Marsh Data Head Ties AI Payoff to Reuse as 95% of Pilots Fail

95%试点折戟，Marsh数据主管指AI回报取决于复用 全球企业唯有叫停各部门的重复建设，并在各业务单元推行可复用的智能体架构，才能将激增的AI Token开销转化为底线EBIT。 针对为何95%的企业AI概念验证（POC）最终沦落至“POC坟场”，Kader Sakkaria呼吁组织从零散、事务性的实验转向与底线EBIT紧密挂钩的严谨产品思维。他提出了“复用率”这一指标，用以评估AI模型和工作流在各业务单元之间的共享效率，避免在部门壁垒中重复开发。为实现可持续的规模化，Sakkaria阐述了关键的架构与组织实践，包括湖仓一体数据基线、前置治理管控，以及融合企业内部机制经验与外部顾问专长的“结对工程”人才模式。最后，他提出了一套面向高管的CFO记分卡和五项无需调整组织架构的战术动作，重点强调了明确的双人责任制、价值驱动的阶段关卡以及严格执行的终止标准。 MIT 针对 173 家企业的研究显示，95% 的企业 AI 概念验证（POC）最终沦入“POC 墓地”未能投入生产，仅有 5% 能够成功融入日常业务流并规模化产生价值。 (1:11) 企业董事会的关注点已从盲目实验转向严格的财务问责，要求 AI 项目必须直接贡献于底层 EBIT（息税前利润），以抵消高昂且不断复合累积的 Token 与技术运营成本。 (3:29) 企业在 AI 落地上面临巨大的预期落差：高达 85% 的企业渴望部署 Agentic AI 架构，但目前仅有 23% 真正成功将 AI Agent 规模化集成到企业核心流程中。 (6:56) Kader Sakkaria 提出“复用率”（Reuse Ratio）作为衡量成熟度的关键标尺：复用率低于 20% 的组织处于高成本、重复造轮子的“项目模式”，而超过 50% 的顶级企业则采用“产品模式”实现一次构建、全域复用。 (10:51) 企业推进 AI 过程中常陷入两大陷阱：将治理与权限控制（RBAC/ABAC）后置会带来高昂的“重构税”，而缺乏需求验证的“假定采纳”则会导致耗资数百万美元的工具几乎无人问津。 (13:30)

Enterprise AI initiatives overwhelmingly stall in proof-of-concept stages because organizations build bespoke, fragmented tools instead of reusable capabilities that yield bottom-line financial return. Drawing on enterprise benchmarks, Kader Sakkaria argues that achieving sustainable AI scale requires shifting from transactional project experimentation to an enterprise product model governed by unified data foundations, explicit cost offsets, and measurable reuse. The POC Graveyard and AI Financial Accountability (1:11) Citing an MIT study of 173 enterprise interviews, Sakkaria highlights that 95% of corporate AI proof of concepts (POCs) fail to reach production workflows, while only 39% of the value attributed to enterprise AI POCs trickles down into bottom-line EBIT. Corporate executive boards have moved beyond encouraging open-ended experimentation and now demand concrete proof of financial return. Soft productivity claims and post-amortization adjustments no longer satisfy leadership. Escalating operating expenses make financial justification critical. Sakkaria notes that internal developer teams can rapidly accumulate substantial token consumption costs, citing an example of roughly $60,000 consumed in a single quarter, or an annualized run rate of $240,000. In corporate financial planning, absorbing that level of continuous technology spend requires justifying an offset equivalent to roughly four full-time employees (FTEs), creating a sustainability dilemma if efficiency gains fail to materialize as real bottom-line savings. Fragmented Silos and the Rebuilding Tax (6:56) A wide gap exists between ambition and operational delivery: while 85% of organizations aspire to establish agentic AI architectures, only 23% have scaled AI agents across enterprise workflows. Sakkaria attributes this divide to departmental silos operating under the fallacy that each business unit's needs are entirely unique. While teams justify bespoke builds by citing distinct workflows or consent constraints, departments almost always rely on a shared enterprise data layer. Creating redundant, competing agents inflates ongoing maintenance and Run-the-Business (RTB) operating costs. This fragmentation imposes a severe "rebuilding tax," forcing teams to repeatedly recreate basic components rather than leveraging shared foundations. Furthermore, teams routinely treat essential controls—such as role-based access control (RBAC), attribute-based access control (ABAC), user consent, and data residency—as afterthoughts, resulting in expensive post-implementation rework. Projects also suffer from "assumed adoption," where substantial capital is invested in building tools without validating internal demand, leaving expensive tools with negligible user engagement. Measuring Scalability with the Reuse Ratio (10:51) To counter duplication, Sakkaria proposes tracking the Reuse Ratio, a mathematical metric that places asset reuse frequency in the denominator to measure how many times an AI agent or component is repurposed across workflows and business lines. Organizations with a reuse ratio below 20% operate in transactional "Project Mode," marked by siloed builds, redundant maintenance, and high failure rates. Sakkaria notes that audience polling reflects industry reality, with most organizations sitting at barely 5% reuse. Organizations progress as they move into "Platform Forming and Norming" (20% to 50% reuse) and ultimately reach "Product Mode" (greater than 50% reuse), which defines the top 5% of enterprise performers. Operating in product mode requires engineering core capabilities once and deploying them across multiple business units. This approach leverages unified data pipelines, curtails duplicative software development, and generates compounding P&L leverage. Unified Ownership and Value-Driven Governance (15:31) Scaling enterprise capabilities requires disciplined execution structures rather than disruptive corporate reorganizations. Sakkaria advocates a "pair engineering" model that combines internal personnel—who understand institutional data structures and tribal history—with external modernization consultants. Delivery should be governed by a single accountable ownership duo: a business P&L lead defining commercial outcomes paired directly with a data domain lead responsible for technical execution, avoiding diffused committee decision-making. Governance must also enforce value-driven stage gates equipped with clear "kill criteria" to decommission failing projects early. To establish credibility with finance leadership, Sakkaria outlines an executive CFO Scorecard that evaluates growth, efficiency, and productivity enablers against metrics like reuse, active adoption, reliability, and velocity. Leaders can implement this framework through five zero-reorganization steps: cataloging existing AI assets, calculating the reuse ratio, selecting three high-impact enterprise outcomes, delivering two prioritized outputs, and publishing one explicit kill criterion. Sakkaria concludes that long-term enterprise advantage in AI will not belong to organizations that build the highest volume of bespoke models, but to those with the operational discipline to build robust foundational capabilities once and maximize their reuse across the enterprise.

