栈能直接产生超额回报。而今，随着 Agent 和现成 AI 工具能够极高效地分担硬核编码，招聘考核的第一顺位变成了对投资的热情、对商业世界的底层好奇心，以及将复杂问题用清晰语言沟通表达的能力。针对 Divine 提出的“这是否意味着招聘变得更加主观”，Staraci 完全赞同，认为主观的人性评估在选拔中反而变得更加权重突出。
与此相呼应，投资团队内部正在孕育一种全新的跨界角色——投资团队内部的“AI工程师”。这类人才常驻于核心投资团队内，全权负责团队专属的 AI 技术栈、提示词工程（Prompting）以及与团队投资流程深度契合的代码库构建。这类岗位要求从业者兼具“深厚投资嗅觉”与“前沿技术理解力”，能精准判断什么场景该调用何种技术。对于传统分析师而言，当处理 Excel 和阅读整理新闻的机械工作被完全自动化替代后，其角色定位甚至基金经理（PM）的工作边界，都将面临根本性的重构。
- **硬技能贬值与软实力崛起**：单纯快速写代码的价值被 Agent 稀释，商业直觉、好奇心、求知欲与沟通力成为数据科学类岗位的首要评估指标。
- **核心团队新型编制**：投资团队正配置专属的 AI 工程师，他们深谙金融逻辑又懂技术架构，主导量身定制的 AI 工作流。
---
### [16:05](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=965s) Agent在工作流中的双重定位：信号降噪与自研流程压力测试
针对 Divine 关于“当前工作流程中究竟是 Agent 替人类满世界搜寻想法再由人类把关，还是人类提出假设由机器验证”的提问，Staraci 阐明了 Wellington 目前对于 Agentic 工作流的两大核心应用支柱：
第一重定位是面对庞大覆盖宇宙与剧烈市场波动时的“降噪与过滤”。面对指数级膨胀的新闻与数据，市场充斥着刺耳的噪音。AI 的首要价值是帮投研人员压缩处理信息的时间周期，在海量投资标的中剥离杂音、提炼有效信号（Signal vs. Noise），帮助人类迅速权衡“这条消息是否值得立刻做出反应”。
第二重定位则是将机构专有的投资体系深度“代码化（Codified）”并进行严格的压力测试。Staraci 强调，市场上现在涌现出大量面向基本面投资的第三方通用 AI 工具，如果所有同行都采购相同的第三方系统，投资逻辑便会迅速同质化，进而带来潜在的系统性风险。为了构筑持久的 Alpha，必须构建独特的、高度定制化的自研 AI 支撑系统，将自己的投研框架固化到流程中，既用来发现特色投资机会，也对自身的逻辑假设进行高强度检验。
---
### [18:31](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=1111s) 人类认知与机器推演的交汇：将AI作为假设的压力测试员
在讨论接近尾声时，Divine 抛出了一个锋利的现实难题：如今市场数据是否已经充分到可能让 AI 产生与人类团队完全相悖的宏观判断（例如关于利率走势或资产配置方向）？当人类经验与机器结论发生冲突时，该信任哪一方？
Staraci 明确指出，在目前的演进阶段，机构是将 AI 视作深度的“研究助理”而非最终决策者。人类投资者的不可替代之处，在于拥有横跨多个经济周期的历史经验、模式识别能力以及对过往市场在极端环境下演变规律的深刻记忆。然而，市场环境瞬息万变，“历史往往不会简单重复”。此时，AI（尤其是具备强大推理能力的模型）的核心价值在于充当反向检验工具：以极高的速度对人类主观形成的假设做压力测试，迫使投资人跳出认知盲区，在人机推理的碰撞中检验投资逻辑的韧性。
### [19:47](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=1187s) 作为思维倍增器的 AI：打破认知盲区与人机协同的投资研究
在基本面投资的研究流程中，AI 工具最突出的价值在于充当深度“头脑风暴”的共创伙伴，帮助研究员打破既有思维定势，开拓此前未曾触及的全新视角。尽管 AI 能够显著启发灵感，但团队也必须清醒意识到其底层训练机制所带来的固有偏差（biases）——模型输出的内容往往受制于其训练数据与算法倾向。因此，在享受技术红利的同时，保持对模型局限性的警惕至关重要。
AI 带来的最关键“乘数效应”（multiplier），在于它能够促使投资人进行更深层次、更广维度的推演，从而敏锐捕捉到原本极易被忽略的盲区与细微信号。投资研究的最佳范式绝非完全依赖黑箱，而是将 AI 驱动的深度研究辅助与人类投资经理深厚的“市场经验”（market experiences）紧密结合，实现人机协同的超额认知价值。
在分享的最后，主持人 John Divine 对 Garence Staraci 的精彩洞见以及现场观众的参与表达了诚挚感谢，本场圆桌讨论就此圆满收尾。
```mermaid
flowchart LR
accTitle: 人机协同投资研究工作流
subgraph AI["AI 赋能层"]
A["头脑风暴与灵感拓展"] --> B["挖掘盲区与深度线索"]
end
subgraph Human["人类投研层"]
C["警惕模型训练偏差"] --> D["注入资深市场经验"]
end
B --> E["人机融合决策"]
D --> E
```
- **认知拓展**：AI 帮助投研团队打破思维惯性，开辟过去无法覆盖的研究盲区与新思路。
- **警惕偏差**：清醒认识到模型因训练数据和架构设计而天然存在的输出倾向与偏见。
- **协同乘数**：AI 的深度挖掘能力与人类资深市场直觉的结合，构成了现代基本面投资的核心竞争力。
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=4s) 它能否真正帮我创造 Alpha — Garence Staraci
