Verizon and Banking Leaders Tie AI Tools Directly to Bottom-Line Results

Verizon与银行业高管将AI工具直接与底线业绩挂钩 企业高管认为，人工智能只有与严格的投资回报率指标和标准化治理相结合，而非盲目押注未经证实的生成式功能，才能创造真正的财务价值。 多位企业数据负责人探讨了将AI整合到分析与商业智能工作流中的现实情况、成功经验与潜在挑战。嘉宾们分享了实际应用案例，例如Verizon以客户为中心的Project 624和Truist Bank获得专利的Client Pulse系统，同时强调了恪守ROI准则、加强数据治理以及循序渐进落地而非盲目跟风炒作的必要性。对话还涵盖了分析工具从嵌入式副驾驶（co-pilots）向自主代理（autonomous agents）的演进、员工技能升级策略，以及维护清晰的语义层对确保精准业务决策的关键重要性。 Verizon 通过推出 Project 624 将 AI 深度嵌入数字化渠道、移动 App 与呼叫中心，不仅显著改善客户服务体验，还直接拉动了第二季度财报表现并调高市场预期。 (1:56) Truist Bank 打造专利级平台 Client Pulse，打破传统模型孤岛，将客户洞察与行动建议跨物理网点与数字化渠道进行全域实时协同分发。 (6:50) 专家指出 Generative AI 项目常因缺乏明确的 ROI 衡量标准、基于陈旧或低成熟度数据运行以及缺乏严格治理边界而面临高达 80% 的搁浅风险。 (9:18) 数据与业务领导者应关注顶层财务指标而非技术术语，避免为追逐热门技术盲目跳跃成熟度曲线，倡导通过小步快跑的 MVP 迭代来推进 AI 落地。 (13:33) 现代商业智能（BI）正沿着嵌入式侧边栏 Co-pilot 与自主深度研究 Agent 两大范式演进，企业必须依托强大的语义层与数据治理来防范分散化集成带来的口径冲突。 (21:06)

Enterprise data leaders emphasize that sustainable AI adoption requires disciplined financial metrics, mature data foundations, and incremental organizational progression rather than chasing technical novelties. While embedded assistants and multi-agent platforms are transforming analytics, long-term returns depend on rigorous governance, semantic consistency, and realistic workforce enablement. Aligning Deployments with Measurable Financial Returns (1:56) Elaine Huang highlights Verizon's Project 624, launched on June 24th, which rolled out an end-to-end customer experience architecture spanning digital search, mobile app dialogues, and live call-center agent guidance. She notes the initiative improved customer satisfaction and directly contributed to strong Q2 earnings results. Both Huang and Harveer Singh argue that executive leadership and corporate boards evaluate artificial intelligence strictly on top-line growth and net cost reductions—such as comparing automated workflows to offshore operational costs—rather than technical vanity indicators like cost per token. Singh forecasts that roughly 80% of corporate generative AI use cases will be abandoned by year-end, with only 15% to 20% surviving into ongoing production. He attributes project attrition to uncalculated ongoing model runtime expenses, stale data lakes that fail to reflect shifted consumer behaviors, and inadequate governance safeguards in regulated environments. Drawing a parallel to early enterprise cloud migrations, Singh contends that AI operates as an agility accelerator rather than an inherent cost-saving measure. Navigating Enterprise Data Maturity and Siloed Operations (19:06) Rachel Ibarra and the panel warn against pushing business units across unrealistic developmental leaps, such as moving directly from baseline data collection (stage A) to autonomous agentic systems (stage Z). They stress that technical teams who overlook organizational readiness trigger internal skepticism and resistance, turning potentially transformative initiatives into perceived gimmicks. Successful implementation requires guiding stakeholders through pragmatic, step-by-step milestones (such as moving from stage A to D, then D to H). Siloed operational data poses a concurrent structural hurdle. Singh references Truist Bank's patented Client Pulse platform, which resolved disconnected churn and customer segmentation models by streaming real-time, weighted recommendations directly across branch tellers, ATMs, mobile applications, and point-of-sale systems. Looking beyond internal data warehouses, Ibarra points out that enterprises are also using modern AI tools to ingest and analyze unstructured public web data, blending broader external market signals into their core intelligence pipelines. Architecting BI via Embedded Panels and Governed Queries (21:06) Chitrang Dave rejects recurring claims regarding the obsolescence of traditional dashboards, arguing that business intelligence interfaces are evolving to meet demands for faster delivery. Current adoption centers on two architectures: co-pilot side panels embedded directly within BI tools like Tableau or Domo for natural-language data interrogation, and autonomous deep research agents capable of independently synthesizing raw datasets into comprehensive reports that would otherwise require weeks of manual analyst labor. Panelists caution that decentralized software procurement risks causing analytical discrepancies if business units run localized calculations without an enterprise-wide semantic layer. When underlying metadata is uncurated, querying embedded models risks creating fragile pipelines that output contradictory metrics. To balance access with governance, Brad Ahrens recommends positioning conversational interfaces over pre-built, vetted query templates—such as retrieving specific broker-dealer trades within narrow time frames—prompting non-technical users for necessary parameters without exposing the business to raw SQL errors. Workforce Evolution, Multi-Agent Platforms, and Data Integrity (36:46) Singh highlights an intensifying split in the technical labor market: entry-level computer science graduates encounter steep hiring headwinds, while specialized AI researchers command massive compensation packages. Rather than requiring every employee to master deep data science, he advises junior professionals to build practical fluency with autonomous coding agents to multiply their productivity. Enterprise upskilling is adopting tiered frameworks—such as FINRA's model categorizing staff into general Explorers, intermediate Catalysts who bridge technical and business domains, and dedicated data scientists. Looking to future applications, Huang identifies multi-agent platforms as the next major operational leap due to their ability to execute complex, multi-step business logic across disparate enterprise processes. Dave anticipates specialized agents automating labor-intensive quarterly business review decks using verified enterprise data. However, Singh warns that hyper-realistic generative capabilities require robust anomaly-detection mechanisms, particularly when combined with decentralized networks, to ensure underlying automated insights remain accurate and untampered. While multi-agent systems and conversational query layers offer significant operational acceleration, enterprises remain constrained by data quality, vendor fragmentation, and governance risks. The central challenge moving forward is whether organizations can modernize legacy data infrastructure and establish reliable semantic layers fast enough to prevent generative deployments from producing misleading or unverified corporate reporting.

人工智能在数据分析与商业智能领域的落地正从概念炒作转向注重财务回报与工程治理的务实阶段。多位行业技术高管指出，尽管AI智能体与自动化能赋能业务决策，但企业若缺乏清晰的投资回报率测算、底层语义层治理以及循序渐进的成熟度推进，多数生成式AI项目将面临被淘汰的风险。 业务回报驱动与规模化治理框架 (1:56) 在推动AI实现实际业务成效的过程中，Verizon与Truist Financial展示了系统化落地实践。Verizon的Elaine Huang介绍了6月24日推出的“Project 624”，该项目构建了涵盖搜索、移动应用交互以及人工客服话术辅助的端到端AI客户体验体系，直接助力了公司强劲的第二季度财报表现并促使管理层上调市场预期。为保障交付效率与质量，Verizon建立了严格的治理框架，包括标准化需求分流、业务案例KPI预设、多变量测试、模型再训练反馈循环，以及利用Jira和GitLab的智能体工作流自动化。前Truist银行首席数据官Harveer Singh则分享了获美国银行家协会认可的“Client Pulse”系统，该系统打破传统数据孤岛，结合边缘计算与Salesforce防护机制，在移动端、ATM机和柜员端等全渠道毫秒级分发个性化建议。 然而，盲目推进技术的风险同样显著。Harveer依据行业观察指出，由于忽视生产环境下的真实运行成本、底层数据湖陈旧不匹配，以及受监管环境下的数据防泄漏边界约束，他预计高达80%的生成式AI用例将在当年年底前被废弃，仅有15%至20%能最终存活。他将AI热潮类比于早期的云计算迁移：二者均非廉价的降本捷径，而是加速业务成果的放大器。Elaine也强调，企业高管与董事会关注的是顶线收入增长与净成本节约，而非每Token成本等技术术语；团队切忌为了追求新潮工具而过早引入自主智能体，若底层知识库杂乱，不仅会引发模型幻觉与运行失败，更会招致监管和法律责任，因此必须坚持以最小可行产品（MVP）进行敏捷迭代。 渐进成熟度路径与商业智能架构抉择 (19:06) 管理分析变革必须严密贴合业务团队的数据成熟度。会议指出，技术团队往往热衷于直接跨入机器学习或自主智能体等终局形态（从阶段A跃升至Z），但过度超前于业务方的实际认知水平会引发抵触情绪，使战略技术沦为华而不实的摆设。有效策略应以渐进里程碑为核心，引导业务部门从A推进至D，再由D迈进至H，以此降低组织摩擦。 在系统架构层面，传统仪表板并未消亡，而是在向提升交互效率的方向演化，主要呈现两种路径：一是以Tableau或Domo为代表的“副驾驶模式”，在侧边面板通过自然语言进行实时下钻与查询；二是“自主深度研究模式”，由AI智能体耗费数小时独立分析数据集并合成详尽报告。Avidity Biosciences的Chitrang Dave等人提示，直接采购BI平台原生AI模块能节省人力，但若软件供应商绕过IT向业务部门分散销售，极易导致“冻酸奶综合征”（工具泛滥引发选择瘫痪），更会因分散的局部计算产出与企业大盘相悖的指标。若缺乏健全的企业级语义层与元数据规范，直接叠加嵌入式AI极易演变成产生错误数据的“鲁布·戈德堡机械”。 人才市场割裂与企业分层赋能机制 (30:05) AI对劳动力市场的重塑显现出两极分化的割裂景象。Harveer Singh观察到，一方面计算机与软件工程专业的初级应届毕业生面临严峻的就业困境，入门级岗位大幅收缩；另一方面，科技行业头部企业却频频开出数亿美元的极端薪酬方案竞夺顶尖AI研究学者。面对这一失衡，Singh建议初级与技术人员应深入掌握AI编程智能体（AI coding agents），将自主工具转化为自身的技能杠杆，同时企业内部也应建立持续学习与知识共享的文化。 在企业内部人才体系搭建上，FINRA并未盲目要求全体员工转型为数据科学家，而是构建了三层学习金字塔：最底层的“探索者”负责日常使用AI工具；占比约5%至10%的“催化剂”由业务与数据跨界专家组成，接受高校合作项目深化技能；顶层则保留核心数据科学家团队。这种结构化分层确保了全员技术素养的稳步提升。 对话式分析、智能汇报与外部数据融合 (37:52) 针对未来3至5年分析工具的技术演进，与会嘉宾提出了务实的落地预期。Brad Ahrens指出，近期最具可行性的场景并非强求业务人员编写SQL或成为公民数据科学家，而是在已通过验证的查询模板与元数据之上构建对话式聊天机器人。例如，在调取特定券商在特定分钟内（如苹果公司股票在上午9:30至9:31之间）的交易明细时，机器人可通过交互提示用户补充参数并执行合规查询，兼顾便捷性与数据严谨性。 在报告与外部信息整合层面，Chitrang Dave预测，专业AI智能体将大幅简化季度业务复盘（QBR）等汇报PPT的手工制表流程，基于受信任的内部数据实现报告自动编译。Elaine Huang将多智能体编排视为下一波核心浪潮，用于承接跨业务环节的复杂多步逻辑。针对生成式技术带来的超写实合成内容，Harveer Singh提示需构建以AI为核心的诊断层以侦测系统异常与虚假数据；Rachel Ibarra则总结指出，现代分析正跳出传统单一的数据仓库边界，AI工具使得抓取和处理非结构化公开网络数据变得更为顺畅，从而为企业商业智能注入更多维度的市场外部变量。 本场研讨表明，企业在AI与数据分析领域的竞争关键不在于追逐算法噱头，而在于能否建立扎实的语义治理、端到端运营规范与分步递进的业务变革机制。然而，如何在分散化采购工具与集中式指标管控之间维持平衡，以及如何在日益逼真的合成数据环境中确保分析结果的真实性，仍是横亘在数据领导者面前的核心未决挑战。