Kader Sakkaria在Ai4大会上指出，企业AI面临严重的规模化瓶颈，唯有从分散的项目实验转向以高复用率为核心的平台产品模式，才能将技术投入切实转化为财报上的息税前利润（EBIT）。 95%的项目坟场与EBIT考核压力 (1:11) Sakkaria援引一项针对173家企业的MIT研究指出，高达95%的企业AI概念验证（POC）未能投入生产环境，最终沦为无法产生实际业务价值的孤立试验，仅有5%成功实现规模化。过去一年，董事会对AI的要求已从单纯的跟风试验转向严格的财务回报，要求证明其对EBIT的实质拉动。然而，目前企业AI POC所声称的价值中，仅有约39%能真正反映到企业底层的EBIT数据中。 技术部署的实际成本进一步加剧了盈利考核的紧迫性。Sakkaria以自身研发团队为例，一组接入AI模型的开发人员单季度消耗了约6万美元的Token，年化运行成本达24万美元。在传统财务模型中，这一技术支出需要抵消大约4个全职人力（FTE）的成本才能维持收支平衡。单纯依赖模糊的生产力概念已无法满足董事会，企业必须证明AI投入能带来可持续的成本抵消或利润增长。 部门竖井陷阱与复用率分级 (6:56) 虽然85%的企业渴望建立Agentic AI架构，但实际上仅有23%的企业成功在跨业务流程中扩展了AI智能体。Sakkaria现场调研发现，40至50名参会者中仅有1人曾构建过无需重写即可跨业务部门复用的AI智能体。这种脱节的根源在于部门竖井与“自身业务特殊”的认知误区。各部门常以数据驻留或授权为由自行定制AI，实际上底层依赖的是同一套企业数据层。重复构建不仅导致维护与日常运营（RTB）成本激增，更稀释了整体投资回报。 为此，Sakkaria提出了衡量企业AI成熟度的核心指标——复用率（Reuse Ratio）。复用率低于20%的企业属于“项目模式”，其特点是孤立定制、维护成本高且失败率大；处于20%至50%之间的企业进入“平台形成与规范期”，逐步建立共享标准；而复用率超过50%的企业则进入仅占5%的顶级“产品模式”，能够实现一次构建并在全企业范围内复用，从而形成强大的EBIT杠杆效应。然而现场绝大多数企业自认复用率仅在5%左右，深陷重复构建的税负。 规避POC陷阱的核心能力与配对工程 (13:30) 企业AI常落入两大致命陷阱：一是“重复建设税”，即缺乏统一标准导致团队反复开发底层基础，并在模型上线后才仓促补救基于角色的访问控制（RBAC）、基于属性的访问控制（ABAC）、用户许可和数据驻留合规，造成高昂的返工成本；二是“假设采用”，例如Sakkaria引述前Google员工的案例，工程团队耗资100万至300万美元构建的内部产品最终仅有2名活跃用户，因缺乏真实需求验证而失败。 破解这些困境需要企业建立三项关键能力：首先是湖仓一体（Lakehouse）的基础数据层，在前期嵌入治理标准与合规控制；其次是人才落地的“配对工程模型”（Pair Engineering Model），由于传统企业拥有深厚的业务与数据隐性知识，但缺乏现代架构思维，企业应选拔内部骨干并与高水平外部顾问结对，避免全盘外包带来的缓慢脱节；最后是价值度量体系，以复用频率等客观数学指标追踪资产真实表现，而非盲目铺设孤立工具。 免重组的管理五步法与CFO计分卡 (18:09) 在组织执行层面，Sakkaria借用通用电气前CEO Jack Welch的“单一问责制”原则，主张为AI项目设立单一责任人和统一执行路径。具体而言，应由一名负责P&L成果的业务主管与一名全流程交付的数据领域主管组成双核心组合，摒弃多方扯皮的松散决策。同时，项目阶段评审（Stage Gates）必须严格以业务价值为导向，未能通过增长、效率或复用指标验证的项目必须强制终止，防止无序消耗资源。 为向CFO证明技术投资的合理性并避免引发动荡的组织重构，Sakkaria建议推行包含五大行动的落地计分卡：全面盘点现有AI模型与智能体资产；精确计算各资产在不同部门间的复用率；聚焦于增长、效率或生产力这三项核心业务成果；重点转化并保质交付其中两项优先产出；最后公开发布一项明确的项目终止标准，坚决淘汰无法产生实效的实验，建立良性资本循环机制。 Sakkaria总结强调，未来全球企业在AI领域的胜出者，绝非那些盲目堆砌定制化解决方案或追求模型数量的公司，而是具备工程自律、能够构建统一底座并最大化跨部门资产复用率的组织。唯有将AI作为可复用产品进行运营，才能摆脱低效的POC循环并获得持久的商业价值。

## Background 本内容由 Ai4 于 2026 年 8 月 19 日在 YouTube 平台发布，主讲人为 Marsh 公司全球数据策略主管 Kader Sakkaria。页面介绍指出，该演讲旨在提供一套高管实践框架，帮助全球企业克服组织阻碍、明确治理架构与运营模式，跨越孤立的概念验证实验，构建可衡量业务绩效且可持续规模化的企业级 AI 能力。 ## TL;DR 演讲指出高达 95% 的企业 AI 概念验证（POC）因重复造轮子和无法融入核心工作流而沦为“POC 墓地”，同时企业董事会的关注点已从单纯的生产力转向真金白银的 EBIT 财务回报。为了打破部门烟囱和极高的运维成本陷阱，演讲者提出了核心量化指标——“复用比率（Reuse Ratio）”，将高成熟度组织定义为能跨业务线极致复用底层数据与 Agent 的领跑者。企业需要通过构建统一数据底座、推行内部骨干与外部专家搭档的“结对工程”、设立单一责任人以及价值驱动的阶段门禁来推动规模化。最后，演讲者给出了包含四维指标的高管计分卡与五步落地法，强调未来 AI 竞争的真正赢家是那些能够“一次构建、最大化复用”的组织。 ## Atmosphere & style 现场氛围务实、敏锐且充满冷峻的行业洞察，演讲者以大型跨国企业高管特有的幽默与坦诚开场，迅速拉近了与台下决策者的距离。整场分享节奏紧凑、直击痛点，通过犀利的互动提问与尖锐的成本对比（如开发者 Token 消耗与裁员对冲），打破了生成式 AI 的表面浮躁与虚假繁荣。听感上既有一针见血揭露企业内耗与“重建税”的警醒感，又具备清晰有力的工程与管理落地路径，呈现出极具执行力的高管实战指南风格。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=4s) 麻省理工学院（MIT）调研显示，95% 的企业 AI 实验由于无法融入核心业务工作流而沦为“POC 墓地”的牺牲品，仅有 5% 的企业成功跑通价值闭环。 - [03:29](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=209s) 董事会对 AI 的考核已从早期的表面生产力指标全面转向息税前利润（EBIT）等财务底线，而目前各类 POC 声称的价值中仅有约 39% 能真正归因计入 EBIT。 - [03:29](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=209s) 未加管控的 Token 消耗极易导致成本失衡，单季度人均 6 万美元的 Token 支出若年化相当于数名全职员工的预算，AI 投资必须考量长期的财务可持续性。 - [06:56](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=416s) 企业在 AI 愿景与实际就绪度之间存在严重脱节：高达 85% 的企业渴望建立智能体战略架构，但实际成功实现规模化扩展的仅有 23%。 - [06:56](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=416s) 各业务部门常以“数据驻留合规”或“业务独特性”为由自建孤岛，导致烟囱式重复开发与功能重叠，使得日常运营维护成本急剧飙升并侵蚀利润。 - [10:51](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=651s) “复用比率”（Reuse Ratio）是衡量 AI 成熟度的关键指标，低于 20% 复用率意味着组织仍停留在高维护成本的单点“项目制模式”，而非平台化“产品模式”。 - [12:20](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=740s) AI POC 规模化面临三大陷阱：重复开发底座与权限控制延后的“重建税”、合规流程事后补救的“治理后置”、以及投入巨资研发却无人使用的“假设采用”。 - [15:31](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=931s) 扎实的数据湖仓（Lakehouse）与数据质量、安全访问控制是 AI 规模化的不可或缺的准入门槛，不能脱离数据基座空谈模型与智能体应用。 - [15:31](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=931s) 在人才策略上宜采用“结对工程（Pair Engineering）”模式，将掌握内部业务隐性知识的业务骨干与高水平外部架构顾问结合，以实现高效交付与技能沉淀。 - [18:09](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=1089s) 借鉴单一问责制（Single Accountable Owner），每个 AI 项目必须确立单一责任人、单一业务目标与单一路径，由业务负责人界定 P&L 贡献，数据主管负责端到端技术交付。 - [18:09](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=1089s) 必须将形式主义的项目门禁转变为“价值驱动的阶段门禁”，并设立清晰的“终止标准（Kill Criteria）”，在项目无法实现预期业务价值时果断止损关停。 - [19:47](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=1187s) 向管理层和 CFO 汇报 AI 产出时，应使用涵盖价值（损益表贡献）、采纳度（复用比率）、可靠性与交付速度的高管计分卡，无需进行复杂的组织架构重组。 - [19:47](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=1187s) 企业推进落地可依托轻量级五步法：全面清点资产、计算复用比率、锚定 3 项关键企业成果、全力交付其中至少 2 项业务输出，并发布明确的项目终止标准。 - [22:50](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=1370s) 未来 AI 竞争的真正赢家，绝非盲目开发最多模型或项目的企业，而是具备平台化思维、能够“只构建一次并尽可能多次复用”的组织。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=4s) 开场与目标：走出 95% 的 AI 概念验证墓地 演讲者 Kader Sakkaria（Marsh 公司全球数据策略主管）在开场时幽默地提到，自己在乘飞机前来的途中一直在思考这场演示。作为大型跨国企业的一员，企业内部往往有严格的公关与公关合规限制（global communications），限制了许多在公开麦克风前能聊的内容。他打趣道，自己在讲台上会尽力分享干货，但如果大家想知道更多更真实的细节，欢迎在演讲结束后私下交流，“一旦关掉麦克风，我的发挥空间就会大得多”。他选择跳过那些无聊的个人履历介绍，直奔主题：在 LinkedIn 时代大家随手都能查到个人背景，而今天他希望大家离开会场时，能实打实地带走“一个数字”、“一个比率”以及“一张推动 AI 必备的记分卡”。 ```mermaid flowchart TD accTitle: AI 项目从 POC 到规模化的转化断层 subgraph EnterpriseFailure["95% 陷入困境的组织"] A["大规模开展 AI 实验与 POC"] --> B["孤立的用例开发<br/>(重复造轮子)"] B --> C["无法融入核心工作流"] C --> D["POC 墓地 (POC Graveyard)"] end subgraph EnterpriseSuccess["5% 跑通闭环的组织"] E["统一数据底座与平台化架构"] --> F["构建可复用的 Agent 与框架"] F --> G["深度融入业务工作流与协同"] G --> H["产生可持续的 EBIT 财务价值"] end ``` 随后，他抛出了第一个关键数字：“95%”。现场举手表示听过该数据的人寥寥无几。这项数据源自麻省理工学院（MIT）的一项权威研究，该研究深入访谈了 173 家大型企业机构，调查它们如何在整体上推进 AI 概念验证（POC）。