> [03:03](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=183s) 根本不知道这碗汤是怎么熬出来的 — Garence Staraci
> [03:03](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=183s) 一个非常优秀的实习生，却是一个糟糕透顶的正式同事 — Garence Staraci
> [03:03](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=183s) 我正在进行 VS Code 式编程（I'm VS coding right now） — Garence Staraci
> [06:42](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=402s) 大幅压缩人类提出投资假设到获取事实证据之间的时间跨度 — Garence Staraci
> [06:42](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=402s) 如果 AI 选错了股票，你根本没法解雇它 — John Divine
> [10:03](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=603s) 读懂房间气氛（gauge the room） — Garence Staraci
> [12:20](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=740s) 高IQ向高EQ — Garence Staraci
> [18:31](https://www.youtube.com/watch?v=eaz-x3g0hyU&t=1111s) 历史往往不会简单重复 — Garence Staraci
## Mentioned
Garence Staraci, John Divine, Wellington Management, U.S. News & World Report, Science Equity Research Team, Global Investment Research, ChatGPT, GitHub Copilot, VS Code, Excel, Bloomberg Terminal
## Video description
Garence Staraci - Managing Director, Wellington Management
John Divine - Assistant Managing Editor, Investing, U.S. News & World Report
As investment software evolves from chatbots to autonomous agents, how do fundamental managers protect their competitive edge? We discuss how new workflows impact stock selection and whether widespread tech commoditizes outperformance. Learn how hiring needs are shifting to favor irreplaceable human skills, and how to navigate trust within the black box. Join us to uncover near-term predictions for investment workflows.
[Source](https://www.youtube.com/watch?v=eaz-x3g0hyU)
Source quotes
[0:04] Garence Staraci: 它能否真正帮我创造 Alpha
[3:03] Garence Staraci: 根本不知道这碗汤是怎么熬出来的
[3:03] Garence Staraci: 一个非常优秀的实习生，却是一个糟糕透顶的正式同事
[3:03] Garence Staraci: 我正在进行 VS Code 式编程（I'm VS coding right now）
[6:42] Garence Staraci: 大幅压缩人类提出投资假设到获取事实证据之间的时间跨度
[6:42] John Divine: 如果 AI 选错了股票，你根本没法解雇它
[10:03] Garence Staraci: 读懂房间气氛（gauge the room）
[12:20] Garence Staraci: 高IQ向高EQ
[18:31] Garence Staraci: 历史往往不会简单重复
Source quotes
[3:03] Garence Staraci: great intern, but a terrible colleague
[3:03] Garence Staraci: soup was made
[3:03] Garence Staraci: v-coding
[6:43] Garence Staraci: It's not a technical problem... It's just the human sort of accountability, the judgment aspect into it, the fiduciary aspect to it. How are you going to fire the AI if it makes a bad pick?
