Pinegrove, Sagard Deploy Custom AI to Slash Deal Underwriting Times

Pinegrove与Sagard部署定制AI大幅缩短交易承销时间 机构资产管理机构警告称，若缺乏专有数据管道与严格的治理网关来保护受许可的投资情报，公共AI基础模型将难以奏效。 行业领袖深入探讨了生成式AI与智能自动化在另类及机构资产管理领域的落地应用，涵盖从基层技能提升到自上而下流程再造的广泛实践。重点案例包括从复杂基金文件中自动提取数据、利用LLM整合数百份专家访谈纪要以支持交易承销，以及针对特定角色的Copilot助手。嘉宾在交流中一致强调，企业级AI的价值根本上取决于整洁的数据基础设施、严格的治理机制以及严谨的变革管理。展望未来，专家组预测投资行业将更广泛地转向“量化基本面”（quantamental）范式，并逐步采用能够自主开展项目挖掘与定制化客户投资组合构建的自主智能体。 Sagard 采用双轨战略推进 AI 落地，结合自下而上的员工赋能（通过定制 GPT 培训带来 5% 至 10% 的效率提升）与自上而下的复杂基金法律文件自动化提取流程。 (3:21) American Century Investments 将生成式 AI 布局划分为三大战略支柱：面向特定岗位角色的 Copilot、端到端流程重塑，以及用于超额收益（Alpha）生成的定制模型。 (5:46) SS&C Blue Prism 通过部署能够跨后台系统整合数据并自动起草客户沟通函的智能 Agent，已在内部累计节省了超过 80 万个运营工时。 (8:06) Pinegrove 展示了前台投资尽调的 AI 实践，利用带元数据加权的定制 LLM 处理管道分析了 300 多份专家访谈记录，省去分析师长达三周的繁重人工检索工作。 (9:39) 专家指出公共基础模型受限于合规与权限无法直接开箱即用，资产管理机构必须打通底层数据管道并建立 AI 治理网关，以规避“垃圾进，垃圾出”并确保审计追踪能力。 (22:30)

Asset managers are pairing grassroots upskilling with top-down process automation to accelerate investment due diligence and eliminate operational bottlenecks. However, panel leaders stress that realizing these gains requires robust foundational data infrastructure, strict entitlement governance, and realistic expectations regarding vendor agentic hype. Front-Office Due Diligence and Agentic Deal Sourcing (9:39) Panelists report substantial compression in investment evaluation timelines by applying language models to unstructured qualitative records. Adam Nahari explains that Pinegrove Venture Partners parsed more than 300 expert interview transcripts during an evaluation of a payments provider, using custom pipelines that preserved conversation metadata—such as speaker seniority and interview dates—to compress three weeks of junior analyst review into an automated query matrix. Azhar Dewji notes that Sagard reduced an insurance portfolio company's manual underwriting cycle from multiple hours to five minutes, while internal data engineering workflows for sourcing deals dropped from weeks to hours. Looking forward, Dewji describes configuring autonomous web-crawling agents to scour platforms like PitchBook, Crunchbase, and Hacker News against investment mandate parameters, drafting personalized outreach in a partner's voice while leaving human investors in charge of final engagement. Tarun Sood forecasts that these tools will accelerate "quantamental" investing—blending quantitative machine learning with fundamental equity research—and enable firms to dynamically construct hyper-personalized portfolios tailored to individual client constraints. Operational Process Re-Engineering and Role-Specific Copilots (3:21) Firms are deploying a mix of employee self-service tools and structural workflow overhauls to capture both marginal and enterprise-scale efficiencies. Sagard provisions ChatGPT Enterprise licenses paired with targeted prompt engineering boot camps, yielding 5% to 10% individual productivity gains, while executing top-down automation across complex fund legal paperwork. Its automated document extraction pipeline ingests Limited Partnership Agreements (LPAs), side letters, and subscription documents once, transforming operational expense allocations and audit covenants into structured, queryable data objects. Similarly, Sood categorizes American Century Investments' initiatives into three distinct pillars: persona-based copilots (tailored research models for equity analysts, GitHub Copilot for developers, and Collibra AI agents for data stewards), end-to-end process re-engineering (conversational client reporting and document parsing), and custom alpha generation. On the vendor-practitioner side, Dan Welcel reports that SS&C Blue Prism deployed internal communication agents that query disparate back-end record systems to draft context-aware client correspondence, returning over 800,000 operational hours back to the business. Foundational Data Plumbing and Governance Realities (21:45) The panel rejects the notion that generative AI removes the need for disciplined enterprise data engineering and cataloging. Sood compares enterprise AI deployment to urban planning: language models cannot succeed without sanitized utility lines, clean plumbing, and unified storage, underscoring that "garbage in, garbage out" persists in the generative era. Moreover, Sood warns that out-of-the-box public foundation models are black boxes trained on internet data, meaning asset managers must design internal architectures and conduct vendor bake-offs to contextualize models using proprietary, entitled data without breaching regulatory boundaries. To satisfy compliance and risk teams, Welcel and Dewji emphasize automated technical guardrails. Welcel explains that SS&C Blue Prism engineered an AI Gateway to enforce personally identifiable information (PII) detection, data masking, and audit trails before queries reach external models, while cautioning firms against over-engineering simple tasks with complex agentic frameworks ("using a chainsaw to cut down a flower"). Dewji adds that AI is simultaneously being deployed inward to inspect corporate databases, automatically generating catalog metadata and flagging sensitive columns to prevent data leaks. Developer Productivity and Infrastructure Cost Optimization (35:38) Generative tools are fundamentally altering engineering economics and cloud infrastructure management across investment platforms. Sood notes that developer web traffic to legacy programming communities like Stack Overflow has cratered as engineering teams universally adopt tools like GitHub Copilot, driving higher productivity expectations and contributing to a visible softening in open software engineering job requisitions. At the data infrastructure layer, Nahari shares that Pinegrove deployed an AI optimization tool (Zipr) to tune Databricks compute workloads, achieving an immediate operational cost reduction of nearly 40%. Nahari emphasizes that as the benchmark duration for LLM task execution doubles roughly every seven months, continuous compute and DataOps optimization will become essential to maintaining cost-effective institutional deployments. The panel concludes that while technological barriers to analyzing unstructured data have fallen, organizational friction persists around compliance, data privacy entitlements, and vendor marketing hype. Sustainable enterprise scale requires combining top-down executive mandates and robust data plumbing with hands-on AI literacy programs that anchor deployment in practical, role-specific workflows.

在由Ai4主办的研讨会上，来自Sagard、American Century Investments、SS&C Blue Prism及Pinegrove Venture Partners的行业高管指出，人工智能正在深刻重塑资产管理的前台投资决策与后台运营流程；然而，机构若要获取持续的超额收益与效率红利，不仅依赖通用大语言模型，更取决于底层数据基础设施的成熟度、严格的合规治理网关以及结合具体场景的端到端流程重组。 组织落地范式：角色赋能与端到端流程重构 (3:21) 资产管理机构在推行人工智能时，普遍采取分层推进策略以平衡局部效率与全局变革。Sagard的实践兼顾自下而上的员工赋能与自上而下的流程改造：一方面向全员提供ChatGPT Enterprise账号并开办定制GPT训练营，促使员工自主提升约5%至10%的个人生产力；另一方面针对复杂繁琐的基金法务流程，搭建自动化文档抽取管线，将合伙协议（LPA）、附随函件及认购协议等数百页非结构化法律文本转化为可复用的结构化数据对象，彻底消除了跨部门调取合规与审计条款的重复性人工查阅。 American Century Investments则通过三层架构部署生成式AI：一是基于岗位画像的赋能，涵盖基本面分析师的研究Copilot、软件工程师的GitHub Copilot以及数据管理人员的Collibra智能代理；二是端到端流程重构，包括自动化文档提取管线与对话式动态客户报告接口；三是超额收益生成（Alpha Generation），将传统机器学习与现代生成式模型相结合，辅助投资团队验证交易假设。在运营自动化层面，SS&C Blue Prism在内部Middle和Back-office部署了协同大语言模型的通信智能体，实现跨后台系统自动提取数据并回复客户咨询，累计为企业节省了超过80万工时。 投研尽调提速与投资组合赋能实践 (9:39) 在核心前台投资与尽职调查环节，生成式AI显著压缩了海量非结构化文本的研读周期。Pinegrove Venture Partners在评估一家支付领域的标的企业时，面对超过300份专家访谈记录（长达数千页的非结构化定性材料），没有按照惯例派遣初级分析师耗费三周逐页阅读，而是构建了定制化的LLM摄取管线。该管线系统性提取并保留了专家职务、资历背景及访谈时间戳等关键元数据，确保大语言模型在针对具体承保问题及竞对基准分析时能够加权历史背景，输出高置信度的定性分析结果。 除了内部投研，AI赋能已延伸至投资组合的投后价值创造。Sagard通过半天密集的虚拟工作坊帮助被投初创企业的工程与产品团队消除AI技术门槛，并在1对1的专项运营优化中，协助一家保险科技被投公司将耗时数小时的手工核保流程压缩至5分钟，同时保持并提升了核保准确率。针对一级市场的项目搜寻，Sagard进一步预测，自主网络爬虫智能体已能够依据具体投资策略，自动扫描Crunchbase、PitchBook及技术论坛，筛选标的并以合伙人语气起草个性化接洽邮件，将投资经理从机械性搜索中解放出来。 底层基础设施与数据治理合规网关 (22:30) 与会专家一致反驳了“LLM能够天然吞吐非结构化数据因而无需前期数据清洗”的市场误解。American Century Investments指出，在生成式AI时代，“垃圾进，垃圾出”的问题更为严峻；分析师产出的分散非结构化数据必须经过脱敏、清洗，并与现有数据湖统一整合。其高管以城市规划作比喻：若缺乏健全的电力和供水管网等市政设施，地表上的摩天大楼便无法立足；同理，没有经过治理与清洗的数据管道，语言模型也无法真正发挥商业价值。 机构投资高度依赖敏感的专有数据与持牌特许数据，这些数据无法直接输入公开发布的基础大模型。为此，SS&C Blue Prism开发并商业化了AI Gateway（AI网关），作为连接模型与数据的防护栏，在提示词调用外部LLM之前自动执行个人身份信息（PII）检测、数据脱敏与访问权限审计，确保端到端的可追溯性。Sagard的实践同样证实，利用AI辅助开发工具可将过往耗时数周的数据接入工程缩短至数小时，且AI能自动化巡检数万张数据库表并生成元数据定义，主动化解合规与数据泄漏风险。 研发降本增效与量化基本面长期演变 (35:38) AI对底层数据运维与技术团队的生产力重组已产生实质影响。American Century Investments观察到，随着工程人员普遍采用AI代码助手，Stack Overflow等传统编程问答社区的网络流量出现断崖式下跌，研发效率的倍增正伴随着初级软件工程岗位的招聘放缓。在基础设施运维（DataOps）方面，Pinegrove通过引入AI计算优化工具Zipr调优Databricks算力配置，直接降低了约40%的云端计算成本。此外，大语言模型处理复杂任务的有效时长正以大约每7个月翻一番的速度呈指数级增长。 展望未来行业格局，American Century Investments认为资产管理将全面转向“量化基本面”（Quantamental）投资范式，即量化模型与AI深度嵌入基本面选股，并推动高度定制化投资产品的普及，使零售和机构客户能够按需生成排除特定行业或匹配个性化风险因子的定制指数与基金。SS&C Blue Prism则提醒机构保持架构理性，避免用“电锯修剪花朵”，切勿将智能体（Agentic AI）视为即插即用的万灵丹，而应紧扣业务规模与实际ROI匹配合适工具。 与会高管总结认为，资产管理行业的专业人员正不可逆转地转向“AI辅助决策者”，技术重心已从前沿模型的概念验证转移到合规落地。然而，如何在极其严格的监管、数据权限约束与模型黑箱之间求得平衡，以及如何在大规模推行中化解组织内部的变革阻力，仍是各家机构构建长效竞争壁垒的关键考量。