这项研究揭露了一个极其残酷的现实：95% 的企业 AI 工作最终都死在了“POC 墓地”（POC graveyard）里，根本无法走入生产环境，更无法无缝融入机构的核心业务工作流中来实现辅助增强并释放业务价值。演讲者向在场所有人提出核心设问：那真正成功的 5% 组织究竟做对了什么？在座的各位必须时刻在脑海中审视一个问题——“你的团队到底能否做到构建一次，就能在整个 AI 领域内反复复用？” - **MIT 调研样本**：基于对 173 家企业机构的深度访谈。 - **95% 的沉没率**：95% 的企业 AI 概念验证停滞在实验阶段，沦为 POC 墓地的牺牲品。 - **5% 的成功法则**：仅有 5% 的企业成功将模型与智能体工程化落入业务流。 --- ### [03:29](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=209s) 董事会提问转向：从表面生产力到真金白银的 EBIT 考量 企业面临的评价体系与考核球门（goalpost）已经发生了根本性位移：衡量标准已经从最初单纯的“生产力指标”，全面转向“对组织真正的底层财务影响”。在生成式 AI 浪潮初期，几乎所有企业的董事会都在追问类似的问题：“你们在 AI 方面做了什么？我们拥抱 AI 了吗？进展如何？”高管们为了交差，纷纷上马各类自认为极具前景的 POC 和用例。然而仅仅过了一年，董事会的提问方向发生了 180 度大转弯：“你们花了这么多钱，到底给组织创造了什么真实的商业价值与财务影响？” 为了说明这种成本与产出失衡的严峻性，演讲者分享了上一季度的真实经历：团队为一部分研发人员开放了 AI 工具权限，结果在短短一个季度内，平均每位开发者消耗了大约 6 万美元的 Token 费用。若按此粗略年化推算，单人一年的 Token 开销将高达 24 万美元左右。在崇尚成本管控的企业环境里，这笔开支相当于需要裁减约 4 名全职员工（FTE）才能填平成本。“消灭 4 名员工的成本换来这些 Token 开销，这到底值不值？”他在午餐时还与老东家加拉格尔（Gallagher）的前同事热烈探讨：这种投资回报能否长期维持？目前各类概念验证所创造的价值中，实际能直接归因并体现在息税前利润（EBIT）上的仅占 39% 左右。企业高层现在紧盯的不是软件折旧摊销后的数字，而是在摊销前，AI 到底为企业的 P&L（损益表）和息税前利润带来了多少实质拉动。 - **Token 消耗与人力对冲案例**：单季度人均约 $60,000 Token 消耗，年化约 $240,000，相当于约 4 名 FTE（全职员工）的人力预算。 - **价值归因低迷**：调研表明仅有 39% 由 POC 声称带来的价值能够真正计入 EBIT 底线。 - **董事会核心考量**：不再满足于尝试性尝试，而是要求直接汇报对 P&L 损益表和 EBIT（息税前利润）的净贡献。 --- ### [06:56](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=416s) 雄心与落地脱节：智能体规模化鸿沟与烟囱式开发 AI 技术落地的另一个核心矛盾是“采用愿景（Adoption）严重超前于组织的准备就绪度（Readiness）”。从行业统计数据来看，高达 85% 的受访企业在愿景层面极具野心，希望建立起全面的 Agentic AI（智能体 AI）战略架构；但现实中真正能将智能体应用落地并实现规模化扩展（scaling）的组织仅有区区 23%。绝大部分团队的实际操作，依然是陷入单一孤立的用例式敲打，根本没有把 AI 当作一种贯穿工作流的端到端自动化流程来打造。 ```mermaid flowchart TD accTitle: 部门数据割裂与烟囱式 Agent 开发困境 subgraph DataLayer["底层统一数据资产"] BaseData["企业底层统一数据源 (Single Data Layer)"] end subgraph DepartmentSilos["部门烟囱与借口"] BaseData -. 借口: 数据驻留 / 授权限制.-> DeptA["业务部门 A"] BaseData -. 借口: 自身业务具有独特性.-> DeptB["业务部门 B"] BaseData -. 借口: 独立合规流程.-> DeptC["业务部门 C"] end subgraph RunCostTrap["运维成本黑洞"] DeptA --> AgentA["自建 Agent A"] DeptB --> AgentB["自建 Agent B"] DeptC --> AgentC["自建 Agent C"] AgentA & AgentB & AgentC --> RunCost["高昂运维成本 (Run the Business Cost)<br/>彼此竞争，侵蚀 EBIT"] end ``` 演讲者着重指出：“AI 并不能单纯依靠使用量的增加来实现规模化扩展。”如果研发出的智能体缺乏跨部门复用性，高昂的维护成本将彻底拖垮整个商业模型。他在现场做了一个举手互动，询问在座四五十位企业精英中，有谁能够拍着胸脯保证自己团队开发的单个 Agent 在不经代码重写或重构的情况下，就能在企业内部其他业务单元被直接复用。全场只有一位参会者举起了手。演讲者苦笑道，会后一定要和这位同仁深入请教。这种普遍困境的根源在于：每个业务单元都标榜“我们的业务独一无二”，并习惯拿“数据驻留”（data residency）和“数据授权同意”（data consent）当作挡箭牌。表面上各自筑墙自建体系，但剥开外衣看，底层往往依赖的是同一套基础数据层。这种各自为政导致了企业内部充斥着功能重叠、彼此竞争的 Agent，不仅未能放大价值，反而让“业务日常运营维护成本”（Run the Business Cost）急剧飙升，最终在财务损益表审查面前全盘溃败。 - **雄心与现实鸿沟**：85% 的企业渴望落地 Agentic AI 框架，但实际成功扩展智能体的仅有 23%。 - **现场调研实况**：现场约 40-50 位企业高管中，仅有 1 人实现了智能体跨业务部门的免重写无缝复用。 - **部门壁垒借口**：各部门频繁以“数据驻留合规”、“客户授权协议”和“业务独特论”为由各自采购搭建，形成严重资源内耗。 --- ### [10:51](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=651s) 核心破局指标：定义“复用比率”（Reuse Ratio） 为了摆脱低效的恶性循环，演讲者在自身组织（Marsh）内部推行并确立了一个关键管理量化指标——“复用比率”（Reuse Ratio）。如果在座的管理者今天只能带走一个实践工具，那必然是建立起对自己企业复用比率的审视机制。 ```mermaid flowchart LR accTitle: 企业 AI 成熟度复用比率阶梯 Low["低于 20% 复用率<br/>【项目制模式】<br/>单点交付、高维护成本、烟囱系统"] Mid["20% - 50% 复用率<br/>【平台规范化阶段】<br/>形成标准平台、规范共享能力"] High["高复用率 (前 5% 企业)<br/>【产品化平台模式】<br/>Agent 与数据跨部门规模化共享"] Low --> Mid --> High ``` 演讲者将企业的复用比率清晰地划分为三个成熟度梯队： 1. **复用率低于 20%**：意味着该组织本质上仍然在以传统的“项目模式”（Project Mode）运作，而非“产品模式”（Product Mode）。每一次接到需求都在从零重新造轮子，导致技术资产无法沉淀，运维压力随用例增加呈线性甚至指数级上升。 2. **复用率处于中间过渡区间**：表明底层平台正在经历“形成与规范化”（Forming and Norming）阶段，共享架构初见雏形，正在逐步发挥支撑作用。 3. **高复用率梯队**：正是开篇提到的那 5% 卓越领跑企业。它们在底层基础设施、核心数据模型、智能体编排框架上做到了极致复用，唯有达到这个成熟度，AI 投资才能真正释放规模化杠杆效应。演讲者在部分尾声邀请台下观众对照该比率评估自身企业当前所处的位置，为接下来的评测框架与后续拆解埋下伏笔。 ### [12:20](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=740s) 概念验证中的复用比率与三大核心陷阱 现场观众互动调查显示，绝大多数企业在 AI 资产复用率上处于极低水平，仅有极少数人能达到 20% 以上，当前行业现实普遍徘徊在 5% 左右。企业亟须建立对“复用比率（Reuse Ratio）”的量化理解。通过一套将“特定智能体（Agent）或流程在各业务线被复用的次数”作为分母的数学公式，管理层可以清晰追踪 AI Bot 在整个组织各垂直业务领域的实际采用度与复用情况。 ```mermaid flowchart TD accTitle: AI POC 三大核心陷阱 subgraph Traps["AI 概念验证（POC）三大陷阱"] T1["重建税 (Rebuilding Tax)<br/>重复开发底座，治理与 RBAC/ABAC 后置"] T2["合规后置成本<br/>上线后才处理数据同意、角色与数据驻留"] T3["假设采用 (Assumed Adoption)<br/>耗费数百万美元研发却仅有极少数用户使用"] end Traps --> G["POC 走向项目墓地 (Graveyard)<br/>无法实现工业化与规模化"] ``` 在 AI 的概念验证（POC）历程中，普遍存在三大阻碍规模化的陷阱： - **重建税（Rebuilding Tax）**：不同团队使用相同的数据，却各自重复开发底层架构，且将治理与管控视作事后补救，而非前期内建能力。基于角色或属性的访问控制策略（RBAC / ABAC）常被延后处理，导致后期在数据同意（Consent）、数据角色和数据驻留（Data Residency）合规上面临严重的返工。 - **治理后置（Governance as an Afterthought）**：合规与治理未能在规划初期融入整体组织能力，被迫在实施后进行高成本的二次重构。 - **假设采用（Assumed Adoption）**：脱离实际需求盲目投入。曾有 Google 前员工透露，团队耗费一两百万甚至三百万美元研发的产品，在公司内部最终仅有两名实际用户。这种缺乏真实采用率的 POC 终将走向“项目墓地（Graveyard）”，无法实现工业化扩展。 --- ### [15:31](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=931s) 规模化落地的核心能力：数据基座、复合人才与价值衡量 要推动 AI 走出孤岛并实现规模化，企业必须协同演进多项核心能力。首要支柱是建立扎实的数据底座（Data Foundation），这是不可或缺的准入门槛（Table Stakes）。企业应重点关注数据湖仓（Lakehouse）中的数据标准、数据质量保障以及安全控制框架，而非单纯纠结于结构化与非结构化数据的形态差异。 ```mermaid flowchart LR accTitle: 结对工程人才赋能模式 subgraph Pairing["结对工程 (Pair Engineering) 模式"] A["内部骨干 (Champions)<br/>沉淀业务基因与数据隐性知识"] B["高水平外部顾问<br/>注入现代化技术思维与架构方法"] A <-->|技能互补 / 赋能培训| B end Pairing --> C["兼具业务深度与现代架构的落地交付"] ``` 在人才与组织采纳（Talent and Adoption）方面，大中型企业普遍积累了深厚的业务隐性知识（Tribal Knowledge），清楚掌握数据分布与历史遗留问题（"where the skeletons are hidden"），但往往缺乏现代化的技术思维。传统的全盘外包模式会导致高昂的业务学习曲线。更高效的解法是采用类似“结对工程（Pair Engineering）”的机制：选拔内部核心业务骨干（Champions）并进行技能提升，将其与高水平外部顾问紧密搭档，既保留了机构知识，又注入了先进的工程实践。 - **数据底座建设**：在 Lakehouse 中固化数据质量指标与访问控制策略。 - **人才结对赋能**：内部知识持有者与外部顶尖专家结对，缩短认知磨合期。 - **价值闭环衡量**：坚持“无度量即无价值”，将复用比率等量化指标作为追踪 AI 资产业务贡献的起点。 --- ### [18:09](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=1089s) 单一责任人机制与价值驱动的阶段门禁 借鉴通用电气（GE）前 CEO 杰克·韦尔奇（Jack Welch）提出的“单一问责制”（俗称 "Single throat to choke"），AI 项目必须确立单一责任人（One Accountable Owner）、单一目标（One Outcome）以及单一实施路径（One Single Path）。组织架构上应明确由一位业务负责人界定该价值流对损益表（P&L）的具体贡献，同时由一位数据领域主管（Domain Lead）对内部端到端的系统构建与交付负全责，坚决避免多头负责导致的推诿与目标割裂。 ```mermaid flowchart LR accTitle: 价值驱动的阶段门禁与终止机制 subgraph StageGates["价值驱动阶段门禁 (Value-Driven Stage Gates)"] G1["阶段 1"] --> G2["阶段 2"] G2 --> G3["阶段 3"] G2 -.->|未达业务预期| K["触发终止标准 (Kill Criteria)<br/>果断关停无价值项目"] end ``` 在推进流程中，传统的阶段门禁（Stage Gates）极易沦为形式主义的“打钩勾选（Checkbox）”过场。必须建立“价值驱动的阶段门禁（Value-Driven Stage Gates）”，设置明确的控制与监测指标。门禁不仅用于推进项目，更核心的作用在于提供清晰的“项目终止标准（Kill Criteria）”，一旦发现项目无法产生既定业务价值，能够迅速果断地关停（Kill），避免资源持续浪费。 --- ### [19:47](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=1187s) 无需重组的高管计分卡与五步落地法 向管理层和 CFO 汇报 AI 业务价值时，应聚焦于能够直接映射到财务损益表（P&L）的指标，主要涵盖三大核心驱动力：增长赋能（Growth Enabler）、效率赋能（Efficiency Enabler）和生产力赋能（Productivity Enabler）。向管理层展示 AI Agent 的产出时，直接引入“复用比率（Reuse Ratio）”这一核心数据，即可直观体现资金投入的合理性、组织采纳度以及实际产生的业务价值，无需进行剧烈的组织架构重组（Reorganization）。 | CFO 汇报计分卡核心维度 | 衡量焦点与业务映射 | |:--- |:--- | | **价值 (Value)** | 对应损益表（增长 / 效率 / 生产力）的实际经济收益 | | **采纳 (Adoption)** | 跨部门使用广度与智能体复用比率（Reuse Ratio） | | **可靠性 (Reliability)** | 数据质量标准、安全管控与合规运行表现 | | **速度 (Velocity)** | 数据产品与 AI 能力端到端交付及迭代周期 | 为帮助团队在不破坏现有组织架构的前提下快速见效，可落地执行轻量级五步法： - **清点资产**：对现有全部 AI 资产与数据产品建立完整清单（Inventory）。 - **计算复用比率**：量化各智能体与流程的复用次数及使用范围。 - **锚定三大企业成果**：挑选 3 个能够直接体现业务价值的企业级结果。 - **聚焦突破输出**：选取其中至少 2 个成果全力推进落地并交付业务输出。 - **发布终止标准**：设立并公开 1 套明确的 Kill Criteria，依托价值门禁机制坚决裁撤无效项目。 ### [22:50](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=1370s) 总结陈词：AI 时代的赢家在于构建可复用能力 演讲在对企业未来 AI 战略的核心洞察中进入尾声。Kader Sakkaria 强调，在接下来的几年里，能够在激烈的竞争中脱颖而出的企业，绝不是那些“开发了最多 AI 模型或项目”的公司，而是那些真正做到“只构建一次，并尽可能最大化复用”的企业。这种以平台化思维和可复用能力为核心的战略，才是让企业摆脱重复造轮子、实现规模化效益与持续商业价值的关键所在。 演讲者以这句发人深省的核心观点（punchline）作为全场演讲的精炼总结，向在场的所有企业高管与技术实践者表达感谢并结束了分享。 - **核心结语**：“未来几年真正的赢家，不会是那些构建了最多 AI 的企业，而是那些能够‘只构建一次并尽可能多次复用’（build it only once and reuse as much as possible）的企业。” ## Notable quotes > [00:04](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=4s) 一旦关掉麦克风，我的发挥空间就会大得多 — Kader Sakkaria > [00:04](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=4s) 你的团队到底能否做到构建一次，就能在整个 AI 领域内反复复用？ — Kader Sakkaria > [03:29](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=209s) 消灭 4 名员工的成本换来这些 Token 开销，这到底值不值？ — Kader Sakkaria > [06:56](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=416s) AI 并不能单纯依靠使用量的增加来实现规模化扩展。 — Kader Sakkaria > [06:56](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=416s) 我们的业务独一无二 — Kader Sakkaria > [15:31](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=931s) where the skeletons are hidden — Kader Sakkaria > [18:09](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=1089s) Single throat to choke — Kader Sakkaria > [22:50](https://www.youtube.com/watch?v=mAKNAKCF9DU&t=1370s) 未来几年真正的赢家，不会是那些构建了最多 AI 的企业，而是那些能够‘只构建一次并尽可能多次复用’（build it only once and reuse as much as possible）的企业。 — Kader Sakkaria ## Mentioned Kader Sakkaria, Marsh, MIT (Massachusetts Institute of Technology), Gallagher, Google, Jack Welch, GE (General Electric), Agentic AI, POC Graveyard, Reuse Ratio, Lakehouse, RBAC / ABAC, Pair Engineering, Stage Gates, Kill Criteria ## Video description Kader Sakkaria - Global Head of Data Strategy, Marsh Why do so many promising AI pilots fail to create lasting enterprise value? The challenge is rarely the technology alone—it is the organization’s ability to align strategy, data, governance, operating models, talent, adoption, and measurable outcomes. This session will present a practical executive framework for moving beyond isolated experimentation and building sustainable enterprise AI capability. Drawing on real-world transformation experience across global enterprises and the public sector, I will explain how leaders can create the foundations needed to scale AI responsibly and convert innovation into business performance. Attendees will learn how to: * Diagnose the organizational barriers preventing AI from scaling * Connect AI investments to growth, productivity, risk reduction, and customer value * Establish clear ownership, governance, and an operating model for enterprise execution * Build the data, talent, and adoption capabilities required for sustainable results * Measure AI success through business outcomes rather than technical activity Executives, business leaders, and transformation practitioners will leave with a practical blueprint they can use to evaluate their current AI maturity, prioritize the capabilities that matter most, and turn AI ambition into measurable enterprise value. [Source](https://www.youtube.com/watch?v=mAKNAKCF9DU)

[0:04] Kader Sakkaria: 一旦关掉麦克风，我的发挥空间就会大得多 [0:04] Kader Sakkaria: 你的团队到底能否做到构建一次，就能在整个 AI 领域内反复复用？ [3:29] Kader Sakkaria: 消灭 4 名员工的成本换来这些 Token 开销，这到底值不值？ [6:56] Kader Sakkaria: AI 并不能单纯依靠使用量的增加来实现规模化扩展。 [6:56] Kader Sakkaria: 我们的业务独一无二 [15:31] Kader Sakkaria: where the skeletons are hidden [18:09] Kader Sakkaria: Single throat to choke [22:50] Kader Sakkaria: 未来几年真正的赢家，不会是那些构建了最多 AI 的企业，而是那些能够‘只构建一次并尽可能多次复用’（build it only once and reuse as much as possible）的企业。