## Background 该内容由行业峰会频道 Ai4 于 2025 年 8 月 27 日发布在 YouTube 平台上。视频记录了一场汇聚 Verizon、Avidity Biosciences、Truist Bank、FINRA 及 Cardinal Group Management 等企业高管的专家圆桌论坛，探讨如何利用 AI 赋能数据分析与商业智能（BI）、自动化数据处理流程，并解析企业在规模化落地 AI 面临的战略、治理与投入产出挑战。 ## TL;DR 论坛深入剖析了 AI 与商业智能融合的实战路径，警告企业警惕脱离业务根因与数据准备度的盲目跟风，断言多数缺乏明确 ROI 的 GenAI 用例将面临淘汰。Verizon 与 Truist 等企业分享了通过端到端客户体验重塑（如 Project 624 和 Client Pulse）实现商业价值的成功实践，强调建立统一语义层、数据安全护栏与敏捷反馈闭环的重要性。在技术选型与团队建设方面，嘉宾们提出了防范供应商“冻酸奶综合征”的小规模验证策略，以及 FINRA 的三层人才培养架构。展望未来 3 到 5 年，多智能体（Multi-Agent）协作、自动化业务报告生成以及基于成熟查询库的自然语言交互将成为重塑 BI 的核心力量。 ## Atmosphere & style 整场对谈氛围务实、坦诚且充满幽默感，多位跨行业的资深高管以一线实战经验对冲了行业狂热炒作。嘉宾们频频抛出诸如“冻酸奶综合征”、“最糟糕的鲁布·戈德堡机械”等生动贴切的比喻，既有对业务端盲目索要“闪亮新玩具”的犀利吐槽，也有对工程架构、数据治理与人才转型的深层关怀，让观众在轻松自如的交流节奏中汲取极具落地参考价值的战略清醒剂。 ## Key takeaways - [01:58](https://www.youtube.com/watch?v=P7vWXznFK-E&t=118s) Verizon 通过 Project 624 打造端到端 AI 客户体验，覆盖数字交互、人工协同与闭环反馈，直接提升了 NPS 并在财报中验证了商业回报。 - [06:50](https://www.youtube.com/watch?v=P7vWXznFK-E&t=410s) Truist 推出专利“客户脉搏（Client Pulse）”技术，打破孤立预测模型孤岛，在各业务触点实现毫秒级的实时决策下发。 - [09:18](https://www.youtube.com/watch?v=P7vWXznFK-E&t=558s) 警惕 GenAI 伪需求与盲目跟风，预计业内 80% 的 GenAI 用例将被淘汰，立项必须追问业务根因、核算推理 ROI 并确保底层数据质量就绪。 - [11:41](https://www.youtube.com/watch?v=P7vWXznFK-E&t=701s) AI 与云计算类似，核心价值在于提升业务敏捷度与加速价值产出，而非盲目削减成本；在推进前须建立严格的数据隔离与安全护栏（Guardrails）。 - [13:33](https://www.youtube.com/watch?v=P7vWXznFK-E&t=813s) 推进 AI 落地应坚持业务价值优先（关注收入增长与成本缩减而非 Token 成本），切忌技术方案走在问题前面，坚守 MVP 敏捷迭代原则。 - [18:42](https://www.youtube.com/watch?v=P7vWXznFK-E&t=1122s) 企业普遍面临数据治理与 AI 演进的双轨推进挑战，管理者应引导业务团队小步快跑，避免脱离组织接受能力强行跃迁。 - [20:41](https://www.youtube.com/watch?v=P7vWXznFK-E&t=1241s) 传统 BI 仪表盘正向两大新范式演进：集成自然语言交互问答的副驾驶模式（Co-pilot），以及能自主进行深度分析挖掘的研究智能体模式（Deep Research）。 - [22:49](https://www.youtube.com/watch?v=P7vWXznFK-E&t=1369s) 警惕 AI 工具选型中的“冻酸奶综合征”，面对井喷的供应商应通过小规模 PoC 进行低风险验证，重点考量成熟度与数据安全性。 - [25:12](https://www.youtube.com/watch?v=P7vWXznFK-E&t=1512s) 面对原生 BI 工具的黑盒挑战，企业应优先构建统一的企业级语义层、集成 GenAI 数据市场，并推行定制化数据素养培训。 - [28:21](https://www.youtube.com/watch?v=P7vWXznFK-E&t=1701s) 在缺乏健全元数据（Metadata）的前提下盲目开放自然语言查数据极度危险，企业应优先将 AI 工具赋能专业分析师以保障数据准确性。 - [30:05](https://www.youtube.com/watch?v=P7vWXznFK-E&t=1805s) FINRA 建立“探索者”、“催化剂”（懂业务与 AI 的桥梁人才）和“专家层”的三层培养架构，避免全民盲目学习手写代码。 - [33:22](https://www.youtube.com/watch?v=P7vWXznFK-E&t=2002s) 职场人需掌握与 AI 高效对话的方法论成为“AI 低语者”（AI Whisperer），在业务调研与高管沟通中充分借助增强工具提效。 - [35:12](https://www.youtube.com/watch?v=P7vWXznFK-E&t=2112s) AI 不会完全取代数据分析岗位，但会促使技术人才市场分化，熟练掌握编程智能体（Coding Agents）将成为技术人员的核心竞争优势。 - [37:52](https://www.youtube.com/watch?v=P7vWXznFK-E&t=2272s) 未来 3 到 5 年，基于成熟查询库的对话式 Chatbot、Agent 自动化生成季度汇报材料（QBR）、多智能体平台以及融合非结构化公开网络数据将成为 BI 变革的核心趋势。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=P7vWXznFK-E&t=4s) 嘉宾背景介绍与开场 论坛开场，主持人邀请台上的各位嘉宾进行简短的自我介绍，说明各自的姓名、职位，并用一句话概括人工智能（AI）如何对其所在领域的分析与商业智能（BI）工作产生深远影响。 * **Elaine Huang（黄伊莱恩）**：现任 Verizon 企业数据与分析枢纽（Enterprise Data Analytics Hub）的高级总监，全面统筹公司的商业数据与 AI 解决方案。 * **Chitrang Dave**：现任 Avidity Biosciences 的 AI 与分析负责人，深耕生命科学、制药与医疗保健行业多年，此前曾在多家医疗器械公司担任相关领导职务。 * **Harveer Singh**：在报名参会时担任 Truist Bank（美国金融信托银行）消费银行部的首席数据官（Chief Data Officer）。他提到自己近期完成了一次从企业高管到创业者的有趣转型，目前是一家由区块链驱动的无线技术公司的创始人，并在 Solana 区块链上推出了奖励计划。他在发言中指出，AI 在帮助他完成从传统大型机构高管到初创企业创始人的日常转型与落地执行中发挥了关键作用。 --- ### [01:58](https://www.youtube.com/watch?v=P7vWXznFK-E&t=118s) Verizon 的 Project 624：通过端到端 AI 重塑客户体验 在进入正式讨论时，主持人向 Elaine 提问：业界虽然充斥着 AI 变革分析领域的舆论炒作，但真正的变革往往只发生在具体的业务场景中。Elaine 分享了 Verizon 在 2024 年 6 月 24 日正式上线的重磅项目——**Project 624**。她向台下观众互动提问有多少人是 Verizon 的客户，并坦言电信行业在历史上往往面临客户满意度不高、净推荐值（NPS）垫底的困境。为了从根本上改变这一痛点，Verizon 推出了这套由 AI 全面赋能的客户体验升级体系。 ```mermaid flowchart TD accTitle: Verizon Project 624 AI客户体验流转与闭环架构 subgraph S1["数字交互端 (Digital)"] A["用户进线咨询"] --> B["AI 引导个性化搜索与推荐"] B --> C["App 内拟人化 AI 智能助手"] end subgraph S2["人工坐席协同 (Agent Assist)"] C -->|"转接人工"| D["动态上下文背景注入"] D --> E["实时对话引导与挽留策略"] end subgraph S3["价值闭环与迭代 (Feedback Loop)"] E --> F["多变量实验与 ROI 归因"] F --> G["模型持续重训与迭代"] G -.-> B end ``` Project 624 覆盖了客户生命周期的每一个触点。从最初的数字渠道接触开始，AI 便介入并引导高度个性化的搜索与套餐推荐，化解客户的疑问与不满；在官方 App 内部，用户可以与 AI 助手进行极其自然的拟人化多轮对话；当客户选择从数字渠道转接到人工服务时，AI 会在坐席后台实时提炼并推送该客户的完整历史背景与动态对话引导，极大地提升了解决效率，改善了客户满意度并降低了流失率。该项目不仅直接助力 Verizon 在第二季度财报中交出了超出华尔街预期的亮眼成绩，还成功向董事会和高管层证明了 AI 投资的商业回报。 Elaine 强调，该项目能够在极度严苛的时间窗口内成功上线，核心在于三大维度的标准化与工程执行力： - **业务指标强挂钩**：在业务立项（Business Case）阶段就严格定义 KPI 与价值衡量标准，结合多变量测试（Multivariate Testing）清晰剥离并量化 AI 带来的增量业务收益。 - **闭环反馈机制**：构建端到端的集体数据反馈回路（Feedback Loop），将一线实际交互数据持续回流用于模型的再训练与调优。 - **智能体自动化交付（Agentic Automations）**：将开发流程高度标准化，从统一的需求接入（Unified Intake）到自动创建 Jira 工单、再到 GitLab 代码流转，全流程自动化显著缩短了 AI 功能从研发到推向市场（GTM）的周期。 --- ### [06:50](https://www.youtube.com/watch?v=P7vWXznFK-E&t=410s) 金融场景的实时“客户脉搏”与全渠道触点联动 Harveer 顺着 Elaine 的分享补充了金融服务行业的落地经验。他指出，传统金融机构过去普遍面临数据孤岛（Data Silos）问题。咨询公司和数据团队虽然耗费多年将各类数据汇总并建立了无数孤立的预测模型（如客户流失预测模型、RFM 价值模型等），但始终无法回答最根本的“那又如何（So what?）”——这些孤立模型产生的决策在各个触点各自为战，无法形成统一合力。 为了解决这一问题，Truist 研发并获得了专利技术——**客户脉搏（Client Pulse）**，该成果在同年早些时候斩获了美国银行家协会（American Bankers Association, ABA）的大奖。该架构不再单独看待单点模型，而是实时综合评估客户当前的整体状态与意图（即“脉搏”），并针对该状态赋予动态权重，直接将其转化为各个交互渠道中的可执行动作。 无论是客户在网点柜台（Teller Machine）、自动取款机（ATM）、销售终端（POS），还是在手机 App 或面对理财经理，系统都能在客户触达的毫秒级时间内做出针对性的决策推送。Harveer 感叹道，过去在终端应用中上线一个微小的变更可能需要耗费数年时间，而如今借助边缘端内嵌 AI 的应用与完备的安全护栏（Guardrails，如 Salesforce 等生态中的治理层），企业能够快速实现数据联邦化与渠道实时决策下发，极大提升了客户体验与业务响应敏捷度。 --- ### [09:18](https://www.youtube.com/watch?v=P7vWXznFK-E&t=558s) 警惕 GenAI 伪需求：从“追逐热点”回归“业务根因与数据准备度” 针对主持人提出的“AI 虽能发掘传统分析遗漏的深层洞察，但往往伴随着高昂成本、架构复杂性及业务端信任危机，该如何平衡先进性与可用性、可解释性”的问题，Harveer 直接向台下的业务线代表抛出了一个尖锐的现实痛点。 他幽默地调侃道，台下参会的业务代表回去之后往往会成为技术团队的“噩梦”——因为业务人员在行业峰会上看到了炫酷的 GenAI 演示，回去就会兴奋地要求“我也要这个”，迫使技术团队手忙脚乱地去拼凑方案。但真正的核心秘诀在于：**“你们到底想达到什么业务成果（Outcome）？”** Harveer 尖锐地指出，目前大多数生成式 AI（GenAI）用例在立项时最大的硬伤在于**没有人追问“为什么（Why）”**： 1. **脱离真实业务痛点**：许多团队一上来就急于寻找可以使用 GenAI 的场景，却不问为什么要解决这个问题、当前企业是否有解决该问题的意愿与资源、它是否符合公司的核心优先级。 2. **缺乏 ROI 衡量标准**：大量用例根本没有核算过在生产环境中持续运行这些复杂大模型的推理与算力成本，完全脱离了投入产出比（ROI）。他大胆断言：“我敢打赌，今年业内 80% 的 GenAI 用例到年底都会被废弃，只有大约 15% 到 20% 能够真正存活并投入实际生产。” 3. **底层数据准备度严重不足**：除了业务层面的“为什么”，另一个致命瓶颈是数据底座是否就绪。许多企业连基础数据的质量、多样性与收集频率都无法保障。Harveer 直言，自己曾接手管理过一些数据湖，里面的数据甚至已经陈旧、停滞了数年，但相关团队却依然狂热地叫嚷着“让我们在上面跑一个生成式 AI 用例吧”。如果不提前理清业务根因并夯实数据质量，盲目跟风上线 AI 只会带来巨大的资源浪费。 ### [11:41](https://www.youtube.com/watch?v=P7vWXznFK-E&t=701s) 数据治理、安全防线与 AI 投资回报率（ROI）的现实考量 在探讨客户行为预测时，理解数据的价值和时效性至关重要。客户的偏好和互动方式随时间不断变化，针对已经流失或不再活跃的客户构建的模型往往具有短暂的生命周期。因此，企业在启动分析前，必须明确业务目标，并严格审视底层数据的成熟度。 此外，数据治理与安全性在强监管行业中尤为严苛。在金融等监管环境下，特定客户互动场景对数据的使用有着严格的法律和合规限制。例如，内部持有的某些敏感数据绝不能直接用于营销场景。企业必须厘清数据边界，严防数据混淆（intermingling），同时还要审慎评估外部采购数据的可信度，以确保向客户提供的建议合规且值得信赖。在推进任何 AI 场景之前，建立坚固的安全护栏（guardrails）是防范未来风险的前提。 谈及投资回报率（ROI），回顾云计算的发展历程，当年行业普遍向客户推销“上云能节省数亿美元成本”，但实际落地后发现成本并未显著降低，甚至持平或更高，云带来的核心价值是敏捷度而非单纯的廉价。AI 同样如此——长期来看它不一定会降低成本，但能够让企业以极快的速度达成业务成果。衡量 AI 项目的成效应聚焦于速度与价值产出，而非盲目的成本削减。 ```mermaid flowchart TD accTitle: AI与数据项目的战略护栏及落地原则 subgraph Governance["治理与风控体系"] A["明确业务目标与数据成熟度"] --> B["合规边界与数据隔离"] B --> C["前置构建安全护栏<br/>(Guardrails)"] end subgraph Implementation["敏捷落地与价值验证"] C --> D["对齐商业指标<br/>(收入增长/成本缩减)"] D --> E["MVP 敏捷迭代<br/>(小步快跑/灰度实验)"] E --> F["AI 快速产出价值<br/>(重在速度而非单纯省钱)"] end ``` - **数据隔离与防混淆**：严格划分合规与营销数据的边界，禁止将不可用于营销的受规管数据混入商业推广模型。 - **外部数据信任度**：引入第三方数据源时，必须全面验证其准确性与合规性，确保客户咨询建议的稳健。 - **ROI 认知对齐**：AI 与云计算类似，主要优势在于缩短交付周期、加速业务结果输出，而非立即节省绝对预算。 --- ### [13:33](https://www.youtube.com/watch?v=P7vWXznFK-E&t=813s) 业务价值优先、抵制技术诱惑与敏捷迭代 从实际落地经验来看，推进 AI 产品与项目开发需要恪守三大核心原则： 首先，**在项目伊始就必须精准定义业务价值**。技术团队可能对纯技术概念着迷，但业务领导层、CFO 和 CEO 关心的绝不是“每 Token 成本（cost per token）”，而是切实的商业指标——项目能带来多少增量收入？能为公司贡献多少业务增长？相比离岸外包团队，采用智能体（Agentic）AI 方案运营呼叫中心能降低多少运营成本？在立项之初就与高管层、董事会确立公认的成功衡量指标（KPIs），能避免后期在投入产出上的无休止争论。 其次，**切忌让技术方案走在问题前面**。技术团队往往渴望尝鲜市场上最新的“闪亮新玩具（shiny new toys）”，但领导者必须平衡团队的技术冲动与实际业务痛点。解决一个具体问题可能只需传统方案、混合方案，甚至完全不需要 AI。从两年前人人盲目追求 GenAI（生成式 AI），到现在言必谈 Agentic（智能体），很多团队低估了数据质量的复杂性： - 知识库和数据若未彻底清洗，模型的输出结果将缺乏一致性与稳定性； - 一旦部署到生产环境，如何确保模型绝不发生幻觉（hallucination）； - 如何建立代码监控与负责任的 AI（Responsible AI）治理体系，避免企业面临潜在的法律诉讼风险。 最后，**切勿试图一口吃成胖子（Don't try to boil the ocean），坚持敏捷化与实验驱动**。始终坚守产品管理的基本原则：从最小可行性产品（MVP）起步，实施灰度发布（throttle rollout），在持续的小步迭代中建立反馈闭环（feedback loop），稳步沉淀长期价值。 --- ### [18:42](https://www.youtube.com/watch?v=P7vWXznFK-E&t=1122s) 数据成熟度与 AI 演进的双轨博弈 在现场互动中，当被问及“有多少企业是先把数据、BI 平台和数据文化做到卓越状态，再去攻克 AI”与“有多少企业是数据治理与 AI 同时双轨并进”时，绝大多数观众举手认同正在同时推进两项任务。 这种“双轨起舞”的现状要求管理者必须精准评估企业在数据与 AI 两条成熟度曲线上的实际位置： $$ \text{演进步伐：从阶段 A 稳步推向阶段 D} \quad \longrightarrow \quad \text{切忌从 A 点直接强行跃迁至 Z 点} $$ 技术和数据人员往往容易沉浸在终极愿景中（例如直接上线复杂的机器学习或 Agentic 选项），但若脱离业务团队的接受能力强行推行，不仅会遭遇来自业务方的巨大阻力，更会导致项目沦为毫无实际商业影响力的“噱头（shiny object）”。务实的方法是聚焦于阶段性推进（如 A 到 D、D 到 H），稳步引导业务部门前行。 --- ### [20:41](https://www.youtube.com/watch?v=P7vWXznFK-E&t=1241s) 传统 BI 的嵌入式升级与智能体深度调研新范式 面对当前主流 BI 工具（如 Domo、Tableau 等）纷纷内置 AI Agent 的趋势，企业在“直接使用嵌入式 AI”与“彻底重构为 AI 原生（AI-first）分析架构”之间面临抉择。 针对业内长期存在的“仪表盘即将消亡（death of dashboards）”论调，实际情况是仪表盘不会消失，而是会变得更加智能化。传统的仪表盘开发周期长达两周到两三个月，业务部门早已对此失去耐心。未来 BI 分析正向两大模式演进： 1. **副驾驶协同模式（Co-pilot Model）**：在传统仪表盘侧边集成一个交互式问答面板。用户可以直接就仪表盘呈现的图表或底层数据进行自然语言提问与即时探索，这一高度交互的模式正迅速成为各类分析工具的标配。 2. **深度调研智能体模式（Deep Research / Agentic Model）**：代表了 AI 从“搜索”到“研究”再到“深度推理研究”的演进。面对海量数据集，系统可以调动 AI 智能体自主分析数小时，自动挖掘出人类分析师可能需要花费两周时间才能梳理出的洞察，并自动生成一份结构完备的研究报告。 ### [22:49](https://www.youtube.com/watch?v=P7vWXznFK-E&t=1369s) 供应商繁杂与“冻酸奶综合征”：评估 AI 嵌入工具的审慎之道 随着各类分析软件争相嵌入人工智能，企业所面临的决策链条正变得愈发复杂。过去团队只需讨论“是否使用了正确的数据做决策”，如今却演变成“是否在正确的数据上匹配了正确的模型与工具”。不同供应商分别向 IT 部门或业务部门进行销售，某些基于局部小数据集的边缘计算与集中式平台计算甚至会得出大相径庭的结论。面对当前 AI 供应商井喷的局面，企业管理者步入展厅时会发现屏幕上有一半以上的赞助商自己从未耳闻，根本无法迅速判断其优劣。 针对这种选型困境，Harveer Singh 提出了生动的“冻酸奶综合征”（Frozen Yogurt Syndrome）比喻：当你走进一家冻酸奶店，面对上百种配料和口味随意混搭，最后端出来的往往是一份让人难以下咽的产物。