[6:43] Garence Staraci: How are you going to fire the AI if it makes a bad pick?
[9:11] Garence Staraci: gauge the room
[18:31] John Devine: square that circle
[18:31] Garence Staraci: history doesn't repeat
Original transcript
[0:04] I'm John Devine, assistant managing editor of
investing with US News and World Report. Um,
[0:09] and I'll let Gance uh introduce himself here in a
second. uh but just to talk about what we're going
[0:14] to speak about today. We will be speaking about
the evolution of AI and fundamental investing
[0:20] sort of different eras that it's gone through uh
where the human in the loop is uh hiring um and
[0:27] maybe if there's some commoditization of alpha
that's occurring and I'm sure some other things
[0:32] as well. So, uh, with that, uh, Gas, if you want
to, uh, just tell me a little bit about yourself,
[0:38] what you do for Wellington, um, and then how your
use of AI has changed over time with like I guess
[0:45] we talked about before some, uh, three different
eras of of AI, if you will, of pre-hat GBT, chat
[0:54] GBT, and then today's era, and how uh, you've been
implementing AI throughout those eras. Thank you,
[1:02] John. Hi everyone. Um so I'm Gahon Stasi. I'm
managing director at at Wellington management
[1:08] um at Wellington. Sort of in a nutshell um I lead
the science equity research team uh which is uh
[1:15] a cross- sector team um that invest as part
of Wellington's um global investment research
[1:20] which is basically the core at the core of the
investment platform um of Wellington management
[1:25] and it's um it's a rather sort of unique team
because it's a team that blends a lot of different
[1:30] profiles within sort of Wellington. So we have
people with a more fundamental investing sort of
[1:35] um traditional profiles, data scientists,
engineers. So we have really this blend
[1:40] of sort of profiles within our team. And and the
reason why we have that is because we have a very
[1:45] um sort of data intensive process and AIdriven
now process. So by data intensive I mean um really
[1:52] from traditional data which you could think of all
the fundamentals pricing data and so on but also a
[1:57] lot of alternative data uh across all the sectors
that we sort of we cover and um and now um sort of
[2:04] lot of AI sort of agentic sort of solutions that
we develop as part of of of our process. So it's a
[2:09] it's a sort of rather unique um sort of blend of
of profiles and uh in a way of a different type
[2:14] of process very datadriven um and um and then so
basically we've it's interesting from the personal
[2:21] standpoint I was sort of an exquant so I was a
quant sort of equity researchers and I drifted
[2:25] towards more fundamental investing over the years
and and it's been a very um interesting to see
[2:32] um really being at the junction between sort of
traditional fundamental investors data scientists
[2:37] and tech and more tech engineers to see how over
time over the past few years the eras you just
[2:42] mentioned to see really like the evolution over
the years or towards AI from sort of skepticism to
[2:48] excitement to doubts to really is it here to help
me generate alpha and so on so it's like it's an
[2:54] evolving process uh but it's you know uh it's been
interesting to witness okay and then um let's see
[3:03] so what era would you say we're in now are we in a
semi-agentic era Where are we in I guess probably
[3:09] not full agentic, right? I mean, how are you using
agents in your and your workflow? Now, I think
[3:15] maybe I'll take a step back, John, if you don't
mind talking about sort of the areas that you just
[3:18] mentioned because I think it's you want to give a
bit of a preamble towards bailing it to this. Um,
[3:23] so it's it's interesting because it's it's I I
tend to classify it as like the pre sort of chipt
[3:29] era. Uh, the chachi when it came and now really
the agentic piece which is sort of nowadays. Uh
[3:34] and it's been um interesting to see that one point
I could emphasize it's not only what AI could do
[3:40] but also who could use AI because there's always
an interesting feedback loop between who are the
[3:44] users and that feeds itself into what are the
solutions that we build that feels itself or