[1:11] Kader Sakkaria: What is that your team can build once and then it can be reused again and again within the AI space? [3:29] Kader Sakkaria: What are you doing about AI? Are you adopting AI? What is happening? [3:29] Kader Sakkaria: What is the business value that impact that you're actually creating to the organization by spending all this money? [18:09] Kader Sakkaria: I don't know how many of you guys have seen like Jack Welch during his GE days... is what is a 'single throat to choke' is what he would say, but I'm kind of making it more politically right: one accountable owner, and one outcome that we're all working towards, which actually leads to one single path. [22:50] Kader Sakkaria: The ones that actually build it only once and reuse as much as possible

[0:04] So I was kind of thinking about this  presentation as I was coming in my plane ride. [0:10] And a lot of you guys I'm pretty sure that you're  part of corporate or being part of the corporate. [0:15] So your global communications dictate what you can  or cannot say as well. So I'm going to do my best [0:22] to kind of give you more insights, at least some  tidbits that you can take away with. Um but happy [0:27] to talk after this. So when I'm off the mic um I  think I have flexibility to give even more better [0:32] details than what I can do on my mic. So wanted to  wanted to give you that heads up. So um I'm going [0:39] to slip, you know, all these who I am and the  boring facts of what I did and all that. I don't [0:44] think we can all look it up. LinkedIn days, right?  I want to focus on two things, right? Uh today. So [0:53] what I want you guys to kind of leave with  is one number, one ratio that you can use, [1:01] and one scorecard that you can actually leverage  as far as driving AI is concerned. Okay. So 95%. [1:11] So how many of you guys have  heard this number in this [1:19] There you go. At least I see few hands going up.  So this is this is from a MIT study, right? That [1:25] was actually done for AI. And it's about what  some of the AI um proof of concept lot of the [1:36] organizations have been doing and how they've been  kind of looking at AI in in total totality, right? [1:42] But if you look at the headline on this number,  the headline is very thin, right? If you actually [1:48] look on under the covers, the thing lot of us  are very analytical, statistical type of people. [1:53] We ask like, "Okay, where does that number come  from, right? So, it actually comes from about 173 [2:00] interviews that MIT did with a lot of corporate  organizations and then came back with this number [2:07] which says 95% of what the AI work that's actually  being done is still kind of dies in the POC [2:16] graveyard rather than that being productionalized  and being used as part of a institution workflow [2:23] and augmented enough to be able to create  value. So, that's the biggest challenge a lot [2:28] of these organizations, specifically corporate  organizations, are actually having, right? So, [2:34] there's one question that I would always like  to kind of have in in in my mind, right? Like [2:39] what that is is what is that 5% of the  organizations that are actually doing well [2:46] where the next the rest of the 95% is actually  failing on, right? So, as we go through [2:53] I want you to look through one question in  your mind, right? Um that I want you to keep [3:01] thinking about this and I'll as I go through  my presentation, I'll try my best to answer [3:06] some of these questions at least answer to  that question I'll build on is what is that [3:12] your team can build once and then it can be reused  again and again within the AI space, right? That [3:18] is one thing I want you to think about a little  bit more as as we go through this presentation. [3:26] Okay. [3:29] Okay. So, the goal post  literally has actually moved from [3:34] from all of our perspective, right? It's actually  moving in from a productivity measure, right? [3:40] To more of what that impact is actually happening  to the organization itself. And if you look at [3:48] pretty much every one of the organizations in the  boards, they were actually questioning on is what [3:55] are you doing about AI? Are you adopting AI? What  is happening, right? That was the main dialogue [3:59] and discussion. Pretty much every board was asking  and the leadership behind that is was like, okay, [4:05] let me do something about it, right? And then we  all basically started creating POCs, use cases [4:11] that we thought is going to be the best to kind  of be able to run and and and implement. But [4:16] what has happened in the last year, the same board  is actually asking a different question, right? [4:23] Is what is the business value that impact that  you're actually creating to the organization [4:28] by spending all this money? And I'll I'll give you  an example. So but the last one quarter, we were [4:37] actually looking at some of the developers that  we actually opened up um AI to. And on an average, [4:45] some of these developers have been using $60,000  worth of token. I'm not It's not an exact number. [4:50] It's It's an average that we are trying to  balance out. Right? And so if you extrapolate [4:55] for that for the whole year, right? So that's  about $240,000 worth of token like we'll end up [5:01] having the developers spending. So that means,  right? For all of us like we are actually asked [5:06] to manage our cost, meaning eliminate people from  here within the organization to create that value. [5:13] So on an average, it's going to be about four  FTEs, right? That you're going to be using that [5:19] to replace. Now, the question is, is that worth  that four FTE replacement? And literally at lunch, [5:26] I was talking to some of my colleagues