面对新技术，企业最明智的做法是像在店里“小口品尝”一样，先进行小规模验证，切忌在确认其真正契合自身预算、能解决实际问题且获得企业长期支持之前盲目深投。相比于冒着数据泄露或交付失控风险去追逐不成熟的最新工具，供应商的成熟度、背后的团队实力以及过往的成功案例，往往是更安全可靠的考量指标。 - **多维决策压力**：模型选择、工具生态与数据源的多样化导致企业决策链条呈级联式复杂化。 - **分散与集中之争**：针对局部数据子集的工具计算与企业集中式处理平台可能存在结果口径不一致的权衡。 - **“小步试吃”策略**：在全面推广前通过小规模 PoC（概念验证）验证工具成熟度与数据安全性，避免资源浪费。 --- ### [25:12](https://www.youtube.com/watch?v=P7vWXznFK-E&t=1512s) 自研还是外购？利用语义层与数据素养破解原生 BI 工具的“黑盒”挑战 在探讨商业智能（BI）与数据可视化工具的 AI 能力时，“自研还是外购”（Build vs. Buy）是企业绕不开的经典命题。Elaine Huang 明确倾向于外购现成方案，尤其是现有的成熟可视化平台。如今的技术演进使得企业只需将本地部署的 BI 升级或迁移至云端，即可直接利用其原生嵌入的 AI 能力，无需在底层算法和工程人力上进行沉重投资。然而，外购绝非万灵药，通用工具无法覆盖所有的边缘场景（Edge Cases），一旦企业需要进行过度定制化改造，外购的成本优势便会迅速瓦解。 ```mermaid flowchart TD accTitle: 语义层与数据治理架构 accDescription: 展示企业如何通过语义层、AI数据市场与素养培训解决原生BI黑盒问题 subgraph Governance ["统一治理与语义层"] A["底层多元数据源"] --> B["企业级语义层 (Semantic Layer)<br/>标准指标与口径归一化"] B --> C["集成 GenAI 的数据市场<br/>(Data Marketplace)"] end subgraph Personas ["多角色赋能"] C --> D["业务决策者"] C --> E["数据分析师"] F["定制化数据素养培训"] -.-> D F -.-> E end subgraph Tools ["原生 BI / AI 工具"] D --> G["原生 BI 可视化与 AI 模型<br/>(规避黑盒偏差，输出一致结果)"] E --> G end ``` 从企业首席数据、分析与 AI 办公室（CDO/CAIO Office）的集中治理视角来看，原生 BI 工具最大的痛点在于其黑盒特性——模型深嵌在封闭生态中，管理层难以洞察底层逻辑，更难以保证不同业务部门获取到一致的结果。为了攻克这一难题，必须强化三项基础设施建设： 1. **构建企业级语义层（Semantic Layer）**：对指标定义与业务逻辑进行标准化与合理化规整，确保所有工具基于统一语义提取数据。 2. **在数据市场（Data Marketplace）中嵌入生成式 AI**：引导不同画像的业务用户快速、准确地对接合规数据，避免繁冗的入职与审批流程引发业务抵触。 3. **推行分层数据素养计划（Data Literacy Programs）**：针对不同角色定制培训，让员工透彻理解数据背后的业务逻辑，从而在调用原生 AI 工具时具备辨别异常与解释结果的能力。 --- ### [28:21](https://www.youtube.com/watch?v=P7vWXznFK-E&t=1701s) 元数据缺失的风险与赋能分析师：避免成为“鲁布·戈德堡机械” 主持人在讨论中补充了来自一线的真实痛点：当前业务部门已不再有耐心为了一张报表或看板等待两三周甚至数月，然而在缺乏健全元数据（Metadata）的前提下，贸然向业务人员开放“自然语言查数据”极其危险。如果仅凭一句简单的“我们到底有多少客户”，系统在底层混乱的定义中可能会拼凑出极其离谱的错误答案。 缺乏元数据支撑的 AI 问答就像是在运转“世界上最糟糕的鲁布·戈德堡机械”（The world's worst Rube Goldberg machine）——看似精巧繁复的自动化流程层层传导，最终输出的却是一个完全失真的荒谬数字。为了兼顾效率与安全，诸多企业采取了折中战略：优先将强大的 AI 生产力工具交由专业数据分析师使用，大幅缩短看板与分析报告的交付周期；而在面向全员的业务端，则审慎评估风险，优先在容错率较高、非高风险决策的场景中落地 AI。 --- ### [30:05](https://www.youtube.com/watch?v=P7vWXznFK-E&t=1805s) FINRA 的三层人才培养架构：从“探索者”到业务与技术的“催化剂” 作为美国金融业监管局（FINRA）的高级副总裁，Brad Ahrens 在迟到入场时幽默地打趣道：“抱歉，我的生成式 AI 智能体目标跟我本人的目标不太一致，等做完强化学习它下次就能按时到了。”随后，他深入分享了 FINRA 在监管全美经纪交易商生态中，面对 AI 浪潮所推行的人才技能转型与文化管理路径。 ```mermaid flowchart LR accTitle: FINRA 三层人才培育体系 accDescription: 展示从探索者、催化剂到数据科学家的三层人才培养与连接模型 A["探索者 (Explorers)<br/>日常消费 AI/分析成果的业务人员"] -- "5%~10% 选拔<br/>高校联合培训计划" --> B["催化剂 (Catalysts)<br/>懂业务监管规则 + 具备数据感知力"] B <--> C["专家层 (Experts)<br/>数据科学家 / 数据分析师 / AI 工程师"] subgraph Leadership ["管理层对齐"] D["领导层专属培训"] -.-> B D -.-> C end ``` FINRA 并不盲目追求将全体员工培养成“公民数据科学家”（Citizen Data Scientist），而是将组织人才划分为清晰的三层结构并实施差异化培养： - **探索者（Explorers）**：处于体系一端的广大日常业务人员，主要消费高级分析与生成式 AI 产出的结果。 - **催化剂（Catalysts）**：从探索者中选拔出 5% 至 10% 的骨干，通过与知名高校联合开发的体系化课程进行深造。该项目不要求学员手写 Python 代码，而是着重建立对高级分析与 AI 能力的认知与敏锐度。他们凭借深厚的监管业务背景，成为连接一线业务场景与技术专家的关键桥梁。 - **数据科学家与分析专家（Data Scientists & Analysts）**：处于另一端的技术核心，专注于构建高质量模型并与“催化剂”协同落地解决方案。 此外，FINRA 专门设立了针对领导层的培训项目，旨在帮助管理层形成清晰的 AI 认知，从而筛选出高价值的用例待办列表（Backlog）。Brad 特别警示，组织必须高度警惕仅仅因为技术流行而到处滥用生成式 AI 所造成的“高效的低效”（Efficient Inefficiencies），技术落地的核心永远是创造实质价值。 ### [33:22](https://www.youtube.com/watch?v=P7vWXznFK-E&t=2002s) 成为“AI低语者”与职场准备的新标准 Harveer Singh 结合个人经历分享了 AI 工具如何改变日常工作习惯。他在印度长大，自认在语言和表达上总会存在难以完全消除的短板，且个人的知识库终有耗尽的时候，因此必须借助增强型工具来提升自己。在当今时代，如果去参加会议却没有做好准备，已经没有任何借口可言——因为只需一条简单的 Prompt 就能在几分钟内掌握基础信息。 他特别提到了一个正在流行的新概念——**“AI 低语者”（AI Whisperer）**。这并不要求每个人都成为顶尖科学家，而是掌握一种与 AI 顺畅对话的方法论： - 在清晨与老板开会前，利用 AI 梳理沟通思路，即便结果不理想也能找到得体、专业的表述方式。 - 在区块链创业过程中面对大量未知项目与交易所时，利用 AI 在 5 到 10 分钟内快速做完背景调研、掌握市场动态。 - 他建议甚至在孩子上大学之前，就应该培养他们成为“AI 低语者”的技能。 ```mermaid flowchart LR accTitle: AI Whisperer Workflow A["输入 Prompt / 业务意图"] --> B["AI 辅助快速调研 & 话术优化"] B --> C["5-10 分钟掌握行业脉络"] B --> D["高效准备高管会议 / 业务汇报"] ``` --- ### [35:12](https://www.youtube.com/watch?v=P7vWXznFK-E&t=2112s) 数据团队的人才需求、分化与技能升级 Elaine Huang 对 AI 是否会取代数据分析岗位提出了务实看法。她认为 AI 不会彻底消除数据分析领域的职位与需求，但未来团队所需的人员规模可能不像过去预期的那么多。熟练编写 Python 和代码的能力依然必不可少，工具会让开发周转速度大幅提升；更重要的是团队必须具备开放心态与创新创造力。在技术迭代极快的当下，无论是基层员工还是团队管理者，唯有建立持续学习和拥抱新技术的文化，才不会被 AI 浪潮淘汰。 Chitrang Dave 则指出了当前技术人才市场出现的奇特“两极分化”现象： - **初级市场遇冷**：计算机科学（CS）与计算机工程专业的应届毕业生就业环境极其严峻。 - **顶尖人才争夺白热化**：扎克伯格（Mark Zuckerberg）等科技巨头开出数亿美元的薪酬抢夺顶尖 AI 研究员。 面对这种令年轻人感到迷茫的局面，Chitrang 给出的核心建议是**精通编程智能体（Coding Agents）**。具备深厚技术底色的人一旦正确使用这些 Agent 工具，就会获得强大的“超能力”。 --- ### [37:52](https://www.youtube.com/watch?v=P7vWXznFK-E&t=2272s) 圆桌总结：未来 3 到 5 年最具变革性的 AI 趋势与能力 在对谈的收尾环节，各位嘉宾围绕“未来 3 到 5 年内对数据分析与商业智能（BI）最具颠覆性的 AI 趋势”进行了圆桌展望： ```mermaid flowchart TD accTitle: Future AI Trends in Analytics and BI subgraph QueryLayer["查询与交互"] B["Brad: 对话式 Chatbot 覆盖成熟查询库"] end subgraph AutomationLayer["自动化与协同"] C["Chitrang: Agent 自动生成 QBR 汇报材料"] E["Elaine: 多智能体平台 (Multi-Agent) 全面落地"] end subgraph DataTrustLayer["数据与信任体系"] H["Harveer: AI 赋能区块链共识与内容真实性校验"] R["Rachel: 拥抱非结构化/公开网络数据"] end ``` - **Brad Ahrens（FINRA）——成熟查询之上的对话式交互**：无需强求业务人员成为“平民数据科学家”或去学 SQL。未来 1 到 2 年内，企业可以针对经过实战检验的高质量查询库和完备元数据搭建 Chatbot。用户只需用自然语言提问（如“查询某券商在 9:30 至 9:31 期间交易苹果股票的所有记录”），Chatbot 会自动匹配对应查询，并通过反问引导用户确认股票代码、时间范围和券商编号等参数。 - **Harveer Singh ——内容真伪辨别与共识架构增强**：AI 极大地降低了内容生产门槛，但辨别真实与虚假变得异常困难。在区块链领域，单纯的“共识机制”如果建立在全盘错误的基础上也会失效，AI 可以作为共识机制之上的新层级，用于捕捉异常并指出共识可能出错的地方。 - **Chitrang Dave ——智能体自动化生成业务报告**：企业往往需要耗费数周时间准备季度业务复盘（QBR）PPT。未来 Analytics 与 AI 智能体将能够直接调用企业内部真实事实数据，全自动生成汇报材料与图表，真正实现 BI 辅助业务决策的初衷。 - **Elaine Huang（Verizon）——Agentic AI 与多智能体（Multi-Agent）平台**：下一轮技术热潮将围绕 Agentic AI 展开。多智能体平台具备极广的应用场景，能够几乎嵌入到每一个自动化用例中，为企业带来显著的生产力与效率跃升。 - **Rachel Ibarra（Cardinal Group Management）——突破数仓边界，吸纳海量公开外部数据**：传统 BI 极度依赖数仓内部清洗完善的结构化数据。未来 AI 将帮助企业爬取并利用广阔的公开网络数据，虽然这些数据不再如数仓那样整洁纯粹，但多元数据源的融合将为业务分析打开全新的维度。 ## Notable quotes > [09:18](https://www.youtube.com/watch?v=P7vWXznFK-E&t=558s) 你们到底想达到什么业务成果（Outcome）？ — Harveer Singh > [09:18](https://www.youtube.com/watch?v=P7vWXznFK-E&t=558s) 我敢打赌，今年业内 80% 的 GenAI 用例到年底都会被废弃，只有大约 15% 到 20% 能够真正存活并投入实际生产。 — Harveer Singh > [09:18](https://www.youtube.com/watch?v=P7vWXznFK-E&t=558s) 让我们在上面跑一个生成式 AI 用例吧 — Harveer Singh > [13:33](https://www.youtube.com/watch?v=P7vWXznFK-E&t=813s) 切勿试图一口吃成胖子（Don't try to boil the ocean） > [22:49](https://www.youtube.com/watch?v=P7vWXznFK-E&t=1369s) 当你走进一家冻酸奶店，面对上百种配料和口味随意混搭，最后端出来的往往是一份让人难以下咽的产物。 — Harveer Singh > [30:05](https://www.youtube.com/watch?v=P7vWXznFK-E&t=1805s) 抱歉，我的生成式 AI 智能体目标跟我本人的目标不太一致，等做完强化学习它下次就能按时到了。 — Brad Ahrens > [33:22](https://www.youtube.com/watch?v=P7vWXznFK-E&t=2002s) 在当今时代，如果去参加会议却没有做好准备，已经没有任何借口可言——因为只需一条简单的 Prompt 就能在几分钟内掌握基础信息。 — Harveer Singh > [35:12](https://www.youtube.com/watch?v=P7vWXznFK-E&t=2112s) AI 不会彻底消除数据分析领域的职位与需求，但未来团队所需的人员规模可能不像过去预期的那么多。 — Elaine Huang ## Mentioned Verizon, Avidity Biosciences, Truist Bank, Cardinal Group Management, FINRA, Elaine Huang, Chitrang Dave, Harveer Singh, Rachel Ibarra, Brad Ahrens, Project 624, Solana, Jira, GitLab, American Bankers Association, Salesforce, Domo, Tableau, Mark Zuckerberg, Apple ## Video description Chitrang Dave, Head of Data, AI, & Advanced Analytics, Avidity Biosciences Harveer Singh, Former Chief Data Officer, Consumer & Small Business Banking, Truist Bank Rachel Ibarra, Senior Director of Data and Analytics, Cardinal Group Management Elaine Huang, Senior Director, Enterprise Data & Analytics, Verizon Brad Ahrens, Senior Vice President, Advanced Analytics, FINRA Abstracts In this session, panelists will explore how AI is transforming data analytics and business intelligence (BI), enabling organizations to extract deeper insights, improve decision-making, and gain a competitive edge. Learn how AI-powered analytics tools are automating data processes, providing predictive insights, and enabling real-time analytics. Panelists will share success stories of integrating AI into BI platforms and discuss the opportunities and challenges of scaling AI-driven analytics across organizations. Discover how to maximize the value of AI in your data analytics strategy. [Source](https://www.youtube.com/watch?v=P7vWXznFK-E)

[9:18] Harveer Singh: 你们到底想达到什么业务成果（Outcome）？ [9:18] Harveer Singh: 我敢打赌，今年业内 80% 的 GenAI 用例到年底都会被废弃，只有大约 15% 到 20% 能够真正存活并投入实际生产。 [9:18] Harveer Singh: 让我们在上面跑一个生成式 AI 用例吧 [13:33] 切勿试图一口吃成胖子（Don't try to boil the ocean） [22:49] Harveer Singh: 当你走进一家冻酸奶店，面对上百种配料和口味随意混搭，最后端出来的往往是一份让人难以下咽的产物。 [30:05] Brad Ahrens: 抱歉，我的生成式 AI 智能体目标跟我本人的目标不太一致，等做完强化学习它下次就能按时到了。 [33:22] Harveer Singh: 在当今时代，如果去参加会议却没有做好准备，已经没有任何借口可言——因为只需一条简单的 Prompt 就能在几分钟内掌握基础信息。 [35:12] Elaine Huang: AI 不会彻底消除数据分析领域的职位与需求，但未来团队所需的人员规模可能不像过去预期的那么多。

[9:18] Harveer Singh: I bet 80% of these use cases will be abandoned by the end of this year. Only like 15 to 20% will make it. [9:18] Harveer Singh: Why are we doing this?

