McDonald's, McKesson Embed Governance into AI to Shift from Compliance

麦当劳与麦克森将治理深度融入AI，推动合规模式转型 企业领袖得出结论：若不通过自动化控制、跨职能协作与聚焦特定用例将数据治理切实融入日常工作流，AI应用落地就难以成功。 全球数据领袖分享了企业组织如何演进数据治理策略，以支持可扩展且合规的AI举措。专家研讨涵盖了夯实数据基础的关键必要性、管理非结构化数据与并购复杂性的技术手段，以及AI用例库和控制框架的落地实施。与会嘉宾还探讨了如何将数据治理从官僚化的行政负担转变为占比达75%的业务赋能工具，以及如何借助领域专用AI智能体与多智能体协同编排来加速数据获取。 数据治理已成为企业推进 AI 战略的核心基石；来自麦克森（McKesson）、麦当劳（McDonald's）与 PASHA Holding 的高管强调，AI 转型需结合自上而下的平台愿景与自下而上的全员工具普及。 (0:20) 非结构化数据和企业并购是当前治理的主要挑战，企业正借助 Data Mesh 架构将数据质量与管理职责直接交由业务领域负责人承担，而非单靠中心化数据团队。 (17:04) 企业正通过部署内部 AI 用例库、风险跟踪门户和自动化元数据管理，推动治理从传统人工维护转向自动化智能管理，消除算法“黑盒”疑虑并保障跨业务安全共享。 (23:18) 数据治理正从繁琐的合规审查转向业务赋能，目标是将精力从传统的 75% 行政管理扭转为 75% 运营执行，并通过内置于数据目录的 AI Agent 加速日常数据调用。 (31:00)

Enterprise data leaders argue that successful artificial intelligence adoption requires repositioning data governance from an administrative compliance hurdle into an automated, operational baseline across business units. Embedding AI Across Global Operations (0:20) Large organizations are actively shifting AI from isolated experiments into foundational operating models. At McKesson—a Fortune 9 healthcare distributor generating over $400 billion in revenue with 80,000 employees—AI investments center on distribution center robotics, service delivery timing, and oncology research to improve cancer outcomes. McDonald's, spanning 41,000 restaurants across 110 countries, is integrating AI into core operational strategies to drive business-line efficiencies while evaluating effects on crew members and customers, though teams must adapt their rollout to navigate stricter European regulations compared to more relaxed U.S. rules. PASHA Holding, a diversified Caucasus-based group spanning banking, fintech, retail, and construction, built a regional B2C digital ecosystem connecting financial tools, marketplaces, and public transit. CDO Valeh Nabiyev explains that the conglomerate couples data-driven readiness with a dual framework: top-down platform standardization combined with bottom-up employee enablement to integrate AI tools directly into everyday standard workflows. Governing Unstructured Repositories and Domain Ownership Managing unstructured data—such as internal PDFs, recipe cards, localized store menus, and training materials—poses major validation hurdles before feeding machine learning models. Mergers and acquisitions exacerbate these frictions by introducing disjointed legacy repositories. McDonald's addressed this by aligning cybersecurity, legal, privacy, and information governance teams to establish pragmatic classification and retention guidelines, starting first in the U.S. market before expanding internationally. At PASHA Holding, governance structures apply identical standards to structured and unstructured information. Implementing a Data Mesh architecture ensures that decentralized business domains—such as customer or transactional units—remain accountable for their own data stewardship and quality, preventing central data teams from becoming single points of failure. Balancing Governance Controls with Business Enablement Enterprises pair risk-mitigation frameworks with operational enablement tools to govern model lifecycles transparently. McDonald's deployed three internal control applications: a global Use Case Library, a Risk & Lifecycle Tracking Portal, and an Open Assessment Interface to evaluate third-party vendor risks and eliminate 'black box' development. Operationally, the company identified roughly nine AI use cases, relying on automated metadata curation to supply contextual grounding for algorithms. In highly regulated verticals, platform layers allow cross-entity data sharing without creating massive, insecure data dumps. PASHA Holding deployed a federated broker layer incorporating data minimization and anonymization to comply with strict banking secrecy and medical privacy laws. Similarly, McKesson automates third-party metadata curation, allowing staff to leverage standardized markers like national pharmacy IDs while keeping domain specialists focused on clinical analysis rather than manual tagging. Shifting Governance from Administrative Drag to Velocity Mature governance frameworks aim for a target operational state of 75% operational execution and 25% administrative overhead, though immature programs frequently find this balance inverted due to heavy policy drafting and stakeholder education. To bridge this gap, leaders emphasize removing technical 'governance speak'—such as abstract discussions of metadata and lineage—and instead framing initiatives around tangible business problems like fixing broken data feeds or resolving customer inquiry errors. Because many business stewards handle data duties on the side of their primary jobs, automated tooling is critical to maintain velocity. McDonald's global lead John Tucker notes that governance teams act as business enablers by deploying chatbots inside data catalogs and automating SQL query verification across enterprise chat platforms like Slack and Teams, streamlining metric alignment and data access without manual bottlenecks. Transitioning governance into an automated operational asset hinges on domain-level accountability and invisible, automated tooling. However, organizations continue to balance regional regulatory divergence and the friction of business stewards managing data duties alongside their primary roles.

企业在人工智能时代的成功不仅取决于先进模型的引入，更依赖于将底层数据治理从繁琐的行政合规转向以业务赋能为导向的自动化运营支撑。来自麦克森（McKesson）、麦当劳（McDonald's）和PASHA Holding的治理主管指出，建立统一的风险管控架构、明确业务域对非结构化数据的责任，以及运用AI智能体消除技术沟通壁垒，是确保AI规模化落地的核心支柱。 企业AI战略布局与跨域治理基石 (0:20) 在不同行业生态中，数据治理均被视为AI规模化应用不可妥协的底座。麦克森公司（McKesson）作为年营收超4000亿美元、拥有约8万名员工的财富榜第9位医疗健康企业，其数据治理主管Rachel Smith介绍，公司正将AI与自动化重点投向物流供应链机器人、客户履约提速以及肿瘤与临床研究三大前沿领域，并强调AI的技术突破始终始于基础数据质量。跨国连锁快餐麦当劳在全球110个国家和地区经营着4.1万家餐厅，其全球数据治理负责人John Tucker表示，团队正将AI直接嵌入核心战略与日常运营，在向最高管理层汇报整体AI战略的同时，必须应对全球法规的显著分歧，例如美国相对宽松的监管与欧洲严苛的合规标准之间的张力。 在高加索地区拥有多元化业务的PASHA Holding则展现了生态系统型企业的治理路径。其首席数据官Valeh Nabiyev表示，集团构建了连接银行、零售、电子钱包及公共交通的B2C数字化生态，其AI转型并非零散的用例集合，而是采取“自上而下”确立统一愿景、平台与核心能力，与“自下而上”为全员配备日常AI工具相结合的双轨框架，以确保AI能力深度内嵌至企业常规流程。 破解非结构化数据瓶颈与并购摩擦 非结构化数据的高效管理被与会专家视为现代企业喂养AI模型时面临的最大现实挑战。与结构清晰的数据库表不同，企业内部充斥着大量PDF文档、生产线提取物、本地餐厅菜单、操作配方卡和培训手册等原始文本。如果缺乏针对性发现工具与校验流水线，极易导致未经核验的异常数据渗入机器学习模型，而企业并购（M&A）活动中遗留系统与异构模式的强行并入则进一步放大了治理阻力。 针对此类资产，麦当劳采取跨部门协同策略，联动首席信息安全官、网络安全、信息治理及法务隐私团队，将刚性安全条例转化为适应性操作指南，并在餐厅规模最大的美国市场先行试点数据分类标准后再向全球铺开。PASHA Holding则选择在Data Mesh（数据网格）架构下推进治理，Nabiyev指出，无论结构化还是非结构化数据，均统一划归业务领域（如客户或交易单元）承担质量所有权，由产生和运营数据的业务线扮演数据保管人角色，避免了由中央数据团队包揽全部责任而产生的权责脱节。 双轨落地：模型风控与自动化元数据赋能 企业推进AI应用时呈现出风险防御与业务提速并行的双轨趋势。在风险防控与监管合规维度，麦当劳部署了三大基础AI控制应用：覆盖全集团的用例库（Use Case Library）、模型风险与生命周期追踪门户，以及公开评估界面。这套透明化的评估框架打破了算法黑箱，能够横向监控模型各阶段风险与第三方供应商隐患，有效遏制各区域市场的重复开发。 在业务赋能与运营提速维度，麦克森改变了传统依赖人工的治理模式，引入自动化元数据智能模型进行第三方数据编目与特征标注。这一方案让一线员工能够基于标准化国家药房标识符等行业数据进行高效检索，将业务团队从枯燥的打标签工作中解放，专注于高价值的临床研究业务。而在受制于银行保密法、保险合规与医疗隐私法规的PASHA Holding，企业通过构建具备脱敏和数据最小化机制的联邦数据中介层，借助Collibra或Informatica等工具打通受控的安全调取通道，既满足了严苛法规，又使各子公司业务部门获得了自主分析与跨业务集成的自服务能力。 治理角色转型：重塑运营比例与语言体系 数据治理正经历从低效官僚审核向关键业务运营动能的身份转型。专家指出，成熟企业的理想工作分配比例应为75%用于业务运营执行、25%用于行政事务；然而，大多数发展中企业的现状恰好倒挂，绝大部分资源被用于员工教育、说服管理层和政策起草。为了扭转这一失衡，治理团队必须彻底摆脱合规“警察”的刻板形象，转变为通过高质量数据驱动业务加速的使能者。 阻碍协同的主要症结之一在于治理团队惯用的抽象术语（如数据血统、元数据管理）容易使高自主权的业务条线产生疏离感。多位负责人强调，团队必须杜绝“治理行话”，转而聚焦于修复损坏的数据流或排查客户查询错误等具体业务痛点。针对业务部门兼职数据管理员精力受限的现状，麦当劳正通过部署集成在数据目录中的AI聊天机器人，以及在Slack和Teams中实现自动化SQL查询校验，以领域专属智能体协助新员工查找数据并降低技术摩擦，化解跨部门标准拉齐的协作阻碍。 当前企业数据治理已从理论构建全面转向实操攻坚，但在多法域监管冲突加剧、兼职业务管理员治理负荷沉重，以及海量非结构化数据处理等现实约束下，全自动元数据编目与领域驱动责任制的有效融合仍处于持续验证阶段。如何在大模型快速迭代的节奏中保持跨域风险可控，仍是各行业组织长期探索的核心课题。

## Background 本内容由 Ai4 频道于 2025 年 8 月 27 日发布在 YouTube 平台上。该视频为一场圆桌讨论，页面简介指出，专家们聚焦于数据治理在确保 AI 项目成功中的关键作用，深入探讨了在保障数据质量、数据隐私、安全与合规的同时，如何打破数据孤岛、降低风险并建立对 AI 驱动决策的信任。 ## TL;DR 来自麦克森（McKesson）、麦当劳（McDonald's）以及高加索多元化集团 PASHA Holding 的三位数据高管，深入分享了各自在医疗、跨国餐饮和综合生态中落地 AI 与数据治理的实战经验。讨论揭示了数据治理正在从传统的“行政合规把关人”转向以加速洞察为核心的“进攻型业务赋能者”，并探讨了利用元数据自动化及多 Agent 架构提升效率的路径。针对非结构化数据治理难题，嘉宾们主张沿用领域所有权和数据网格等成熟模式，拒绝另起炉灶。最后，专家们一致建议企业切忌全盘铺开或陷入工具选型内耗，而应从小处着手，绑定高商业价值的具体用例，通过跨职能协同推动组织变革。 ## Atmosphere & style 整场对话务实、严谨且充满一线实战经验的碰撞，嘉宾们摒弃了空洞的技术黑话与行业炒作，以坦诚甚至略带自嘲的幽默直面治理痛点。麦克森负责人的干练犀利、麦当劳总监的稳健条理与控股集团 CDO 的宏观战略思维相互交织，主持人穿针引线的节奏流畅自然。讨论从早期的试错教训层层推进至前沿的多 Agent 与元数据智能化落地，既有对监管与风险的敬畏，又洋溢着以治理驱动业务增长的进攻性与笃定感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=4s) AI 治理的核心支柱是数据治理，AI 模型的性能和可信度完全取决于底层输入数据的质量与血缘追踪能力。 - [02:06](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=126s) 面对 AI 浪潮，各行业推进策略各具特色：McKesson 侧重供应链自动化与肿瘤科研，麦当劳聚焦全球门店赋能与合规平衡，PASHA Holding 则致力于打造多业务数字生态与双向推进机制。 - [06:41](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=401s) 现代企业董事会对数据可信度与合规提出更高技术审计要求，端到端的数据血缘追溯和彻底删除能力成为管理层应对监管质询的底气。 - [10:18](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=618s) 企业推进 AI 转型需注重全员 AI 素养教育与心理安全感建设，通过普及 Microsoft Copilot 等工具赋能日常办公，并强化负责任使用原则。 - [12:00](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=720s) 数据治理应从被动的“关卡与防御型”转向主动赋能的“进攻型”模式，核心使命是缩短业务获取见解与数据的时间并提高数据质量。 - [15:10](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=910s) 数据治理是一项业务运营而非纯理论框架，采用“轴辐模式”（Hub and Spoke）让中心制定标准、业务端落地执行，才能真正嵌入企业肌理并构筑网络安全基石。 - [17:04](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=1024s) 非结构化数据治理成为 AI 时代的最大痛点，企业应优先聚焦受监管和涉及隐私的高风险数据，并在模型构建阶段严格把关数据来源与合规性。 - [20:50](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=1250s) 面对非结构化数据无需重复造轮子，应依托数据网格（Data Mesh）理念，将数据所有权与质量责任严格落实到产生该数据的业务领域（Domain）。 - [23:18](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=1398s) 麦当劳通过建立 AI 用例库和透明评估流程实现治理管控闭环，同时利用 AI 加速元数据管理与语义整理，为业务提供更深度的上下文理解。 - [25:25](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=1525s) McKesson 探索元数据智能化，在不打扰业务的前提下自动抽取业务上下文，并训练专属治理 AI 模型以颠覆传统人工治理模式。 - [30:53](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=1853s) 数据治理的理想状态应是 75% 的业务运营赋能与 25% 的行政管控，团队应避免使用生涩的治理术语，用解决现实痛点的方式引导业务协作。 - [33:03](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=1983s) 多 Agent 协同架构（领域 Agent 与跨域 Agent 互通）结合 Slack/Teams 等日常办公入口，大幅加速了业务人员获取数据与上下文的效率。 - [36:29](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=2189s) 数据治理从根本上是一场业务变革，集团中枢应专注于“铺设轨道”（标准与政策），各业务单元负责本地化适配与运行反馈。 - [38:31](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=2311s) 开展治理切忌贪大求全，应从小处着手，绑定高商业价值的单一用例（如客户终身价值 CLV）并行交付，以显性 ROI 赢得管理层与业务端信任。 - [41:16](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=2476s) 治理平台的工具功能往往大同小异，应避免选型内耗，成功的关键在于专业人才能力以及跨网络安全、法务、架构与数据团队的“联合战线”协同。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=4s) 欢迎致辞与嘉宾阵容介绍 在 Ai4 大会现场，The Ravit Show 创始人兼主持人 Ravit Jain 宣布圆桌论坛正式开始。