[3:48] how does our investment process changes how do
we pick stocks like all of that sort of fills
[3:52] into himself and and that accelerated depending
on who was using it. So if you think about sort
[3:57] of the pre sort of um sort of chipt era a lot
of the call AI a lot of like sort of the more
[4:04] machine learning and LLPdriven skills were really
the one needed and they were really helped by the
[4:08] the the data scientist uh the quants really um
using you know a lot of NLP use a lot of LP sort
[4:14] of sentiment use a lot of like even like early on
Q&A in terms of public filings the pre-sh GPD sort
[4:20] of prototypes that that existed out there uh but
that was really helped by those people because the
[4:25] technical expertise statistical expertise, the CS
expertise to do it. And that was mainly invisible
[4:29] to the traditional fundamental investors that
were just consuming it uh without necessarily
[4:33] understanding and without necessarily like how
the the soup was being made, but they were just
[4:37] sort of understanding the end product. And then
came Chpt and then this is when things started
[4:42] entering into the broader investment landscape,
the really sort of fundamental investing landscape
[4:45] where Chad GPT came and suddenly everyone starts
having like a research assistant feel like they
[4:49] have a research assistance with them. So you
would start by doing very simple things on on
[4:53] the fundamental investing side like sort of simple
summarizations of public documents. Then it would
[4:59] be uh very sort of drafting memos. Um very sort
of simple task uh but knowing now I have someone
[5:05] next to me uh helping me and and but you still had
some bottlenecks. I would I would frame it like
[5:11] you still had issues like hallucinations of some
of this model. you still had issues like lack of
[5:16] memory uh internal knowledge that would really I
would argue preventing you from feeling like you
[5:22] really had like really like a vertical step up in
terms of AI within the fundamental process. Um,
[5:27] you know, I used to joke back then that then then
the GPT felt like Ford for on the fundamental side
[5:31] felt like, you know, you had a a bit of a great
intern but a terrible colleague basically working
[5:35] with you. That's kind of how the feeling what it
was. And then after that came the agentic piece.
[5:41] And then there's two things that really changed
is really like the people now that couldn't code
[5:45] start really embracing it. And that's why what
I was talking about who could use AI. That's the
[5:49] next leg. And and that that's interesting because
we're still witnessing this. People who have no
[5:54] coding expertise, no coding background. They
start having um some of these tools now. And and
[5:59] they start feeling like they can build dashboard,
they can build screeners very quickly by just vi
[6:03] coding really. Um you know people get the smaller
people get instead of VS code let's say and they
[6:09] get github copilower and then suddenly they start
v coding and you hear the word say I'm VS coding
[6:13] right now you know when you know what VS code is
is kind of interesting always you know brings you
[6:17] a bit of a smile but people start really like
embracing it and then you have the agentic
[6:20] piece now which is now it really fits into the
core of the investment process you start really
[6:25] having like uh you know pulling fings your Excel
modeling your financial modeling it touches upon
[6:31] really like um you know feedback loop between what
you're doing and feeding back what you're doing
[6:36] and so on. So that's really sort of like a yeah.
So is there a is there a goal uh is there sort of
[6:43] an end vision for the workflow that you guys are
are working towards uh when incorporating agents
[6:49] into your workflow like how how do you see that
evolving and what's the sort of end end version of
[6:54] it? Sure. Um the piece what's interesting is I'm
going to talk mostly I guess about the traditional
[6:59] research about trying to figure out how to pick
stocks but there's something that I kind of really
[7:02] quickly emphasize want to emphasize is it's also
touching the entire really the agentic piece is
[7:06] really touching the entire investing workflow.