[0:04] All right, thanks so much everyone for [0:06] joining us today. Uh before we get into [0:09] the good questions here and start [0:11] talking about data and AI, um let's go [0:14] around and get some [0:16] do not turn my head away. Okay. Uh let's [0:19] go around for some quick introductions. [0:21] So please folks share your name, your [0:24] role and just in one sentence how your [0:28] work in analytics and BI is being [0:30] impacted by AI. Hello, my name is Ela [0:33] Hang. I work for Verizon. Um I'm a [0:36] senior director in Verizon Enterprise [0:39] Data Analytics Hub. I'm managing all the [0:42] commercial data and AI solutions. [0:47] >> Hi everyone, my name is Chitrang Di. I [0:50] am at a company called Avidity [0:51] Biosciences uh leading AI and analytics [0:54] and I've been in life sciences and uh uh [0:57] this pharma healthcare space for a long [1:00] time in my career. Uh used to work at a [1:02] medical device company a couple of [1:04] different ones uh before this. [1:06] Uh good afternoon my name is Harvir [1:08] Singh. Uh when I signed up for the [1:11] conference I was the chief data officer [1:13] of consumer bank for truest financials. [1:17] uh I'm not uh interestingly [1:20] today what I'm going to be talking about [1:21] is a little bit about what I'm doing in [1:24] my current role as well as the founder [1:25] of a wireless technology company uh [1:28] which is powered by blockchain and our [1:31] we just launched a rewards program on [1:33] Solana. So u I don't know if something [1:36] some of the some of the founders are [1:38] here as well. It's been a very [1:39] interesting journey for me to move from [1:42] you know corporate corporate executive [1:44] to a founder. So, we'll be sharing some [1:47] some of the roles because AI has really [1:49] helped me do some of my stuff uh as as [1:53] we talked about it. [1:56] >> Awesome. Thanks, everyone. All right, so [1:58] this first question is for Elaine. Um, [2:01] there's a ton of hype around AI [2:03] transforming analytics. We all talk [2:05] about it, but real change tends to [2:07] happen in specific areas of the [2:09] business. So can you share one concrete [2:11] example where AI either fundamentally [2:14] changed a business outcome or how teams [2:16] work and what made that project [2:18] successful? [2:20] >> Sure. Um I think I can share many. Uh [2:24] how many of you are Verizon customers? [2:27] If you can raise your hands. Okay, [2:29] great. Good turnouts. Thanks for your [2:32] business. Right. Um being a telecom [2:35] customer, you probably uh have a lot of [2:37] frustrations or historically used to [2:41] have a lot of frustrations about the [2:43] customer experience, right? On average, [2:46] I think telecom industry has is known to [2:51] have the lowest MPS scores compared to [2:55] some of the other industries. So um if [2:59] you are tracking Verizon news on June [3:02] 24th we have successfully launched a [3:05] project called the project 624 and that [3:08] project is aimed to launch a series of [3:14] uh customer experience enhancement [3:16] capabilities that's powered by AI. So [3:20] all the way through from at the [3:22] beginning when customers started [3:23] contacting us at the digital channel [3:26] using AI to help guide the personalized [3:30] uh research and provide personalized [3:34] offers, answer questions, help ease [3:37] customers frustrations and all the way [3:40] through to within the app you can have a [3:43] human-like dialogue with your uh AI [3:47] assistants. And when customer wants to [3:50] be transferred from digital channel to [3:53] human assisted channel, then we offer AI [3:58] empowered capabilities to help guide our [4:01] core center agents to provide dynamic [4:04] conversations and a lot of backgrounds [4:07] about that customer's information so [4:10] that we can help improve customer [4:12] satisfaction, reduce churn and this [4:16] initiative was very successful. As a [4:19] result, we had a great Q2 earning uh [4:24] results announced to the street and uh [4:28] not only we um proved our investment in [4:31] AI was successful, but also we were able [4:34] to raise our expectations to the street. [4:37] So that was a great example. But all of [4:40] those were delivered and a really really [4:43] tight time frame. So that really [4:46] requires us to use laser focused [4:49] execution and not only we standardized [4:53] our framework build a process to manage [4:56] prioritize our AI use cases but also we [5:01] um add on top of that some because of [5:04] those standardizations and all the way [5:06] from the rigors we added on identifying [5:10] KPIs measurements at the business case [5:13] level that help us to be able to really [5:17] focus on uh capturing business values ac [5:20] across all the processes at every step [5:23] in our framework but also we're able to [5:26] leverage our experimentations [5:28] multivariant testings to clearly [5:31] articulate the incremental values that [5:34] were brought through those AI [5:36] capabilities and eventually we create a [5:39] collective feedback loop to help retrain [5:42] the model, perfect the model so that we [5:44] have a perfect ecosystem and all those [5:47] were tied back to our business value not [5:51] just from the technical investment [5:53] perspective and that really helped us to [5:56] prove to our board to prove to our [5:59] senior leadership team that all those AI [6:03] investments were successful. So that [6:06] allowed us to remain laser focused and [6:09] then keep reinventing and prioritize our [6:12] AI capabilities along the way. All those [6:16] standardizations also enable us to add [6:19] some agentic automations which all the [6:22] way from unified intake to creating juro [6:26] ticket to use our um GitLab code and [6:30] help us to speed up that go to market [6:33] process when we launch those AI [6:35] solutions that enabled us to really be [6:38] able to deliver those solutions in a [6:41] very tight u time frame. So that was a [6:44] very successful example recently we just [6:47] launched. [6:50] >> Right. Anyone else have anything to add [6:51] or [6:52] >> Yeah, I can I can add a little bit. So [6:54] one of the experiences in the past uh [6:58] that has really seen very interesting [7:00] outcome is we've all collected you know [7:03] financial institutions all collect data [7:06] in different places. Uh it's all siloed. [7:09] We've all like you know I was in [7:11] consulting before and we've also helped [7:13] other financial institution bring all of [7:15] the data together but one of the things [7:16] that was always lacking is the so what [7:18] right you have you build all these [7:20] models you have a churn model I have an [7:22] RFM model I have I'm predicting this I'm [7:24] predicting that all of those decisions [7:26] were taken in a in a silo depending on [7:29] the channel the outbound channel [7:31] depending on how the customer is [7:32] interacting with you one of the things [7:34] that has really helped which is what we [7:37] were implementing uh back in where I was [7:39] was to start looking at this as [7:42] collectively as a pulse the client pulse [7:45] and this is where you know we actually [7:48] got an award at the American Bankers [7:49] Association earlier this year uh the [7:52] technology is patented by truest it's [7:55] more around how do you measure the pulse [7:58] of the client what weight do you put on [8:00] that pulse and how do you drive the [8:04] outcome of that pulse to actionable [8:07] outcomes in the channels whether Whether [8:08] it's with the bankers, whether it's on [8:11] the machines that they interact with, [8:12] and it doesn't have to be just the app. [8:14] It could be an ATM machine that they [8:15] interact with. It could be a teller [8:16] machine or it could be a teller machine [8:18] that interact with or it could be a [8:19] point of sale that can interact with. So [8:21] they you have opportunities at every [8:23] single point do something about it. Back [8:27] in the days, the technology was not [8:28] there. But I think with all the edge u [8:32] applications applications which have AI [8:35] embedded into it with the right guard [8:37] layers Salesforce a great example with [8:38] the right guardrails you can start [8:40] actioning on those uh data points uh [8:44] across the board. So look at some of the [8:46] data being federated some of the actions [8:48] and decisions being being uh done in in [8:51] the channels when the customer is [8:53] interacting with you real time u to make [8:56] it very very actionable. So that I think [8:58] you know was very very powerful to see [9:01] otherwise it took like years to roll out [9:04] a simple change in a end application [9:08] with DAI technologies all can be plugged [9:11] in and pushed very very fast in the [9:13] channels making the customer's life [9:15] easier. [9:18] >> Okay great thank you both. Um so this [9:22] next question is for Harveier. So we [9:24] know that AI can surface um some [9:27] insights that traditional analytics [9:29] might miss but that tends to introduce [9:32] more costs um complexity and maybe even [9:35] confusion for business users. So in your [9:38] experience where have you faced those [9:40] tradeoffs between pushing advanced AI [9:42] capabilities but also keeping things um [9:45] usable, explainable and trusted? Yeah, [9:49] for all the people that are sitting here [9:50] who are who represent the business side [9:52] of the house, raise your hands. [9:55] You you're going to be a nightmare for [9:57] your tech teams when you go back because [9:59] you're going to all