## Background 本内容由 Ai4 频道于 2025 年 8 月 26 日发布在 YouTube 平台上。视频记录了一场题为《AI 驱动的资产管理创新：增强决策与提升效率》的专题圆桌论坛，汇聚了来自 Pinegrove Venture Partners、American Century Investments、Sagard 及 SS&C Blue Prism 的多位数据与 AI 负责人，深入探讨 AI 在资产管理领域重塑数据分析、优化投资策略与提升运营效率的落地实践。 ## TL;DR 圆桌讨论围绕资管机构在前台投资决策与中后台运营中如何系统化落地 AI 展开。嘉宾们分享了通过全员普及 ChatGPT、非结构化长文档与专家访谈自动化解析、定制化 Alpha 生成以及智能自动化节省数十万工时的真实案例。同时，专家们深入探讨了投后赋能、数据治理与合规门槛（如 PII 过滤与数据权限管理），并对当前过度炒作的 Agentic AI 提出了理性审视，强调健全底层数据基础设施才是决定 AI 规模化成功的关键基石。 ## Atmosphere & style 整场对话务实、严谨且充满一线从业者的专业洞见，丝毫没有浮于表面的空泛宣讲。主持人节奏把控得当，四位技术与投资高管在“前台投研赋能”与“中后台合规治理”之间形成了极佳的攻守呼应。讨论既有剖析实际提效案例时的兴奋感，也有谈及数据权限壁垒、技术选型与概念炒作时的清醒与克制，让听众深刻体会到顶级资管机构在拥抱前沿 AI 浪潮时兼顾创新与风控的沉稳底色。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=SoanrcutwpQ&t=4s) 资管机构在 AI 布局上兼顾前台投资决策（Alpha 生成、项目挖掘、尽调）与中后台运营（文档结构化、治理合规、智能客服），实现攻守平衡。 - [03:22](https://www.youtube.com/watch?v=SoanrcutwpQ&t=202s) Sagard 采用“自下而上”（ChatGPT Enterprise 全员普及与训练营）与“自上而下”（长篇 LPA/Side Letter 法律文档单次解析结构化入库）的双轨 AI 战略。 - [05:46](https://www.youtube.com/watch?v=SoanrcutwpQ&t=346s) American Century Investments 确立了 AI 落地的三大支柱：基于角色画像的岗位赋能（Persona-based）、前后台流程重构以及定制化 Alpha 收益生成。 - [08:06](https://www.youtube.com/watch?v=SoanrcutwpQ&t=486s) SS&C Blue Prism 将大语言模型与智能自动化深度融合，通过通信智能体跨系统自动生成个性化回复，内部累计节省超过 80 万工时。 - [09:39](https://www.youtube.com/watch?v=SoanrcutwpQ&t=579s) Pinegrove 自研定制化 LLM 管道穿透数百份专家访谈纪要，借助严格的元数据管理，将尽调提取分析周期从三周大幅压缩。 - [11:33](https://www.youtube.com/watch?v=SoanrcutwpQ&t=693s) 在投后价值赋能方面，Sagard 通过实战工作坊与嵌入自动化层，协助保险被投企业将核保耗时从数小时降至 5 分钟。 - [16:12](https://www.youtube.com/watch?v=SoanrcutwpQ&t=972s) 在尽调与另类数据整合中，资管机构面临“自研 vs 采购”的抉择，需在前沿通用大模型与传统金融数据巨头（如彭博、FactSet）自建模型之间进行严谨打靶比对。 - [19:02](https://www.youtube.com/watch?v=SoanrcutwpQ&t=1142s) 资管机构必须将基础模型引入内部安全边界，在严格遵守数据权限（Entitlements）的前提下使用私有数据微调，消除公网模型的黑盒与越权风险。 - [22:30](https://www.youtube.com/watch?v=SoanrcutwpQ&t=1350s) 数据基础设施与治理是 AI 落地的核心基石；大模型犹如上层的“摩天大楼”，必须依托底层清洗通畅的数据管道与 AI 网关等“供排水系统”才能稳定运行。 - [26:18](https://www.youtube.com/watch?v=SoanrcutwpQ&t=1578s) 提升全员 AI 素养需深入一线开展实战研讨，通过注入历史投资备忘录等内部专有语料构建定制 GPT，从而建立强烈的业务认同感。 - [29:16](https://www.youtube.com/watch?v=SoanrcutwpQ&t=1756s) 行业应警惕对 Agentic AI 的过度炒作与“即插即用”营销话术，企业级落地必须注重系统底层架构的健壮性与可扩展性，坚持适度匹配工具原则。 - [33:23](https://www.youtube.com/watch?v=SoanrcutwpQ&t=2003s) AI 显著反哺数据工程与安全治理：多源数据摄取从数周压缩至几小时，并能秒级识别 PII 敏感字段并配置访问隔离与脱敏。 - [35:38](https://www.youtube.com/watch?v=SoanrcutwpQ&t=2138s) 开发与运维生态正在经历范式转移：AI 编程助手成为基线工具，同时 AI 调优引擎在 DataOps 运维中可将计算集群成本降低近 40%。 - [37:45](https://www.youtube.com/watch?v=SoanrcutwpQ&t=2265s) 风险投资（VC）全流程正迈向端到端智能体协作：Agent 可自主全网抓取筛选标的并草拟冷邮件，人类投资人则保留在核心逻辑判断与人际沟通的闭环中。 - [40:45](https://www.youtube.com/watch?v=SoanrcutwpQ&t=2445s) AI 驱动基本面分析与量化投资深度融合（Quantamental），并大幅降低构建门槛，推动面向个人投资者的超个性化（Hyper-personalized）定制投资组合普及。 - [42:32](https://www.youtube.com/watch?v=SoanrcutwpQ&t=2552s) 消除合规风控部门阻力的关键在于建立透明的数据链路追踪与权限控制，并由高层自上而下确立战略共识以推进组织变革。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=SoanrcutwpQ&t=4s) 圆桌开场与嘉宾背景介绍 圆桌主持人兼 Pinegrove Venture Partners 的合伙人及首席数据与 AI 官 Adam Nahari 首先明确了本次讨论的核心议题：探讨各家机构如何利用 AI 及更广泛的数据科学技术来增强投资决策能力，并提升中后台运营的整体效率。随后，四位来自另类投资、传统资产管理和智能自动化领域的领军人物依次进行了自我介绍。 Sagard 的数据基础设施主管 Azhar Dewji 介绍道，Sagard 是一家专注于风险投资、私募股权、私人信贷、房地产及医疗健康等另类资产类别的全球资产管理公司；在加入 Sagard 之前，他曾于多家科技公司从零搭建数据平台，近期曾任职于 Shopify 主导数据治理战略。American Century Investments 的首席数据与 AI 官 Tarun Sood 随后发言，指出 American Century 是一家总部位于堪萨斯城并拥有全球业务的中型资管机构；他的职责涵盖数据与 AI 的“攻守两端”——“进攻端”聚焦于通过用例拉动收入增长与提升效率，“防守端”则专注于数据治理及确保满足严苛的监管合规要求。SS&C Blue Prism 的解决方案咨询总监 Dan Welcel 介绍称自己在公司任职超过六年半，主要在销售端协助客户识别并在自动化智能体（Automation Agents）中落地具体业务场景。最后，Adam Nahari 介绍了 Pinegrove Venture Partners 的业务布局：作为一家涵盖信贷二期市场（Credit Secondaries）、母基金（Fund of Funds）和跟投（Co-investments）的风险投资与成长型投资平台，Pinegrove 致力于将数据科学与 AI 置于投资决策的“刀刃”（tip of the spear）之上，贯穿投资假设制定、项目挖掘、尽职调查到被投企业价值赋能的全生命周期。 ```mermaid flowchart TD accTitle: 资管机构 AI 应用定位与全景结构 subgraph S1["投资与决策前台"] A["投资假设制定与项目挖掘"] B["海量专家访谈与尽调抽取"] C["Alpha 策略与投资组合优化"] end subgraph S2["中后台运营与客户交互"] D["基金合规文档单次解析"] E["智能客服与多系统联动"] F["数据治理与监管合规 (Defense)"] end S1 --> G["提升投资胜率与回报"] S2 --> H["释放人力与运营提效"] ``` --- ### [03:22](https://www.youtube.com/watch?v=SoanrcutwpQ&t=202s) Sagard 的双轨 AI 战略：自下而上赋能与基金文档结构化抽取 Azhar Dewji 详细阐述了 Sagard 内部采用的“自下而上（Bottom-Up）”与“自上而下（Top-Down）”相结合的 AI 推进路径。在全员共识中，“如果你不使用 AI，你就会被时代甩在身后”。在自下而上的层面，公司为每位员工配备了 ChatGPT Enterprise 企业版许可，并打通其日常工作邮箱和内部文档系统；不仅如此，Sagard 还针对各个业务团队举办专门的实战训练营（Bootcamps），指导员工构建定制化的 Custom GPT。Azhar 强调，单靠开箱即用的 ChatGPT 固然不错，但只有将其针对特定业务场景定制时，才能释放出成倍的价值。自下而上的普及预期能够为每位员工带来 5% 到 10% 的效率提升。 然而，更具颠覆性的变革来自于自上而下的流程重构。Azhar 以基金法律文档的提取为例：在日常运营中，机构沉淀了大量的有限合伙协议（LPAs）、旁白信（Side Letters）和认购协议等动辄长达数百页的核心文件。以往每当业务部门需要确认某项特定条款（如年度审计要求或运营费用的分摊规则）时，初级分析师都必须先找到文件，在数百页密密麻麻的条款中反复翻找核对。为此，Sagard 构建了专属的文档解析与抽取数据管道：所有长篇文档仅需被系统接入并处理一次，关键字段和条款就会被结构化转化为全公司统一、可信、随时可调用的数据对象，极大消除了重复检索的低效与摩擦。 - **全员工具链覆盖**：为每位员工提供集成邮箱与知识库权限的 ChatGPT Enterprise。 - **场景化定制培训**：通过团队专项训练营推广 Custom GPT，预计为日常工作带来 5% 至 10% 的效率增益。 - **文档单次入库机制**：对长达数百页的 LPA、Side Letters 等法务与合规文档进行一次性摄取并转化为受信任的结构化对象。 --- ### [05:46](https://www.youtube.com/watch?v=SoanrcutwpQ&t=346s) American Century Investments 的三大 AI 落地支柱 Tarun Sood 分享了 American Century Investments 推进 AI 落地的系统性框架。该机构将所有探索与上线的场景划分为三大业务支柱（Buckets）： 第一是**基于角色画像的 AI 赋能（Persona-based AI）**。该策略专注于端到端重塑机构内部各个具体岗位的作业流。例如，针对从事前端基本面研究的金融分析师，为其定制专属模型或 Copilot 工具以优化投研流程；针对软件研发人员，全面引入 GitHub Copilot 等代码生成助手以加速工程交付；而在数据团队内部，则积极采用 Collibra 等工具内置的 AI 智能体来精简与加速数据治理工作。 第二是**前后台流程的重构与再造（Process Re-engineering）**。除了与 Sagard 类似、正在积极开展原型的非结构化文档提取外，团队重点布局了客户报告领域的对话式 AI（Conversational Reporting），探索通过自然语言交互方式向投资人输出定制化、实时的投研与业绩报告。 第三是**定制化 Alpha 收益生成（Custom Alpha Generation）**。当投资团队提出具体的投资假设时，数据与技术团队会结合传统的 AI/ML 算法与生成式大模型进行深度验证与回测，直接利用前沿技术驱动投资创新并捕获超额 Alpha。Tarun 表示，目前机构在上述三大领域均已有多个用例进入实际生产环境。 --- ### [08:06](https://www.youtube.com/watch?v=SoanrcutwpQ&t=486s) SS&C Blue Prism：结合智能自动化与大模型节约数十万工时 Dan Welcel 从软件服务商兼内部实践者的独特视角剖析了 AI 的应用价值。SS&C Blue Prism 不仅对外提供软件，更在内部深度“自产自用（Dogfooding）”，从而保持高度的敏捷性并能及时调整应用策略。公司将大语言模型与自身成熟的智能自动化技术（Intelligent Automation）相结合，重点攻关中台和后台运营，并深度赋能客户交互团队。 在具体的客户沟通场景中，系统部署了由大模型驱动的通信智能体：当收到客户的入站咨询时，智能体能够自动理解诉求意图，跨底层多个业务系统调取关联数据，将其组织为 LLM Prompt 并调用大模型推理，最终毫秒级生成极具针对性和个性化的答复内容。Dan 透露，仅凭这一项客户交互智能体应用，SS&C Blue Prism 在内部就已经为业务部门累计夺回并节省了高达 **800,000 个工时**。这一显著成果不仅体现了自上而下的推动决心，也印证了 AI 在降低内部运营成本与提高客户满意度上的双重红利。 --- ### [09:39](https://www.youtube.com/watch?v=SoanrcutwpQ&t=579s) Pinegrove：通过定制 LLM 管道穿透数百份专家访谈辅助投资决策 针对前面嘉宾聚焦于中后台和运营场景的探讨，主持人 Adam Nahari 补充了一个 AI 直接介入前台投资决策与项目核保（Underwriting）的典型案例。在近期评估一家支付服务提供商及其与数家竞品对比的交易过程中，Pinegrove 的投资团队获取了 300 多场专家访谈纪要，累积形成了多达数千页的高价值非结构化文本。 在传统的工作模式下，一名分析师往往需要花费整整三周时间埋头通读所有专家访谈，才能提炼出支持投资决策的核心洞见。而 Pinegrove 自研了定制化的 LLM 数据处理管道，能够将这些海量非结构化文本全面解析，并针对项目核保中至关重要的特定业务问题进行精准问答与交叉比对。Adam 特别强调，构建此类 LLM 管道的关键在于“定制化元数据管理”：必须准确捕捉访谈元数据——例如受访专家的身份背景、资历级别以及对话发生的具体时间节点等。缺少这些关键维度的结构化支撑，大模型就无法准确评估行业动态演变，更无法为投资团队提供真正具备可信度的决策支撑。 ### [11:33](https://www.youtube.com/watch?v=SoanrcutwpQ&t=693s) 赋能被投企业：Sagard 借助 AI 推动投后价值创造与核保流程提效 在讨论完投资机构自身沉淀数据与机构知识库后，主持人 Adam Nahari 将话题转向另一个关键维度——投后价值创造（Portfolio Value Creation），并邀请 Sagard 的数据基础设施负责人 Azhar Dewji 分享实践经验。Azhar 表示，Sagard 旗下的风险投资业务将为被投企业提供实际价值视为核心使命，而 AI 的爆发为此带来了巨大机遇。目前初创公司普遍面临一种困境：虽然所有人都知道必须拥抱 AI，但对于小型团队而言，在保持敏捷交付的同时落地具体战术往往令人望而生畏。为此，Sagard 采取了系统化的赋能路径。 Sagard 首先推行了半天的线上实战工作坊（Bootcamps）。这些会议面向被投企业的工程师、产品经理及核心开发人员，带领他们经历“构思创意（Ideate）—利用 AI 工具快速原型设计（Prototype）—实际构建可用产品（Build）”的完整闭环，旨在提升员工技能，帮助他们理解 AI 如何在日常工作中提供辅助。 除了通用的赋能工作坊，Sagard 还深入被投企业开展一对一诊断，梳理其日常运营中最耗时的核心痛点。Azhar 分享了一个代表性案例：一家主营保险业务的被投企业，其原本的保单核保（Underwriting）流程极其漫长繁琐。Sagard 团队与其共同梳理业务流后，针对性地在其业务管线中嵌入了特定的 AI 自动化层。最终，该企业的核保处理时间从原本的“数小时”大幅压缩至仅需“5 分钟”，在实现数倍提效的同时还显著提升了核保流程的流畅度与准确率。 ```mermaid flowchart LR accTitle: Sagard 投后企业保险核保流程 AI 优化对比 subgraph Traditional["传统核保流程"] T1["人工收集多方数据"] --> T2["逐项人工审核与评估"] --> T3["耗时数小时输出核保结论"] end subgraph AIAssisted["AI 赋能改造后流程"] A1["自动化数据流接入"] --> A2["AI 管道自动化特征抽取与初审"] --> A3["5 分钟内完成精准核保"] end ``` ### [14:37](https://www.youtube.com/watch?v=SoanrcutwpQ&t=877s) 赋能模式探讨：战略指导与实操开发的边界 针对 Sagard 的赋能模式，Adam 进一步追问：在协助被投公司时，团队更多是停留在战略规划与方案设计层面，还是会亲自“上手写代码”（Hands-on keyboard）参与工具的实际开发与交付？ Azhar 说明，截至目前 Sagard 仍然更侧重于战略赋能而非直接代写代码。团队通过内部的行业与 AI 专家帮助被投企业精准定位痛点并设计解决方案雏形，但最终的工具构建与代码落地仍交由被投企业团队自主完成，因为这些系统最终需要由他们长期维护和运营。不过他也提到，未来随着业务演进，Sagard 可能会在实施层面进行更深度的介入。此外，针对 Adam 提出的“是否协助被投企业进行 AI 人才招聘、撰写岗位描述（Job Description）及画像分析”等组织能力建设问题，Azhar 坦言目前尚未正式展开，但鉴于 Sagard 自身在人才寻访与大型组织构建方面的丰富经验，这无疑是未来极具潜力的拓展方向。 ### [16:12](https://www.youtube.com/watch?v=SoanrcutwpQ&t=972s) 尽职调查与另类数据整合：自研与采购（Build vs. Buy）的抉择 随后，话题转向如何利用 AI 赋能投资尽职调查（Due Diligence）与投资研究。American Century Investments 首席数据官 Tarun Sood 指出，传统资产管理行业向来高度依赖标准数据体系——如证券行情、投资组合头寸、基准指数（Benchmarks）和复合投资组合等；而在大语言模型（LLM）与各类 Copilot 辅助工具的加持下，最大的变革在于机构能够以极快的速度接入、汇总并合成海量的“另类数据”（Alternative Data），极大缩短了投研人员从原始数据聚合到形成洞察的周期。 在具体落地工具时，资产管理机构正面临严峻的“自研还是采购”（Build vs. Buy）决策，尤其是在投资组合分析领域： - **前沿大模型原生厂商**：如 OpenAI（ChatGPT 方案）、Perplexity、Anthropic 等，正在快速推出针对企业与分析场景的开箱即用方案。 - **传统金融数据巨头转型**：彭博（Bloomberg）、FactSet 等机构长期服务于资产管理行业，如今正基于自身沉淀的专有金融数据底座，训练并推出行业专用的定制化大模型。 Tarun 强调，作为首席数据官，核心任务是对原生 AI 创新与传统数据供应商的大模型产品进行严谨的选型与实测打靶（Bake-off），从而决定最佳的技术采购或自研路径。 - **核心提效环节**：大幅压缩另类数据采集、清洗与聚合的流转周期，加快投研分析师与基金经理的决策节奏。 - **技术选型维度**：需横向比对通用底层大模型（如 OpenAI、Anthropic、Perplexity）与垂直金融大模型（如 Bloomberg、FactSet 自建模型）在资产分析场景下的准确性与适配度。 ### [19:02](https://www.youtube.com/watch?v=SoanrcutwpQ&t=1142s) 数据协调、权限管理与私有化模型的落地挑战 对于另类数据的整合，Adam 提出了两项现实挑战：其一是跨多源异构数据的归一化与协调（Harmonization），例如当研究员想调取某家特定公司（如 Stripe）的全景信息时，系统能否顺畅打通各渠道数据；其二是多源数据融合后的洞察可信度，系统必须能清晰识别不同数据源的局限性与潜在偏差，确保分析师提问时能获得可靠答案。 Tarun 对此表示认同，并补充了资产管理行业面临的另一个底层阻碍——**私有数据与数据使用权限（Data Entitlements）**： > “你可以引入 OpenAI 的基础模型，但它们主要是基于公开互联网数据训练的。而在投资界，我们拥有大量的私有数据，而且这些数据涉及严格的权限授权。你不能直接把公共模型拿来即插即用。” Tarun 解释道，公共大模型的训练数据边界本质上受限于公开网络，且对机构而言运作机制仍存在一定的“黑盒”特性。资产管理机构必须将基础模型引入内部安全边界，在严格遵守数据权限（Entitlement）的前提下，利用机构自有的专有与私有数据进行定制化训练与微调，同时明确界定谁有权限分析哪些受控资产组合数据。 ### [21:45](https://www.youtube.com/watch?v=SoanrcutwpQ&t=1305s) 基础设施的迷思：大模型时代是否还需要底层数据清洗？ 在讨论的尾声，Adam 抛出了一个在生成式 AI 时代广受争议的话题：部分观点认为，既然现代大语言模型具备强大的非结构化数据解析能力，那么传统上耗时耗力的数据清洗、格式标准化以及跨数据流聚合等底层数据工程工作，在生成式 AI 时代是否已经变成了“可被省略的历史包袱”？带着这一核心问题，Adam 将麦克风交给了 SS&C Blue Prism 的解决方案咨询总监 Dan Welcel，请他从底层基础设施与企业实际准备度的角度进行深入拆解。 ### [22:30](https://www.youtube.com/watch?v=SoanrcutwpQ&t=1350s) 数据基础设施与数据治理：AI 落地的核心基石 在探讨 AI 落地应用时，Dan Welcel（SS&C Blue Prism 解决方案咨询总监）强调集中化的底层基础设施是支撑 AI 系统和应用存储、收集与使用数据的核心基础。如果底层数据存在孤立、分散或重复复制的现象，将会严重拖慢企业的创新步伐，阻碍 AI 方案的最终实现。除了数据基建，另一大关键支柱是“数据治理模型”（Governance Model）——即如何精准控制数据访问权限、实现全面的可审计性与可追溯性。 为了解决市场上缺乏成熟合规管控方案的痛点，SS&C Blue Prism 早期在内部自主研发了一套 AI 网关（AI Gateway），如今已将其推向市场。Dan 解释道，虽然公司采取由上而下的策略全面将 AI 融入产品与解决方案中，但若缺乏底层治理模型，就无法为内部利益相关者和外部客户提供他们所期望的安全屏障与信任感。 