本场讨论的主题是“在 AI 时代驾驭数据治理”（Navigating Data Governance in the AI Era）。主持人指出，无论台前还是幕后，大家都在热烈探讨 AI 治理，而数据治理正是支撑 AI 治理最坚实的核心支柱。在正式深入议题前，三位来自不同大型企业的高管分别介绍了各自的企业背景与职责： 首先发言的是 McKesson（麦克森公司）全球企业数据治理与第三方数据解决方案负责人 Rachel Smith。她笑着向现场观众提问：“在座有多少人听说过 McKesson？”看到台下举起不少手，她感到十分欣慰，因为在很多场合大众往往对这家幕后巨头并不熟知。Rachel 介绍道，McKesson 是美国乃至全球规模最大的医疗健康企业之一，在财富 500 强中高居第 9 位，拥有约 80,000 名员工，年营业收入超过 4000 亿美元。 紧接着是 McDonald's（麦当劳）全球企业数据治理总监 John Tucker。他统领这家全球极负盛名的快餐巨头的数据治理实践。麦当劳在 110 个国家和地区拥有超过 41,000 家餐厅，在如此庞大的全球网络中应对 AI 时代的数据治理，充满复杂性与挑战。 最后做自我介绍的是高加索地区大型多元化控股集团 PASHA Holding 的首席数据官 Valeh Nabiyev。Valeh 坦言台下可能很少有人听过 PASHA Holding，它立足于阿塞拜疆及高加索地区，旗下业务极具多元化，涵盖金融服务、银行、保险、电商平台（Marketplaces）、金融科技（Fintech）以及建筑等诸多核心支柱产业。 - **McKesson**：财富第 9 位（Fortune 9），全球约 80,000 名员工，营收超 4000 亿美元。 - **McDonald's**：全球 110 个国家和地区布局，门店数达 41,000 家。 - **PASHA Holding**：阿塞拜疆及高加索地区领军多元化集团，横跨银行、保险、电商与工程建筑。 --- ### [02:06](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=126s) 医疗、餐饮与跨界集团的 AI 落地前沿 主持人 Ravit 顺势抛出首个核心问题：在 AI 与数据治理的交织演变下，各个行业巨头正在发生怎样的变革？来自不同领域的企业又是如何部署与推进实际举措的？ McKesson 的 Rachel Smith 首先分享了医疗健康领域的最新进展。她指出，医疗被视作 AI 突破与技术革新的核心前线。McKesson 的投入重点分布在三个维度：其一是供应链与物流仓储，全面引入自动化与机器人技术提升分发中心效率；其二是通过精准调度与配送速度，持续提升客户服务满意度；其三是在肿瘤学（Oncology）与临床科研领域的深入应用。McKesson 作为全美最大的肿瘤学服务支持者之一，正借助 AI 开展癌症护理与临床结果优化研究，践行其“改善人人医疗成果”的企业核心宗旨。 ```mermaid flowchart TD accTitle: AI Application Scenarios Across Enterprises subgraph MCK["McKesson 医疗健康"] M1["供应链自动化与机器人"] M2["物流时效与客户体验"] M3["肿瘤科研与癌症治疗方案"] end subgraph MCD["McDonald's 餐饮运营"] D1["门店运营与员工赋能"] D2["全业务线流程提效"] D3["全球合规与策略落地"] end subgraph PSH["PASHA Holding 综合生态"] P1["B2C 一体化数字生态圈"] P2["客服中心与精细化运营"] P3["全员赋能与双向推进"] end ``` 麦当劳的 John Tucker 阐述了餐饮业的切入点：当前团队正致力于将 AI 全面融入企业顶层战略与前线运营中。麦当劳不仅关注 AI 如何重塑顾客体验与门店员工日常作业，还在深入挖掘如何为各业务线实现效率跃升。目前的节奏一方面是加紧业务线战略落地并向 C 级高管层（C-suite）汇报，另一方面必须审慎应对不同法域的监管差异——相较于美国本土目前较为宽松的环境，欧洲市场的 AI 与隐私法规明显严苛得多，全球化企业必须因地制宜灵活应对。 PASHA Holding 的 Valeh Nabiyev 则回顾了其正在构建的高加索地区首个端到端数字一体化 B2C 生态体系。该生态无缝串联了银行、电商平台、电子钱包（e-wallet）以及公共交通系统，每日服务本国数以百万计的用户。在谈及 AI 时，Valeh 表现得十分冷静与务实：集团目前在 AI 领域尚处于探索初期，但由于前期扎实推进了数据转型与“数据驱动决策”文化的建设，已经打下了坚实的企业数据治理底座。他们现阶段的 AI 用例主要聚焦于客户关怀（Customer care）和运营降本增效，但他特别强调：“AI 转型绝不仅仅是堆砌几个 AI 用例（AI transformation is not about AI use cases）。”PASHA 采取的是“自上而下”（由高管层共享使命、愿景、技术平台与核心能力）与“自下而上”（赋能每位基层员工在日常流程中善用 AI）相结合的推进策略。 --- ### [06:41](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=401s) 监管收紧与董事会关切：数据可信度是根基 Ravit 谈到，大型企业要把 AI 落实到 100% 全面应用并验证成效往往需要漫长的周期，他随即询问各位嘉宾：在技术变迁、监管风向以及企业文化演进三股力量冲击下，各自组织内部感受到了哪些深层次的变化？ Rachel Smith 直截了当地指出：“AI 的起点始终是数据（It starts with the data）。”尽管在技术大会上常能听到关于去监管化（Deregulation）的声音，但在医疗金融等强监管行业，有关隐私合规的法规——包括 HIPAA、PHI（受保护健康信息）、PII（个人可身份识别信息）以及加州 CCPA 等——丝毫不曾松动。现在的核心拷问是：数据究竟在哪里？由谁在访问？第三方的外部数据是否严格遵守了合同授权条款？ Rachel 观察到一个显著转变：现在的企业董事会（Board of Directors）高度懂技术，而不再是外行。过去十年中发生的诸多企业数据泄露风波，让如今的董事会常常向 CIO 和技术高管抛出极为尖锐的技术审计问题： - 我们的数据底账与血缘全景到底在哪里？ - 是否具备端到端的数据可追踪性（Traceability）？ - 当加州 CCPA 等法规赋予用户“删除个人信息”的权利时，系统能否真正逆向追溯到底层源系统，确保数据被物理性彻底清除？ Rachel 坦言，自己大量的时间都是直接与首席执行官（CEO）坐在一起，向他展示系统能力、数据脉络与合规支撑，从而让 CEO 在面对外界质询与法律监管时能够成竹在胸。 同时，Rachel 尖锐地剖析了数据治理在过去十余年间的曲折历史。数据治理概念由来已久，但许多机构的努力最终宣告失败，原因就在于没有将它视作一门严谨的“跨技术与业务复合学科”（Discipline），反而任命了不懂数据架构设计的人去牵头。对于如今大语言模型（LLM）的兴起，她坚信关键在建立“数据信任”（Data Trust）——必须配备懂业务规则、验证逻辑与数据清洗转换的人才，才能防止 LLM 产出不可控的内容。 --- ### [10:18](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=618s) 全员普及教育：从恐惧替代到驾驭 Copilot 麦当劳的 John Tucker 从组织文化的角度探讨了这轮变革。他指出，随着 AI 走入聚光灯下，甚至连首席人力资源官（CHRO）也开始深入介入 AI 议题，重点放在提升“AI 素养”（AI Literacy）与消除员工恐慌情绪上。 当各种降本增效的 AI 案例接连上马时，绝大多数员工的第一反应往往是焦虑：“我的岗位是不是保不住了？”为此，管理层必须投入巨大精力展开全员通识教育，明确向团队传递安全感：AI 的介入是为了替大家剥离繁琐重复的基础工作，腾出更多精力去承担更高阶的战略思考与业务创新。 为了将 AI 真正置于每个人手中，麦当劳已经在全公司范围内广泛推广并部署了 Microsoft Copilot。从日常会议记录的自动听写转录、邮件起草，到常规业务流程的自动化串联，员工开始切实体会到工具带来的红利。John 强调，当前的重心正落在如何“负责任地使用 AI”（Responsibly using AI），指导员工明晰何时该用、何时不该用，以及如何以最佳且合规的方式发挥其价值。 - **组织战略协同**：CHRO 深度介入，从人才发展与心理安全感维度普及 AI。 - **全员工具赋能**：规模化铺设 Microsoft Copilot，将 AI 嵌入日常办公场景。 - **负责任原则**：明确安全边界与合规红线，确保技术赋能不偏航。 --- ### [12:00](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=720s) 进攻型数据治理：打破“关卡”与“阻碍者”标签 PASHA Holding 的 Valeh Nabiyev 带来了独具一格的治理哲学反思。他认为，数据治理本身在此轮 AI 浪潮中必须经历蜕变。在传统认知中，很多企业将数据治理视作单纯的“阻碍性职能”（Barrier function），这个部门存在的意义似乎就是不断对业务需求说“不”，或者刻板地划出 GDPR 与法规禁区。 Valeh 将数据治理清晰地划分为两类： 1. **防御型数据治理（Defensive Data Governance）**：聚焦于安全风控、权限管控与合规限制。 2. **进攻型数据治理（Offensive Data Governance）**：主动赋能业务，提升整体效率。 PASHA Holding 选择有意识地将战略重心向“进攻端”倾斜。Valeh 指出，这绝非忽视合规与安全，而是明确将数据治理部门的使命定位于“持续缩短业务获取见解的时间与获取数据的时间”（Decreasing time to insight and time to data），同时大幅提升数据质量。在这个时代，“垃圾进，垃圾出”（Garbage in, garbage out）这句老话依然是绝对真理。如果底层数据未经过滤与治理，直接投喂给 AI 将引发巨大的灾难。 Valeh 进一步强调，单靠几块可视化仪表盘（Dashboard）绝不可能凭空创造出真正的数据驱动文化；同理，跑几个孤立的 AI 试点也绝非真正的 AI 转型。必须搭建稳固的基础能力平台，辅以流程和终身学习机制，让使用 AI 或构建机器学习模型像日常分析数据一样自然。 --- ### [15:10](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=910s) 数据治理即业务运营：嵌入组织毛细血管 Rachel 对 Valeh 的“进攻型治理”理念表示高度赞同，并作了进一步升华。她指出，许多企业把数据治理当成高高在上的概念、原则或理论框架（Framework）去谈论，这种做法本身就是极其危险和错误的：“数据治理是一项真实的业务运营活动（Data governance is a business operation）。” ```mermaid flowchart LR accTitle: Hub and Spoke Governance Model subgraph HUB["治理枢纽 (Hub - 中心平台)"] H1["企业统一标准"] H2["元数据管理系统"] H3["治理工具链与协议"] end subgraph SPOKE["业务辐射端 (Spoke - 业务单元)"] S1["业务逻辑与规则落地"] S2["AI 与大模型场景使能"] S3["日常运营与价值转化"] end HUB <--> SPOKE ``` 数据治理必须深深扎根在企业的日常运营肌理中，具备明确的执行能力载体，例如元数据管理（Metadata management）。Rachel 指出，数据治理不仅是每个人的日常职责，更是网络安全团队的第一道防线。缺少了数据治理提供清晰的数据底账，网络安全（Cybersecurity）的所有防护措施在落地时都会“面朝下摔得粉碎（falls flat on its face）”。因此，她日常与企业首席信息安全官（CISO）保持着极其紧密的联动。 在组织架构实践上，McKesson 历经两年半的沉淀，采用了“轴辐模式”（Hub and Spoke Model）： - **枢纽（Hub / 中心层）**：由中央团队统一输出顶层标准、治理协议与基础工具支撑。 - **辐条（Spoke / 业务端）**：将治理能力派驻并渗透进各业务单元内部，因为“真正的商业奇迹只会在一线业务中迸发”。 Rachel 在发言尾声分享了一个生动的职场轶事：她曾对自己的主管开玩笑说：“我的终极工作目标就是‘让自己失业’（My job is to work myself out of a job）。”当时主管听后大惑不解，而 Rachel 解释说，事实正是如此——当组织内部不再有人把“数据治理”作为一个独立而特殊的行政口号到处念叨，而是自然而然地将其内化为业务运营的标准动作去赋能 AI 和业务提效时，数据治理才真正取得了最高维度的成功。 ### [17:04](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=1024s) 非结构化数据的治理挑战与合规防线 主持人 Ravit Jain 将话题从跨部门协作与防御机制切入到数据形态与组织技能上。他指出，随着 AI 的演进，企业内部究竟在用什么样的数据喂养模型，直接决定了对系统的信任度，而处理非结构化数据更是带来了极其复杂的治理门槛。Rachel Smith 开玩笑回应道，Ravit 简直像是收了别人的钱专门抛出这个痛点问题一样。她坦言，在大多数企业中，非结构化数据绝对是治理工作面临的最大难题，要真正掌控它非常令人望而生畏。 Rachel 表示，McKesson 在这一领域投入了大量精力，优先聚焦于受强监管或涉及隐私的高风险数据。虽然不便透露具体的技术产品名称，但团队已经部署了一系列工具和流程来主动探测与管控这些非结构化数据。当前企业想要构建的绝大多数 AI 模型，其底层几乎都依赖海量的非结构化数据，因此全员教育至关重要。员工必须清晰理解其中的合规风险；同时，即便数据治理团队不直接拥有“AI 治理”的全部管辖权，也必须在模型开发和 AI 产品落地的过程中建立验证机制，防止有人直接从生产线上随意截取数据，从而制造出本应受到严格管制的衍生内容。此外，频繁的企业并购（M&A）与多元业务部门的整合，让原本就复杂的非结构化数据治理雪上加霜。 - **高风险优先**：首要精力集中在具有监管合规要求（Regulated）和隐私（Privacy）约束的非结构化数据上。 - **全员教育与风险认知**：让普通员工理解非结构化数据流转中的潜在法律与合规风险，树立红线意识。 - **验证机制**：在 AI 模型构建阶段把好数据关，确保输入模型的数据来源合法，而非从业务生产线非正规抓取。 - **并购（M&A）复杂性**：业务单元不断整合带来了多套异构体系，使原本相对容易处理的结构化数据治理也受到牵连。 --- ### [19:33](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=1173s) 麦当劳的全球非结构化数据分类与留存实践 面对庞大的非结构化数据版图，麦当劳的企业数据治理总监 John Tucker 分享了快餐帝国的独特挑战。在麦当劳，遍布全球的食谱卡片（Recipe Cards）、菜单设计以及各类新研发产品，构成了规模惊人的非结构化资产。最大的难点不仅在于数据本身格式杂乱，更在于“根本搞不清楚这些信息究竟散落在哪里”。 为了破解这一难题，John 的团队与麦当劳的首席信息安全官（CISO）深度合作，制定了一套贯穿全公司的全局数据分类规范。这套规范最初是以强制性政策（Policy）的形式推行，随后演进为更具实操性的指导方针（Guideline），帮助各团队理清文档的敏感级别与保留周期（Data Retention）。这项庞大工程并非治理团队单打独斗，而是整合了数据治理团队、网络安全团队、专门负责留存合规的信息治理团队（Information Governance），以及隐私合规与法务团队。麦当劳采取了逐个国家推进的策略，首先从拥有最大餐厅体量的美国本土市场试点落地，随后逐步向欧洲等各大国际市场推广。 ```mermaid flowchart TD accTitle: 麦当劳非结构化数据跨部门治理框架 subgraph Governance["协同治理生态"] CISO["CISO 网络安全团队"] DG["数据治理团队"] IG["信息治理团队<br/>(数据留存)"] Legal["法务与隐私合规"] end subgraph Policy["规范落地"] Class["数据分类指南<br/>(敏感级别评估)"] Ret["生命周期管理<br/>(留存时限标准)"] end subgraph Rollout["全球分步推广"] US["第一阶段: 美国市场<br/>(最大体量试点)"] Global["第二阶段: 国际市场<br/>(逐国推广适配)"] end Governance --> Policy Policy --> Rollout ``` --- ### [20:50](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=1250s) 避免重复造轮子：PASHA Holding 的领域所有权与数据网格 PASHA Holding 的首席数据官 Valeh Nabiyev 提出了不同视角的务实观点。他认为在面对非结构化数据时，业界没有必要去“重新发明轮子”（Reinvent the bicycle）。数据治理的核心底层架构并不应该因为数据是结构化还是非结构化而发生根本性逆转。在 PASHA Holding，他们依托数据网格（Data Mesh）架构运作，各独立子公司在集团旗下分别作为独立的业务领域（Domain）运作，例如涵盖客户数据、交易数据等。 Valeh 指出，尽管他们目前深耕于金融服务、银行和保险领域，核心数据以结构化为主，但要支撑 AI 转型就必须同样管好非结构化数据。前几年行业内热烈探讨的领域所有者（Domain Owners）、领域管家（Domain Stewards）和领域托管人（Domain Custodians）等角色，在非结构化数据治理中依然完全适用。顶多是针对非结构化形态增设特定领域或辅助角色，但核心治理理念绝对不变——必须由产生并运营该数据的业务部门（Business Units）来承担数据所有权（Ownership）。Valeh 强调：“只要大家依然把数据团队视作唯一要对数据质量（无论结构化还是非结构化）负责的人，那么这套治理就注定无法成功。” - **数据网格架构**：以业务领域（Domain）为中心划分边界，各业务单元对自身产生的数据负全责。 - **治理模型通用性**：反对因非结构化数据而另起炉灶建立全新治理体系，主张沿用已成熟的领域管家与所有权体系。 - **摆脱治理背锅侠**：数据质量的责任必须落在业务端，中心化数据团队绝不能单独包揽一切数据问题。 --- ### [23:18](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=1398s) 麦当劳的 AI 落地用例：管控闭环与治理提速 聊到具体的实战用例时，John Tucker 将麦当劳的探索归结为两条腿走路：一是为全局 AI 战略设置防护控制（Controls），二是通过 AI 加速数据治理自身的日常运作。在管控侧，麦当劳今年上线了三个非常接地气的用例： 1. **AI 用例库（Use Case Library）**：对全集团正在开发或评估的所有 AI 模型和用例建立统一目录。 2. **合规透明度与复用追踪**：梳理各用例在数据生命周期中所处的位置与潜在合规风险，打破跨国信息孤岛（例如西班牙团队想做某项探索时，可以一眼看到美国团队是否已有成熟成果，避免重复造轮子）。 3. **透视评估流程（Assessment Transparency）**：打破传统工单系统（如 ServiceNow）如同“黑盒”一般的申请体验，公开透明地展示评估指标与业务价值诉求，重点审查第三方应用所集成的底层模型及其内在风险，并将这些分类结果对全员开放。 在日常治理加速侧，随着业务用例和外部技术的井喷式爆发，团队必须跟上节奏。麦当劳梳理出了约 9 个赋能用例，一部分用于协同隐私和安全审查，另一大部分则集中于**元数据管理与语义整理（Metadata Curation）**。John 强调，业务上下文（Business Context）是决定 AI 成功与否的关键；AI Agent 或各种落地应用对企业业务的理解越深入，给出的产出结果才越精准。 --- ### [25:25](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=1525s) McKesson 的元数据智能化与治理颠覆 Rachel Smith 接过话茬，分享了 McKesson 围绕第三方数据（Third-Party Data）开展的智能化探索。为了在提供业务自主权的同时建立防线（Guardrails），团队在内部构建了“数据智能”（Data Intelligence）机制。最具代表性的突破在于：他们成功对海量数据集进行了自动化编目，并在**不强制占用业务人员宝贵时间**的前提下，自动提炼并补齐了业务上下文。 Rachel 激进地指出，“数据治理本身也亟需被颠覆”。过去治理体系极度依赖人工设定的数据所有者和数据管家，虽然流程中依然需要保留人工介入（Human-in-the-loop），但治理团队通过数据目录和元数据管理已经积累了海量的数据智能，完全足以开始自己训练并构建专属的治理 AI 模型。在诸如药品代码（Pharmacy IDs）等高度标准化、跨行业的医疗业务场景中，技术完全可以自动完成匹配与治理。治理团队的目标不是取代业务部门，而是放大（Amplify）他们的能力，将员工从“在办公桌边角料时间里处理琐碎治理事务”的痛苦状态中解放出来，专注于高价值的研究和解决方案交付。 - **无痛业务编目**：利用智能化工具从元数据中自动抽取上下文，打破“必须拽着业务部门开会才能定标准”的传统低效模式。 - **自我颠覆的治理模式**：从被动审批转向主动构建治理专用 AI 模型，与大型科技供应商合作实现元数据自动化。 - **赋能而非替代**：通过自动化接管常规重复工作，让业务专家回归业务本身，实现跨部门共赢。 --- ### [27:53](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=1673s) PASHA Holding 的联邦数据中台与自助式服务 Valeh Nabiyev 探讨了跨多元成熟度实体的平台化实践。PASHA Holding 旗下各成员企业的数字化与治理成熟度参差不齐，因此催生了差异巨大的分析型用例——有的企业在开发基于消费者商超消费行为的替代性信用评分模型（用于替代单一征信评分放贷），有的则在运用 AI 进行智能市场分析。 为了从底层支撑这些异构需求，PASHA Holding 数年前构建了一个中心化的联邦数据平台（Central Federated Platform）。Valeh 特别澄清，这绝非传统的中心化数仓或数据湖去强行控制一切，而是一个**对齐各方的联邦式数据中间人（Broker）**。由于涉及极其严格的银行业保密法、保险安全法以及医疗保密法规，平台内置了多层安全控制与数据最小化工具。