learn about some [10:01] stuff and you're going to tell them, "I [10:02] want this." And they'll scramble to [10:04] figure out how to make it happen for [10:05] you. But then again, this is where the [10:08] the sauce lies, right? What is it? What [10:11] is the outcome that you guys are trying [10:13] to get to? Because that is the question [10:15] nobody asks. We the first step is can we [10:19] do a use case? Why? Why do you want to [10:22] do a use case? What is the problem that [10:24] you are trying to solve? Do we have the [10:26] appetite to do that now? What are your [10:29] business priorities? So, one of the [10:30] things that I found missing in most of [10:32] the Gen AI use cases within I can talk [10:35] about financial services is the why. We [10:38] don't ask why enough. We are not asking [10:40] why enough. We are not tying any Gen AI [10:42] use cases to ROI because we don't even [10:44] know how to measure it because we don't [10:46] know how much it's going to cost to run [10:47] these models in in those production [10:49] environments. A lot of this is just [10:51] you're seeing and then you see some [10:53] output in a conference and you're very [10:55] happy about it and then you go and start [10:56] to build something. [10:58] I bet 80% of these use cases will be [11:01] abandoned by the end of this year. Only [11:04] like 15 20% will make it. But if those [11:06] 15 20% will make it, put the thought [11:10] behind the why now before you actually [11:12] start implementing it. [11:15] Why is just one of the parameters. The [11:17] other parameters in my mind is is your [11:20] data ready for it? Which means is the [11:23] quality of the data ready for it? Is the [11:26] the type of data that you are collecting [11:28] has variety in it? Is the frequency of [11:31] the data that is being collected is is [11:33] available. [11:35] I have run data leaks where the data is [11:37] stale for many years and they're like [11:39] yeah let's run genu geni use case on top [11:41] of that data what is the value the [11:44] customer behavior that you're trying to [11:46] predict for customers that have long [11:48] gone or they're not even interacting or [11:49] they're not there people's choices [11:51] change their interactions change so some [11:53] of these are short-lived right so you [11:56] have to start thinking of what is it [11:58] that you're trying to do and then start [11:59] to look at that data around it which is [12:02] very important so the maturity of data [12:03] is important [12:05] And then the governance part of it, [12:07] governance security is extremely [12:08] important. The [12:11] coming from a regulated industry a lot [12:14] of times we we cannot use certain things [12:16] in certain types of uh uh interaction [12:20] with our customers. I can't use uh some [12:24] of the data that I have for marketing [12:26] use cases. Where are those boundaries [12:28] and how do you prevent the intermingling [12:30] of that data as well? how much of the [12:32] data that you are sourcing from outside [12:33] that you have trust in it to be able to [12:36] advise the customer as well. So there is [12:38] there is these guard rails that you need [12:39] absolutely need now in order for you to [12:42] be safeguarded in the future. So before [12:45] you actually start any use case if if [12:48] not done build those guardrails now [12:50] before you go down that path and then [12:53] finally the ROI it's you know it is [12:57] cloud was sold to most of us and I was [13:01] one of them I used to sell this to my [13:03] clients is like this is great you know [13:04] it's going to save you hundreds and [13:06] millions of dollars uh in cost has it in [13:10] my opinion no it's at least the same or [13:13] more It's definitely made life easier, [13:15] but it's definitely not cheap. AI is the [13:18] same way. It's may not may not be [13:21] cheaper over a period of time, but it'll [13:23] get you the outcome fast. And that's [13:24] what you need to think through when you [13:26] are measuring these these um uh AI use [13:29] cases that you're building. [13:32] >> Yeah. [13:33] >> Yeah. So, I I think I just to add to uh [13:37] to the points, right? Um maybe three [13:39] main things um based on my personal [13:42] learning. The first one is really uh [13:46] focus on identifying your business value [13:49] at the beginning of your project or at [13:52] the beginning of your product [13:54] development. I know throughout the years [13:57] uh we've been hearing oh value creation [14:00] business value you'll be thinking what [14:03] does it matter right your technical team [14:06] does not care but as a business leader [14:09] as a product owner it matters a lot [14:12] because if you go talk to your business [14:15] stakeholders who has the funding that [14:17] you wanted to you know have them invest [14:20] on your product you have to be able to [14:23] talk a language and justify what matters [14:26] to them and none of your CEO CFOs going [14:30] to care about well when you started [14:33] talking about cost per token they do [14:35] your CTO may still do but what's mostly [14:40] important and matters to your business [14:41] owner is how much revenue you bring in [14:44] to the door how how many growth you can [14:46] help me to increase this year what's the [14:49] cost saving I can make if I run the core [14:54] centers based on a you know a gentic [14:56] solution based on your AI solution [14:58] versus an offshore solution how you [15:02] justify identify those value are going [15:05] to be super important to to decide how [15:09] much funding you're going to get whether [15:11] you are a private owner of your AI [15:13] product or you work for a company right [15:15] and all those uh have to be determined [15:19] at the beginning especially how you [15:21] define and measure the success [15:23] I know it's a painful process. Trust me, [15:26] I personally had gone through this for a [15:29] long time and sometimes still [15:31] struggling. But if you do this right at [15:34] the beginning, you'll be a much easier [15:37] journey down the road. So you don't have [15:39] to go back and forth talk about to [15:41] stakeholders and and try to argue what's [15:43] what's right, what's wrong, right? And [15:45] it'll be much easier later on if you [15:48] have agreed upon KPIs for success [15:50] measure at the geo even with your senior [15:53] leadership team with your board of [15:55] directors with your CFOs you wanted to [15:58] lay out up front say here's how we're [16:00] going to measure our AI initiative [16:03] success and second thing I wanted to say [16:06] is you know don't put your technical [16:09] solution way ahead of time so you [16:12] neglect what exactly the problem you're [16:15] trying to solve. Focus on what problems [16:18] you're trying to solve because a lot of [16:20] times you're going to go through a lot [16:23] of technical shiny new toys. Your [16:26] technical team are going to be really [16:28] keen to get their hands dirty, start [16:30] playing with all the new technologies, [16:33] right? But is that the right solution? [16:35] Maybe yes, maybe no. Maybe you need a a [16:39] combination, a hybrid approach to solve [16:42] that particular solution. or maybe you [16:44] don't even need AI to solve that poll to [16:47] solve that particular problems right as [16:50] a leader you have to be able to uh [16:53] juggle different desire and do that [16:56] balancing game so really focusing on [16:58] what problems you are trying to solve [17:00] helps like couple of years ago it's all [17:03] about geni everything you talk your [17:05] business leaders who doesn't even do [17:07] codings may come to you say I want a gen [17:09] solution but you're going to be able to [17:12] ask is GNI the right solution for your [17:15] problems maybe not right aentic is now [17:19] very hot everyone is talking about we [17:21] want aentic this agentic that but can [17:24] you really do agentic at the beginning [17:26] you may not a lot of people [17:29] underestimate the data quality issues [17:31] they started they think we can do [17:33] agentic because we already have all this [17:36] knowledge base we cleaned up all the [17:38] data really when you start doing that [17:41] validation you'll know your model result [17:43] is not persistent. How you guarantee the [17:46] hallucination will never happen once you [17:48] deploy to the production. How do you [17:50] monitor your code and make sure you have [17:52] that responsible AI governance in place [17:55] so you comfortably enough to launch a [17:58] commercial solution that you would not [18:00] put your risk yourself and your team [18:02] under risk for any potential lawsuit. [18:05] Right? So all those questions you have [18:06] to think about. And then lastly is [18:09] really just don't try to boil the ocean. [18:12] Be agile. Well, a lot of people talk [18:14] about agile. That really is, you know, [18:16] baby steps. Do MVP first. Iterate. You [18:20] were going to need iterations and