from  Gallagher, um the previous company I was with, is [5:32] can this sustain, right? How much of that value  can be sustained through this whole process? [5:38] So that is what the challenge lot  of these organizations right now is. [5:43] The gap that the enterprise is trying to figure  out is this money that we are spending in AI, is [5:50] it going to be worthwhile for us to sustain what  the value for the P&L impact is going to happen? [5:56] So, every one of the board, now their question has  become is like, how is this trying to get added [6:02] to the overall bottom line, which is the  EBIT, right? Like earnings before income tax [6:08] is what everybody's looking at that impact rather  than not after you amortize your software, but [6:12] even before that, what does that happening and how  is the value being getting created at as part of [6:18] the organization? So, that's the key driver that  that's that that everybody is trying to push for. [6:24] There There's, you know, a bunch of stats that I  can actually go through, right? Like, you know, [6:29] specifically if you look at it, 39%, right?  Actually are the ones of the value that the [6:35] POCs have been created is literally attributing  towards the EBIT data. So, it is a small sliver [6:43] of an impact that actually creating for the  money that each of the organizations are [6:47] actually building and spending money on. Okay. So,  adoption Now, how does adoption work, right? So, [6:56] adoption literally has outrun the  readiness that we've all been focused on. [7:01] So, if you look at look at the golden, you know,  the bars, the bar charts versus the blue chart, [7:06] right? So, there's a bunch of aspiration, which  is about 85% of the, you know, organizations want [7:12] to be energetic AI and that's the framework that  they want to build in. But then, if you look at [7:18] the number of agents that's actually scaling  is 23%, right? There's a huge gap between [7:24] what the aspiration is, which is where the  things are as far as the agents are concerned, [7:29] right? How the agent framework is actually  working. So, the the true gap actually is the [7:35] ambition, right? That we're all working towards  in trying to be able to adopt some of these work [7:41] more as an automated process within your workflow  rather than just trying to make it more of [7:48] a use case driven approaches. That's what we end  up doing as far as proof of concept is concerned. [7:54] So, the key is the goal post is moving. People  are all at least the board is asking the question [8:01] where what it is doing to our P&L and then the  third is like how is the adoption happening within [8:06] the organization, right? So, those at least  are the ones that the question is slowly now [8:11] I'll kind of build the answer around the question  that I talked about what the 5% are doing. [8:16] So, if you start thinking in your mind how  does your organization fit into these scales, [8:22] start thinking about it, right? And I'm  going to have a quick quiz at the end of [8:26] end of this session. Um you don't have to reveal  which organization you're from, but I would like [8:31] to kind of get some numbers from you uh to  to see where where where you guys fall under. [8:38] Okay. So, AI does not scale with  usage, right? Because like I mentioned, [8:45] we all create all these POCs of values that  we think arbitrarily we are trying to create, [8:51] but if there's no reuse of those agents or the use  cases that you're building on, then it is going [9:00] to be a significant challenge to the organization  to justify the cost structure, right? So, just to [9:08] show of hands, right? How many of you can say that  you created one agent but like it's being used [9:14] reused across your different organization without  even touching a rewriting or or that agent. [9:23] Interesting. Okay, so one out of  maybe about 40-50 people here. Um [9:29] I need to talk to you after this. I would love to  kind of understand how you guys are doing that. Um [9:35] part of what we all kind  end up doing is, right? Each [9:39] department or each unit wants to create their  own AI because if you ask the department, [9:44] we are unique, right? There's a lot of people hide  behind data residency, data consent, and then they [9:51] basically hide behind those, but if you look at  the underlying data is a single data layer, right? [9:57] But then each unit is actually trying to build  their own structure around this. And so, that is [10:02] what I mean by like where you have the base data,  but you're creating value through not through use, [10:08] reuse of it, but you're actually creating value  by trying to build your own. But what ends up [10:12] happening is that's when you're going to have  multiple competing agents doing the same thing. [10:17] You're going to have more cost that's actually  going to drive your maintenance structure, [10:21] what we call the run the business cost, is going  to go significantly higher. So, remember the P&L [10:26] question I asked? Now, everybody's going to  tie it back and saying, "How is this kind of [10:30] impacting my P&L as well?" So, there's going  to be a significant, you know, thought process [10:36] needs to get put around this of not just building  once, but how do you actually build and, you know, [10:42] once and try to reuse within the enterprise  organization organization as much as you can. So, [10:51] the