随后，Tarun Sood（American Century Investments 首席数据官）从数据质量和“垃圾进，垃圾出”（Garbage in, garbage out）的经典准则切入，指出在 AI 时代这一问题变得更为复杂。投资分析师会产出海量非结构化数据（如客户与公司分析），如何将这些非结构化数据与现有的结构化数据清洗、净化并整合到一个统一平台（无论是数据湖还是数据仓库），是实现高效数据管理的前提。只要实现数据的物理集中或逻辑整合，并配以扎实的数据治理和元数据（Metadata），查找和利用数据就会变得非常容易。 Tarun 还分享了一个生动的比喻：就像在城市中建造摩天大楼一样，若没有通畅的电力、供水和排水等基础公共市政设施（Plumbing），摩天大楼就根本无法正常运作。同样地，大语言模型（LLM）想要真正发挥威力，企业必须先打通底层数据管道并完成数据清洗净化。 ```mermaid flowchart TD accTitle: 数据基础设施与治理模型架构 accDescr: 展示从非结构化与结构化数据输入，经过基础设施管道和治理网关，最终赋能上层大语言模型应用的完整架构流程。 subgraph DataSources["数据输入层"] A1["非结构化数据 (分析师备忘录/报告)"] A2["结构化数据 (市场/财务数据)"] end subgraph CorePlumbing["数据基础设施与管道 (城市供水电系统)"] B1["数据清洗与净化 (Sanitization)"] B2["集中存储 (数据湖 / 数据仓库)"] B3["利用 GenAI 自动生成与管理元数据"] end subgraph GovernanceLayer["安全治理层 (AI Gateway)"] C1["访问控制与权限审计"] C2["敏感信息 (PII) 实时检测过滤"] C3["调用追踪与合规审计"] end subgraph LLMApp["上层业务赋能 (摩天大楼)"] D1["大语言模型与智能体 (LLM / Agents)"] D2["投研备忘录辅助 / 自动化决策支持"] end DataSources --> CorePlumbing CorePlumbing --> GovernanceLayer GovernanceLayer --> LLMApp ``` - **元数据自动化**：American Century Investments 正在探索利用生成式 AI（GenAI）本身来自动定义和补全数据湖中存储的所有元数据，从而让企业内部的“数据产品”更加完备。 - **城市管网类比**：大语言模型是耸立的“摩天大楼”，而数据管道、数据清洗和治理则是地下的“供排水系统”，缺少后者，上层建筑就无法稳定运作。 --- ### [26:18](https://www.youtube.com/watch?v=SoanrcutwpQ&t=1578s) 提升全员 AI 素养与落地定制化 GPTs 针对主持人提出关于员工培训与定制化 GPT 工具的实践经验，Azhar Dewji（Sagard 数据基础设施负责人）分享了 Sagard 在推动全员 AI 素养（AI Literacy）方面的策略。Azhar 指出，尽管互联网上存在大量 AI 教程，但在当前模型眼花缭乱（数十种模型并存）的环境下，业务人员往往会感到无所适从、不知道针对什么场景选用什么模型。因此，团队在起步阶段采取了“手把手”辅导的方式，直接帮助技术熟练度不高的员工跨越门槛。 在具体推行上，Sagard 放弃了空洞的概念灌输，而是采取深入一线的研讨会（Boots on the ground workshops）形式。技术人员会花一个下午的时间与具体的业务团队坐在一起，针对他们手头的实际日常工作展开讨论，演示 LLM 或 Chatbot 如何实实在在地提升工作效率。Azhar 认为，当员工亲眼看到 AI 在真实业务场景中发挥作用，而不是听到抽象的效率收益承诺时，才能建立起最强烈的业务认同感（Buy-in）。 在构建定制化 GPT（Custom GPTs）方面，Azhar 阐述了两个核心要点：一是针对不同员工或特定业务纵向领域（Vertical）设计精准的提示词（Prompt）；二是为 GPT 注入充足的上下文参考文档。以“撰写与解析投资备忘录（Investment Memos）”的定制 GPT 为例，团队为其输入了公司现有的大量投资备忘录语料库，让模型理解标准排版格式、关键评估指标和核心问题结构。通过这一共同构建的过程，业务人员也理解了底层运作逻辑与数据要求，未来便能独立针对新需求开发自己的专属 GPT。 - **实战工作坊**：深入各业务部门开展半日专题研讨，用真实工作流替代抽象理论演示。 - **投资备忘录定制 GPT**：通过挂载历史投资备忘录语料库，使模型能够自动遵循内部格式规范并精准抓取核心数据，同时辅助撰写新的投资备忘录。 --- ### [29:16](https://www.youtube.com/watch?v=SoanrcutwpQ&t=1756s) 警惕智能体（Agentic AI）与即插即用方案的过度炒作 当主持人询问在投资领域哪些 AI 方向存在过度炒作（Overhyped）时，拥有 20 多年行业经验的 Tarun 结合他在展厅现场的见闻直言：当前对“智能体 AI”（Agentic AI）的炒作最为明显。Tarun 回顾了过去二十年技术浪潮的演进历程——从最初的商业智能（BI）与数据仓库，到后来的云计算浪潮，再到数据科学，直至如今的 AI 浪潮。 Tarun 明确表示，Agentic AI 确实具备极高价值和应用前景，但他提醒业务和 IT 部门的管理者保持理性。很多人参加行业展会后，满脑子都是各种新概念，误以为只要采购一家供应商的“即插即用”（Plug-and-play）解决方案，就能一举解决所有数据难题、网络瓶颈或算力挑战。 Tarun 强调，从业者必须时刻专注于技术基本面：如果缺乏合理的系统架构和解决方案设计，任何智能体方案都无法在部门乃至企业级规模实现扩展和稳定运行。针对特定单点用例表现优异的方案，未必具备向全企业推广的扩展性（Scalability），技术实践者必须保持清醒，切忌盲目跟风概念炒作。 - **技术演进四部曲**：商业智能/数据仓库 $\rightarrow$ 云计算 $\rightarrow$ 数据科学 $\rightarrow$ 人工智能（及当前的 Agentic AI）。 - **避坑核心建议**：警惕供应商的“无缝即插即用”营销话术，优先评估方案在底层架构上的健壮性与在企业级（Enterprise）层面的可扩展性。 --- ### [31:50](https://www.youtube.com/watch?v=SoanrcutwpQ&t=1910s) AI 在数据工程与合规治理中的实际应用 在探讨 AI 是否反哺数据治理和数据工程本身时，Dan Welcel 介绍了 SS&C Blue Prism 在这方面的实践。由于公司服务于医疗保健和金融服务等受到高度严格监管的行业，数据合规是重中之重。通过在 AI 网关中引入自动化智能机制，系统可以在任何数据发送给大语言模型之前，自动执行个人身份信息（PII）检测与脱敏过滤，确保客户数据的绝对安全与合规。 同时，Dan 也对技术选型提出了务实的看法。他肯定了 Agent 技术的价值，并提到内部也在积极使用，但用了一个形象的比喻提醒大家：“你不需要总是拿一把电锯去剪一朵花（You don't always need a chainsaw to cut down a flower）。”企业应该将 Agent 视为工具箱中的工具之一，始终坚持“对合适的任务使用合适的工具”（The right tool for the right job），避免在简单场景下过度使用重型复杂工具。 - **敏感数据前置拦截**：在请求进入大语言模型前，通过 AI 网关实现自动化 PII 识别与拦截。 - **适度原则**：避免技术滥用，在需要轻量解决方案的场景中切忌盲目套用复杂的 Agentic 架构。 ### [33:23](https://www.youtube.com/watch?v=SoanrcutwpQ&t=2003s) AI 赋能数据工程与治理：从 Deal Sourcing 到 PII 识别 在工程实践中，为了寻找投资标的（Deal Sourcing）往往需要从多个不同来源摄取数据。在五年前，这类数据摄取与整理流程通常需要耗费数周时间；而如今借助与 AI 协同的现代编程工具，几个小时内即可搞定。高质量数据进入大语言模型（LLM）的门槛大幅降低，使投资机构能够始终处于前沿位置，敏锐捕捉那些原本可能遗漏的小众优质标的。 在数据治理与安全合规方面，AI 的表现同样引人注目。如果让人工去逐一核对成千上万张数据表以排查敏感信息，既枯燥又极易出错。将数据输入 AI 后，模型不仅能理解内容并自动生成精准的表描述，还能深入字段级进行分析，识别包含个人身份信息（PII）的列并建议进行权限遮蔽。这种自动化治理从源头上防范了数据泄露和违规风险，为机构构建起坚实的安全防护网。 - **数据摄取提速**：多源数据整合从过去的数周缩短至几小时。 - **自动化字段治理**：AI 自动标注数据表描述，秒级定位 PII 敏感字段并配置访问隔离。 --- ### [35:38](https://www.youtube.com/watch?v=SoanrcutwpQ&t=2138s) 开发者生态巨变与 DataOps 成本优化 软件与数据工程领域的日常工作方式正在经历范式转移。行业调研显示，随着工程师全面拥抱 GitHub Copilot 等 AI 编程助手，Stack Overflow 与 Reddit 等传统技术问答社区的网络流量出现了断崖式下滑。如今，无论是工程师自身还是管理层，都将使用 Copilot 视作基线生产力工具。这一趋势提升了单兵作战能力，但也引发了软件工程招聘市场的降温与岗位需求减少。 在 DataOps 运维方面，AI 驱动的优化工具同样展现出惊人的投资回报率。例如团队在 Databricks 集群性能与成本调优中引入 AI 工具（如 Sync Computing / 优化平台），直接将计算与运维成本降低了近 40%，证明了下一代 AI 工具在底层基础设施优化上的实际商业价值。 ```mermaid flowchart LR accTitle: AI 时代工程与运维生产力闭环 subgraph Dev["开发与编码"] A["GitHub Copilot / 编程助手"] --> B["代码自动生成与补全"] end subgraph DataOps["数据运维与治理"] C["AI 调优引擎"] --> D["Databricks 成本降低 ~40%"] end subgraph Org["组织级收益"] B --> E["交付周期从周级压缩至小时级"] D --> E end ``` --- ### [37:45](https://www.youtube.com/watch?v=SoanrcutwpQ&t=2265s) 自动化 VC 投资工作流与智能 Agent 落地 评估大语言模型能力演进的指标显示，LLM 能够独立执行任务的时长上限大约每 7 个月翻一番，呈现出指数级增长态势。针对未来投资机构的新场景，风险投资（VC）全流程的端到端自动化成为了最具落地前景的方向之一。 传统的 VC 流程包含找项目、触达、准备投资材料、投后数据跟踪等阶段。在 Agent 架构下，投资团队可以部署自主智能体，根据特定的投资策略（例如偏远细分市场中的特定农业科技公司），全网抓取 Crunchbase、PitchBook、Google 搜索乃至 Hacker News。智能体筛选出符合画像的标的后呈报给投资人；投资人选定目标后，Agent 会以投资人个人的口吻与语调自动起草冷邮件（Cold Email）并协调安排会议。人类只需保留在决策闭环（Human in the loop）中把控核心投资逻辑与人际沟通，从而将精力从繁琐重复的信息搜集解脱出来。 ```mermaid flowchart TD accTitle: AI Agent 驱动的 VC Deal Sourcing 工作流 subgraph Discovery["1. 标的发现与筛选"] A["设定策略画像 (如特定 Agritech)"] --> B["Agent 全网抓取 (PitchBook/Crunchbase/HN)"] B --> C["输出候选标的清单"] end subgraph Interaction["2. 人机协同与触达"] C --> D{"人类投资人挑选 (3家)"} D --> E["Agent 以投资人语调起草邮件"] E --> F["自动安排会议并同步日程"] end subgraph Investment["3. 决策与投后"] F --> G["深度尽调与关系建立 (人类主导)"] end ``` --- ### [40:45](https://www.youtube.com/watch?v=SoanrcutwpQ&t=2445s) 宽客基本面化与超个性化投资产品 AI 的持续渗透将从两个核心维度重塑资产管理业务模式： 1. **基本面量化（Quantamental）的崛起**：量化分析与传统基本面分析的边界进一步消融，基本面分析师依托 AI 的海量非结构化数据处理能力，正在快速向量化深度演进。 2. **超个性化（Hyper-personalized）投资产品**：以往只有超高净值或机构客户才能享受定制化投资组合。随着底层技术成本下降与生成能力的普惠化，资管机构能够轻松针对个人客户的偏好，剔除特定证券、打包专属资产，一键构建超个性化的定制基金与投资策略。 在决策速度上，未来的资产管理将实现决策周期的极致压缩。AI 与智能体负责提供速度与规模，人类则专注于最擅长的人性洞察与信任构建，形成流畅的人机协作决策体系。 --- ### [42:32](https://www.youtube.com/watch?v=SoanrcutwpQ&t=2552s) 落地阻力与推进策略：合规信任、AI 素养与高层共识 在企业内部推行 AI 工具时，最大的摩擦点往往集中在隐私保护与合规风控部门。将数据接入 LLM 之前，必须在第一天就建立起极其清晰的数据链路追踪（Data Line of Sight）与细粒度权限控制体系。要消除合规部门的疑虑，关键在于建立深层次的跨部门信任与制度保障。 此外，变革管理（Change Management）与全员 AI 素养（AI Literacy）是决定落地成败的另一大支柱。技术迭代速度极快，单靠自下而上的探索往往难以跟上节奏；通过高层自上而下（Top-down）确立明确战略，将使用 AI 提升效能、融入产品写入组织目标，能有效凝聚共识并扫除落地障碍。 ## Notable quotes > [03:22](https://www.youtube.com/watch?v=SoanrcutwpQ&t=202s) 如果你不使用 AI，你就会被时代甩在身后。 — Azhar Dewji > [03:22](https://www.youtube.com/watch?v=SoanrcutwpQ&t=202s) 单靠开箱即用的 ChatGPT 固然不错，但只有将其针对特定业务场景定制时，才能释放出成倍的价值。 — Azhar Dewji > [19:02](https://www.youtube.com/watch?v=SoanrcutwpQ&t=1142s) 你可以引入 OpenAI 的基础模型，但它们主要是基于公开互联网数据训练的。而在投资界，我们拥有大量的私有数据，而且这些数据涉及严格的权限授权。你不能直接把公共模型拿来即插即用。 — Tarun Sood > [22:30](https://www.youtube.com/watch?v=SoanrcutwpQ&t=1350s) 就像在城市中建造摩天大楼一样，若没有通畅的电力、供水和排水等基础公共市政设施，摩天大楼就根本无法正常运作。同样地，大语言模型想要真正发挥威力，企业必须先打通底层数据管道并完成数据清洗净化。 — Tarun Sood > [26:18](https://www.youtube.com/watch?v=SoanrcutwpQ&t=1578s) 当员工亲眼看到 AI 在真实业务场景中发挥作用，而不是听到抽象的效率收益承诺时，才能建立起最强烈的业务认同感。 — Azhar Dewji > [29:16](https://www.youtube.com/watch?v=SoanrcutwpQ&t=1756s) 很多人参加行业展会后，满脑子都是各种新概念，误以为只要采购一家供应商的‘即插即用’解决方案，就能一举解决所有数据难题、网络瓶颈或算力挑战。 — Tarun Sood > [31:50](https://www.youtube.com/watch?v=SoanrcutwpQ&t=1910s) 你不需要总是拿一把电锯去剪一朵花（You don't always need a chainsaw to cut down a flower）。 — Dan Welcel > [31:50](https://www.youtube.com/watch?v=SoanrcutwpQ&t=1910s) 始终坚持对合适的任务使用合适的工具（The right tool for the right job），避免在简单场景下过度使用重型复杂工具。 — Dan Welcel ## Mentioned Adam Nahari, Tarun Sood, Azhar Dewji, Dan Welcel, Pinegrove Venture Partners, Sagard, American Century Investments, SS&C Blue Prism, Shopify, OpenAI, Anthropic, Perplexity, Bloomberg, FactSet, Databricks, Sync Computing, GitHub, GitHub Copilot, Collibra, ChatGPT Enterprise, Crunchbase, PitchBook, Hacker News, Reddit, Stack Overflow, Stripe ## Video description Adam Nahari, Partner, Chief Data and AI Officer, Pinegrove Venture Partners Tarun Sood, Chief Data Officer, American Century Investments Azhar Dewji, Data Infrastructure Lead, Sagard Dan Welcel, Director, Solution Consulting, SS&C Blue Prism Abstracts This session will explore how AI is transforming asset management by improving data analysis, optimizing portfolio strategies, and enhancing risk management. Panelists will discuss real-world examples of AI-driven tools that help asset managers gain deeper insights into market trends, automate decision-making, and improve operational efficiency. Learn how AI is reshaping investment strategies and portfolio management through predictive analytics and automated processes. The panel will also cover best practices for integrating AI into existing workflows and the challenges faced when balancing human expertise with AI-driven insights. [Source](https://www.youtube.com/watch?v=SoanrcutwpQ)

[3:22] Azhar Dewji: 如果你不使用 AI，你就会被时代甩在身后。 [3:22] Azhar Dewji: 单靠开箱即用的 ChatGPT 固然不错，但只有将其针对特定业务场景定制时，才能释放出成倍的价值。 [19:02] Tarun Sood: 你可以引入 OpenAI 的基础模型，但它们主要是基于公开互联网数据训练的。而在投资界，我们拥有大量的私有数据，而且这些数据涉及严格的权限授权。你不能直接把公共模型拿来即插即用。 [22:30] Tarun Sood: 就像在城市中建造摩天大楼一样，若没有通畅的电力、供水和排水等基础公共市政设施，摩天大楼就根本无法正常运作。同样地，大语言模型想要真正发挥威力，企业必须先打通底层数据管道并完成数据清洗净化。 [26:18] Azhar Dewji: 当员工亲眼看到 AI 在真实业务场景中发挥作用，而不是听到抽象的效率收益承诺时，才能建立起最强烈的业务认同感。 [29:16] Tarun Sood: 很多人参加行业展会后，满脑子都是各种新概念，误以为只要采购一家供应商的‘即插即用’解决方案，就能一举解决所有数据难题、网络瓶颈或算力挑战。 [31:50] Dan Welcel: 你不需要总是拿一把电锯去剪一朵花（You don't always need a chainsaw to cut down a flower）。 [31:50] Dan Welcel: 始终坚持对合适的任务使用合适的工具（The right tool for the right job），避免在简单场景下过度使用重型复杂工具。

