Edward Jones Uses Tactical AI Sprints to Secure Enterprise Buy-In

Edward Jones利用战术性AI冲刺赢得企业支持 为小型业务部门快速部署自动化能够培养出一线倡导者，从而为多年期企业级AI项目争取高管预算。 立足于Edward Jones在北美的业务规模，演讲者介绍了一种双轨制AI投资组合模型，该模型平衡了多年期转型项目与短期战术性成果，以培养一线倡导者并争取高管层面的支持。会议从“人员、流程与产品”的视角探讨了规模化扩展，强调技术与商业兼备的“双语”沟通能力、AI工厂框架，以及专有数据资产相较于通用商品化模型的价值。讨论还重新定义了数据就绪状态，指出先进AI需要长期的纵向行为历史数据，而非静态的当前交易记录。为确保项目成功投入生产，会议最后提出了促进跨职能共情的切实战术，包括不设议程的咖啡交流以及淡化数据科学与MLOps团队之间的角色边界。 Edward Jones 依托覆盖北美的约 20,000 名财务顾问与 800 万客户所沉淀的数据资产，坚持立足企业 1 年、2 年和 5 年战略路线图推进 AI 立项，优先破除业务结构性障碍而非追逐技术新潮。 (2:30) 团队推行双轨创新组合策略，将 6 至 24 个月的长期转型项目与针对繁琐流程的敏捷冲刺相结合；POC 的成功衡量标准不仅在于技术验证，更在于培育出能主动争取组织资源的业务代言人。 (2:30) 在围绕人员、流程与产品扩展 AI 能力时，企业需要培育跨越技术与商业边界的“双语”人才，并在基础大模型日益商品化的背景下，将专有数据与知识资产转化为不可替代的核心壁垒。 (6:04) AI 数据就绪度要求超越仅维护当前状态记录的传统数据治理，必须结构化存储随时间演进的行为变化轨迹（如捕捉授权委托书的历史连续变更），以支持老年金融欺诈识别等复杂分析。 (9:50)

Strategic AI leadership requires aligning multi-year enterprise goals with immediate business buy-in generated through targeted, tactical solutions. Drawing on enterprise deployments at Edward Jones and EY, Dr. Joseph Simonis and Sasha Ostojic argue that durable AI value relies on proprietary data architecture, process re-examination, and cross-functional empathy rather than commoditized models. Dual-Track Portfolios and Strategic Buy-in (2:30) Enterprise AI roadmaps typically span one-, two-, and five-year strategic horizons, yet transformational initiatives require 6 to 24 months to deliver results. At Edward Jones—which supports approximately 8 million clients across every U.S. county with nearly 20,000 advisors—Dr. Joseph Simonis manages this timeline gap through a two-speed portfolio. His team pairs long-term investments with short tactical sprints targeting acute operational pain points, such as automating a quarterly 40-hour manual process for a four- or five-person unit. Simonis observes that while the direct enterprise return of automating localized tasks is modest, delivering fast relief converts frontline teams into active internal champions. In partnership models that depend on consensus, technical validation alone is insufficient; proof-of-concept success is measured by cultivating vocal stakeholders who advocate for resources, staffing, and executive support during prolonged enterprise rollouts. Scaling Through the Three Ps Framework (6:04) EY's Sasha Ostojic frames the operational barriers of scaling AI across people, process, and product. For people, organizations must cultivate bilingual leadership by equipping technical staff with commercial literacy while teaching business units foundational AI concepts. For process, firms must establish AI Centers of Excellence and automated factories to avoid reinventing workflows, while fundamentally questioning whether legacy procedures should exist at all rather than merely using AI to accelerate an eight-hour flawed task into one hour. On the product dimension, Ostojic stresses that commercial foundational models are increasingly commoditized across sectors. Consequently, sustainable competitive advantage no longer stems from the model layer, but from how effectively an enterprise curates proprietary data assets and structures unstructured institutional knowledge. Structuring Data for Pattern Recognition (9:50) Traditional operational data integrity—ensuring records are accurate, secure, and accessible—is inadequate for sophisticated machine learning. Simonis explains that AI readiness requires preserving and organizing longitudinal data specifically to surface behavioral patterns and anomalies over time, rather than storing only current transactional states. To illustrate, Simonis highlights Power of Attorney (POA) designations on retirement accounts. While operational governance simply verifies that an active POA designation is valid, predictive models aimed at detecting senior fraud depend on identifying frequent, month-over-month modifications. If an organization discards historical revisions and presents only the current state, fraud patterns become undetectable. Ostojic adds that managing unstructured knowledge requires equivalent rigor through certified taxonomies and governance standards. Building Cross-Functional Empathy Across Delivery Teams (12:33) Deploying machine learning into production demands coordination across disparate groups, including business stakeholders, data scientists, data engineers, and MLOps specialists. To prevent transactional friction, Simonis institutes mandatory 30-minute, one-on-one kickoff chats without business agendas, directing team members to discuss personal backgrounds, families, and career trajectories. Establishing interpersonal rapport mitigates conflict during inevitable technical hurdles, shifting defensive interdepartmental disputes into collaborative problem-solving among familiar colleagues. Simonis also encourages cross-functional empathy by blurring operational duties, having MLOps engineers participate in executive business presentations and inviting data engineers into modeling workflows. Simonis and Ostojic conclude that scalable AI depends primarily on cultural trust, workflow elimination, and longitudinal data design rather than algorithm selection. However, the discussion leaves open how organizations should formally measure the financial return of foundational knowledge taxonomies and cultural alignment initiatives alongside standard software metrics.