reuse ratio. So, this is one thing that I have  started looking at within my organization to start [11:01] at least a level of measuring that we can actually  start doing, right? Because the biggest problem [11:07] within the organization that we all struggle  through is build once, reuse it as much as you [11:13] can, right? Because that is going to be one  of those. So, if you can actually take one [11:19] thing away from this presentation when you walk  through is think about the reuse ratio within [11:25] your organization, right? So, if your reuse is  under 20%, that means you're actually operating [11:34] on more of a product mode not on project mode  not on a product mode as an organization itself. [11:40] So if you're in the middle somewhere, so the  platform is actually forming and norming it's [11:44] kind of helping through it. But then you're above  that means now you're in that top 5% that I talked [11:49] about, right? Which is the organization that's  literally reusing the data and the agent and the [11:56] framework as much as possible. So show of hands  like how much do you think I mean I don't know [12:04] your company and all that stuff so I'm not going  to use it against you. But how how many of you [12:10] are up in the start of the journey or how many of  you guys are on the mature part of this journey? [12:20] Someone who who's all under 20%?  Okay. Someone in the middle? [12:27] There's few and who are on the other  side? Okay. I see this which is [laughter] [12:33] Okay. So at least there you go. So the majority  still at this even just a survey here shows that [12:38] 5% right now. There's only one person that's  saying that at actually at that which is which [12:43] is a true statement, right? No, which is what the  reality is right now where the organizations are [12:48] all kind of driving towards is how do you actually  understand the reuse ratio and figure out a way [12:56] to look at it. And I use a mathematical formula  maybe after this meeting if you want I'm happy [13:01] to share the mathematical formula with with  with you because one of the key thing that [13:05] you have to put in in your denominator is how  many times this particular agent is being used [13:11] or like in a particular process is being reused.  Happy to share the formula with you which I use [13:16] within my organization across all our business  verticals to see how well the adoption is and [13:22] how well the usage and the reusage of of of these  bots are happening within within the organization. [13:30] Okay, the three traps that actually help AI  in our POC journeys, right? So, number one [13:41] is the rebuilding tax, which is what I was  talking about like you know, not everybody [13:45] tries to use the same you know, baseline. They use  the same data, but they try to kind of build it on [13:50] on it and then they also put on controls and  governance as an afterthought. It's never thought [13:57] through up front. Um I would be interested in  kind of trying to learn a lot more people because [14:03] your R back or A back policies that you want  to apply on your data and all that becomes more [14:08] of an afterthought process rather than being  built as part of your overall organizational [14:14] capability itself. Um that is something that  I'm at least within our org we are trying to [14:18] push really hard to kind of see how does the  adoption happen on the governance side. So, [14:24] it has to be part of the journey and when  you're doing your thinking rather than post [14:29] implementation then you say, "Well, I need to kind  of look at consent of the data. I need to look at [14:35] the data roles that are associated and then  the type of residency that's built into it [14:39] and see if that is going to be something that  we have to go and do the rework. And then the [14:44] last one is the assumed adoption, right? So,  if you look at the organizations is and I was [14:53] actually talking to somebody who used to work in  Google, which was very interesting conversation. [14:58] And so, his statement that he actually made was  there's a bunch of people who will actually will [15:03] get together, they'll actually build a product and  there'll be only two users using it. They'll spend [15:07] a word of million dollars or three million dollars  on to it, but there's two users using. Literally, [15:12] I'm talking about just within Google, not outside  Google itself, right? So, imagine that level of [15:17] adoption is what you're going to get then your POC  is is to go to the graveyard. It's not going to go [15:22] to a level of reusability where you can  industrialize the process itself to scale. [15:31] So, one of the areas that we are actually looking  at as the six capabilities that can actually, [15:36] you know, needs to move together. I'm not going to  go through every one of them, but there's a couple [15:41] of ones that I'm going to highlight in for you  guys, right? So, one is establishing a very strong [15:47] data foundation. Like, that is something that's  actually a a a kind of a baseline, table stakes, [15:53] whatever you want to call it. That is where you  need to focus on is how do you actually create and [15:58] what kind of data stand, you know, standards and  data qualities that you can imply. I'm not talking [16:03] about structured, unstructured, like the type that  you want to have, but it's like once you have your [16:07] data in your