So from trading to risk management to portfolio
[7:12] constructions to even on the client side sometimes
in terms of like RFPs and so on. So there's really
[7:17] like this whole step where this is really going
into every single piece of the investment process
[7:21] and you have some um you know especially in the
trading side where you have a lot of very lot of
[7:26] data that's a very important piece where it's very
ahead sometimes even of research nowadays on the
[7:31] research side um you really have now a notion
where uh you have a full-blown assistant next
[7:38] to you that just pull all the recurring task all
the ties that take time the grunt work a little
[7:42] bit is getting more and more fully done in a fully
agentic which in some sense sort of compress the
[7:49] the timeline between human questioning and to the
evidence that you get. So it really compress that
[7:55] time frame. Um and then it just shift that time
frame into human judgment. I I'll call it this
[8:01] way in terms of human evaluation of what we get
as opposed to really doing the grunt work. So this
[8:06] is really a process going um really on its way. We
can see that coming. It's not fully there yet, but
[8:11] you know it's coming reasonably soon. Um the thing
that it's not and that alo something that I would
[8:17] argue it's um will not be something it's on the
record but it will not be something that is going
[8:23] to pick stocks fully on its own. Sure we have some
quantum out there trying it but but I guess here
[8:28] it's not a technical problem. Sure you can get
that going and imagine a future where it's fully
[8:32] there. It's just the human sort of accountability
the judgment aspect into it. the fiduciary aspect
[8:39] to it being able to say well this this is the
AI how are you going to fire it if it makes a
[8:43] bad pick you know exactly right but I mean that's
actually like sort of very deep sort of problem
[8:47] that you you know in in that sense like how do
you justify that even um with your clients it's
[8:52] like a lot of innovations where I don't know like
whether you call it sort of u tax preparations or
[8:57] financial advice where at the end of the day
sure you can think about the whole technical
[9:00] aspect there but at the same time at the end and
human has to be there you want a human so that I
[9:05] don't think you know anytime soon it's we'll get
the full uh full sort of automations from A to
[9:11] Z. Okay. And let's talk a little bit about um
commoditization um as you know a lot of these
[9:16] models are somewhat close to each other um and you
know the endgame is alpha here is outperformance.
[9:24] So how do those two things tie together? Do you
worry about sort of everybody chipping away at
[9:30] each other's alpha and then there is none.
Um how does that play in? That's a very deep
[9:36] philosophical question that that is very going to
be the the big question of the next few years. So
[9:41] you can have a lot of thoughts but uh time will
tell. So I make I guess a broad um general sort
[9:48] of comments there um is that you know it's as I
look at it like in many other techn technological
[9:56] innovation the past whether it's like the XL the
Bloomberg even the email to some extent you can
[10:00] go back you can argue that there's a notion that
there there's this new tool coming in compressing
[10:03] a lot of the some of the edge that people have is
now getting sort of compressed so I do think that
[10:09] um like any other technological innovation there's
pools of alpha that are going to shrink slash
[10:13] disappear um and others gonna get built sort
of somewhere else. It will not disappear fully.
[10:18] It will just move somewhere else. Um has the
ecosystem embraced that and how it moves. So as
[10:24] an example, I think that there's there there used
to be on a fundamental side an an alpha edge going
[10:30] into the person that could really uh crank through
the information quickly, get the information, the
[10:37] relevant information, crank through it, understand
the implications, put that into the financial
[10:41] models or update the models, try to really go back
to conclusions out of it and come up relatively
[10:45] quickly with an answer, what should we do with
this stock, with this company and so on. So that
[10:49] part is and people who have superior access of
information in some sense, people have superior
[10:54] tech stack in some sense who could do all that
faster than others. Um that is a legit I think
[10:59] source of alpha has been and that part is going
to be really really compressed if you imagine a
[11:03] world where all that is fully agentic and clean.
Um and so what I call process alpha it's going to
[11:09] sort of sort of go away a little bit. Um and then
you know and then it becomes the big philosophical
[11:14] questions of where all that sort of moves right
some people will tell you that like there's a new
[11:18] um sort of um you know longer term horizon type of
strategy people because they compress time frame
[11:24] so much now that we're processing the information
in the news alpha is going to be pushed on a
[11:28] longer time horizon that's one thing but then you
can tell yourself if everyone goes there where's
[11:31] the alpha there it's going to compress there at
that moment I think that at the end of the day
[11:36] um to to really answer your questions I think
there's certain key component that are going
[11:39] to be more important than ever when it comes to
alpha. Um I think that there's a notion that your
[11:44] own internal data um your access to management,
your ability to just um you know your ability to
[11:53] just understand differentiated viewpoints, try
understand inflection, try to understand quality
[11:57] of management stuff that are hard to quantify a
little bit where the human aspect will come in and
[12:02] where you're going to need to have the right
intuition, the right emotional intelligence,
[12:05] whatever you want to call it to really be able to
gauge the room and understand where the narrative
[12:09] is wrong. Uh I think the human is more than
ever will be at the center of the next sort
[12:14] of generation of alpha in a world where everything
else will be fully sort of um aspect and the idea
[12:20] of customization as well. That's a a beautiful
segue into my next question which is about the
[12:25] human in the loop actually um and what you're
looking for in somebody that you're hiring now
[12:30] and how that's changed over time and uh maybe
what skills are more valuable today than they
[12:36] were yesterday and vice versa. That's actually a
very um very relevant question and very important
[12:42] because that is changing I think a lot right now.