安永（EY）咨询顾问 Sasha Ostojic 与爱德华琼斯（Edward Jones）AI 创新主管 Joseph Simonis 博士指出，企业 AI 战略的成效取决于能否将其紧密锚定在商业蓝图中，通过兼顾短期微赢与长期变革的双轨组合、重构面向模式识别的历史数据架构，以及建立跨职能同理心来持续兑现商业价值。 双轨制创新：以战术微赢获取长期战略支持 (2:30) 在服务约 2 万名财务顾问以及覆盖全美各县约 800 万个人客户的业务规模下，Simonis 强调，AI 举措的立项绝不能盲目追逐新奇技术，而必须首先对准企业 1 年、2 年与 5 年的战略路线图，在诊断出阻碍战略达成的结构性痛点后，再评估现有数据资产与建模能力。针对全企业级变革通常需要 6 至 24 个月周期这一现实，团队实施了双轨推进策略：在稳步推进长期转型项目的同时，持续穿插周期短、见效快的战术型项目。 这类战术项目的关键价值在于培育基层业务同盟。团队主动寻找饱受重复性繁琐工作困扰的小型业务单元（例如每季度需投入 40 小时进行手工流程的 4 至 5 人小组），虽然自动化此类任务直接产生的企业级投资回报率相对有限，但能迅速缓解业务痛点并积累好感。Simonis 认为，概念验证（POC）的成功标志不仅在于技术可行性验证，更在于能否将受益的业务团队转变为主动的内部倡导者；在合伙制企业结构中，这些一线业务支持者将在技术团队向管理层争取 18 个月以上的大型项目预算与编制时，提供决定性的支持力量。 规模化三要素：人才、流程与数据资产 (6:04) Ostojic 借助 Marcus Lemonis 的“人员、流程与产品”（People, Process, Product）框架，系统梳理了 AI 规模化落地中的关键障碍与应对路径。在人员维度，部门孤岛与沟通脱节是常见阻碍，企业需要重点培养“双语型”领军人才，既提升技术人员的商业思维与沟通能力，也训练业务部门掌握与 AI 协作的通用语言。 在流程与产品维度，Ostojic 提醒企业防范重复制造轮子以及“将低效甚至错误的既有流程自动化”的陷阱。企业应当建立 AI 卓越中心和具备复用能力的“AI 工厂”，在引入技术前先从本质上审视该流程是否具备存在的必要，而非仅仅将一项耗时 8 小时的陈旧流程压缩至 1 小时。此外，由于主流商业基础模型正日益趋于同质化，通用的算法模型已难以构成竞争护城河，企业的核心差异化优势完全取决于其独有的专有机构数据与非结构化知识管理资产。 数据就绪度：从单点状态转向时序行为分析 (9:50) 谈及数据资产，Simonis 强调仅满足传统运营标准（如受信任、安全与可访问）的数据对于高级 AI 而言远远不够。传统数据治理往往侧重于记录当前最新状态（Current State），但 AI 应用的就绪度则要求企业必须专门为了长期的模式识别与异常检测，完整留存并结构化时序行为数据。 Simonis 以退休账户的授权委托书（POA）场景说明了这一差异：常规业务合规仅需确保当前的 POA 记录合法有效并绑定至账户；但在防范老年人金融欺诈的 AI 场景下，核心的风险警示信号是该账户在数月内被频繁更改 POA 归属。若企业仅对外提供最新单点数据而归档了历史变更记录，AI 算法便无法识别潜在的欺诈行为模式。Ostojic 补充表示，非结构化数据的知识管理是当前数据就绪度的核心，领先机构正在推行形式化的知识分类法与知识认证标准，其严谨程度已逐步看齐成熟的关系型数据治理。 打破协同壁垒：以人际连接建构同理心 (12:33) 要将生产级 AI 转化为切实的商业回报，需要业务负责人、数据科学家、数据工程师与 MLOps 专家之间的紧密配合。Simonis 提出，单纯在指导委员会中按职能指派成员并不足以消除摩擦，团队成员必须作为具象的人建立人际连接，而非退守在部门筒仓之后。 为此，Simonis 规定跨职能团队在项目启动时必须进行 30 分钟的一对一非正式交流（通过视频或面对面），且明确不得设定任何业务议程，仅谈论成长背景、家庭、宠物与职业经历。这种人际熟悉度在后续出现技术交付阻碍时能够显著降低沟通摩擦，促使成员更具建设性地协作，避免将人际冲突上升为部门对立。此外，团队还通过有意模糊工作边界来深化协同，例如让 MLOps 工程师参与准备面向高管的业务演示汇报，并鼓励数据工程师直接参与核心模型开发。 两位专家的分享表明，AI 负责人的核心角色正从算法攻坚转向对组织协同、业务互信与时序数据体系的系统性重构；但在淘汰非必要旧流程时如何平衡组织惯性阻力，以及知识认证标准在多元监管环境下的适用边界，仍是企业在持续规模化进程中面临的现实挑战。

## Background 该视频由 Ai4 于 2025 年 9 月 14 日发布在 YouTube 平台上。内容记录了安永（EY）AI 财富管理解决方案负责人 Sasha Ostojic 与爱德华琼斯（Edward Jones）数据科学高级经理 Joseph Simonis 的对谈，围绕如何将前沿 AI 构想转化为切实业务价值展开，深入探讨了高价值 AI 项目的筛选、规模化挑战与跨团队协作文化。 ## TL;DR 分享结合爱德华琼斯服务全美 800 万客户的庞大数据底座，阐述了 AI 战略落地的核心路径。在项目排布上，团队推行“长期战略大项目”与“敏捷轻量级小痛点”并行的双轨策略，将 PoC 的成功标准定义为“赢得一线业务合伙人的拥护”；在规模化推进中，通过“人员、流程、产品（3P）”框架将企业重心从单纯模型算法转向专属数据与动态知识管理（如防范老年欺诈的时间序列洞察）；最后，分享强调跨职能落地的核心在于人文同理心，倡导通过“无议程咖啡聊”与职能边界模糊化打破部门隔阂。 ## Atmosphere & style 整场对话轻松务实、节奏明快，开场以“挡在大家去吃午饭路上”的幽默自嘲迅速破冰拉近距离。Sasha 凭借深厚的咨询视角提供严谨宏观的 3P 框架，而拥有数学博士背景的 Joe 则以生动的业务一线案例与细腻的管理心法与之默契呼应。交流既有严密的数据架构与欺诈识别逻辑，又充满关注团队成员心理的人文温度，呈现出兼具战略高度与实操细节的精彩探讨。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=FLGW3O332kk&t=4s) 爱德华琼斯拥有约 20,000 名财务顾问并覆盖全美所有县级行政区，海量客户与交易数据构成了其开展 AI 创新的核心底座。 - [02:30](https://www.youtube.com/watch?v=FLGW3O332kk&t=150s) AI 创新不能仅因技术新颖而盲目跟风，而必须严格锚定企业未来 1 至 5 年的战略转型目标与现实阻碍。 - [02:30](https://www.youtube.com/watch?v=FLGW3O332kk&t=150s) 大周期战略项目应与基层轻量级 AI 痛点解决相结合，通过小胜（Small Wins）在漫长交付期内维持势头与团队士气。 - [02:30](https://www.youtube.com/watch?v=FLGW3O332kk&t=150s) PoC（概念验证）成功的终极标准不仅在于技术指标的高低，更在于能否将一线业务人员转化为坚定的项目拥护者与共识盟友。 - [06:04](https://www.youtube.com/watch?v=FLGW3O332kk&t=364s) 规模化 AI 需依托 3P 框架（人员、流程、产品），破除跨团队沟通断层并要求业务与技术团队共同推进技能升级（Upskilling）。 - [06:04](https://www.youtube.com/watch?v=FLGW3O332kk&t=364s) 企业应建立 AI 工厂与卓越中心（CoE）沉淀可复用资产，并将 AI 作为重塑遗留流程的契机，而非简单为旧流程打补丁。 - [06:04](https://www.youtube.com/watch?v=FLGW3O332kk&t=364s) 基础模型已高度商品化，企业的核心差异化壁垒在于专有数据与业务认知，推动数据治理向知识管理（Knowledge Management）演进是关键。 - [09:39](https://www.youtube.com/watch?v=FLGW3O332kk&t=579s) 数据 AI 就绪度要求从静态合规记录转向支持动态模式洞察，保留时序演变数据对于老年欺诈识别等复杂业务场景至关重要。 - [12:26](https://www.youtube.com/watch?v=FLGW3O332kk&t=746s) 跨职能协作的基石是将同事视为具体的人而非冷冰冰的职能标签，通过无议程的 1 对 1 交流可建立深厚的团队心理同理心缓冲。 - [12:26](https://www.youtube.com/watch?v=FLGW3O332kk&t=746s) 适当模糊数据科学家、工程师与业务方之间的职能边界，让团队成员体验彼此的岗位挑战，能显著提升端到端 AI 交付凝聚力。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=FLGW3O332kk&t=4s) 会议开场与爱德华琼斯的业务及数据底座 会议由来自安永（EY）的 AI 与数据咨询专家 Sasha Ostojic 主持。Sasha 在安永主要负责与金融服务机构合作，利用人工智能推动业务转型。他在开场时指出，本次分享将全面探讨 AI 领导者角色的演进，涵盖优先级排序、规模化挑战、跨团队协作以及前沿趋势等核心议题。随后，Sasha 隆重介绍了对谈嘉宾——爱德华琼斯（Edward Jones）的数据科学高级经理及 AI 与数据创新实践负责人 Joseph Simonis（Joe）。Joe 不仅拥有数学博士学位，还手握多项技术专利。Joe 上台后打趣道，大家的手持麦克风运转正常，虽然自己正挡在大家去吃午饭的路上，但“千万别去想肚子有多饿”，保证会带来一场兼具趣味与干货的交流。 为帮助在场听众建立背景认知，Joe 介绍了爱德华琼斯的业务体量与运作机制。作为一家专注于通过投资和储蓄帮助客户实现财务目标的金融服务巨头，爱德华琼斯拥有约 20,000 名财务顾问，业务遍及美国与加拿大，服务着约 800 万个人客户。令人瞩目的是，就在几个月前，该机构正式达成了客户网络覆盖全美每一个县（every county across the United States）的里程碑。面对数以万计的顾问、海量客户以及庞大的交易流，公司坐拥极为丰富的北美投资者数据宝藏。Joe 的核心使命就是将这些深厚的数据资产与机器学习、统计学及 AI 技术相融合，赋能公司业务并协助客户达成终极财务目标。 - **机构网络规模**：约 20,000 名财务顾问，遍布美国和加拿大全境。 - **客户与地域覆盖**：服务约 800 万个人客户，全美所有县级行政区均有其客户分布。 - **创新团队定位**：处于实验（Experimentation）与落地执行（Execution）的交汇点，致力于用统计、ML 和 AI 释放海量财富管理数据的业务价值。 --- ### [02:30](https://www.youtube.com/watch?v=FLGW3O332kk&t=150s) AI 项目优先级排序与 PoC 的真实成功标准 在探讨如何筛选和排布 AI 创新项目时，Joe 明确强调，自己的工作起点绝非因为 AI 技术“新颖炫酷”（new and shiny）就盲目引入，而是严格锚定公司的战略目标。他会首先审视公司未来 1 年、2 年乃至 5 年的发展雄心与转型方向，剖析达成这些目标路上的关键阻碍，再评估现有的数据和工具能够提供何种支撑，以此梳理出服务于重大转型的“创新项目组合”。然而，这类涉及深层变革的核心项目通常周期漫长，往往需要 6 个月到 18 或 24 个月。 为了在漫长周期中维持势头，Joe 采取了“大项目推进与小创新穿插”的双轨策略。他的团队会主动寻找业务流程存在微小痛点的基层团队——例如一个四五人的小组每三个月就要耗费 40 个小时在一项令人头疼的例行流程上。团队会为他们量身定制轻量级的 AI 解决方案。这类小工具从全公司层面上看或许没有庞大的 ROI，但却能为业务团队省下宝贵的 40 小时，并为数据科学团队赢得宝贵的“敏捷胜利”（small wins）。 这种小胜策略深刻重塑了概念验证（Proof of Concept, PoC）的价值评估标准。在合伙人架构的爱德华琼斯内部，任何执行方案都需要跨部门达成广泛共识。当业务人员切身感受到 AI 的实际效用时，他们会从旁观者转变为坚定的同盟与倡导者（advocates）。因此，在 Joe 看来，判断一个 PoC 是否真正成功的标志，不在于算法指标有多高，而在于“是否为自己争取到了业务合伙人站在同一战线”。正是这些在业务一线建立的信任背书，让管理层愿意在面对耗时 18 个月的大型项目时，放心地将资源调配给创新团队。 ```mermaid flowchart TD accTitle: 双轨制 AI 项目组合管理 accDescr: 展示自上而下的战略大项目与自下而上的敏捷小项目如何形成飞轮效应 subgraph StrategicTrack ["战略级重大转型（周期：6-24个月）"] A["公司战略目标 (1/2/5年愿景)"] --> B["识别长期业务与组织障碍"] B --> C["大型 AI 架构与解决方案交付"] end subgraph QuickWinTrack ["敏捷级痛点突破（周期短/小团队）"] D["下沉业务一线挖掘微小痛点"] --> E["开发轻量级定制 AI 工具"] E --> F["解放基层生产力 (获取 Small Wins)"] end F --> G["赢得业务部门拥护者 (Advocacy & Buy-in)"] G -->|"自下而上反哺支持与资源"| C C -->|"实现业务深层变革"| H["沉淀为组织长期 ROI"] ``` --- ### [06:04](https://www.youtube.com/watch?v=FLGW3O332kk&t=364s) 规模化 AI 的 3P 框架：人员、流程与产品 面对爱德华琼斯在大规模落地 AI 过程中的经验，Sasha 结合安永跨行业的咨询洞察，系统性地提出了规模化 AI 的“3P”框架——人员（People）、流程（Process）与产品（Product）。