[0:04] Adam Nahari: the tip of the sphere [8:06] Dan Welcel: eating your own dog food [11:33] Azhar Dewji: Everyone knows that you need to use AI, but tactically using it at small startups can be a little bit intimidating. [19:02] Tarun Sood: plug and play with a public model [24:00] Tarun Sood: Think of like a city, right? And if you want to have a skyscraper in the city, and you want to build multiple skyscrapers, if you don't have sound utilities—if you don't have good electricity, if you don't have good water and all of that and good plumbing—then that skyscraper is not going to be very successful... Similarly, if you're going to have language models and really get the value out of those models, you still need that sound infrastructure in order for that plumbing and the data pipelines to be sanitized and clean. [26:18] Azhar Dewji: boots on the ground [31:50] Dan Welcel: AI agentic agents are great, and there's certainly a place and we're using them internally, but you don't always need a chainsaw to, you know, cut down a flower, right? So, it's kind of a tool in the tool bag, and we want to make sure we're using the right tool for the right job. [41:47] Dan Welcel: AI-assisted decision makers

[0:04] Hi everybody. So, uh, where I was hoping [0:06] that we could focus our conversation [0:09] today was in really understanding how we [0:11] as a group have thought about leveraging [0:14] AI and maybe data science more broadly [0:17] in order to enhance investment [0:19] decisionmaking within our respective [0:22] organizations and then also maybe on the [0:25] back office as well how we've thought [0:27] about leveraging AI in order to make our [0:30] organizations more efficient. Um, so [0:33] without further ado, maybe we can start [0:34] with a quick round of introductions. Do [0:37] you want to get us started? [0:39] >> Yeah, sure. Hey everyone, my name is [0:41] Azar Devji. I work at Cigard. Cigard is [0:45] a global asset management company uh [0:47] specializing in alternative asset [0:49] classes. So we invest in venture [0:51] capital, private equity, equity, sorry, [0:54] private credit, uh real estate, [0:55] healthcare, and a few other lines. Prior [0:58] to my time at Cigard, I've worked at a [1:00] number of uh technology companies [1:02] building out data platforms from the [1:03] ground up. Most recently, I was at [1:05] Shopify working on the company's data [1:08] governance strategy. [1:11] Well, hi everyone. My name is Tun Sud [1:13] and I'm the chief data and uh AI officer [1:16] at American Century Investments. So, we [1:18] American Century is a midsize asset [1:21] manager based in Kansas City. We also [1:23] have offices all over US and uh [1:26] globally. Um and in my remmit I look at [1:30] data and AI both from an offense and [1:33] defense perspective. If you think about [1:35] offense mostly on the you know use cases [1:38] focus on Revjan or efficiency and on the [1:42] defense side more mostly on the [1:43] governance and you know we are in a [1:45] regulator environment. So making sure [1:48] that regulators are happy with us. So [1:50] that's a little bit about me. More than [1:51] happy uh looking forward to the panel. [1:54] >> Perfect. Hi everyone. I'm Dan Welsaw. [1:56] I'm director of solution consulting at [1:58] SSNC Blue Prism. Been with the company a [2:01] little over six and a half years. So uh [2:03] in my role I typically work on the sales [2:05] side helping our customers understand [2:07] where to uh implement what use cases um [2:10] within you know automation agents. [2:13] >> And I'm Adam Nihari. I work at Pinerrow [2:15] Venture Partners. Pinrow is a venture [2:17] capital and growth um investment [2:20] platform with a credit secondaries, fund [2:24] of funds and co-investments [2:27] practices. We're really focused on [2:29] leveraging data science and AI across [2:32] the investment life cycle, thesis [2:35] development, sourcing, due diligence, [2:37] and Porco value creation. [2:40] um bringing data science and AI really [2:43] to the tip of the sphere of how we make [2:46] better investment decisions. Uh so I I [2:49] thought we could start our conversation [2:50] today maybe with each um person on the [2:54] panel quickly talking about how you're [2:57] thinking of leveraging AI internally [3:00] within your organization. what tools or [3:04] platforms you've already built or um [3:08] adopted internally. What do you what [3:10] have you seen in terms of um [3:15] where have you been able to generate the [3:17] most ROI? So, uh maybe we can start with [3:21] you Azar. [3:22] >> Sure thing. Um so, at Cigard, the way [3:25] that we approach AI is from the lens of [3:28] both pushing it from the bottom up and a [3:30] top down perspective. And what I mean by [3:33] that is the bottom up. Every employee [3:36] across the guard is very much encouraged [3:38] to utilize AI in their day-to-day jobs. [3:41] We kind of view it as transformational [3:42] as everyone here does. And if you're not [3:44] using AI, you're getting left behind. So [3:46] for us, that's you know a standard chat [3:48] GBT enterprise license. It's a [3:51] connection to their email, all their [3:53] documents so that they can use chat GBT [3:55] to get insights from their information. [3:58] And then we've also actually gone a [3:59] little bit further to really empower [4:01] them and that's by doing targeted boot [4:03] camps for each team to help them build [4:06] custom GPTs. So chat TV by themselves or [4:10] by itself great tool but when you really [4:13] lever uh or leverage a custom GPT for [4:15] specific use cases then you really start [4:17] seeing additional value. And so on the [4:20] bottoms up approach, we're really [4:21] looking at efficiency gains of hopefully [4:23] like, you know, 5 to 10% per employee, [4:26] which is meaningful, but we want to see [4:28] a much more transformational approach to [4:30] using AI across our business, which is [4:33] where the top-down initiatives are going [4:35] to play a huge role. So an example of [4:37] that is finding out or finding areas [4:40] where right now processes are a little [4:42] bit slow, they're inefficient, or they [4:44] could be redone entirely. A really great [4:46] example of that is in document [4:48] extraction across our funds. We have [4:52] LPAs, we have uh side letters, we have [4:54] subscription agreements, we have all [4:56] this information that's stored in these [4:58] hundreds of pages long documents. And [5:00] each time a specific um value is needed [5:04] from the document, whether it's the [5:06] yearly audit requirements or how we [5:08] treat operational expenses or anything [5:10] along those lines, it involves an [5:12] associate going to the document. First [5:14] of all, finding it, crawling through [5:16] maybe a hundred pages, finding the [5:18] specific piece of information, reporting [5:20] that back. This happens multiple times [5:22] per document, per fund. So, what we've [5:25] done is built out document extraction [5:27] pipelines where for each of these [5:29] documents, we ingest them one time. We [5:31] get all the information in objects that [5:33] anyone can reference and anyone can [5:35] trust. And it really smooths out the [5:37] process for finding, retrieving [5:39] information, and making sure it's [5:41] trusted across the business. [5:46] uh so I I can go next. So the way we are [5:49] looking at uh leveraging AI is uh [5:53] three-fold right. So I would say we've [5:56] bucketed our use cases in these three [5:59] buckets. Number one is uh what we call [6:02] persona based AI. So if you think about [6:04] any role within our firm how can we [6:08] optimize it? So if it's a financial [6:10] analyst front end uh you know [6:13] fundamental research. So how can we [6:15] optimize that um that role and how can [6:18] we give them capabilities whether [6:20] they're custom models or co-pilots um so [6:24] we are enabling that uh similarly if [6:26] it's a financial analyst or a software [6:29] engineer how can we enable GitHub [6:31] co-pilot or regular co-pilots for them [6:34] so that codew writing becomes easier and [6:37] also within my teams we are we are [6:39] looking at uh tools like Cal Libra and [6:42] their their AI agents and others [6:45] to streamline data governance for [6:47] example. So those are some uh some ways [6:50] on the persona based AI we are looking [6:53] at leveraging you know the geni models. [6:57] Uh the second use case we think about is [7:00] looking at any process as a backend [7:03] process or uh you know front-end process [7:06] and how can we re-engineer with the [7:08] power of AI. So if you think about um [7:10] you know u document extraction one of [7:13] the things that Ozar mentioned we are [7:15] looking into that very closely and uh we [7:17] are prototyping with that uh we're also [7:20] leveraging u you know client reporting [7:22] is one area where we are looking at [7:24] conversational reporting conversational [7:26] AI so that's something that we are [7:29] actively looking at uh and then the [7:32] third is uh of course custom use cases [7:35] so if you think about alpha generation [7:38] if you have a custom hypothesis then we [7:41] go and leverage either a traditional a [7:43] IML model or a genai model and how can [7:46] we leverage that uh to and drive not [7:50] only innovation but also an offer. Uh so [7:53] those are those are three buckets I [7:55] would say and that uh we are actively uh [7:58] we have several use cases in production [8:01] but uh we are actively looking at them [8:03] in these three use cases. [8:06] >> Nice. Yeah, and that's kind of a fun and [8:08] interesting question for me to answer at [8:10] SSNC Blue Prism because not only do we [8:13] provide software and services, we get to [8:15] use our software internally, right? So, [8:18] it gives us