各机构如果需要跨公司调取数据，数据需统一流经该平台并按规范交付。在与生态伙伴（如赞助商 Dataiku 等）完成能力复制后，团队进一步推出了“自助式服务”（Self-service）用例。业务人员现在可以直接进入标准数据目录平台（无论底层是 Informatica 还是其他工具），按需查看并评估数据质量，自主发起数据申请，数据获批后直接进入其专属分析沙箱供其进行跨表关联与高级建模，成为了推动业务转型的核心发动机。 --- ### [30:53](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=1853s) 数据治理的属性之争：运营手段还是行政合规？ 在讨论接近尾声时，主持人 Ravit 将话题引向管理本质：数据治理到底更偏向于行政管控（Administrative）还是日常运营（Operational）？ Rachel Smith 给出了一个鲜明的观点：“它在理想状态下应该是 **75% 的日常运营和 25% 的行政管控**。”然而现实中，大多数企业由于数字化成熟度不足，情况往往恰恰相反。许多团队把绝大多数精力浪费在疲于奔命的“合规说教与全员扫盲”上。治理团队不得不花时间去说服拥有极大自主权的业务主管，甚至必须刻意告诫团队“千万别对业务部门讲治理行话（Governance Speak）”。比如与其空谈元数据血缘，不如直接问业务：“你发出的这个外部数据接口如果被质询，你能立刻说清它的源头在哪里吗？”当对方回答“不知道”时，他们才会真正理解治理的运营价值。 John Tucker 对此深表赞同。他透露麦当劳在过去三年半的治理体系建设历程中，同样经历了这个转型阶段。从早期 50/50 的拉扯状态，麦当劳正在逐步向 75% 的运营赋能迈进。早期的治理工作就像是在“赶鸭子上架”，把所有精力都耗费在查验员工是否遵守政策、控制项是否到位；而现在的核心重心已经彻底扭转为——**如何真正为业务赋能并提速**。 > “我过去常跟团队强调：请千万别跟业务部门讲那些治理术语黑话。帮助他们理解痛点最有效的方法，就是问他们现实的问题——比如他们维护的数据接口出了故障时，他们能否说出数据的源头到底在哪里。” —— Rachel Smith ### [33:03](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=1983s) 多 Agent 架构赋能数据访问与上下文共享 在探讨数据治理如何赋能业务时，John Tucker 阐述了将治理能力工具化并融入日常工作流的具体路径。过去数据治理往往陷入“赶猫式”的拉扯，耗费大量精力去协调各类指标和标准；但一旦转向组织赋能，核心挑战就变成了如何加快数据获取速度。对于刚入职的新员工，智能 Chatbot 不仅可以直接回答基础的数据疑问，还能引导他们定位到数据目录（Data Catalog）中的具体资源。治理团队还在探索通过 AI 辅助校验 SQL 查询语句，确保分析人员访问的是正确的数据库和数据源，并将这些能力无缝部署在员工最常用的 Slack 或 Microsoft Teams 等协作工具中。 针对主持人 Ravit Jain 关于是否引入 Agent（智能体）的追问，John Tucker 透露团队目前正在孵化 9 个 Agent 用例。整体架构采取“分域治理与跨域贯通”的模式：在微观层面，为每个业务域（Domain）配备专属 Agent；在宏观层面，设立一个跨域 Agent 负责串联各个领域，解答涉及多源数据集关联的复杂问题（例如将客户数据集与交易数据集进行交叉匹配）。由于底层数据具备了更深度的语义上下文，这种 Agent 架构显著缩短了业务端获取洞察的时间（Time to Insights）。更重要的是，这些 Agent 之间能够互相调用与通信，若企业其他业务场景需要上下文数据支持，可以直接调用治理平台的 Agent 接口获取信息。 ```mermaid flowchart TD accTitle: Multi-Agent Governance Architecture accDescr: Architecture showing specialized domain agents connected to a cross-domain governance agent and user interfaces. subgraph Interface ["交互与协作终端"] Slack["Slack / Teams 交互入口"] Catalog["数据目录 Data Catalog"] end subgraph GovernanceAgents ["治理与上下文 Agent 架构"] CrossAgent["跨域协同 Agent<br/>(处理跨数据源关联与治理逻辑)"] AgentA["客户域 Agent"] AgentB["交易域 Agent"] AgentC["其他业务域 Agent"] end subgraph External ["组织其他业务用例"] OtherApps["外部业务 Agent / 应用调用"] end Slack <--> CrossAgent Catalog <--> CrossAgent CrossAgent <--> AgentA CrossAgent <--> AgentB CrossAgent <--> AgentC OtherApps <--> CrossAgent ``` - **9 个 Agent 孵化用例**：覆盖特定业务域的专属 Agent 以及跨领域全局 Agent。 - **工作流即时集成**：将 SQL 校验与数据目录查询直接嵌入 Slack / Microsoft Teams。 - **Agent 间协同互通**：治理 Agent 作为上下文提供方，支持跨部门其他 AI 系统的 API 调用。 --- ### [35:26](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=2126s) 跨行业经验与高标准治理框架的灵活性 Rachel Smith 对多 Agent 在工作流中的构建深度表示赞赏，并联想到快餐连锁行业与强监管行业在合规审批流程上的差异，认为这为敏捷构建提供了便利。对此，John Tucker 分享了其职业路径带来的独特治理视角：他早年深耕于金融服务业，随后经历汽车行业，如今来到快餐连锁（QSR）领域，但他在构建底层数据与 AI 治理框架时，依然沿用金融行业严格的隐私政策和合规标准。 虽然餐饮行业除美国食品药品监督管理局（FDA）相关规定外整体监管较为宽松，但团队在架构设计上坚持多维度设防，力求打造具备高弹性的治理框架。“即便当前身处去监管化（Deregulation）的环境中，谁也无法保证三四年后政策会发生怎样的改变。”这种以金融级标准搭建但保持高度灵活性的体系，使企业能够在政策与市场风向变化时迅速调整转向。 --- ### [36:29](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=2189s) 治理即变革：控股集团的“铺轨”与跨业务赋能 针对数据治理的本质定位，Valeh Nabiyev 明确指出，数据治理“既非纯粹的运营性事务，也非简单的行政职能，从根本上说它是一场业务变革（Transformational）”。他以控股集团（PASHA Holding）的运作模式打了个生动的比方：总部核心团队的职责是“铺设铁轨”，而铁轨上运行着各式各样的“火车头”——旗下子公司业务形态各异，成熟度参差不齐，从高度数字化的电商业务到强监管的传统银行业务皆有分布。 ```mermaid flowchart LR accTitle: Central vs Local Governance Model accDescr: Central team builds policies and rails while business units adopt and provide feedback. subgraph Central ["集团总部数据中枢"] direction TB Rails["铺设治理轨道 (Policies & Frameworks)"] ThirdParty["第三方 AI Agent 接入治理与安全机制"] end subgraph BusinessUnits ["多元业务板块 (Locomotives)"] direction TB Bank["银行业务<br/>(高成熟度/强合规)"] ECom["电商业务<br/>(高敏捷/数字化)"] Others["其他子公司<br/>(成长中业务)"] end Rails -->|"适配与落地"| Bank Rails -->|"适配与落地"| ECom Rails -->|"适配与落地"| Others Bank -.->|"实践反馈"| Rails ECom -.->|"实践反馈"| Rails Others -.->|"实践反馈"| Rails ``` 为了应对这种复杂性，总部团队专注于提炼 Know-how、设计标准化政策、流程机制以及攻克前沿课题（如第三方 AI Agent 接入时如何确保其接触企业数据过程中的可控与合规），而不直接插手各业务线的日常运营。各子公司本土团队则负责将通用框架本地化适配，并在落地过程中向总部反馈偏差。“无论是谈论数据驱动的决策文化，还是 AI 驱动的全面转型，数据治理都是这场变革的核心燃料和主力引擎。” --- ### [38:31](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=2311s) 专家建言：从高价值用例切入与避免工具陷阱 在被问及给企业高管及数据治理从业者的建议时，Valeh Nabiyev 首先强调了“从小处着手，切勿贪大（Never start big）”的原则。试图对企业所有数据进行一刀切式的全盘治理从一开始就行不通，因为直接证明数据治理自身的货币化投资回报率（ROI）一直是业内最大痛点。高管们常问：“我们计划在数据治理上砸数百万美元，能换回什么回报？” 打破这一僵局的有效策略是通过具体的业务用例来反哺治理。Valeh 以控股集团构建生态系统时的“客户终身价值（Customer Lifetime Value, CLV）”项目为例：团队将数据质量与治理规则作为交付 CLV 用例的并行模块一同上线，直观展示了治理如何赋能业务增收。一旦赢得管理层和业务端的信任，后续向全集团推广和规模化扩张便水到渠成。 此外，Valeh 提醒从业者警惕“工具选择陷阱”。市面上的数据治理平台与工具功能大同小异，很多团队往往耗费过多时间纠结于工具选型。“最关键的不是你选择了哪款工具，而是你的团队具备怎样的能力；拥有一支成熟专业的团队，无论用什么工具都能打胜仗。” - **启动策略**：绝不要试图治理所有数据，优先绑定高商业价值的单一用例（如 CLV）。 - **价值证明**：将数据治理指标与业务项目成果捆绑交付，以显性价值获取高管支持。 - **工具认知**：平台功能高度同质化，避免陷入选型内耗，核心在于团队的人才能力与执行成熟度。 --- ### [41:16](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=2476s) 变革管理与跨职能协同：构建 AI 治理的“联合战线” John Tucker 将“用户采纳（Adoption）与变革管理”列为治理成败的首要关键。再优秀的工具和流程，若无法被业务团队真正用起来便毫无价值。针对开启 AI 治理之旅的企业，John 提出了两点核心建议： 1. **组建跨职能协同阵营**：治理团队必须深度联合网络安全（Cybersecurity）、隐私合规、企业架构、工程研发以及 AI Hub / 数据科学团队，共同制定护栏与控制策略。 2. **坚持业务赋能导向**：主动思考如何运用 AI 解决长期困扰业务的痛点，用实际解决问题的能力为治理项目赢得业务认可。 Rachel Smith 完全赞同上述观点，并回顾了自身早期的教训：团队初期曾试图包揽一切，结果“什么都没解决好”，花费了很长时间才重新获得业务部门的买单。她强调“治理即增长（Governance is growth）”，治理不仅关乎合规，更是业务可持续发展与收入增长的“保险单”。即便是不受严苛监管的行业，在处理控股公司协同或第三方数据交互时，也绕不开商业伙伴协议（BAA）等法律约束。 Rachel 再次呼应了 John 的跨职能协同观点，将治理比作需要全村合力的“集体村落（Collective Village）”。网络安全、企业架构、合规、隐私与法务团队必须紧密绑定，清晰划分各个职能的职责泳道（Swim lanes），明确每个环节的责任归属（Accountability & Responsibility）以及知会对象，确保业务在合规与创新的双轨上并行加速。 ## Notable quotes > [02:06](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=126s) AI transformation is not about AI use cases — Valeh Nabiyev > [06:41](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=401s) It starts with the data — Rachel Smith > [12:00](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=720s) Decreasing time to insight and time to data — Valeh Nabiyev > [15:10](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=910s) Data governance is a business operation — Rachel Smith > [15:10](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=910s) falls flat on its face — Rachel Smith > [15:10](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=910s) My job is to work myself out of a job — Rachel Smith > [20:50](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=1250s) 只要大家依然把数据团队视作唯一要对数据质量（无论结构化还是非结构化）负责的人，那么这套治理就注定无法成功。 — Valeh Nabiyev > [30:53](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=1853s) 我过去常跟团队强调：请千万别跟业务部门讲那些治理术语黑话。帮助他们理解痛点最有效的方法，就是问他们现实的问题——比如他们维护的数据接口出了故障时，他们能否说出数据的源头到底在哪里。 — Rachel Smith > [35:26](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=2126s) 即便当前身处去监管化（Deregulation）的环境中，谁也无法保证三四年后政策会发生怎样的改变。 — John Tucker > [36:29](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=2189s) 无论是谈论数据驱动的决策文化，还是 AI 驱动的全面转型，数据治理都是这场变革的核心燃料和主力引擎。 — Valeh Nabiyev > [38:31](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=2311s) Never start big — Valeh Nabiyev > [38:31](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=2311s) 最关键的不是你选择了哪款工具，而是你的团队具备怎样的能力；拥有一支成熟专业的团队，无论用什么工具都能打胜仗。 — Valeh Nabiyev > [41:16](https://www.youtube.com/watch?v=fWUrJhc_M0Q&t=2476s) Governance is growth — Rachel Smith ## Mentioned Ai4, Ravit Jain, The Ravit Show, Rachel Smith, McKesson, John Tucker, McDonald's, Valeh Nabiyev, PASHA Holding, HIPAA, PHI, PII, CCPA, GDPR, Microsoft Copilot, ServiceNow, Dataiku, Informatica, Slack, Microsoft Teams, FDA ## Video description Rachel Smith, Global Head of Enterprise Data Governance & Third-Party Data Solutions, McKesson Ravit Jain, Founder & Host, The Ravit Show John Tucker, Director, Enterprise Data Governance, McDonald's Valeh Nabiyev, Chief Data Officer, PASHA Holding Abstracts This panel will focus on the critical role of data governance in ensuring the success of AI initiatives. Experts will discuss best practices for managing data privacy, security, and regulatory compliance while maintaining data quality across your organization. Learn how strong governance frameworks enable organizations to mitigate risks, eliminate data silos, and build trust in AI-driven decision-making. Join the conversation to discover how to implement effective governance strategies that align with your business goals and the evolving AI landscape. [Source](https://www.youtube.com/watch?v=fWUrJhc_M0Q)

[2:06] Valeh Nabiyev: AI transformation is not about AI use cases [6:41] Rachel Smith: It starts with the data [12:00] Valeh Nabiyev: Decreasing time to insight and time to data [15:10] Rachel Smith: Data governance is a business operation [15:10] Rachel Smith: falls flat on its face [15:10] Rachel Smith: My job is to work myself out of a job [20:50] Valeh Nabiyev: 只要大家依然把数据团队视作唯一要对数据质量（无论结构化还是非结构化）负责的人，那么这套治理就注定无法成功。 [30:53] Rachel Smith: 我过去常跟团队强调：请千万别跟业务部门讲那些治理术语黑话。帮助他们理解痛点最有效的方法，就是问他们现实的问题——比如他们维护的数据接口出了故障时，他们能否说出数据的源头到底在哪里。 [35:26] John Tucker: 即便当前身处去监管化（Deregulation）的环境中，谁也无法保证三四年后政策会发生怎样的改变。 [36:29] Valeh Nabiyev: 无论是谈论数据驱动的决策文化，还是 AI 驱动的全面转型，数据治理都是这场变革的核心燃料和主力引擎。 [38:31] Valeh Nabiyev: Never start big [38:31] Valeh Nabiyev: 最关键的不是你选择了哪款工具，而是你的团队具备怎样的能力；拥有一支成熟专业的团队，无论用什么工具都能打胜仗。 [41:16] Rachel Smith: Governance is growth

[0:43] Rachel Smith: when you talk about AI, it starts with the data [2:08] Rachel Smith: advance healthcare outcomes for all [33:03] John Tucker: herding cats