AI 项目绝不能在单一部门的孤岛中推进，它本质上是一项横跨数据科学、系统工程、业务前端、合规风控乃至销售条线的跨职能工程。 在**人员（People）**维度，规模化的阻碍通常源于跨团队的沟通断层。Sasha 指出，在 AI 项目研讨中，业务人员与技术专家常常陷入“各说各话、无法共鸣”的尴尬境地。破解该瓶颈需要双管齐下：一方面需要既懂业务逻辑又能把握底层技术细节的复合型领导者；另一方面必须推进全员技能升级（Upskilling）——过去强调技术人员要“懂业务”，而如今同样关键的是“业务人员必须掌握 AI 语言”，从自上而下与自下而上两个层面打破认知壁垒。 在**流程（Process）**维度，首要原则是避免“每次都重新发明轮子”。许多企业在完成一个出色的 PoC 后，往往在下一个项目中推倒重来。解决之道在于提炼可复用的框架与流程资产，诸多领先机构正通过建立企业级的“AI 工厂”（AI factories）与“AI 卓越中心”（Centers of Excellence, CoE）来沉淀通用标准。而在微观层面，对于传统遗留业务流程（legacy processes），AI 的应用绝不仅是“将 8 小时的流程压缩至 1 小时”，更需要站在全局重新审视：该流程现有的各个环节能否被重构与重构象？流程本身是否还有存在的必要？AI 应当作为重塑业务的契机，而非简单为旧流程打补丁。 在**产品（Product）**维度，Sasha 提出了一个核心论断：“基础模型正在高度商品化（models are commoditized）”。对于绝大多数主营业务并非 AI 研发的企业而言，算法模型本身不再构成真正的竞争壁垒。企业的唯一差异化优势在于其“专属企业数据”与“行业业务认知”。要最大化这一壁垒，传统的单纯数据管理必须向“知识管理”（Knowledge Management）演进，将分散的企业非结构化数据提炼转化为结构化的组织知识。 --- ### [09:39](https://www.youtube.com/watch?v=FLGW3O332kk&t=579s) 数据与知识就绪度：从静态数据质量到动态模式洞察 针对“数据是核心差异化壁垒”的论断，Sasha 将话题引向数据与知识的“AI 就绪度”（AI readiness）。Joe 对此深表赞同，并直言业界常说的“垃圾进，垃圾出”（garbage in, garbage out）。他强调，在爱德华琼斯，数据就绪不仅要求数据可信（trusted）、可访问（accessible）以及受保护（protected）——鉴于财富管理涉及极为敏感的个人财务隐私，安全防护更是重中之重。更关键的是，组织必须以“正确的方式”使数据变得可用。 为了具体说明何为“以 AI 视角存储和提供数据”，Joe 举了一个生动的行业案例： - **退休账户与授权代理人（POA）机制**：每位投资者在退休账户上通常都会指定委托代理人（Power of Attorney, POA），以防突发状况下有人能接管账户。从传统业务和静态数据质量的角度看，金融机构只需确保“在特定时间点，系统里记录的 POA 信息是准确且合规的”，这就完成了数据质量要求。 - **分析与欺诈识别视角的质变**：但在防范针对老年人的财务欺诈（Senior Fraud）场景中，仅有当下的静态记录远远不够。老年欺诈的一个典型预警信号（red flag）是：客户账户上的 POA 在数月之内频繁发生变更（flipped again and again month over month）。这往往意味着客户的家庭关系或身心状况发生了重大异常，需要机构及时介入关怀并核查“您还好吗？”。 - **AI 系统的数据诉求**：如果历史变更记录被直接归档冷存、算法无法直接调取比对，那么 AI 系统就无法感知到时间序列上的异常模式。只有当整个组织意识到，必须将数据及其演变历史以支持模式识别的方式保存并提供给新工具时，企业才算真正进入了“AI 就绪”状态。 Sasha 补充道，这种演进正推动着业界从传统的“内容管理”迈向系统性的“知识管理”。许多企业正比照传统的数据治理流程，为非结构化知识资产建立知识认证标准（Knowledge Certification Standards）与知识分类体系（Knowledge Taxonomies），以此构建真正的 AI 知识壁垒。 --- ### [12:26](https://www.youtube.com/watch?v=FLGW3O332kk&t=746s) 跨职能协作与团队同理心建设 对话的最后，两位领导者将目光重新聚焦于“人”这一最核心的驱动要素。Joe 指出，想要将 AI 系统真正推向生产环境并创造可观的 ROI，必须依赖业务人员、数据科学家、数据工程师以及 MLOps 工程师的紧密咬合。然而，仅仅把这些职能代表召集在同一张会议桌前是远远不够的。 Joe 提出，团队协同的关键在于：“会议桌两旁的每个人，看对方时必须看到的是一个鲜活的人，而不是一个冰冷的职能标签（see another human being, not another function）”。为此，在爱德华琼斯组建跨职能 AI 团队之初，Joe 会推行一项硬性文化举措： - **无议程一对一咖啡聊（1-on-1 Coffee Chats）**：无论在线上 Zoom 还是线下，团队成员必须与组内每一位同事安排 30 分钟的一对一交流。 - **去工作化的交流规则**：这场对话严格不设工作议程（No agenda），大家只需随性交流宠物、孩子、近期的家庭生活，或是过往的工作趣事。其核心目的就是了解彼此作为“人”的一面。 - **化解摩擦的心理效应**：Joe 指出，当项目推进遇到挫折与阻碍时，这种连接会产生微妙的同理心缓冲——“你不再是对着抽象的 MLOps 职能发火咆哮，你面对的是活生生的同事 Lindsay，这会让你意识到冲她发火并不合适”，团队自然会倾向于以更具建设性的方式协同沟通。 此外，Joe 还大力倡导“模糊职能边界”以培养团队间的同理心：放手让 MLOps 工程师去为业务方撰写汇报 PPT，允许数据工程师参与模型的构建过程。通过体验彼此岗位的工作挑战，团队成员不仅深化了对端到端流程的理解，更锻造出强大的凝聚力，从而为高质量交付复杂的 AI 项目打下坚实的基础。