the ability to, you know, [8:20] stay agile. It allows us to kind of [8:22] test, you know, what we're using, pivot [8:25] if we need to. Um, so kind of one of [8:27] those AI tools that we get to use is our [8:29] own AI powered um automation agents, [8:32] right? So we're combining AI and um [8:35] intelligent automation to improve um you [8:39] know middle office back office [8:40] operations right so we service our [8:42] clients using our own software um you [8:46] know in um customerf facing teams we're [8:49] deploying LLMs um generative AI as well [8:53] where we can actually use our um [8:55] customer communication agents to get [8:58] back to customers much quicker and [9:00] efficiently than we used to in the past. [9:02] So, you know, uh, understanding an [9:04] inbound inquiry, um, having an agent [9:07] then go query multiple systems, um, [9:10] gather data, build an LLM, run a prompt, [9:13] take that output, and then send a very [9:15] personalized and targeted, you know, um, [9:18] correspondence back, right? So, um, even [9:21] that one agents, we're already seeing [9:23] 800,000 hours returned back to our [9:25] business. So, you know, it's a it's a [9:26] definite top-down approach for us as [9:29] well. are encouraged to use these tools [9:31] both you know internally to make our [9:33] jobs more efficient but you know how do [9:34] we make our customers um happier as [9:36] well. [9:39] >> Uh a lot of the examples that we [9:42] mentioned um [9:46] didn't yet talk about insight generation [9:50] that directly contributes to investment [9:52] decisionmaking. So maybe I'll just add [9:54] an example uh of that type. We were [9:58] recently looking at a deal where we were [10:01] evaluating [10:02] a payments provider [10:06] against several competitors and we were [10:09] able to get our hands on over 300 expert [10:13] interviews. That's thousands and [10:16] thousands of pages of text. [10:18] Historically, an analyst would have to [10:22] go through all of those expert [10:24] interviews. that would probably take [10:26] them three weeks to do and try and [10:30] generate insight out of those expert [10:32] interviews. But now with LLMs, we've [10:36] built custom LLM pipelines that allow us [10:40] to ingest all of that unstructured [10:43] information [10:45] um and generate and answer specific [10:48] questions that make a difference to our [10:54] underwrite of these investment [10:57] opportunities. So that's just an example [11:00] of how we're act using AI to actually [11:02] make better investment decisions. And we [11:05] found that um the LLM pipeline [11:08] customization is an important piece of [11:11] this. So for example, when you're [11:12] looking at e expert interviews um meta [11:15] data such as who are the speakers, when [11:19] did the conversation take place, um [11:23] those kinds of things are really [11:25] important. Um, and we've also looked at [11:28] document ingestion and uh making sure [11:33] that we're building an institutional [11:35] store of knowledge with all of the [11:38] information and all of the deals and [11:39] companies and funds and GPS and LPs that [11:43] we've ever looked at. Um, so I I think [11:47] another area that we haven't talked [11:48] about yet which will be interesting to [11:51] explore is portfolio value creation. And [11:54] I know Azar that you've done a lot of [11:56] work there. Maybe you could uh say a few [11:59] words about that. [12:00] >> Yeah, for sure. So at Sigard, we [12:03] obviously have a branch in venture [12:04] capital and it's a it's a strong mandate [12:07] for us to really provide true value for [12:10] our portfolio companies and using AI [12:13] there is tremendous opportunity to [12:15] provide that value. So we've done that [12:17] in a few ways to date. One of them is by [12:21] driving these workshops where we invite [12:23] a number of our portfolio companies to [12:25] attend virtually and we really walk them [12:27] through um a like a half-day session [12:30] where we ideulate on different ideas. Uh [12:33] we prototype them using various AI tools [12:36] and then we actually build something [12:38] again using AI tools. These sessions are [12:40] aimed at both engineers, product [12:42] managers and any other builders that are [12:44] interested to attend uh at the portfolio [12:47] companies. And the real goal there is to [12:49] one upskill a lot of those employees and [12:53] two help them understand how AI can [12:56] supplement them in their day-to-day life [12:58] and their day-to-day job. As of as of [13:01] today, everyone knows that you need to [13:02] use AI, but tactically using it at small [13:05] startups can be a little bit [13:07] intimidating because you want to move [13:09] fast, but you don't necessarily know how [13:11] to do that in all situations. So [13:13] providing boot camps like this is a [13:15] great opportunity to get them to [13:18] understand how they can leverage that on [13:19] their day-to-day. Another example is [13:22] working more one-to-one with portfolio [13:24] companies where we talk to them, we uh [13:27] understand their current business, which [13:30] we obviously already do, but specific [13:32] pain points that they're experiencing on [13:34] a day-to-day. So what is um a problem or [13:38] a situation where right now you are [13:40] spending a large portion of your time [13:42] and there's an opportunity to maybe make [13:44] that a little bit better. A key example [13:46] of that is one of our portfolio [13:48] companies. They are in the insurance [13:50] industry and underwriting for insurance [13:53] is a process that can take an extremely [13:55] long time but has a lot of uh [13:58] opportunities to be automated to make it [14:00] a lot faster. So we sat down with them, [14:03] we understood the use case and we uh [14:06] provided them with specific examples of [14:08] where AI could be layered into their [14:10] pipelines to make this process a lot [14:12] faster, a lot smoother and hopefully [14:15] more accurate as well. So working with [14:18] that team, we managed to get their [14:19] underwriting process down from multiple [14:21] hours down to five minutes, which is a [14:24] tremendous game for gain for them [14:25] because now they can operate a lot more [14:27] quickly. and workshops like that are [14:30] something that we're trying to do as [14:31] often as possible to make sure that [14:33] we're bringing a lot of value to our [14:35] portfolio companies. [14:37] >> It sounds like you provide a lot of [14:39] support in terms of uh strategy and [14:43] maybe planning. Do you ever support your [14:47] portfolio companies with hands-on [14:50] keyboard uh tool development or [14:54] execution or do you um always focus more [14:58] on the strategy? [14:59] >> Yeah, to date we haven't done too much [15:02] hands-on keyboard. I think our goal is [15:05] to empower them as much as possible and [15:08] we uh we obviously utilize our own [15:11] industry like AI experts that we have to [15:14] identify their specific problems and uh [15:17] come to the conclusion of what the [15:19] solution might be but at the end of the [15:21] day we want them to build those [15:23] solutions themselves because they're [15:24] going to be operating it. So there might [15:26] be an opportunity in the future for us [15:28] to be a little bit more involved in that [15:29] process, but as of now it's more about [15:31] empowering them and getting them to [15:33] build the products. [15:34] >> What what about um supporting them in [15:38] terms of uh hiring, finding the right [15:41] people, helping them uh design a job [15:45] description, what is the right think [15:47] through what is the right profile for a [15:49] job. Do you get involved in those types [15:51] of areas? [15:52] >> We haven't dived into that yet. I think [15:54] that's an area that we're definitely [15:55] interested in pursuing because at an [15:58] organizational at our scale, we have a [16:00] lot of experience and understanding of [16:02] how to form some of those job [16:04] descriptions, how to source the right [16:05] talents. So, we haven't got there today, [16:07] but I think it's definitely an [16:09] opportunity for us in the future. [16:12] >> Duran, maybe you can tell us a bit more [16:15] about um we've talked about Porco value [16:17] creation. Maybe you can talk to us a bit [16:20] about using AI to enhance due diligence [16:23] and investment research. [16:26] >> Sure. Sure. Um so so you know if you [16:28] think about asset management right um [16:31] there's definitely a lot of uh data that [16:34] is required uh whether it is traditional [16:36] data that we are used to uh which is [16:39] securities portfolios benchmark [16:41] composits I can go on and on right um [16:44] and uh then you know leveraging that on [16:47] your traditional models that you've been [16:49] using for a long time but now with the [16:52] power of AI and co-pilots and LLMs I [16:55] think One of the things that is changing [16:57] is you can take a lot of alternative [16:59] data and leverage that and synthesize [17:02] that much faster. So one of the things [17:05] that you're streamlining is the entire [17:07] data aggregation cycle itself. So how [17:10] can you smoothen that cycle and how can [17:12] you fashion the cycle right? So number [17:14] one that is an area that we're focusing [17:16] on a lot which is how do we synthesize [17:19] different data uh and get that to our uh [17:23] you know analysts and portfolio managers [17:25] faster. So that's one area. The second [17:29] is