[0:04] Welcome everyone. Uh we are here at AI4 [0:06] and I'm super excited. I'm your host and [0:09] founder of the Robert show. I'm here [0:11] today to discuss about navigating data [0:14] governance in the AI era. I have here [0:16] with me the great enterprise leaders [0:18] who've been doing great things in the [0:20] data governance space and of air. [0:22] Obviously we were chatting a lot about [0:23] AI governance how data governance has [0:25] been a strong pillar. Uh but before [0:28] getting into you know the discussion I [0:30] think a quick intro would be fantastic [0:32] I'm pretty sure. Maybe Rachel you can go [0:34] first. [0:34] >> Awesome. Thank you. Good afternoon [0:36] everyone. I'm Rachel Smith and I lead [0:38] the enterprise data governance for um [0:40] McKessan and thirdparty data solutions. [0:43] Um how many of you have heard of [0:44] McKessen before? [0:46] >> Oh good. There are some in here. There's [0:48] a lot. Awesome. A lot of times I most [0:50] people haven't heard of McKesten. We are [0:52] the largest healthcare global healthcare [0:54] um company in the US for Fortune 9 [0:58] approximately 80,000 employees and we [1:00] are about 400 billion plus revenue. So [1:03] pretty big company. Glad to see there's [1:05] some folks here with the recognition. [1:07] >> Fantastic. John. [1:09] >> Sure. Uh John Tucker. I lead the data [1:11] governance practice for McDonald's [1:13] globally. Um we are one of the largest [1:16] quicks service restaurants with 41,000 [1:18] restaurants around in 110 countries. So [1:22] uh it's been very uh interesting [1:25] navigating the governance in AI here for [1:27] sure. [1:28] >> We obviously don't know McDonald's right [1:31] question [1:31] >> John we love it. Um Val, [1:34] >> hello everybody. Um my name is Val. I am [1:38] CEO in Pasha Holding. [1:40] Maybe nobody of you know what's a Pasha [1:43] holding. Uh Pasha holding it's a big [1:47] diversified holding in Azerban. It's [1:50] Kosus region and we are making mostly [1:54] financial services, banking, insurance, [1:58] marketplaces, fintech, [2:01] construction, very diversified business. [2:04] >> Fantastic. Yeah, thanks for those [2:06] introductions. is uh quickly wanting to [2:08] deep dive into you know what's happening [2:10] in the AI and data governance how is it [2:12] changing in different organizations I [2:15] know you all come from diverse [2:16] organizations here Rachel how does it [2:18] work in your organization what's [2:19] happening [2:20] >> there's a lot happening uh we we are [2:22] investing uh in AI and innovative [2:25] solutions when you think of our um you [2:28] know distribution centers uh supply [2:30] chain whether it's automation robotics [2:33] that's one area the other area for us is [2:35] we look at you know ways in we can serve [2:37] and delight uh our customers. And so [2:40] just in terms of of time, speed to [2:43] delivery and getting things uh timed the [2:45] right way. And the last one is uh we are [2:48] obviously one of the biggest providers [2:49] from an oncology perspective. We're [2:51] using it for research and how can we [2:54] make improvements and you know we think [2:56] about cancer care and outcomes which one [2:58] of our biggest uh purposes is how do we [3:00] advance healthcare outcomes for all. So [3:03] uh lots of exciting things happening uh [3:05] in healthcare and we and we heard this [3:07] throughout the conference that [3:08] healthcare is going to be one of the [3:10] biggest frontiers when it comes to AI [3:11] improvements and breakthroughs. [3:13] >> Totally I agree when it comes to [3:14] regulated industries and data kind of [3:16] plays the most important role may be [3:18] healthcare finance and other regulated [3:20] industries. John, what's happening at [3:22] McDonald's? [3:23] >> I would say at McDonald's we're in that [3:25] space where we're now starting to think [3:27] of how do we embed AI into all of our [3:29] strategies and how we run our [3:30] operations. So how does that impact our [3:32] customers, our crew members, but also [3:34] how do we have efficiency gains within [3:37] our various different lines of business [3:38] as well. So right now there's a scramble [3:41] of putting that strategy together in the [3:43] business and then presenting that back [3:44] to the sea level sweep, but also as you [3:47] think of regulations now they're a [3:48] little bit more laxed than the US but [3:50] you know in the European region [3:52] obviously a little bit more stringent. [3:53] So we are trying to navigate that [3:55] accordingly across the global [3:57] enterprise. [3:58] >> Okay. uh what's happening [4:01] >> in Pasha holding uh I mean before going [4:04] to the AI we are now focused on the [4:07] building B2C ecosystem [4:09] >> be ecosystem it's the ecosystem [4:13] >> uh of the our companies [4:16] uh how our ecosystem said it's the first [4:20] digital integrated ecosystem in co [4:23] causes [4:24] uh and we are trying to uh make our [4:30] customers life easier by integration and [4:32] sim seamless integration of all our [4:35] banking marketplace [4:37] uh a wallet [4:39] >> uh public transportation all together I [4:42] mean and million customers daily using [4:44] our services in in our country [4:47] >> and coming to here from AI perspective [4:50] we are on the very beginning in AI [4:52] perspective uh we are we did a lot of [4:55] things from [4:57] data transformation y the datadriven [4:59] decision making culture creation in our [5:01] assets uh it's strong and robust uh [5:05] foundation al I mean data governance [5:08] already built and we will not forget [5:12] data governance in AI era as well I can [5:15] definitely say that yeah and we are now [5:17] doing mostly use cases in customer care [5:20] domain in operations and effectiveness [5:24] increase uh use cases, AI use cases, but [5:28] we we recognized and we understand that [5:31] AI transformation it's not about AI use [5:34] cases. [5:35] >> Mhm. [5:35] >> And for us, we trying now to to make it [5:40] in the way of top down and bottom up on [5:42] the best way. Top down is sharing [5:45] mission, vision, platforms, [5:46] capabilities. [5:48] bottom up it's a making every people in [5:51] our organization enabling every people [5:53] in our organization to use AI on their [5:57] daily life and and make it integ make AI [6:00] integral part of our organizational [6:02] processes [6:03] >> totally those are fantastic insights I [6:05] kind of also feel like we're just [6:06] scratching the surface when it kind of [6:08] you know we start talking about AI I [6:09] know there's a lot of content that goes [6:12] out and we've been hearing so many [6:13] different things different leaders [6:14] different uh companies as well you know [6:17] talking about data governance uh and the [6:20] AI governance as well. Uh but it for you [6:24] know I'm pretty sure for enterprises as [6:27] big as yours it just takes a while for [6:30] yall to you know say that okay this is [6:32] really working or we have really [6:34] implemented AI 100% and everyone's using [6:36] it's just it's just going to take a [6:39] little while. Uh but just just a [6:41] follow-up question around this was [6:43] around how do you all see the tech shift [6:48] happening the regulation or even the [6:50] culture what's the change there uh that [6:54] you see in your organizations maybe [6:56] Rachel [6:57] >> so I'll start so the foundational pieces [7:00] are critical [7:02] >> you talk about AI it starts with the [7:04] data [7:05] >> uh to answer your question in terms of [7:06] the shifts uh we see a lot of things and [7:08] I've heard in this conference [7:10] deregulation but for regulated [7:12] industries especially when it comes to [7:14] your privacy your HIPPA data uh PHI PII [7:18] CCPA you name all of those there's no [7:20] change there and so everyone wants to [7:23] know where is my data who's using my [7:25] data and for me at the enterprise level [7:28] the board is very interested most boards [7:30] corporate boards now have individuals [7:32] that are super technical um if you look [7:34] back in the last decade or so um I won't [7:37] name companies but we all know them [7:39] where they've had issues, right? And so [7:42] a lot of questions the board started [7:43] asking of CIOS is like, well, where is [7:45] your data? How are you able do you have [7:48] traceability? Do you know exactly who's [7:51] using that data? Uh third party data [7:53] comes with contractual restrictions as [7:54] part of that. And so what's shifting is [7:57] most companies are moving including us [7:59] to modernize your data your your [8:02] landscape end to end. Data governance [8:04] has been around uh for a long time. [8:06] However, data governance has seen many [8:08] many iterations and unfortunately in a [8:10] lot of places failed efforts because [8:12] companies did not make the necessary [8:14] investments to put individuals with um [8:17] it's truly a discipline. It is somewhat [8:18] of a technical as well as a business [8:21] discipline, right? So, putting someone [8:23] who doesn't understand um you know how [8:25] you go through sort of data architecture [8:28] uh and building things the right way is [8:30] is not a recipe for success. And so what [8:33] I've seen is that I spend a lot of time [8:36] at the C uh the table um with our CEO [8:40] showing him the capability, showing him [8:41] where his data is, showing him what [8:43] we're doing and how that's going to [8:44] ensure the questions if we were to get [8:47] them, you know, how he could sit with [8:49] comfort and be able to answer those [8:51] questions and that we also can comply. [8:54] um in places where you have and right [8:56] now this is becoming you know you think [8:58] of the federal level is one aspect but [9:00] at the state level they all have very [9:02] different requirements and so California [9:04] was one of the first ones to see CPA [9:06] when someone wants to remove their data [9:08] can you actually go and see back to the [9:10] source system and ensure that that data [9:12] has been removed so that really is the [9:15] shift in the conversations um from an AI [9:17] perspective the question is for me it's [9:20] all about data trust and so every data [9:23] that is being the data that's being [9:24] used. Can we trust it? And and that's [9:27] not as easy as saying yes or no, right? [9:29] And so have we actually built the things [9:31] in place and we're working with the [9:32] business units today. Do we have the [9:34] people in place that can build those [9:35] right business rules, validation, you [9:38] know, data transformation that needs to [9:39] happen across the building whether it's [9:41] a large language model um that's being [9:43] deployed etc. So it is uh it's a fun [9:46] time to be in this space. It has [9:48] definitely shifted from when I started [9:50] you know 15 so years ago and I think it [9:53] will continue to evolve. [9:54] >> Yeah totally uh I agree in terms of you [9:57] know when we kind of think about like in [10:00] the governance space at least trust kind [10:01] of plays a very important role and even [10:03] when you're kind of even working with [10:05] the AI uh the data quality kind of [10:08] matters a lot uh and for enterprise [10:10] leaders that should be number one [10:12] priority for sure. John what do you [10:13] think uh how's it how's the change [10:15] happening what what are you thinking [10:17] about the shift [10:18] >> yeah I think one of the shifts we have [10:20] seen is now that AI is at the forefront [10:22] you do get like even your chief human [10:25] resource officer now talking about it [10:26] and thinking about what's the literacy [10:28] around that and how do we educate the [10:30] organization to make sure that everybody [10:32] feels safe as we start to deploy many [10:34] different use cases because obviously [10:35] you're going to have efficiency gains [10:38] and most are going to start worrying [10:39] about hey what happens to my role now we [10:42] need to ensure them that hey now we'll [10:44] have more time to do other things more [10:46] strategic things and think about how we [10:48] can innovate within the business. So a [10:51] lot of our focus right now has been on [10:53] more so on the education side of it. So [10:55] as you mentioned earlier you know [10:56] putting AI in everyone's hands. We've [10:58] deployed, you know, Microsoft copilot [11:00] around the organization. We've opened [11:02] that up for every employee to start [11:04] using and leveraging. And they are [11:05] starting to see, hey, we can do some [11:07] pretty cool things, whether it's, you [11:08] know, transcription, whether it's doing [11:10] something with the emails or just a [11:12] general business process that you [11:14] typically would do yourself. Now, you [11:16] can have AI start to leverage that, [11:19] leverage that in your AI's use cases. [11:21] But I believe that the education has [11:23] been very important in the literacy [11:25] around how we're using it, the best way [11:26] to use and leverage it and when to use [11:28] and leverage it and just ensure that [11:30] everyone feels safe. Yeah, I think [11:32] responsibly using AI is like the keyword [11:35] your I kind of feel like uh and for [11:39] McDonald's obviously it's a huge company [11:41] a lot of employees lot big data team out [11:43] there and u making their job easier with [11:47] the tool is the key and if you all can [11:49] do that with AI nothing better than that [11:51] Pasha what do you think about you know [11:53] the shift that's happening in uh how do [11:56] you kind of see it internally at Pasha [11:58] as well [12:00] >> uh what we what we thing because my [12:02] again going before going to AI it's [12:06] important to to mention that data [12:08] governance also changes. [12:10] >> Yeah. [12:11] >> Yeah. So and I will say from perspective [12:14] of data governance [12:16] conditionally I'm dividing data [12:18] governance to their offensive data [12:20] governance and defensive data [12:22] governance. Yeah. And if most of [12:25] organizations considering data [12:27] governance as something barrier function [12:29] who will say no or function who will say [12:34] which GDPR rules or other regulation [12:36] rules should be applied [12:38] uh in our case data governance focused [12:42] on their offense side and uh we made it [12:46] intentional. Yeah, it's not meaning that [12:48] security and compliance things is not [12:50] important but from focus perspective [12:53] when your data governance function [12:55] mission is there enable organization and [12:59] focused on the decreasing continuously [13:01] of time to insight time to data this [13:04] insight can be generated by the way with [13:06] AI also yeah I mean uh and increasing [13:08] data quality which coming important in [13:11] AI topic as well preventing this garbage [13:14] in garbage out yep [13:17] I think nobody forgot about this [13:19] concept. Yeah. And data govern I feel [13:21] that data governance particularly in [13:22] this conference also I mean like [13:25] something like low. [13:28] So [13:30] um but yeah coming back to the question [13:33] mostly offense data quality increasing [13:36] data quality and decreasing time to [13:38] insight is very important for us and [13:41] with coming of the AI which shift comes [13:45] and uh I think important shift is there [13:50] uh that our organizations now should [13:53] consider this AI as part of their daily [13:56] life And to consider this as part of the [13:59] daily life, we should have robust data [14:01] governance. I'm not speaking about AI [14:03] governance even. Yeah, robust processes [14:06] inside uh of organization etc. Long [14:09] story short, I'm think that the very [14:12] important part is they're building this [14:14] platform of capabilities. [14:16] >> Yep. [14:16] >> And continuous learning because in AI [14:19] topic something will not be different [14:22] will be like in the data transformation [14:24] as well. I mean when we build datadriven [14:26] decision making it was about creating [14:29] dashboard. Okay we create dashboard it [14:32] be datadriven after creating dashboard. [14:34] No really no but how we made it to be [14:38] datadriven by uh making this agenda of [14:43] organization [14:45] creating platforms governance processes [14:49] people's capabilities and it's a normal [14:51] part of life okay you need to create [14:52] dashboard you can create ML model or you [14:55] can you can integrate AI [14:58] >> yeah I think those are various ways how [15:01] you know different organizations are [15:02] kind of you know focusing on different [15:04] uh aspects of data governance and AI [15:07] governance. Uh and good to hear from, [15:09] you know, different [15:10] >> I would like to touch on something that [15:11] he said as well. Um and I love that when [15:13] you said you know it's data governance [15:16] most organizations it becomes this thing [15:18] right and I like to say if you're [15:20] talking about data governance as a [15:21] principle and a framework then that in [15:24] itself is problematic. Data governance [15:26] is a business operation and that's what [15:28] we've done in our space. It should be [15:30] embedded in your organization. it it's [15:33] not something this thing you should talk [15:35] about. There's actual capabilities that [15:37] come along with that, right? When you [15:39] think of your metadata management and so [15:41] I believe a lot of companies now are [15:43] starting to see that shift and ensuring [15:45] that the business understands it because [15:48] quite frankly governance is a part of [15:49] everyone's role. It also is the [15:51] adjacency. One of my colleagues is here [15:53] from our cyber space and so our cyber [15:55] security team um you know working with [15:58] the cyber team is sort of first line of [16:00] defense but without this data governance [16:02] cyber everything they're doing fails it [16:04] falls flat on his face so I spent a lot [16:06] of time with our CISO so I think it's [16:08] really important especially for those [16:10] that are on very very journeys we [16:12] started this at the enterprise level [16:14] about two and a half years ago it does [16:16] need to be top down as you said but it [16:18] al also has to be bottoms up but top [16:20] down in the sense that the executive [16:22] table has to understand it and believe [16:24] in it and then empower the organizations [16:26] to do you know you have your spoke we [16:28] have a hub and spoke model so from the [16:30] center making sure your protocols and [16:32] sort of enterprise standards and tools [16:34] are there but from an organization [16:36] perspective the business is where the [16:37] magic really happens so I think that's a [16:39] really important point is that it is not [16:42] it's no longer uh hopefully in most [16:44] organizations uh people are walking [16:46] around talking about data governance in [16:47] fact I used to joke with my leader that [16:50] you know my job is to work myself out of [16:51] job and he laughed at me. He said, "Why [16:53] do you say that?" I said, "It really [16:54] is." And I have because I'm not talking [16:56] about governance. We're actually [16:58] centered in the business and you know uh [17:00] building capabilities that's helping [17:01] with all the AI enablement helping with [17:04] our cyber security organizations and [17:06] then with privacy and the compliance and [17:07] legal teams makes their jobs easier when [17:09] there are things that we have to go out [17:11] and defend. [17:12] >> Totally. I think uh those are fantastic [17:14] points Rachel and good addition to you [17:16] know what I said. uh I I want to also [17:19] shift gears a little bit towards you [17:20] know data and skills. Uh it becomes [17:23] interesting when we kind of you know [17:24] obviously talking about various shifts [17:26] that are happening. Uh but internally uh [17:29] it also depends a lot on how what data [17:33] like trust you mentioned or what data [17:35] are we feeding and then at the same time [17:37] the skill set. So what's the main [17:40] challenge in working with both [17:42] structured and unstructured data and [17:44] usually in your organization? And I'm [17:46] pretty sure uh there are various [17:47] challenges but in unstructured data it [17:49] just becomes a lot more kind of curious [17:52] to know a little bit. [17:53] >> Wow you Vinnie must have paid you my so [17:56] um unstructured is probably the biggest [17:58] challenge for most organizations [18:00] especially for us getting your hands [18:02] around it. The education piece is very [18:04] important and so um and it can be pretty [18:07] daunting. So it's really understanding [18:10] you know going after things where it's [18:12] uh regulated or privacy you know uh type [18:16] data is where we've spent a lot of our [18:18] time. uh we've put some tools in place [18:20] and processes uh leveraging I don't want [18:23] to name products here but where we are [18:26] going after that unstructured [18:27] unstructured data uh to make sure we [18:30] have that and I think for most companies [18:32] that's the biggest challenge because [18:34] most AI models that you want to build [18:36] likely are going to have a huge amount [18:38] of unstructured data