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=FLGW3O332kk&t=4s) every county across the United States — Joseph Simonis > [02:30](https://www.youtube.com/watch?v=FLGW3O332kk&t=150s) new and shiny — Joseph Simonis > [02:30](https://www.youtube.com/watch?v=FLGW3O332kk&t=150s) small wins — Joseph Simonis > [06:04](https://www.youtube.com/watch?v=FLGW3O332kk&t=364s) models are commoditized — Sasha Ostojic > [09:39](https://www.youtube.com/watch?v=FLGW3O332kk&t=579s) garbage in, garbage out — Joseph Simonis > [09:39](https://www.youtube.com/watch?v=FLGW3O332kk&t=579s) flipped again and again month over month — Joseph Simonis > [12:26](https://www.youtube.com/watch?v=FLGW3O332kk&t=746s) see another human being, not another function — Joseph Simonis > [12:26](https://www.youtube.com/watch?v=FLGW3O332kk&t=746s) No agenda — Joseph Simonis ## Mentioned Sasha Ostojic, Joseph Simonis, 安永（EY）, 爱德华琼斯（Edward Jones）, AI, ML, PoC, CoE, MLOps, POA, Zoom ## Video description Sasha Ostojic, AI Wealth Management Solutions Leader, EY Joseph Simonis, Senior Manager Data Science, Edward Jones Abstracts In a world where innovation is the norm, the challenge lies in translating cutting-edge AI concepts into tangible business outcomes. This session invites AI leaders to explore the critical journey from visionary ideas to impactful execution. We will delve into strategies for identifying high-value AI initiatives, overcoming scaling challenges, and fostering a culture of collaboration across teams. With a focus on harnessing AI not just as a tool, but as a transformative force, this session will equip leaders to drive sustainable growth and competitive advantage across any sector. Join us to discover how to lead with purpose and make AI a cornerstone of your organizational strategy. [Source](https://www.youtube.com/watch?v=FLGW3O332kk)

[0:04] Joseph Simonis: every county across the United States [2:30] Joseph Simonis: new and shiny [2:30] Joseph Simonis: small wins [6:04] Sasha Ostojic: models are commoditized [9:39] Joseph Simonis: garbage in, garbage out [9:39] Joseph Simonis: flipped again and again month over month [12:26] Joseph Simonis: see another human being, not another function [12:26] Joseph Simonis: No agenda

translating cutting-edge AI concepts into tangible business outcomes [0:04] Sasha Ostojic: navigating prioritization, scaling hurdles, organizational collaboration, and industry trends [2:30] Joseph Simonis: scattered quick-turnaround tactical projects alongside long-term efforts [2:30] Joseph Simonis: A proof of concept is not merely a technical validation; it succeeds when it creates active business champions who advocate for broader AI integration. [6:04] Sasha Ostojic: Three Ps framework: People, Process, and Product [6:04] Sasha Ostojic: Develop bilingual leadership; upskill technical talent in commercial fluency while training business units to speak AI. [9:50] Joseph Simonis: AI readiness requires structuring and retaining data specifically for analytical pattern detection over time rather than simple transactional storage. [12:33] Joseph Simonis: you're not yelling at MLOps, you're yelling at Lindsay.