you know we are always making this [17:32] decision on buy versus build which is uh [17:36] whether we want to build a custom [17:38] solution or we want to acquire a [17:40] solution from one of the vendors that [17:42] are out there. And uh the portfolio [17:45] analysis is an area that uh there are a [17:48] lot of opportunities out there. Uh I [17:50] mean if you think about chat GBD has [17:54] come with an offering there. Perplexity [17:56] has an offering. Uh you also have [17:59] offerings from uh Anthropic and others. [18:02] So this is like one area that you have [18:04] to not only if you think about build [18:06] versus buy and you want to buy then you [18:09] have to evaluate all of these different [18:11] technologies on which is probably going [18:13] to be the right technology if you want [18:14] to go down the buy side and not only [18:17] that there are a lot of other vendors uh [18:20] who you've been working for and with for [18:22] a long time who are also creating their [18:25] custom LLMs to with on the data that [18:28] they have uh already on the air that [18:31] they already providing to you. So you [18:33] know the Bloombergs of the world, the [18:36] fax sets of the world. So they're [18:38] creating their own LMS and they're [18:40] offering that as well. So so as a you [18:43] know as a co I would sit down and say [18:45] okay what are the different technologies [18:48] that are out there not only look at the [18:50] innovation that's happening on natively [18:53] and then look at the offerings that the [18:55] vendors are providing and see bake and [18:57] bake them off right and see what is [18:59] going to the best thing for us. [19:02] the the first use case that you [19:04] described around synthesizing [19:07] alternative data. I'd imagine that there [19:11] would maybe be two parts to that. The [19:13] first is aggregating all of these [19:16] different streams of data, harmonizing [19:18] them. So if uh such that if you're [19:22] interested in information about a [19:24] specific company, I'm just going to make [19:26] a company up Stripe and you want to see [19:29] all of the information about that [19:30] specific company across sources, you can [19:34] easily do that. So that that seems like [19:36] it would be a challenge to me. And then [19:39] maybe the second thing is even once [19:41] you've had all of this information in [19:43] one place and uh harmonized [19:46] um generating insight across all of [19:50] these different sources understanding [19:53] limitations [19:54] and areas of potential for each [19:57] different source um so that when a user [20:01] comes and uses this system that you're [20:02] building and asks a question they can [20:05] get a reliable answer. H how have you [20:08] thought about those challenges? [20:11] >> Yeah, I mean you know that's an [20:12] interesting question and I'll add to [20:14] that mix a little bit. So if you think [20:17] about uh a foundational model and the [20:19] technology that they are providing. So [20:21] let's take for example if you again I'm [20:23] just uh hypothesizing here. So if you [20:26] take like for example open eyes models [20:28] right uh and you look at them and you [20:30] say okay they're trained a lot on public [20:33] data but in the investment world we have [20:36] a lot of like private data too and uh [20:39] like you know and that you need to have [20:41] entitlements on that data then it [20:43] becomes complex because you cannot like [20:46] so now you let's say you bring in a [20:48] model from open AI you have to then [20:50] train that internally on the data that [20:53] you have entitlements to and then you [20:57] have and then you of course the public [21:00] side of it open AAI or any of these [21:02] foundational models can probably give [21:04] you but then you have to customize it a [21:06] little bit so that it can be trained on [21:07] the data that you have entitlements to. [21:10] So that's like where the nuances come in [21:12] a little bit that you can't just plug [21:13] and play with a public model. Um but uh [21:16] because public model that is available [21:18] will be all internet and then there are [21:20] also boundaries on the internet on what [21:23] data is available to the underlying [21:25] foundational model which is it'll it's [21:29] still a black box to a bit right so I [21:31] think those are the nuances that we need [21:33] to think through on looking at your [21:35] portfolios and then seeing okay how how [21:37] do we train this on the data that we [21:39] have entitlements on and then who can [21:42] analyze that data [21:45] Some people might think that um since [21:48] LLMs are able to [21:51] ingest unstructured data that might make [21:54] the job of having to clean up and [21:58] standardize and aggregate all of these [22:01] different data streams maybe that's work [22:04] that we don't have to do anymore now [22:06] that we have Gen AI. So Dan, I I'd be [22:11] interested in your perspective on um [22:16] how much work is required at the [22:19] infrastructure level in order to prepare [22:23] our firms for AI so that we're able to [22:26] get the most use out of them. [22:28] >> Yeah, excuse me. I think that's a great [22:30] question. I think it's really [22:31] foundational. I mean, the [22:32] infrastructure's got to be there and [22:34] centralized to support the storing, the [22:36] collection, the use of the data that [22:38] we're going to use in these um these AI [22:41] systems and in applications, right? Um I [22:44] think what we're starting to see across [22:46] our customer base too is like you know [22:48] we we get some of this external [22:49] validation that when things are disparit [22:52] when you do have replication or [22:54] duplication of data it can hinder you [22:57] know the speed at which you can innovate [22:59] and and really bring these AI solutions [23:01] um you know um to fruition. I think [23:04] another kind of important aspect that [23:06] we're starting to see and certainly [23:07] we've seen it internally is you know the [23:10] governance model right how do you [23:12] control the access to the data um how do [23:15] you control the the auditability the [23:17] traceability right and so I think early [23:18] on in our own adoption of AI we just [23:22] didn't find a great solution out there [23:24] so we started to develop our own um AI [23:27] gateway right that we're using [23:28] internally now we've brought it to [23:30] market um but it essentially allows us [23:33] you know the the the ability to have [23:35] this governance model, right? Again, [23:37] it's we have a top- down approach to use [23:39] AI, right? We're building it into our [23:42] products and our solutions, but without [23:44] that kind of base governance model, you [23:46] know, it doesn't give our internal [23:49] stakeholders, our our customers kind of [23:51] the the safety net that they want, the [23:53] the trust. So, I think it's really [23:55] foundational to have the that data [23:57] infrastructure there, but I think the [23:58] governance model as well. [24:00] >> Run, do you have any thoughts on that [24:02] question? Yeah. Yeah. I mean I think to [24:04] me the way I feel it is you know we have [24:07] this old adage right garbage in garbage [24:10] out and I think this is getting more [24:12] complicated in the world of uh AI. So [24:16] because of a lot of unstructured data [24:17] that of course our analysts and others [24:20] will um you know produce on clients and [24:23] companies and now we want to integrate [24:26] that with uh the structured data that we [24:28] have been collecting making sure that [24:31] that data is sanitized and cleansed uh [24:33] and it is in you know one call it one [24:37] place uh you can name it whatever you [24:39] want data lake data warehouse whatever [24:42] uh suits your um you know uh what [24:45] whatever terminology you want to choose [24:47] but I feel like collocation of data if [24:50] you can do that to a certain extent and [24:52] have sound governance [24:54] um and metadata around it then finding [24:57] the data becomes very easy right and I [24:59] think that is one area where even now [25:02] even with like of course you can use [25:04] leverage genai for that as well like [25:07] defining metadata that's one use case [25:09] that we are looking at as to how can we [25:12] define all the metadata that we are [25:14] storing in powered lakes uh and how can [25:17] we use power of AI uh to do that such [25:20] that our data products are more complete [25:22] uh with corresponding metadata and an [25:25] analogy that I always use with my peers [25:28] and uh business uh counterparts on a [25:31] regular basis is think of like a city [25:34] right and if you want to start a uh have [25:36] a skyscraper in the city uh and you want [25:39] to build multiple skyscrapers if you [25:42] don't have sound utilities like if you [25:45] don't have good electricity, if you [25:46] don't have good water and all of that [25:48] and good plumbing then that size scrap [25:51] you know skyscraper is not going to be [25:52] very successful if uh the basic [25:55] communities are not there. So similarly [25:57] I feel like you know if you're going to [25:58] have like language models uh and really [26:02] get the value out of those models you [26:04] still need that sound infrastructure in [26:07] order for you like that plumbing and the [26:10] data pipelines and everything are [26:12] sanitized and clean it's cleansed so [26:14] that you can really leverage the power [26:15] of the LM