many, [18:24] many iterations. Do experimentations, [18:27] you know, stick to your product [18:29] management principles. Do the [18:31] experimentations. Don't do full roll [18:33] out. You can throttle uh throughout the [18:36] way and keep iterating. Collect that [18:37] feedback loop. That will help you a lot [18:39] down the road. [18:42] Thank you. Um, just before we move on to [18:44] the next question, I have a I have a [18:46] question for the audience. How many of [18:48] you feel like your data analytics and [18:51] business intelligence, you know, your [18:54] culture, your platforms at your company [18:56] are in phenomenal shape, really [18:58] advanced, and now you're just tackling [19:01] AI? [19:02] How many of you feel like you're doing [19:04] it at the same time? Data and analytics [19:06] and AI at the same time. Yeah, I get it. [19:10] that is that is very real right now. Um, [19:13] so something that I want to add in just [19:15] from my own experience is that it's it's [19:18] this dance between the two of them and [19:21] figuring out how you're advancing your [19:23] company in both. And and one of the [19:26] things there is is assessing where you [19:29] are on sort of both of those maturity [19:31] curves [19:33] and recognizing, you know, and you [19:36] touched on this, [19:38] recognizing uh where you want to push [19:41] your team forward. And by your team, you [19:44] mean the business team, as opposed to uh [19:47] pushing too hard and getting a lot of [19:49] resistance. And so I think what what [19:52] I've learned in my experience is that [19:54] you may be starting at A and focus on [19:57] how to get them to D because as [19:59] technical people or people in data it's [20:01] so fun to say oh no but there's Z over [20:03] there we're going to have machine [20:05] learning we're going to have agentic AI [20:07] options and so instead figuring out how [20:10] to push forward just enough to say okay [20:14] what what is A to D what is D to H that [20:17] sort of thing as opposed to going for [20:19] the long term because [20:21] pushing folks too hard toward the very [20:24] end, you may be able to see it, but not [20:27] everyone will. And that's where you're [20:28] going to get some of the biggest [20:29] resistance and it's going to turn into [20:32] shiny object [20:34] little bit more than actual business [20:36] impact. So anyway, that's just my two [20:38] cents. Um, all right. This next question [20:41] is for uh Chichang. Many organizations [20:44] are grappling with how to actually embed [20:47] AI into their existing BI tools. Let's [20:49] be honest, I don't know about you all, [20:51] but DOMO has Domo AI agent. I think [20:53] you've got one in Tableau. What does [20:55] that really mean? Um, so versus kind of [21:00] redesigning everything to have an AI [21:03] first analytics capability. So, how are [21:06] you approaching that choice sort of that [21:08] trade-off and what challenges um or [21:10] advantages have you experienced in with [21:12] both? Yeah, this is a this is a great [21:15] question and as you noted rightly uh [21:17] every BI tool out there is building AI [21:19] into the product. So there's definitely [21:22] um a push towards that. Um I'm [21:25] approaching this uh in a couple of [21:27] different ways. So there are a couple of [21:29] different models um and and dashboards [21:32] the the death of dashboards has been [21:34] forecast for a long time. I don't think [21:36] dashboards are going anywhere. Uh they [21:39] will evolve they will uh become smarter. [21:42] uh but at the same time we know a lot of [21:44] you know a lot of you raised hands when [21:46] we said who is from the business right [21:48] there isn't patience for a a a two week [21:51] two month 3 month development cycle for [21:53] a dashboard so people want to move fast [21:56] uh and so I see these two models [21:57] evolving one is this this co-pilot type [22:00] of model where you have your dashboard [22:02] and you have an a panel on the side uh [22:05] where you can ask questions about that [22:07] dashboard or the underlying data behind [22:09] that dashboard and this is highly [22:10] interactive I think see I see that [22:12] across other apps as well. So that's [22:14] that's one model the co-pilot model. Uh [22:17] the second one is uh really exciting. [22:21] This is the kind of and an evolution of [22:24] AI as well from search to research to [22:28] deep research thinking um where if you [22:32] have a data set, can you let a agent [22:35] loose on it uh for a couple of hours and [22:37] and generate insights for you that a [22:40] human may have taken a couple of weeks [22:42] to generate? and not not only that, [22:45] write that into a nice research report. [22:47] And we've seen those and and it's really [22:49] mind-blowing what you can get out of [22:50] these u and so really it's this second [22:54] model gives analysts and and business [22:57] users superpowers to actually go beyond [23:00] asking questions and and surface new [23:02] insight. So those are the two models [23:04] that I'm seeing and that we are actually [23:06] implementing as well. [23:09] >> So maybe I can add something right. So [23:11] like Ranjan said, everyone's offering an [23:14] AI service embedded in it. [23:17] It's good but it's also puts another now [23:19] pressure on top of us when we back in [23:22] the time when we used to talk about okay [23:24] are you using the right data to make a [23:25] right decision. The question now is are [23:27] you using the right model and the right [23:29] data and the right tool to make a [23:31] decision right so it's now just [23:32] cascading effect on top of it because [23:35] what's going to happen is u some vendors [23:38] sell to the technology when some vendors [23:40] sell to the business and the the the [23:44] computation that is happening with a [23:46] smaller subset of data will have [23:47] potentially different results versus [23:50] something that is h having being done [23:52] centralized. So there's a trade-off. You [23:54] have to figure out right that model how [23:55] that is happening. And then again when [23:58] you if you guys walked in and you looked [23:59] at the AI4 uh that big screen with the [24:03] sponsors I have never heard of half of [24:05] the companies as well. So I cannot make [24:08] a judgment whether they are good bad [24:10] what is happening what kind of so for me [24:12] as a as an executive it is very very [24:14] hard to figure out what to use and how [24:16] to use it. There's constant constantly [24:18] new stuff coming and coming. So I call [24:20] it like the u the frozen yogurt [24:23] syndrome. You go to a frozen yogurt, you [24:25] have like hundreds of things, right? And [24:27] you mix and match things and you come up [24:28] with like eh I don't want it, right? So [24:31] it's going to happen over and over [24:32] again. That's why they say they have a [24:34] little tasting and you start tasting a [24:36] little bit. Don't go deep into it unless [24:38] you are absolutely sure this is the [24:39] flavor you like and it fits your budget. [24:42] It solves for it and it is something [24:44] that your enterprise is vested in it. [24:47] Maturity over um sometimes maturity in a [24:51] tool is always good uh than you know [24:54] experimenting with something that you [24:56] might expose your data or you know might [24:58] end up with in a in a in a scenario that [25:00] you don't like. So so yes you might get [25:02] things done faster here but then [25:04] maturity of the tool the team that is [25:06] behind it the what they have done in the [25:09] in the references is definitely going to [25:11] help. [25:12] >> Yeah. So uh I think when it comes to AI [25:15] capability within like a BI tool or some [25:19] other analytics tool um the often battle [25:23] we're thinking is build versus buy right [25:26] do you build your own or do you want to [25:28] buy it off the shelf I personally would [25:32] advocate for a buy solution especially [25:36] when it comes to like a BI visualization [25:39] tool because nowadays is I think the [25:42] technology is so advanced and it's much [25:45] easier you know if you already have an [25:49] in-house BI solution in place right you [25:53] may need to upgrade it migrate to cloud [25:56] but you will be able to use that native [25:58] AI capabilities that's built so [26:00] seamlessly with the tool that you [26:03] already have without having to invest [26:06] heavily on the labor resources and try [26:09] to build your own but uh you still need [26:12] to be very careful cuz that's not [26:14] trouble-free cuz obviously there's no [26:17] one sizefits all. So you cannot expect [26:21] that one particular capability will help [26:23] you solve all the problems. You will [26:25] still have some edge cases that you may [26:28] not be able to use that capability to [26:30] perfect. You may still wanted to [26:33] customize a little bit if you can [26:35] without having to invest heavily, right? [26:38] Otherwise you