as part of that so [18:41] um from an employee population [18:43] perspective you know you really got to [18:45] make sure people understand what the [18:47] associated risk are have policies in [18:50] place um in building some process. I [18:53] don't own AI governance in our [18:54] organization but as part of that is [18:56] they're building models and we have AI [18:58] products etc. It's we have a process [19:01] that we can actually validate that right [19:03] and make sure um that we're not creating [19:06] stuff from unstructured data that [19:08] probably should have originated [19:10] somewhere else and someone pulled it off [19:11] the production line somewhere along the [19:13] way. So it is a big big challenge for us [19:16] and it's one that we continuously uh try [19:18] to address every single day. Structured [19:20] is a lot easier and then also probably [19:23] like a lot of other companies when you [19:24] bring together different business units [19:27] um you know we do a lot of mergers and [19:29] acquisitions as well. So that add on top [19:31] of that just makes it a little more [19:32] complicated. [19:33] >> Yeah 100%. Uh John, how does it work at [19:36] McDonald's? And I'm pretty sure you work [19:38] with a lot of unstructured data all the [19:39] time. [19:40] >> Absolutely. When you think of all the [19:42] recipe cards, menus, the products that [19:44] we create across globally, I mean it is [19:47] is very much a challenge, but also [19:49] understanding where that information [19:50] lives. Yeah, [19:51] >> we have that challenge as well. So, we [19:53] did spend a lot of time with our um CISO [19:56] and developing a data classification [19:58] policy that was a little bit more [20:00] >> I would say became more it was a policy [20:02] became a guideline and that's actually [20:04] kind of helped us you know classify all [20:06] of our documentation. [20:08] >> We're still going through that process. [20:09] we're nowhere near complete, but going [20:11] through the classification of [20:13] understanding the sensitivity level, but [20:15] also how often we should retain that [20:16] information as well. So, really tying [20:18] those things together, it really took a [20:21] a a big team to pull that off with the [20:23] governance team, our uh cyber security [20:26] team, we have a uh information [20:28] governance team, which is more around, [20:30] you know, the retention side of it, but [20:32] um bringing all those folks together [20:33] along with our privacy and legal teams, [20:35] that was uh the big focus for us. So it [20:38] is something we're doing I would say [20:39] country by country. We're taking that [20:41] approach. We started in the US first [20:43] because that's our largest restaurant [20:45] population. Now we're going through all [20:47] of our international markets and taking [20:48] that approach as well. [20:49] >> That's fantastic. [20:52] In our case um if we speak about [20:56] particular unstructured data I would [20:59] like to mention that I don't think that [21:01] in case of unstructured data data [21:04] governance overall data governance model [21:06] should be different. Yeah [21:09] >> I mean in Pasha holding we have the what [21:11] what we have at the moment we have the [21:14] think about it like data mesh. Yeah, [21:17] like each organization each one domain [21:19] of course inside of in each organization [21:21] organization has their domains customer [21:23] transactional etc. [21:24] >> Got it. [21:25] >> I mean it's it's a huge amount of the [21:28] data. Yeah. But mostly structure because [21:30] there now focused a lot on financial [21:32] services, banking, insurance. [21:34] >> Okay. [21:34] >> Most of the data is structured. But of [21:36] course for AI for [21:39] AI transformation for delivering AI use [21:41] cases management and governance of [21:43] unstructured basis important as well and [21:47] do you do you remember that few years [21:49] ago everybody speak about domains I mean [21:51] you are also speaking about this in your [21:53] channel domains [21:55] domain owners domain stewards domain [21:58] custodians right [21:59] >> I think it shouldn't be different why we [22:01] are trying to reinvent bicycle it's my [22:04] personal opinion [22:06] I think that the only thing which should [22:08] change it [22:10] additional creating additional domains [22:13] maybe with additional roles I don't know [22:15] but the overall model should should be [22:17] the same the whole idea is a creating [22:20] ownership of the data in the business [22:22] units which operates I mean the moment I [22:26] mean while we are considering data [22:28] governance and data functions as the [22:30] people who is responsible for their [22:33] quality of structured all un even [22:36] unstructured data I think it's not going [22:38] to fly [22:41] >> yeah no I think uh these are definitely [22:43] great insights and I I was kind of [22:45] thinking about how d how diversified the [22:48] space is as well where every [22:50] organization [22:52] uh you know works differently according [22:55] to the use cases that they have uh where [22:58] for you it could be data mesh for maybe [23:01] John it's a little different and for [23:02] Rachel as well uh but I'm kind of [23:04] curious curious if you all have any use [23:06] case that you would like to also share [23:07] with our audience today. Uh and I don't [23:11] want you to name companies obviously but [23:13] any use case that comes to your mind and [23:15] could be super interesting. Maybe John [23:16] you want to go first. [23:18] >> Yeah I think you know we're looking at [23:19] this in two ways as I said earlier we're [23:21] thinking about how do we embed AI in our [23:23] day-to-day strategy but also my job is [23:26] also how do we ensure we have the right [23:28] controls in place as well. So for us on [23:30] the control side, we've um had three use [23:33] cases we launched this year. Very [23:34] practical. One is just having a use case [23:37] library of all the various different [23:38] models and use cases across the [23:40] organization. [23:41] >> The second was the transparency. So as [23:44] you think of the various policies, laws [23:47] um uh that are coming out, we wanted to [23:49] ensure that what's the risk around all [23:51] those various use cases that we have uh [23:54] cataloged, but also where are they at in [23:56] the data life cycle, right? So you're [23:58] also not re-engineering or having, you [24:00] know, one market in Spain is thinking of [24:02] doing something. We've already done that [24:04] in the US. Can they leverage what's [24:05] already out there and available? So [24:07] those are things that we're thinking [24:08] about. And then the last piece is more [24:10] around the assessment and how that's [24:12] done and having more transparency. [24:14] Usually it's a black box. Many of us [24:15] probably have like service now or [24:17] whatever. They use that that framework [24:19] to you know put requests in. But here [24:22] what we've done is expose here's that [24:24] assessment. Here's the things we need to [24:25] know. what's the business value you're [24:27] looking to get out of this so we can [24:28] start tracking that but also you know [24:31] working if we're using thirdparty uh [24:33] applications trying to figure out what [24:35] are the underlying models underneath [24:37] that what are the inherent risk that we [24:39] need to be aware of and then classifying [24:41] putting certain classifications on those [24:43] and then exposing that to anyone in the [24:45] organization to see and view that. So [24:47] that's on the control side. Now, if you [24:49] think about our day-to-day work in [24:51] governance, we're trying to do things [24:53] faster and quicker so we can keep up. [24:55] Exactly. [24:55] >> These use cases are going to keep [24:56] coming. The technology is changing. So, [24:58] we have roughly I would say nine [25:00] different use cases we've identified [25:02] where one we can help with our privacy [25:05] security teams and then the others is [25:07] just more around the metadata curation [25:09] because that context [25:10] >> is definitely what is going to be key to [25:12] driving success in AI. the more context [25:14] you have and the better it is aligned [25:16] with the business, the better the [25:18] outcome you'll get from the various [25:20] different things you're asking of an AI [25:22] agent or some sort of use case that [25:24] you're deploying. [25:25] >> Yeah, I love it. Uh these are great uh [25:27] use cases that you you've shared. Um [25:29] Rachel, do you have any? We we do um not [25:32] as advanced as where you are but we are [25:35] we are starting well because a lot of it [25:37] is we we do want to give the business [25:39] autonomy um with guardrails and so we [25:42] started sort of retrofitting to what are [25:45] the things that you're doing uh similar [25:46] to what you described from my own team [25:49] one of the use cases we've done is [25:50] around our third party data and so our [25:52] thirdparty data uh one use case we [25:54] wanted to do is to have sort of data [25:56] intelligence internally so that everyone [25:58] could know where that data is right And [26:00] uh we are we actually were able to [26:02] catalog all of this data sets um get to [26:05] business curation without pulling the [26:07] business in because the biggest [26:08] challenge for a lot of us is that [26:09] business context [26:11] >> and uh so one of the things we're also [26:13] looking to do is at least for me when [26:15] you mentioned talent is there's a lot of [26:17] capabilities now I believe governance [26:19] itself needs to be disrupted as well and [26:22] so while we've traditionally thought of [26:23] having data owners and data stewards [26:25] there does need to be that human in the [26:27] loop but there's enough at least for us [26:28] we're amassing enough uh data [26:31] intelligence in the data you know [26:32] through the data catalog and metadata [26:34] management that I believe we can [26:35] fundamentally start to build you know [26:37] models ourselves and we're talking some [26:39] of the some of the bigger companies that [26:40] we're partnering with and how can they [26:42] help us with that um the biggest hurdle [26:45] a lot of times is you know obviously the [26:47] legal and the privacy teams and [26:48] compliance around understanding well how [26:51] are we going to validate right but I [26:53] believe that there's enough of um you [26:56] know intelligence when you think of like [26:58] pharmacy IDs and things like that. I [27:00] mean that's pretty standard across the [27:02] space. And so those are some of the [27:04] areas that we're looking to simplify and [27:05] then also from data owners and data [27:07] stewards just building some of those [27:09] going you know learning from them and [27:11] understanding how we can introduce that [27:13] in technology as well. And so those are [27:16] some of the opportunities that we're [27:18] looking for, not to displace, but also [27:20] to amplify. And ultimately, you want the [27:22] business to be, you know, doing the [27:24] things that they do best when it comes [27:26] to providing solutions, whether it's [27:28] research or what have you. Uh the [27:30] biggest challenge is you have people [27:31] trying to do a lot of this work off the [27:33] side of their desk. And so whenever you [27:34] can make things easier, is is going to [27:37] be a win-win for everyone. [27:38] >> Yeah. Just one step at a time and you [27:40] you get there. Uh but those are [27:42] fantastic use cases. What do you think [27:44] in your uh case? How do you kind of you [27:47] know identify those use cases? I know [27:50] you all have like a lot of those use [27:52] cases out there. [27:53] >> Yeah. Um it's it very depends what we [27:56] name use case. Yeah. Can use cases which [27:59] we name analytical use cases. There are [28:01] a lot of use cases. We have organization [28:03] with different level of maturity. [28:05] >> Yeah. [28:06] >> Reality and governance and atmosphere [28:08] inside of the organization. So I [28:10] understand you very well. So and based [28:14] on the maturity organizations are [28:16] delivering several use cases. Yeah. I [28:18] mean analytical use cases based on [28:20] behavioral scoring models uh to to to [28:24] to give loans not based only credit [28:27] bureau rate but also based on spendings [28:30] of people in the markets. I don't know [28:34] AI powered uh market analysis tools. uh [28:38] I don't know there are a lot of use [28:40] cases on analytical side if to speak [28:43] about data governance and data use data [28:46] management use cases we build there a [28:49] few years ago we build a central [28:50] platform uh in the holding and it's not [28:53] like data warehouse or central data lake [28:56] together all data and control everything [28:59] uh from the center no our model is [29:01] federated but aligned yeah we we build a [29:05] platform with act as like broker okay [29:07] one organization ask another [29:08] organization data. The other [29:11] organization put data on this platform [29:13] and we made it because of a lot of [29:15] regulations. I said yeah we have banking [29:17] insurance I mean in banking you has bank [29:20] secrecy law in insurance you have [29:22] insurance security law medical secrecy [29:24] law a lot of laws we made I mean several [29:26] layers of security [29:28] >> minimization minimization tools and we [29:31] enabled in organizational daily [29:34] generation of this daily usage. It was [29:36] their face number one when we said that [29:38] okay now organizations and and we built [29:41] let's say clones together I mean by the [29:44] way we did we did it with together with [29:46] data I mean it's one of the sponsor of [29:48] this event and we replicated this and [29:51] built inside of each organization their [29:53] own capabilities okay now analytical use [29:56] cases continue what we can do next our [29:58] next use case was a self-service and [30:01] self-service is there [30:03] uh I mean What I mean self-s service is [30:06] a very broad term. Yeah. Uh we we we [30:08] made it for data acquisition. Now [30:10] organization can just enter to the [30:12] platform to the classic data catalog. It [30:14] can be any of the catalog in one [30:17] organization can be informatic in [30:19] another. [30:20] There are a lot of the tools here and [30:23] just picking up the required data [30:26] qualify if the quality is enough and [30:29] just requesting this data and this data [30:32] comes to his area where he can start yep [30:35] >> doing simple things I mean joining of [30:37] the two tables or more difficult things [30:41] and it's all happening on the on the on [30:43] the same platform I think it's a most [30:46] important use case yeah use case for [30:48] building other use cases which which is [30:51] really transformative. Yeah. So, [30:53] >> yeah. Fantastic. Uh I since we're here [30:56] and I know keeping time in mind, uh I [30:58] want to touch base on a little bit of [31:00] the governance and the responsible uh AI [31:03] as well. Uh how do you define data [31:05] governance administration? Is it more [31:08] administrative or operational? Uh maybe [31:11] Rachel, you can start first. What do you [31:13] think? [31:14] it is 75% operational 25% administrative [31:17] but in most organizations it's probably [31:19] the reverse of that because they're [31:21] depending on as you mentioned in [31:22] maturity they're spending so much time [31:24] to educate [31:26] >> and build so I believe the question the [31:28] answer should be it's definitely [31:30] operational but in theory and and in [31:32] practice it's really going to depend on [31:34] the maturity of the organization yeah [31:36] where they are because a lot of [31:38] organizations are really spending a lot [31:39] of time just educating educating [31:42] educating and then influencing as well [31:44] because a big part of governance is [31:46] you're still working with business [31:47] leaders who probably have a lot of [31:49] autonomy and um need to be convinced as [31:52] to why they need to understand as you [31:55] mentioned earlier you know where my [31:57] metadata is and and quite frankly they [32:00] don't understand it and what I used to [32:02] say to my team is please don't use [32:04] governance speak with them because it's [32:06] you know helping them to understand okay [32:08] well you got you had this outgoing data [32:10] feed um you know you got an inquiry [32:12] about it can you tell where the source [32:13] of that data is and they're likely to [32:15] say I don't know and right so then you [32:17] can talk to them about well why this is [32:19] important but in most places I believe a [32:22] lot of companies have made a gained a [32:24] lot of ground just because of the AI you [32:26] know cycle that we're in but at the end [32:28] of the day it should technically be an [32:30] operational function [32:32] >> yeah uh what do you think John [32:34] >> yeah I think you know we've been on our [32:35] journey for about three and a half years [32:37] and just you know standing up governance [32:39] as a foundational capability and I agree [32:41] with you I do believe it's a 50/50 or [32:44] you get to that 75% mark on the [32:46] operational side and I think we're [32:48] starting to make that transition there. [32:50] But I think as far as you know making [32:52] that transition it's more focused on how [32:55] we are enabling the business before it [32:57] was just making sure are we adhering to [32:59] the policies do we have the right [33:00] controls and it's kind of like hurting [33:03] cats and we've all [33:04] >> if you've been in this profession it's [33:06] very much that and getting people to [33:08] agree and align on what are the various [33:09] mat uh metrics and standards that need [33:12] to be out there. But when you think [33:14] about enabling the organization now [33:16] you're starting to solve those [33:17] challenges around how do you get data [33:19] access faster? How do you uh you know if [33:22] you're trying to you're new to a company [33:24] you're trying to get really good [33:25] knowledge on a data set. How do you have [33:28] that chatbot answer some very basic [33:30] questions but also point them to your [33:32] data catalog where that information can [33:34] actually be found or even how do you get [33:36] to the point where if you want to check [33:37] a SQL query [33:39] >> you know there there's a lot of tools [33:40] out there today that do really good jobs [33:42] at this but making sure you have the [33:44] right SQL query are you hitting the [33:45] right database are you hitting the right [33:46] source so those are things that we're [33:48] thinking about and then how do we deploy [33:50] that whether it's through Slack or Teams [33:52] or something like that what majority of [33:54] our users are using [33:55] >> just a follow-up question you do you all [33:57] also use agents for something like this [33:59] where [33:59] >> so we the nine use cases that we are [34:02] developing are really agents. So one you [34:05] talked about domains earlier we are [34:07] thinking about doing this in a smaller [34:09] capacity where you have one agent per [34:11] domain and then we have one agent that [34:13] cuts across all of them to answer some [34:15] of those questions around you know you [34:17] may be joining data across many [34:18] different data sets from your for us it [34:20] would be like our customer data set [34:22] versus our transactional data set right [34:24] >> trying to answer some of those basic [34:25] questions. So we are thinking about that [34:27] from a governance perspective. But now [34:30] that we understand a lot more of our [34:31] data and having those contextual agents [34:33] out there and available, we think that [34:35] that's something that can definitely [34:37] streamline time to insights or speed to [34:40] insights as they say uh with the [34:42] business as well. [34:43] >> And does it also help uh agents to [34:46] actually interact with each other at the [34:48] same time? [34:48] >> Yes. Yeah. And so we are thinking yes, [34:50] we're embedding it in our governance [34:52] platform. However, we're