models. [26:18] Azar, you it sounds like um [26:23] doing work on AI literacy was a focus [26:28] for you guys. Um could you tell us a bit [26:31] more about how you've thought about [26:33] that? You also mentioned custom GPTs. I [26:36] think it'll be interesting to learn more [26:38] about um how you've thought about [26:41] training folks to develop those. if you [26:43] can give us a few examples of custom [26:46] GPTs that your teams have gotten value [26:49] out of, I think that could be [26:50] interesting as well. [26:53] >> Yeah, so in terms of um just the [26:55] literacy in general, I think there's [26:58] there's a ton of resources that already [27:00] exist out, you know, on the internet [27:02] today for how to get started with a lot [27:03] of these tools, but that can still be [27:06] daunting in its own right, especially in [27:08] a world where there's, you know, dozens [27:10] of models at this point. um and you know [27:13] which one do you use for which use case. [27:14] So trying to be a little bit more [27:17] explicit at the start or holding the [27:19] hands of uh employees who maybe didn't [27:21] have as much uh proficiency in AI I [27:24] think was uh something that was really [27:25] helpful for us to get things off the [27:27] ground. And in terms of building the [27:29] custom GPTs and also this education a [27:32] lot of that was driven by boots on the [27:34] ground um workshops with individual [27:37] teams. So spending an afternoon with [27:39] them discussing how uh leveraging LLMs [27:43] or chat bots can help them with their [27:45] jobs today. And I think that's where you [27:48] start seeing the uh biggest buy in when [27:51] it's not just some hypothetical um [27:54] efficiency gain that maybe they can get [27:56] if they know how to properly use some [27:58] some chatbot or some some AI tool, but [28:02] seeing it in action um and having that [28:04] presented to them. So with the GPTs [28:07] specifically, obviously you want to [28:09] layer in the right prompt based on what [28:11] they're trying to achieve. Uh and that's [28:13] going to be very specific based on the [28:15] individual uh employee or the the [28:18] individual um vertical that they're [28:20] tackling. But you also want to leverage [28:22] or layer in the individual or certain [28:25] documents that are going to provide the [28:26] right context to the GPT to answer [28:29] specific questions. So an example is um [28:32] if we wanted to build a GPT for uh for [28:37] parsing in or not parsing but uh reading [28:39] investment memos and getting the right [28:41] data from them and uh potentially [28:43] writing new investment memos. Obviously, [28:45] you want to provide it with a decent [28:47] corpus of existing investment memos so [28:49] that the GPT understands what's the [28:52] right format, what are the key pieces of [28:54] information that are uh important to ask [28:56] for. And then it's working with the [28:58] employee to make sure that they [29:00] understand generally what's going on, [29:03] why it's important, and why certain data [29:06] is needed to build that GPT. So, [29:08] tomorrow they can build their own custom [29:10] GPT for the next use case. [29:16] Are there areas where you guys feel AI [29:19] is overhyped [29:21] specifically [29:23] in terms of getting value out of it in [29:27] the investment context? Maybe we can [29:30] start with you Trent [29:33] over hyped. [29:35] I mean, I just walked the tech expo [29:37] floor and I saw the number of companies [29:40] that have agent uh AI uh offerings. So, [29:45] I would probably say that uh you know uh [29:48] I mean I've been in this industry for [29:51] I'll age myself a little bit. I've been [29:53] in this industry for more than 20 years [29:55] and um I feel like we had a wave of [29:59] initially BI warehousing. Then we moved [30:01] to the cloud computing wave and the data [30:03] science wave and now we are in this AI [30:06] wave right uh and I feel like you know [30:09] agentic AI is the new you know buzzword [30:11] and also a lot that is of course there's [30:14] definitely a lot of merit to it in this [30:16] conference itself I've uh attended [30:18] multiple talks um and there's definitely [30:21] a lot of value uh with agentic AI but uh [30:25] you know as I feel like you know there [30:28] are a lot of you know my peers in [30:30] business and in IT who go and attend [30:33] these conferences and then we come back [30:36] with with a lot of ideas and feel like [30:40] uh we will have a vendor solution come [30:41] in and that will plug and play [30:44] seamlessly and I think that is the hype [30:47] sometimes that we have to be careful of [30:50] because you still have to you know focus [30:53] on the fundamentals right if you don't [30:55] have a sound architecture if you don't [30:56] have a sound uh solution design that [31:00] solution is not going to scale and work [31:03] uh because uh if you just put like you [31:05] know if you just put an agent to AI [31:07] architecture let's call it and say oh [31:09] yeah I'll just bring that in now and [31:11] everything it it'll solve all my data [31:14] problems or all my network problems or [31:16] compute problems that's not going to be [31:18] the case right so so I think that's [31:20] where uh we as practitioners have to be [31:24] mindful of on yes a solution may work [31:28] really well for a use case but how [31:32] scalable is that solution for the [31:34] enterprise? How scalable is a solution [31:36] for uh not only enterprise but for a [31:39] department those are things that uh [31:41] practitioners have to be careful of and [31:43] not get caught up in the hype of uh you [31:46] know and a lot of architecture out [31:48] there. [31:50] Dan, have you seen AI help not just on [31:54] the investment side but also in [31:57] improving data governance [31:59] or engineering itself? [32:03] >> Um, yeah, I think so. I mean, um, sorry, [32:08] repeat the question again. But [32:10] >> using AI for, uh, engineering [32:15] or for data governance purposes? [32:18] >> Yeah, for sure. I mean, and that's the [32:20] whole reason behind, you know, we're [32:21] we're one of the vendors out there on [32:23] the on the expo floor, right? And the [32:25] data governance aspect is is huge for us [32:28] because we didn't have that capability [32:30] when we were initially looking at [32:33] solutions, right? So, um I think the the [32:36] AI gateway that we we have now has [32:39] allowed us to really, you know, adhere [32:41] to the right governance policy. Um you [32:44] know, we can do things like PII [32:46] detection, right? before anything gets [32:48] sent to an LLM, making sure we're [32:50] protecting our customers data in the [32:52] right way because we we do serve highly [32:54] regulated industries and healthcare and [32:56] financial services. So, we have to kind [32:57] of think about what is the right tool [33:00] for the right use case, right? So, AI [33:02] agentic agents are great um and they're [33:05] certainly a place and we're using them [33:06] internally, but you don't always need a [33:09] chainsaw to, you know, cut down a [33:10] flower, right? So, it's kind of a tool [33:12] in the tool bag and we want to make sure [33:13] we're using the right tool for the right [33:15] job. I can I can touch on that a little [33:17] bit as well. Um I think uh data [33:19] governance and the engineering aspect [33:21] for AI especially in the financial [33:23] services world there's a lot of [33:25] potential. I think um specifically on [33:28] the engineering side um there's often [33:31] times where you need to ingest sources [33:33] from or multiple different data sources [33:35] to uh get the right information if you [33:37] want to source a deal. If you want to [33:39] find a next company to invest in or [33:41] anything else really, that process, you [33:44] know, five years ago could take anywhere [33:46] from a week to a few weeks to, you know, [33:49] who knows how long. Today, if we're [33:52] leveraging the current uh coding tools [33:54] that exist that are uh working with AI, [33:56] you can do that in a few hours. And so [33:58] the barrier to entry and the barrier to [34:02] good data and data that's can then be [34:05] fed into LLMs is a lot lower than it was [34:08] a few years ago. And so if you're [34:11] leveraging that, then you're really [34:12] presenting or giving yourself the [34:14] opportunity as an organization to be at [34:16] the cutting edge and to make sure that [34:18] you are finding the right deals, finding [34:21] the uh companies that might have fallen [34:22] between the cracks otherwise and make [34:25] sure that you can um you know just be on [34:28] top of things that you might have missed [34:30] previously. And then on the governance [34:32] side, um AI can actually do a really [34:35] great job if you actually give it your [34:37] data to read the data, understand the [34:39] contents of it, and make smart [34:41] suggestions for you. So it can look at [34:44] the data and say, "Okay, this should be [34:46] the right description for the table that [34:48] we're looking at." Or if we're looking [34:50] even more granular than that, we can [34:52] look at specific columns and say, "This [34:54] column has personal identifiable [34:57] information. It should be hidden for the [34:58] vast majority of users." and stuff like [35:01] that is tremendously tremendously [35:04] important as an organization to make [35:06] sure that you're not going to be [35:07] susceptible to any uh data leakage or [35:10] data