And so for the hiring it's it really depends on
[12:47] on who and maybe I'll try to categorize as a data
scientist the fund traditional fundamental sort
[12:53] of investor analyst um and new roles potentially
being created on the data science side of things.
[13:00] uh this is where I guess uh I don't know if it's
the right way of saying it but the age of IQ goes
[13:06] you know now it's EQ versus IQ in a sense that I
think three five years down the years ago which
[13:11] was not too long ago you really wanted people
with the best technical skills the best coding
[13:15] skills because they were still alpha going
back to my earlier point about building great
[13:19] infrastructure great tech stack very quickly
building sort of processes that others couldn't
[13:23] replicate because they have the resources and
the means to do it so you were really trying to
[13:28] identify really um people who could really code
very well very fast and the rest came after
[13:35] um and now this is changing really I think now
in terms of the data sense sort of roles we see
[13:40] as companies the first thing we gauge really
it's a little bit about curiosity it's about
[13:44] passion for investing it's a little about ability
to communicate very well ability to just really be
[13:50] um you know curious about the investing landscape
try to come up with differentiated thoughts you
[13:55] try to gauge that first and foremost and the
the the the hard sort of tech technical skills
[14:00] come a little bit after because you can understand
that you can use sort of agents now and so on or
[14:04] solutions that we have that can help with this. So
that is that's interesting because it seems like
[14:08] hiring is becoming even more subjective then in
terms of the yeah I would agree a lot of things
[14:13] are going to become you know more a little bit
more sort of subjective in that sense where you
[14:17] know the human evaluation will come a lot even
stronger um and and and that's so you're totally
[14:23] right on that front on on the on the new roles
I think that there's something we see emerging
[14:28] in in in within investment teams is there's people
now um I don't know if we can call it AI engineers
[14:34] or but people who are the core investment teams
and and they own the the the tech the AI stack
[14:40] inside those investment teams. So people that are
responsible within the team to really build sort
[14:45] of the AI processes, the prompting, the relevant
libraries that fits within the investment process
[14:49] of the team and for that you need two things
uh probably you need to have an understanding
[14:54] of the technology, understanding what to use
for what and understanding of the financial
[14:57] aspect of it to understand what is the relevant
pro is the relevant approach you want to take
[15:02] how do you want to build this. So people that
are really are fundament who are sort of have
[15:07] an investing investing acumen if you want and
people who can actually really understand the
[15:12] technology and build sort of the AI stack that's
sort of a little bit new and on the traditional
[15:17] analyst side the traditional sort of fundamental
analyst that's a big open questions and going
[15:21] back to some other points is like um in a [clears
throat] world where a lot of the traditional work
[15:27] that analysts have been doing um the Excel
aspect of it the the news processing thing
[15:33] um what happens to these roles when when you know
AI is at the stage where it's fully clean and you
[15:38] can really sort of put all of that into a fully
AI automated process um you know it's it's what is
[15:44] going to be the impact on that even on portfolio
managers down the line of of of a place where the
[15:48] technology is really almost there you can really
rely on it at really almost a complete level um
[15:54] that's that's really going to be the the questions
and and you know there's different sort of answers
[15:59] but but that's that's an evolving process as Well,
so today when we're talking about workflow, do you
[16:05] have um agents set up that go and scour the, you
know, the investment landscape for ideas and then
[16:13] you have humans vet those ideas or is it the other
way around or is it a little bit more nuanced than
[16:19] that? Um tell me about how you think about that.