thinking about [34:54] how does this connect to other agents [34:56] that might be out there that need to [34:57] call [34:58] >> and ask for that same information. So, [35:00] we're constructing them in a way where [35:02] if you have another use case across the [35:04] organization that requires that [35:06] contextual information, you can call our [35:08] agent and it provides the information to [35:10] you as well. [35:11] >> That's awesome. I also you know [35:13] sometimes wonder where we are in the [35:15] world where now okay agents interact [35:17] with each other but at the same time [35:19] there will be an a data governance agent [35:21] actually looking at all the data what's [35:24] been you know in the workflow right so [35:26] >> that's what I wanted to ask too because [35:27] then the build right the engineering [35:29] phase of that so your data you probably [35:32] don't have a lot of the regulatory [35:34] requirements I'm sitting here like wow [35:36] that's awesome but we have a lot more um [35:39] you know getting through some of the the [35:41] reviews that would be required to do [35:42] that. Um, next [35:44] >> coming from a financial service [35:46] background and I've been in automotive [35:48] and now in you know quick service [35:50] restaurant industry I'm still thinking [35:52] in my days of financial service. So [35:54] everything we build is with that same [35:56] context on the privacy side policy side. [35:59] So we still [36:00] >> while we don't push it [36:01] >> as much in our space because we're not [36:03] highly regulated other than food and [36:05] drug administration [36:06] >> I I would say that we still look at it [36:08] from all those different angles. So we [36:11] make our framework very flexible. So if [36:13] we need to pivot quickly, we can do [36:14] that. And I think in this space right [36:18] now, you know, there is deregulation, [36:19] but things could change in four years [36:21] from now. Yeah. [36:22] >> You never know. So we're making sure [36:24] that our framework allows for those [36:25] things and that we can accommodate it. [36:27] >> That's fantastic. Great insights. So [36:29] John, well, what do you think? [36:32] I'm thinking that it's neither [36:33] operational administrative [36:36] and from our perspective [36:37] transformational [36:39] uh and [36:42] think about it like uh I'm holding [36:45] companies and we we central team are [36:49] building rails where we and by building [36:52] rails thinking about that there will be [36:54] a lot of locomotives going over this one [36:58] is mature another less mature one is the [37:01] e-commerce business in others banking is [37:03] totally different. Yeah. So I we we're [37:07] trying to make it also flexible from [37:09] frame perspective and to be able to [37:12] change it in the road and we we [37:14] establish their uh function with know [37:16] how with central capital which is [37:18] thinking about policies processes [37:20] mechanics. One of the use case I didn't [37:23] mention the third party integration how [37:25] to make this AI agents. Yeah. If you're [37:29] I mean it's you're not building if you [37:31] are just taking AI agent how to make [37:33] sure that the moment it touch your data [37:36] how you will govern h how you will make [37:39] sure interaction there are a lot of [37:41] question and there should be I believe [37:43] that there should be some central team [37:45] cap with capabilities without any [37:48] operations [37:50] focus on the creation policies [37:52] frameworks process and continuously [37:54] updating and doing that and local teams [37:57] is should be in charge of adopting these [38:00] frameworks because I mean they are not [38:01] fits. Yeah, usually they are not fits [38:03] >> for sure [38:04] >> some parts not fits and local teams in [38:07] our case more working on adopting these [38:09] frameworks adopting these policies [38:11] reality and providing their feedback. So [38:14] long story short, I think their data [38:15] governance at all. It's not operational [38:18] administrative. It's transformational. [38:20] It's a feel and main driver of their [38:23] transformation and it's it's no matter [38:26] we speak about datadriven decision [38:28] making culture or AI transformation. [38:31] >> Okay, those are great points. Uh well, [38:33] we'll stick on you. Uh and this is one [38:35] last question for all three of you. uh [38:38] if you all had any advice for uh the [38:42] enterprise leaders you those who are [38:44] working in the data governance space or [38:46] uh are you know wanting to learn more [38:48] about AI governance the implementation [38:52] uh what would be your advice and also [38:56] there's a follow-up question which is [38:58] how do you see the future uh in the data [39:00] and AI governance space so [39:04] we'll let you do the talking [39:07] Okay. Uh maybe some advices will be [39:10] obvious but uh okay never start big. Try [39:14] never try to govern everything [39:17] uh every data you will not able anyway [39:20] you will not able to govern all data. [39:23] Starting small picking up use case [39:28] it will be good if you will be able to [39:30] pick up use case with a high value [39:32] >> be able to you know it's very difficult [39:35] to to show the direct monetary value of [39:38] data governance itself to the business [39:41] >> the usual challenge where data [39:42] executives facing yeah I mean would we [39:45] plan to invest millions to the data [39:47] governance things but what will be [39:49] returned yeah so and it's a good way [39:52] It's a bad news by the way that it's we [39:54] we still have this challenge but the [39:56] good news part is that there is a way [39:59] how to let's say convince the business [40:02] uh is a picking picking up use case and [40:05] building everything use case. Yeah. So [40:08] so we made it with their customer [40:09] lifetime value use case for our [40:11] ecosystem which we're building and build [40:15] data governance things data gold things [40:17] as part in parallel of delivering all [40:20] the use cases. It was like showcasing [40:22] after you you you you have a trust of [40:25] business and management of how to do the [40:28] this uh I think there is no barriers of [40:31] continent and scaling up. Yeah, start [40:34] small pick up the use case. [40:37] Uh another lessons learned may be uh [40:39] when you're trying to select data [40:42] governance tools and data governance [40:44] platforms they're more most of them more [40:47] or less same not stuck as me in their [40:51] selection process. Yeah. So [40:54] the important is not which tool will you [40:56] will select important and the [40:58] capabilities of your people and if you [41:01] have mature people it's no matter which [41:03] tool they will use [41:04] >> and maybe which fits your your need at [41:06] that time so don't get stuck start small [41:09] and just follow the use cases and uh [41:12] you'll be at a good track those are [41:14] great insights val John what do you [41:16] think [41:16] >> I think you know uh from a lesson [41:18] learned adoption is key no matter what [41:21] tool you're using or the process [41:22] processes you're rolling out. I think [41:24] really homing in on change management [41:26] adoption that's going to be what brings [41:28] the success. Um so that's a lesson [41:31] learned for us. But I think from you [41:32] know two points in looking at if you're [41:35] starting your AI journey governance [41:36] journey really partnering with your [41:39] cyber security your privacy your uh [41:42] architecture teams and engineering teams [41:44] and if you have like a AI hub or data [41:47] science group bring them all together [41:49] and talk about how do you want to kind [41:51] of put the controls in place and what [41:53] needs to be there. So that's something [41:54] that we did and then on the I would say [41:57] the business enablement side if you are [41:59] in the governance industry you really [42:01] need to start thinking about how can we [42:02] use and leverage AI to in order to [42:04] enable the business. So start thinking [42:06] of those use cases that can help you [42:08] solve some of those persistent business [42:10] challenges and I definitely believe that [42:12] you'll start to get a little bit more [42:13] traction in your programs. [42:15] >> Good point there. [42:16] >> Yeah, I would agree with everything [42:17] these guys have said. Um I would just [42:19] say that also making sure you're solving [42:22] number one a small something small start [42:23] small and then showcase it um that you [42:27] know how can you scale it because one of [42:28] the things we tried to do at the [42:30] beginning we tried to solve everything [42:31] and we didn't solve anything in the very [42:34] uh start of our journey and so it took a [42:36] while to really get the business to to [42:39] buy in and agree to what we were trying [42:41] to achieve. And so in doing that they [42:43] realize that you know growth uh [42:45] governance is growth and growth means [42:47] that it's an insurance policy for us [42:49] from both a sustainability perspective [42:51] and revenue but as well as from a comp [42:53] compliance perspective. So the questions [42:55] that you're going to get regardless even [42:57] if you're not in a regulated industry [43:00] there are business associate agreements [43:02] that come along with some when you think [43:04] about holding companies or you think [43:05] about um you know third party data. And [43:08] so all of those things will require some [43:10] level of governance whether or not a [43:11] company thinks they have to do it. And [43:13] so by doing so I think that creates sort [43:15] of that trust also allows the business [43:18] to continue in a parallel path. And I [43:20] can't echo enough what John just said [43:22] which is your cyber team your enterprise [43:24] architecture teams your compliance [43:26] privacy legal the collective village is [43:29] super super important so that everyone [43:31] understands the swim lanes [43:32] accountability responsibility and who [43:34] needs to be informed throughout the [43:36] process.