[0:04] All right. Well, good morning everybody. [0:06] Uh, thank you all so much for being [0:07] here. Uh, my name is Sasha Oichich and I [0:10] specialize in AI and data consulting at [0:13] EY where I partner with financial [0:15] service firms to leverage AI to support [0:17] business transformation. Um, so this [0:19] morning we're going to be discussing a [0:21] broad range of topics all focused on the [0:23] evolving role of the AI leader. We're [0:26] going to be covering a broad range of [0:28] topics within that realm. everything [0:29] from prioritization to scaling and [0:32] challenges when it comes to scaling as [0:34] well as fostering collaboration and some [0:36] key trends we're seeing as well. Um so [0:39] for our discussion this morning I'm [0:40] joined by Joe Simones uh who leads the [0:43] AI and data innovation practice within [0:45] Edward Jones. Um in addition to that Joe [0:48] also holds a PhD in math as well as [0:51] several patents to boot. So I really [0:53] could not think of a better person to be [0:55] here with us today to discuss this and [0:56] to really share his insights. So Joe, [0:58] welcome. [0:59] >> Yeah, thank you. I'm very excited to be [1:00] here with everyone. Microphone's [1:02] working, preventing you all from getting [1:04] lunch, so just don't think about how [1:06] hungry you are. Um, we'll have some fun. [1:08] >> Awesome. Well, thanks Joe. So Joe, like [1:10] I um just just maybe to set the stage [1:12] for for all the folks here, right? Can [1:14] you maybe give us a little bit of [1:15] background on Edward Jones, just in case [1:17] uh some of these folks are not familiar [1:18] with it, and then tell us a little bit [1:20] about your role as the AI innovation [1:21] leader at Edward Jones? [1:23] >> Yeah, happily. So, for those who don't [1:25] know, Edward Jones is a financial [1:26] services company. Um, we help clients [1:30] meet whatever financial goals they have [1:32] through investments and and and savings, [1:35] etc. Um, we have roughly 20,000 [1:39] financial advisors spread around the US [1:42] and Canada. Um, we're serving, I think, [1:45] roughly 8 million individual clients [1:47] right now. and those clients. As of [1:51] couple of months ago, um we can now say [1:53] we have a client in every county across [1:55] the United States. So very excited with [1:58] all of those adviserss, all of those [2:00] clients, all of the transactions they're [2:02] doing. We're sitting on a wealth of data [2:05] um about uh investors across the the [2:09] country and Canada. And so my role is to [2:12] take some of that data, mix that with [2:15] the artificial intelligence, machine [2:17] learning, stats, and really help our [2:20] firm do what it's trying to do, which is [2:22] help all of those clients uh meet their [2:25] their end goals. So wherever I can mix [2:27] those together, that's my goal. [2:30] >> Awesome. So thank you very much for [2:31] setting that context. And as the leader [2:34] of AI innovation at the firm, right, [2:36] you're really sitting at the [2:37] intersection of experimentation and [2:40] execution. So, so you when you when you [2:42] step back and you think about your role [2:44] and and your day-to-day within your [2:45] role, right? What are some of the ways [2:48] in which you identify and prioritize AI [2:50] initiatives and when you're thinking [2:53] about this from a proof of concept [2:54] perspective, like what makes a good [2:55] proof of concept? Like what is a [2:57] successful proof of concept? [2:58] >> Great question. Um, [3:01] all of my work really starts at what are [3:04] the strategic objectives of the firm. [3:07] There's no kind of start of AI, it's new [3:10] and shiny. I want to ingest this. It [3:12] starts with what is our company's [3:15] ambition over the next 1 year, 2 years, [3:18] 5 years. What are they trying to [3:21] transform? What are they trying to do [3:22] differently? I start with that portfolio [3:25] and I start to look through and say what [3:28] are the barriers to being able to do [3:29] that and execute that over the next [3:31] couple of years and then from there what [3:34] kind of data and what kind of tools can [3:36] I have or do I have to maybe help [3:40] execute along. So it's really starting [3:43] with the firm's objectives narrowing [3:46] down and that's where I start to [3:47] identify my innovation portfolio for [3:50] some of the big transformative [3:53] efforts of the firm. But those projects [3:55] are going to be anywhere from 6 month to [3:58] 18 24-month projects. So while those are [4:02] going on and they they require great [4:04] transformation, I try to also scatter [4:07] small innovative projects. These are [4:10] things where I go out and find a small [4:12] team that has some tiny barrier. Maybe [4:14] it's a group of four or five people [4:16] doing some business process and that [4:18] business process is just a struggle for [4:21] them. They just don't like doing it. [4:23] Maybe it's something that that takes [4:24] them 40 hours every 3 months, but that's [4:28] 40 hours they could be doing something [4:29] else. Um, and my team will try to find [4:33] an AI solution for that small project. [4:36] It's not a huge ROI when you think about [4:39] firm value, but what it does give is the [4:44] small wins with our product partners [4:48] within the firm, our business [4:50] stakeholders. And when those teams see [4:53] the value of artificial intelligence and [4:55] the value of what we can bring, they get [4:59] excited. They're brought along on that [5:01] journey. And now I have advocates on my [5:04] side when I'm trying to go back to those [5:06] bigger 18-month projects because I have [5:10] someone on the business side saying, [5:11] "Yes, this works for me." Yes, give the [5:14] resources over to Joe and his team to go [5:16] execute that larger project and that [5:18] longer term. So your question about [5:21] what's the value of that PC or when do I [5:23] know it's a success? It's when I've [5:25] armed other people onto my side to [5:28] really get buy in for building AI into [5:31] our our solutions. Um, and some of this [5:34] comes from just how we work at Edward [5:36] Jones. We're partnership. I need to [5:38] wrangle a lot of people to get agreement [5:40] to do any kind of execution and show the [5:42] value and takes a long time. Um, so that [5:45] works for me, but I've been at Edward [5:48] Jones for over 10 years now. So I think [5:50] this is where the partnership with Ian Y [5:52] works well cuz Sasha can come in and say [5:55] yeah that's fine for Y or for Edward [5:58] Jones. What do you see across other [6:01] industries for getting this built in [6:03] scaled? [6:04] >> Yeah. So I mean when we when we think [6:06] about scaling and scaling AI and like [6:09] the challenges within with scaling AI I [6:12] mean the reality is is that it's endless [6:13] right? Um, so maybe just to kind of like [6:15] even anchor us a bit, right? The way I [6:17] kind of like to think about it is across [6:19] like like Marcus Lamonus' three Ps, [6:21] right? Like people, process, and [6:23] product. Um, so in terms of people, [6:25] right, we know that AI initiatives fail [6:28] if they're done in silos, right? So AI [6:31] in and of itself is just inherently a [6:35] crossfunctional [6:36] exercise, if you will, right? It [6:38] requires everything from your data [6:39] science people, your engineers all the [6:42] way to like your business, your [6:43] compliance, potentially sales, right? [6:45] Everybody across the board, right? So [6:47] that's that is a really big challenge [6:49] when it comes to scaling. And so the way [6:50] you kind of tackle that is really in two [6:52] ways, right? One is you obviously need [6:54] strong leaders that can speak both [6:56] languages and speak to all parties [6:57] involved, right? So everybody from the [6:59] business all the way down to the [7:00] technical components in terms of AI [7:02] itself, right? Uh but the second piece [7:04] of that is also upskilling like your [7:07] your firm right more so than ever it's [7:10] absolutely critical that your tech [7:12] people can speak business but now it's [7:15] almost just as important that your [7:16] business people can speak AI right so [7:19] and if you've ever been a part of these [7:21] like conversations when it comes to like [7:23] PC's or like AI initiatives there's [7:25] often times people that are like just [7:27] talking past each other right like and [7:29] and that's so you got to upscale but [7:31] then you also got to have the leadership [7:32] right really from top to bottom and [7:33] bottom to top. Right? So that's from a [7:35] people perspective. From a process [7:37] perspective, you want to avoid re [7:41] reinventing the wheel every single time. [7:43] Right? So you don't want to, you know, [7:46] we we stood up this awesome PC or we [7:48] stood up this awesome AI initiative and [7:49] then now it's completely blank slate. [7:51] Let's do it, you know, again all over [7:53] from scratch, right? So that's where [7:54] trying to come up with repeatable [7:56] processes and frameworks is going to be [7:59] incredibly like critical, right? And so [8:01] what we have seen is folks looking to [8:03] stand up like AI factories, right? As [8:05] well as AI centers of excellence within [8:07] firms, right, to support that framework [8:09] repeatability, if you will. Uh but [8:12] that's kind of more at a macro level [8:14] when we think about process at like an [8:15] individual level. Um we've heard it [8:18] other other you know speakers speak [8:20] about this as well, I'm sure, right? [8:21] Like legacy processes. It's not [8:23] necessarily about taking a legacy or an [8:26] old process and like taking it from 8 [8:30] hours down to 1 hour. That's great. Like [8:32] that's awesome, right? But you should [8:35] think about when you're looking to apply [8:37] AI to improve a process, right? Think [8:39] about it, you know, step back and really [8:40] think about it like what does this [8:42] process look like today? What are all [8:43] the various components of that process [8:45] that can be, you know, rethought, [8:48] reimagined, if you will? Does this [8:49] process even need to exist? Right? Um [8:51] and then AI is just one component of [8:53] that like reimagining process right so [8:55] that's the way I would think about it [8:56] from a process perspective and then you [8:58] know stepping back and thinking about it [9:00] from a product perspective the reality [9:03] is is that models are commoditized right [9:06] like maybe some of you in here are like [9:08] true like AI engineers and researchers [9:11] but the vast majority right are AI [9:13] leaders within firms whose core business [9:15] is not AI right so the differentiator [9:18] there is going to be your firm's data [9:20] and your firm's business, right? So [9:22] that's the way I would really think [9:24] about it and that you know in order to [9:26] to fully leverage your firm's data [9:28] that's where data management yes but [9:30] more so now even knowledge management [9:32] right the idea