Yeah. So it's it's not surprising and it will be
[16:25] part of it a lot of um what I just mentioned like
the the part that the the spirit there first and
[16:30] foremost is to really be able to um automate and
and as much as we can. So when you when you cover
[16:37] such a high sort of investment universe uh and
you see sort of evolving markets nowadays which
[16:42] there's so much volatility going on uh you try to
really benefit reap the benefits of AI and sort
[16:47] of agentic solutions to be able to really um you
know help you filter what's noise versus what's
[16:53] not and that is a very big problem because you
know um you need to um nowadays trying to figure
[16:59] out what is signal versus noise it's it's a very
uh sort of hard problems nowadays when you think
[17:04] about there's so much sort news increasing news
out there and AI helps you compress a time frame
[17:09] at which you can process those news and therefore
potentially react to it. Do you want to react to
[17:14] it? Do you not want to react to it? And can you
rely on AI to really also be fully clean? So I
[17:19] think there's a part of it which is really about
trying to help us in that world of noise and in
[17:23] that world of very large investment sets try
to narrow this down further. Uh so that's where
[17:28] part of our solutions will build on this to try to
spot investment opportunities and there's another
[17:33] layer where we use that to really help us
custom customize and stress test our investment
[17:38] processes. So part of the differentiation that I
mentioned going into alpha is to be able to have
[17:44] um really a unique process uh that is our own that
is fully sort of uh codified in scales and try to
[17:50] really be able to have a full-blown sort of sort
of AI sort of uh supporting system on that because
[17:56] sort of we do think that when we see the rise of
third parties um tools that are actually propose
[18:04] this AR workflows for fundamental investing you
know more and more people will have access to more
[18:10] and more the same thing when it comes to AI as
well and we think like the differentiation on the
[18:13] alpha side is also to help us to have something
differentiated from whatever else is using because
[18:18] you know we think there's also potential risk on
the line on that so I would say it's a lot more
[18:22] about trying to help us narrow down our investment
opportunities on the one hand and on the other
[18:26] hand help us to really codify our investment
process trust test it and get to a point faster
[18:31] where we we can just be really differentiated um
and we don't have much time but I do want to uh
[18:37] ask about um Is there enough data out there at
this point um to where let's say that you know
[18:42] your firm or a group of folks at your firm takes
a certain um view of the markets or where interest
[18:49] rates are going or where asset allocation should
be and then AI maybe has different views on that
[18:55] and how do you square that circle or how much
credence do you put into each you know the human
[19:00] versus the machine belief. Yeah, it's interesting
because it's the way sort of approach that it's
[19:06] it's a lot more as an assistant towards research
at that stage, right? So, so evidently what AI
[19:12] help us to do. We have the exper human experience
towards thinking about different cycles thinking
[19:17] about how all responded in in those conditions
historically to think about how interest rates
[19:22] were used to move. So, we have always of human
sort of market knowledge and pattern recognition
[19:26] that have been built towards investing experience.
But once this is happening and in a world where
[19:31] sometimes the history doesn't repeat then AI
comes here to try to help us a little bit more
[19:36] um stress test and be faster at stress testing
those assumptions. So you know with the ability
[19:42] to just really um help us uh the reasoning aspect
of it that came in with AI more and more reasoning
[19:47] aspect really help us to really brainstorm more
and try to potentially open up to uh really like
[19:53] new ideas that we wouldn't have otherwise. So
that's really what it really helped us on on
[19:58] really like that aspect of of thinking trying to
think differently. So of course we know that with
[20:02] some of these AI solutions you still have sometime
is is the way it's trained and so on. There's
[20:07] still some biases towards what he can tell you
or not. So we're sort of trying to be cognizant
[20:10] of that. But I think that the the the first and
foremost multiplier is our ability to just really
[20:16] um try to use those tools to also try to think
further deeper and try to catch things that
[20:21] wouldn't have caught otherwise. So, it's really a
blend of sort of the AI helping us in our research
[20:25] and us having our preconceived sort of sort
of market experiences. That is evidently like
[20:30] um always an important thing to have. Great.
Well, I think we'll have to leave it there. Gance,
[20:34] uh thank you so much for your time and thank
you guys for coming out. I appreciate it. Go.