lose the merit to to buy [26:40] the solution off the shelf. But sitting [26:43] in the centralized [26:46] um chief data analytics and AI office, [26:49] the challenge I personally encounter [26:52] would be how do you make sure all the [26:55] different AI capabilities that resided [26:58] within that native tool render [27:01] consistent results to different business [27:04] owners. Right? It's hard to control [27:06] because a lot of solutions were so [27:09] embedded in that black box within their [27:11] ecosystem. There's no visibility into [27:14] that AI model behind the scenes. So you [27:17] know what's wrong, what's right and the [27:20] way that um we solve the problem is [27:23] really to uh reinforce build that [27:26] enterprise level semantic layer and [27:29] reinforce the standardization and [27:31] rationalization of semantic BI layer [27:35] informations and we also embedded JAI [27:39] solution within our data marketplace so [27:41] that you can really help different [27:44] personas on boarding to the right data [27:47] at the right timing without uh over long [27:51] processing that people would be [27:52] resistant and also launch a series of [27:55] data literacy programs that are [27:58] meaningful to different personas so [28:00] people fully understand what the data [28:02] behind the scene is. That way when they [28:05] start using those AI capabilities with [28:08] the native BI visualization tool, you [28:11] will not encounter those challenges that [28:14] you cannot explain or cannot control. [28:16] Right? So that would be my personal [28:18] experience to share with you. [28:21] >> Yeah, absolutely. Um something that I'll [28:25] add on, I don't know if the moderator is [28:26] supposed to talk, but I'm going to make [28:28] that executive decision. Um, something [28:30] that that I'll add on in that as someone [28:33] who's going through building AI and BI [28:36] together, you know, you brought up a [28:37] really great point about people don't [28:38] have the appetite anymore to wait 2 or 3 [28:41] weeks or a couple months for a [28:42] dashboard. And at the same time, a lot [28:45] of these tools have uh BI or AI [28:49] capabilities embedded and someone can [28:51] just go say, "Oh, tell me how many [28:52] customers do we have?" Well, as I'm sure [28:55] most of you know, unless you've got [28:56] really good metadata, that is a very [28:59] very scary thing that someone could come [29:01] up with a very bad answer. Um, so one [29:05] thing that I've seen companies do is [29:07] they empower the data analysts with [29:10] really good AI tools to speed up the [29:12] delivery of that dashboard. Um because [29:15] then then you were talking about what [29:17] are the capabilities here with AI and [29:20] saying oh look we're using it to power [29:22] up our data team and then deploying AI [29:25] in other ways in the company that maybe [29:28] feel a little bit safer. Um I you know I [29:31] was talking to someone about AI tools in [29:34] in uh Domo or Tableau and I said you [29:37] know what if you don't have good [29:38] metadata then essentially you were [29:40] powering the world's worst Rube Goldberg [29:43] machine just going going and before you [29:46] know it someone's got a totally wonky [29:48] number. And so it's it's balancing AI [29:52] strategy [29:53] to to power your data and then doing AI [29:56] in other places that maybe don't have [29:58] the the same high stakes mistakes. [30:02] Um, hey Brad, good to see you. [30:05] >> Sorry my uh geni agents goals were not [30:08] aligned with mine. [30:10] >> So a little more reinforcement learning, [30:11] it'll get there. I'll be on time next [30:13] time, but sorry about that. [30:14] >> Glad you're here. Glad you're here. Um, [30:16] this question is for you. So, you know, [30:19] as we've talked about, AI is very much [30:21] changing who does what on analyst teams [30:24] and maybe even who is an analyst. So, [30:27] what kind of shifts are you seeing in [30:29] the roles and skills um required on [30:31] teams as III kind of takes over and how [30:34] are you approaching talent de um [30:36] development and managing that cultural [30:38] change? [30:39] >> Yeah, sure. So, thanks for having me. [30:40] I'm Brad Erenss. I'm senior vice [30:42] president of advanced analytics at [30:43] FINRA. We're a financial services [30:46] regulator. So we regulate uh all the [30:48] broker dealers uh in the United States [30:50] and some that are abroad uh that have [30:52] operations here in the domestic markets. [30:55] Um the way we approach learning now [30:58] across the advanced analytics space [31:00] which now includes generative AI is we [31:02] really take our user population or our [31:04] staff and sort of divide them into three [31:07] different categories. Uh we have [31:09] explorers those are the people that are [31:11] using the outputs of advanced analytics [31:13] and generative AI on a daily basis. Then [31:16] on that's on one extreme and then on the [31:18] other extreme we have data scientists [31:20] and data analysts and what we're really [31:22] trying to do is not make everybody in [31:25] the explorer pool a citizen data [31:27] scientist. That's not going to work. Uh [31:30] that's a very very challenging thing to [31:32] do. So what we do is we look to take a [31:34] group of those explorers, we put them [31:37] through a university-led program that [31:39] we've uh built with a major university [31:42] um and we just try to increase their [31:44] knowledge, fluency, and understanding of [31:46] advanced analytics capabilities. We're [31:48] not teaching them code. They're not [31:50] going to learn Python by themselves in [31:52] this course, but really what we're [31:53] trying to do that is get 5 to 10% of [31:56] that staff population into something we [31:58] call catalysts. Those are people that [32:00] can bridge the gap between day-to-day [32:02] business activities. They're subject [32:04] matter experts. They know the regulatory [32:06] environment. They know their regulatory [32:08] responsibilities and nudge them over so [32:10] they can bridge the gap to our data [32:12] scientists, our experts in generative [32:14] AI, um, and our data analysts to really [32:17] bring solutions right to the level where [32:20] they're at. We we feel that we have to [32:22] meet our people where they're at. So our [32:24] learning programs, we have one for our [32:26] entire staff and then we also have a [32:28] specialized program for leaders that we [32:30] bring them through to really just build [32:33] knowledge fluency and understanding. [32:35] Again, once you inform your leadership [32:38] of what you can do with uh AI or [32:41] analytics, it really, you know, kind of [32:43] the light bulb goes off and you can get [32:45] better solutions, a better backlog of [32:48] use cases and you can actually sort [32:50] through those use cases and really look [32:51] for value. And that's one thing that we [32:53] also stress on with people is try to [32:55] find value in it because we don't want [32:57] to build uh efficient inefficiencies by [33:00] just dumping generative AI everywhere. [33:02] So we're being kind of cautious in the [33:04] way we train around it as well. So one [33:06] of the things that I've uh kind of been [33:09] observing is that yes, a lot of people [33:11] have become like AI experts just because [33:14] they've done some chat GPT and all of [33:16] that stuff. Well, that's not a bad [33:18] thing. Uh I myself use a few of them [33:20] because I'm not a native English [33:22] speaker. I I grew up in India. I feel [33:24] that that's a gap that I always I'll [33:25] have to live with. No matter what I do, [33:27] I'll not be perfect. So I use that to my [33:30] advant I use these tools to my [33:32] advantage. Why? Because everybody else [33:34] around me will become the same. Right? [33:37] uh at certain point my my database of [33:40] knowledge will exhaust and I have to use [33:42] certain augumented uh I would say help [33:46] to make myself better because you go to [33:48] a meeting and you're not well prepared [33:50] in today's age probably is not an excuse [33:53] because don't say you don't have time it [33:56] takes one prompt to learn something very [33:57] very simple right and at least get aware [34:01] before you go so the new term that I've [34:03] been hearing and I think it's very [34:05] interesting is called the AI whisperer [34:06] right? Become a little bit of an AI [34:08] whisperer. Everyone everyone around this [34:10] has we don't have to be scientists. We [34:12] don't have to be this. But having a [34:13] little bit of a whispering methodology [34:15] that can get you get your day faster and [34:18] can you know alleviate some of the [34:20] pressure which is I have a meeting with [34:21] my boss in the morning. I don't know [34:22] what to say to him. Uh the results were [34:24] then weren't that good. Yes, don't make [34:26] an excuse but at least there is a right [34:28] way of saying this. There's a right way [34:29] of you know uh coming up with this [34:31] thing. So what Brad said was at a very [34:34] enterprise era. I'm just using it as you [34:36] know I became a founder of a a a [34:39] blockchain uh and what I found is there [34:43] are so many of them out there that I do [34:45] not know about before I'm meeting [34:46] someone before I'm meeting an exchange [34:47] I'm meeting someone I try to do a little [34:49] bit of research within like 5 10 minutes [34:52] of this it helps me get up to speed on [34:55] what's happening in the market so [34:56] becoming an AI whisperer I think it'll [34:58] be a absolute necessity for everyone [35:02] that's out there and as a matter of [35:04] Right. When your kids are getting into [35:05] like before college, get them to learn [35:09] how to do AI whispering. It'll it'll [35:10] help them. [35:12] >> Yeah. I I I think uh I couldn't agree [35:14] more that uh um when you said we have to [35:18] meet our people where they are, right? [35:21] And I I think maybe my answer is a [35:23] little bit uh contradictory. I I don't [35:26] think AI is going to eliminate a lot of [35:30] rows, positions and needs from the data [35:33] analytics space. What will happen is we [35:36] probably would not need as many as we [35:39] thought it would be down the road, [35:41] right? But we would still need people [35:43] who can write Python, who can code, [35:46] maybe the tool make their job a little [35:48] bit easier. They can, you know, turn [35:49] around much faster. What's more [35:52] important is really to make sure the [35:56] team has a open mindset and they can be [35:59] innovative and creative which I think [36:03] when we're hiring a team of analyst data [36:06] scientists we do look for those trades [36:09] anyway it it's just matter of you know [36:12] nowadays it's more uh it's more [36:15] important than ever right you have a [36:17] very dynamic team that's open and [36:19] embracing new technology technologies [36:21] because we're all learning. We're all [36:23] learning together. Doesn't matter if you [36:25] are individual contributors, you are [36:26] team leaders. Technology changes so [36:29] fast. You just have to keep learning, [36:31] keeping an open mindset. As long as you [36:33] have that culture built, everyone is [36:36] innovative and creative. And I don't [36:39] think anyone will be really retired or [36:42] washed out because of the AI per se. [36:46] >> I just want to add one more point from [36:48] uh you asked about talent. Uh so I I'm [36:50] noticing this interesting. I hire a lot [36:53] of people. Folks reach out to me all the [36:54] time as well. Um and I'm seeing this [36:56] really interesting that the market seems [36:58] terrible for uh entry level [37:01] CS and computer engineering graduates. I [37:04] don't know if you noticed that if you [37:05] know any young people. Um [37:08] at the same time you see these these [37:11] news offers from uh from Zuckerberg for [37:13] for hundreds of millions of dollars for [37:15] AI researchers. So this is a really [37:18] strange market. can be demoralizing for [37:20] young people as well as at the same time [37:23] you know not knowing what's what's going [37:24] on. So um the the advice I give and and [37:27] what I'm implementing against uh with [37:30] with people I work with as well is is um [37:32] get proficient with the with uh coding [37:35] agents uh they really I've seen this uh [37:39] the folks who have the technical depth [37:41] have really get superpowers with uh [37:44] using these tools correctly. So you know [37:46] foster that culture of learning and [37:47] sharing across your organization is what [37:49] I found a lot uh to be very helpful. [37:52] >> Yeah. Yeah. Absolutely. All right. So to [37:55] uh start wrapping this up, let's just go [37:57] around and uh let's talk to everyone [37:59] here. What do we think one one um [38:02] emerging trend or AI capability that we [38:05] think is really going to be [38:06] transformative for analytics and uh BI [38:09] in the next let's say 3 to 5 years? [38:11] Let's start with you Brad. Uh right now [38:14] I think what's what's going to be out [38:16] there probably in the next year or two, [38:18] 3 years is probably too long, uh is [38:20] really just using uh chatbots and [38:22] general conversation against analytics [38:25] queries that are already built. Um, this [38:27] is basically saying that if you've got a [38:29] pool of queries that are already battle [38:32] tested, are ready to use, are accurate, [38:35] uh, why not put a chatbot over the top [38:36] of them with some very strong metadata [38:38] where your people can come in and say, [38:39] "Hey, I want to find in my space it [38:42] would be I want to find all trades in [38:44] Apple that occurred between 9:30 and [38:46] 9:31 for this broker dealer." Right? [38:48] What the what the what the chatbot will [38:51] do is it'll go find the right query. So [38:53] again, we're not creating citizen data [38:54] scientists, but we're going to take them [38:55] to the right query, and then that [38:57] chatbot's going to come back with, okay, [38:59] what symbol do you want? What time frame [39:00] do you want? Which broker do you want? [39:02] And ask them the appropriate questions [39:04] that will bring analytics really, really [39:06] close to the people and they don't have [39:07] to learn SQL. [39:10] >> Uh, so I'm going to say something very [39:12] different, which is I look at content [39:15] and the way the content is consumed and [39:17] the way the content is created. [39:20] For me, creating content has become much [39:24] much more easier. But on the other side, [39:27] it is very difficult to differentiate [39:28] between what is real and what is not [39:31] real. So that will be an interesting [39:33] thing that to see over the period [39:34] because these models are becoming very [39:36] interesting. You know, I ran a model [39:38] that put me on top of a of a building in [39:41] uh in Manhattan with like a one little [39:43] query. So I'm like, damn, that looks [39:45] real, right? So something that is very [39:47] interesting and weaving in my blockchain [39:50] world consensusdriven architectures are [39:53] great but if everybody's wrong that that [39:56] is also wrong. I feel AI can do a little [39:59] bit of better job in building that next [40:02] layer on top of the consensus driven [40:04] architecture to point out anomalies and [40:06] to point out things when the consensus [40:08] could potentially be wrong as well. So [40:10] that's something that to to watch out. [40:12] Um I personally think that there is an [40:14] advancement on top of blockchain uh for [40:16] AI as well which I'm very excited about. [40:19] >> Uh to me it's um really I I said earlier [40:23] dashboards aren't going anywhere. Uh I [40:25] think SQL is is still uh still [40:27] important. Uh um but I I appreciate your [40:32] point there. Uh so I'll take it a step [40:35] further. I what I'm building towards and [40:38] what I'm looking most looking forward to [40:40] to your point is really uh having these [40:42] agents really really be um um augmenting [40:47] human potential meaning um we've been in [40:51] organizations where quarterly business [40:53] reviews the decks for those take weeks [40:55] to prepare um um I would like to get to [40:59] a point where analytics AI and BI is [41:02] really helping generate those decks and [41:04] those reports with factbased [41:07] information from your organization [41:09] automatically [41:11] um and and be able to really facilitate [41:13] the decision making which has been the [41:14] goal of of AI and and and or which has [41:16] been the goal of BI for a long time. [41:19] >> Yeah. So I think the next hype cycle is [41:22] going to be around aentic AI uh [41:25] especially multi- aent platform um [41:28] because it's has such a broad use cases [41:31] and you can technically deploy it in [41:34] every single use case that you can think [41:36] of right automation and when it's done [41:38] correctly you can really see significant [41:42] improvements on your productivity and [41:44] efficiency so it's amazing [41:46] >> all righty um I think mine is uh you [41:50] know right now in in analytics and BI we [41:52] very much focus on our own systems we're [41:54] pulling data from systems we're putting [41:56] it in a warehouse that sort of thing and [41:58] um I think this is really going to open [42:00] up the doors to the data that we have [42:01] access to so more online data publicly [42:05] available data how are we using AI to [42:07] scrape it and then leverage that um you [42:10] know we're not going to have that [42:12] beautiful purity of of a warehouse and [42:14] everything being fantastically [42:15] structured and it it could certainly [42:17] open up the door to analytics And it may [42:19] not be quite as um clean or precise as [42:22] we've as we're used to, but it could be [42:24] a lot more interesting as we incorporate [42:26] more data sources. So that's my two [42:29] cents.