that unstructured data [9:34] managed is knowledge is becoming very [9:36] critical. So, so Joe then wanting to [9:39] kind of bring you in here then like when [9:40] we think about data as like your key [9:42] differentiator right now we're kind of [9:44] starting to pivot into AI readiness uh [9:48] specifically from a data perspective. So [9:50] how do you at Jones ensure that your [9:53] data is AI ready, your knowledge is AI [9:54] ready and what does that mean to you [9:56] specifically? [9:58] >> Very important. Everyone knows garbage [10:00] in garbage out. Um so when I think about [10:02] the data and data readiness of the firm [10:05] um certainly I want it trusted. I want [10:09] it accessible to me. I want it uh [10:12] protected. We have a lot of sensitive [10:14] data. So I want to make sure all of that [10:17] is true. But I think it goes beyond just [10:20] knowing that that the data is trusted [10:22] and accessible. It's accessible in the [10:24] right way. When the organization starts [10:26] to think about how are they storing the [10:28] data and making it available to our AI [10:31] systems. When they start thinking [10:33] differently, then I know I've got them [10:35] kind of at a state where I can use them. [10:37] And really quick example, all of you, I [10:40] hope, have some type of retirement [10:41] account out there. And on that [10:43] retirement account, I hope you have a [10:46] power of attorney. something happens to [10:48] you, someone needs someone else needs to [10:50] be able to step in. From a financial [10:53] firm, it is on us to know that that POA [10:58] is the right person at the right time [11:00] they're needed. That's data quality. But [11:03] from an analytic perspective, that [11:05] doesn't always mean that's enough. Um, [11:09] the financial industry knows that when [11:10] it comes to say senior fraud, one of the [11:13] red flags that can occur is if we see [11:16] someone's POA being flipped again and [11:18] again month over month, maybe there's [11:21] something going wrong in in their [11:23] familial situation and we need to be [11:25] step in and say, "Hey, are you okay?" [11:28] Um, but if I just know who your POA is [11:30] today and all the rest of the data is [11:32] archived somewhere and I can't access [11:34] it, that's no good to me. My AI system [11:36] needs to start seeing the patterns. And [11:37] so when I think of our organization [11:39] saving our data in different ways and [11:42] saying yes, you need to be able to see [11:43] that history and use it in new ways and [11:46] in new tools. I know they're starting to [11:49] to get ready. So that's how I think of [11:52] it. [11:53] >> And and just to kind of bring it back to [11:54] what I was kind of mentioning with [11:56] respect to like data management, right, [11:58] is that when we think about data [12:00] management and shifting to knowledge [12:02] management specifically, right? That's [12:03] where like we're seeing firms really [12:05] start to like take a much more proactive [12:06] approach in that right historically it [12:08] was about content management and now [12:10] we're like moving towards like knowledge [12:11] management very much akin to what you [12:14] see from like traditional data [12:15] management right so standing up [12:16] knowledge certification standards you [12:18] know knowledge taxonomies all of those [12:20] kind of critical components are coming [12:21] into play now um so Joe you know we've [12:26] talked a little bit about you know [12:27] prioritization talked about process and [12:29] and um and data so maybe just the last [12:33] kind of piece here, right? We haven't [12:34] talked about the people and the people [12:36] are the most important piece, right? Um [12:38] so tell us a little bit about how you [12:40] foster collaboration. [12:41] >> Yeah, the collaboration is key to [12:43] getting any of these um AI systems into [12:48] reality and in production so we can get [12:49] ROI. That takes the business, that takes [12:52] our data scientists, that takes data [12:54] engineers, MLOps. So when I bring a team [12:57] together and say, "Hey, all of these [12:59] functions need to come and work [13:00] together." [13:02] It's not good enough to just have the [13:04] functions at the table. [13:06] Everyone at the table has to look across [13:08] the table and see another human being, [13:12] not another function. And so, one of the [13:15] first things I like to do on my teams is [13:16] say, "Everyone on this team, go set up [13:19] some one-on-one coffee chats, whether [13:21] it's on Zoom or in person. Spend 30 [13:24] minutes getting to know every other [13:26] person on your team. No agenda. Just [13:28] show up. Talk about pets, children, care [13:32] going on in your lives. And if you don't [13:34] feel like sharing some of those details, [13:36] talk about past jobs, but figure out how [13:39] everyone else on your team is a human. [13:42] And then as you go through on that [13:45] project, when you hit a speed bump, [13:47] you're not yelling at MLOps, you're [13:49] yelling at Lindsay. And and that doesn't [13:53] feel as good. And so you collaborate [13:55] more in there. And I think that's where [13:58] where the value comes and and blurring [14:00] some of those lines. Let the MLOps [14:02] person build the the presentation for [14:04] the business. Let the data engineer do [14:06] some some model building too. Gain [14:09] empathy for each other on that that [14:11] project. And then your team will be that [14:14] much stronger for delivering the