breaches. Um and it's not a fun job [35:14] for a human to do that to go through [35:16] thousands or tens of thousands of tables [35:18] to identify uh what might be sensitive [35:20] and what isn't. So, if you're able to [35:23] leverage AI to do some of that for you, [35:26] then you're really setting yourself up [35:27] for success to make sure that your [35:29] organization is resilient, you're [35:31] following privacy best practices, and [35:33] you're not going to step on any landmine [35:35] in the future. [35:38] I'll probably add to this, right? I [35:40] mean, I totally agree with both Dan and [35:42] uh Azar U, but what I'll probably add [35:45] is, you know, so you're all already [35:47] seeing this as a proof in the [35:49] marketplace. So if you look at I was [35:51] reading a research that uh the [35:55] web traffic to Stack Overflow and all of [35:57] these websites uh Reddit and others like [36:00] which were uh where people used to go [36:03] for coding help and whatnot has dropped [36:06] dramatically. Now, if you think about [36:09] any AI software engineer, data engineer, [36:12] I'll just club them together, right? Uh [36:14] they're starting with a co-pilot, [36:16] whether it's a GitHub co-pilot or any [36:18] other co-pilot, you choose your mix, [36:20] right? And that's where they start. Uh [36:23] so I I think that is the expectation not [36:28] only from the engineer themselves, but [36:30] also from the you know managers and [36:32] whatnot. uh because these are tools that [36:35] are definitely going to uh the the goal [36:38] is for our developers to be and [36:41] engineers to be more productive, right? [36:43] Uh so uh so I feel like that is [36:46] definitely the area is lowhanging fruit [36:48] right there that if you want to optimize [36:51] uh that is one area that is being [36:53] optimized dramatically and a byproduct [36:56] of that uh is the job market for [36:59] software engineers right is starting to [37:01] soften as you see number of openings uh [37:04] they've definitely decreased I was [37:05] reading the report on that so it's like [37:08] so these are the things that you're [37:10] already seeing where uh definitely [37:13] uh you know this is where AI is [37:15] definitely going to disrupt [37:17] >> and I'll maybe add on the data ops side [37:19] of things we use an AI tool called Ziper [37:23] for datab bricks cost and performance [37:27] optimization [37:28] that has reduced our cost by almost 40%. [37:33] So the opportunity to leverage these [37:36] next generation AI tools also on your [37:39] engineering and data ops um we've found [37:43] to be real [37:45] um a as we think of AI capability [37:50] developing and growing over time if if [37:54] you look at meter for example uh LLM's [38:00] the length at which of task tasks that [38:03] LLMs are able to perform I believe is [38:05] doubling every seven months. So that's [38:08] an exponential rate of growth. As these [38:11] capabilities develop, [38:13] what new use cases have you thought of [38:19] um [38:21] become available to us within our [38:24] investment firms? Uh we didn't practice [38:28] this question. So, I'll I'll I'll ask [38:30] you. Anybody have any thoughts on this [38:32] one? What new capabilities might develop [38:35] over time that aren't available to us [38:37] today? [38:40] I can I can try and tackle this one. And [38:42] it might be a capability that is [38:44] available today. The the market moves so [38:46] fast it's uh it's impossible to keep up [38:48] with everything. Um, I'll stick with uh [38:50] the venture capital example, but if [38:53] we're looking at um a standard venture [38:55] capital um investment flow, what does [38:59] that look like? First, you have to [39:00] source a company. Then you need to [39:03] outreach to the company. You need to [39:05] prepare materials if you actually are [39:07] interested in uh signing that or not [39:09] signing but investing in that company. [39:11] Then you need to over time make sure [39:13] you're getting the right information at [39:15] the right times for the company and uh [39:17] ensuring that your investment looks [39:18] stable. How can that be changed today or [39:22] in the near future with AI and LMS? I [39:26] think uh we're moving to a world where [39:28] most of that process can be automated. [39:32] Uh and I'll say that with the caveat [39:34] that there likely should always be a [39:36] human in the loop. But if we're looking [39:38] at sourcing, I could create an agent [39:40] today that crawls the web to find a [39:44] company that fits my portfolio strategy. [39:46] You know, whether that's like an agrch [39:49] company based in some small market that [39:52] has uh this specific other niche, I can [39:56] send my agent to do that work for me and [39:58] it can look at, you know, Crunchb or [40:00] Pitchbook or just Google search results [40:02] or hacker news or anything else. It can [40:05] find those companies. It can share them [40:07] with me and either it can automatically [40:10] do some outreach on my behalf or I can [40:12] say these are the companies I care [40:13] about. Select the ones you want. I'll [40:16] pick my three that I'm interested in. [40:18] It'll draft that email that uh would [40:20] start that outreach in my voice and my [40:22] tone. schedule the meeting and really [40:24] follow that entire workflow from that [40:26] point where the human can then be uh [40:30] still in the process to make sure that [40:32] the investments are in the right spot [40:34] but also doing other work that might be [40:35] more important on the side that isn't as [40:37] tedious as sourcing companies which can [40:40] be um a headache by itself. [40:45] uh I'll add uh to that uh I think you [40:48] know two areas where I see [40:51] we which will disrupt in asset [40:54] management probably would be one uh the [40:57] rise of quantal investing right uh I [40:59] think uh we have fundamental technical I [41:02] think fundamental is going to be just [41:03] grow more uh as we move forward so [41:07] that's one area uh the second is going [41:10] to be in investor management and asset [41:13] management hyperpersonalized products s [41:15] uh like we're already seeing with a lot [41:17] of advice but uh I think [41:20] hyperpersonalized as in you want to [41:22] create your own custom products whether [41:24] which ones you want to which one you [41:25] want to invest in what securities you [41:27] don't want to invest in how can you [41:29] create a fund around that so with [41:31] technology becoming ubiquitous it will [41:33] become easier for you to create these [41:36] hyperpersonalized products for your uh [41:39] you know your clients so I think those [41:41] are the two areas uh where I feel, you [41:44] know, AI can help us a lot. [41:47] >> Dan, any thoughts on how you might see [41:49] AI reshaping investment strategy over [41:52] the next, let's call it five years? [41:55] >> Yeah, I think I 100% agree with what [41:56] these two have already said, but I think [41:59] um operational agility, operational [42:00] speed, decision-m speed is all going to [42:02] get just compressed down, right? Where [42:05] humans are going to be these kind of AI [42:08] assisted decision makers, right? Where I [42:10] 100% agree with the human in the loop [42:12] aspect, right? where humans are going to [42:14] be even better about the the things that [42:17] we're good about like being human and [42:19] relationships where AI and agents can [42:22] take over the the scale and the speed [42:25] that we need to work in. So I think I [42:27] think that orchestration between the two [42:29] is going to become much more seamless. [42:32] >> We have a minute and a half left. Maybe [42:34] we can do this quickly. Um, I'm [42:36] interested in where you've seen the most [42:39] resistance or hesitation in trying to [42:42] adopt these tools internally [42:44] and how you've thought about addressing [42:47] those challenges. Do you want to go [42:50] quickly? [42:53] >> No, no, you go for it. [42:54] >> Uh, I think, uh, like generally there's [42:56] been a lot of support within our [42:58] organization for adopting these tools. I [43:00] think to keep it brief, the one area [43:02] where there's always a bit of friction [43:04] is on the privacy or compliance side [43:07] where you know you really really really [43:09] need to have solid data line of sight [43:11] and permissioning set up from day one [43:14] because that's the foundation of really [43:16] anything that you're going to do with uh [43:18] LLMs or AI in your company. So I [43:21] wouldn't call it push back necessarily [43:24] but there does need to be specific [43:26] strong buyin and uh trust that's built [43:29] with those teams. [43:31] >> I would say you know change management [43:34] is a big aspect of it right. So we want [43:36] to make sure that uh the we we touched [43:39] upon AI literacy earlier on. So I think [43:42] as long as there is that AI literacy uh [43:45] enabled for the firm then the barriers [43:48] to success uh are significantly removed. [43:53] >> Any final thoughts Dan? Yeah, I mean I [43:56] don't think it's been a hesitation for [43:58] us or a push back, but the the change [44:00] the speed of change has been, you know, [44:02] difficult to keep up with at times. And [44:04] the fact that we've got a top- down [44:06] mandate to use AI tools to make our jobs [44:08] more efficient, embed it into our [44:10] products, you know, organizes us all and [44:13] keeps us aligned towards kind of that [44:14] those common goals.