lakehouse, what kind of a foundation  that you're able to get out of it and the type of [16:13] controls that you can actually put on. The second  that also helps in this is talent and adoption. [16:19] That is one thing that is very evident in at least  in a lot of large corps and mid-size corporations. [16:26] We have a lot of institutional knowledge. We have  a lot of that tribal knowledge of how the data, [16:30] where the data is. And also, we know the skeletons  that's attached behind and where it's been hidden, [16:34] right? But what you don't have is the level of  modernized thinking that needs to kind of get [16:39] injected as part of it. And the immediate  reaction for a lot of us is like, "Oh, [16:43] let's go bring consultants to kind of do it." But  what ends up happening is now with the consultant, [16:48] you're actually trying to go through an entire  learning curve with them on and try to understand, [16:53] let them know what your data is all about and  your business and then try to do it. Now, one [16:58] other easy way that's actually helpful for you to  do is identify champions within your organization [17:04] and then figure a way to upskill and then partner  them with some of these kind of like really high [17:10] um caliber consultants who can actually help drive  the conversation, right? So, a lot of times that [17:15] pair kind of like engineering construct works  pretty well where you have your institutional [17:21] knowledge with some level of upskilling done and  then be able to get another strong consultant [17:26] or contractors, you know, high-caliber people  who can actually help drive the conversation. [17:31] And you know, the last but not least like you  know, I'll quickly go through on this slide is the [17:36] value measurement. The biggest thing on the value  measurement if you look at is if you don't measure [17:43] then you're not tracking anything. You're not  creating value, you're not pushing through. [17:46] So that is one key thing that you have to push  through is how do you actually measure this data [17:52] and then be able to do it. So one of those  measurement I was giving you was reuse ratio. [17:56] So we can use that as part of your journey  that will that will be a good way to start [18:01] at least a start to kind of start finding  it. Okay. So this is one mantra that I use. [18:09] Um, I don't know how many of you guys have seen  like Jack Welch when during his GE days and this [18:14] thing is not changed much is what is a single  throat to choke is what he would say but I'm kind [18:19] of making it more politically right is like one  accountable owner, right? Um, and one outcome that [18:26] we're all working towards which actually leads  to one single path, right? So you need to have a [18:31] proper single business owner who's able to kind of  identify what's the P&L value that this is going [18:36] to create of the value stream and then you have  one person on your data side which is actually our [18:41] domain lead what I would call who's accountable  for end-to-end delivery within your internal [18:46] organization to be able to build and deliver it  together and then so you're heading down the same [18:50] path to say, "Okay, this is a value. This is how  this AI solution is going to be implemented and [18:56] creating value within the organization." So that  is what I would say would be a second point just [19:03] to remember for you as you go through. Right? You  don't want five people saying that I'm accountable [19:08] for. You just need a single person who's  actually accountable for end-to-end delivery. [19:13] So, there's four gates, right? Um so, I  you know, we always talk about stage gates, [19:18] but what ends up happening is stage gates without  proper control just becomes a gate people can [19:23] walk through, right? That's all it is because  they'll say, "Okay, that's just a checkbox that [19:27] I need to do." Rather than, "Hey, let me have  proper outcome and controls that can be managed [19:32] and monitored with this." So, that is one area  that you need to kind of look through as you're [19:37] establishing it not only stage gates, but like  value-driven stage gates that can actually be [19:42] helping driving the conversation  rather than you you doing it yourself. [19:47] Okay. And then the last, remember I said like  I'll give you one ratio, I'll give you a one [19:52] way which is the talent that I talked about  and a quick scorecard by the end. So, this is [19:57] more most at the tail end of it. So, the scorecard  that here's what like I use, right? So, couple of [20:03] things I'll quickly highlight here. Name the P&L  line and the AI initiative that actually moves [20:08] that you know, value of the organization. So, we  always look at three different factors, right? So, [20:14] one is growth enabler, one is efficiency enabler,  and third one is the productivity enabler. [20:18] As long as you can attach to one of them and then  be able to measure and work through is something [20:23] that you need to look at. Um second is like  or not in the order, but the data products, [20:30] how it's consumed, so that way the reusability  ratio that I talked about. That is something [20:34] to measure and then be and put on. So, if  you look at when the type of presentation [20:40] that you're going to be presenting it to your  executives or leadership, right? As part of [20:45] the way that you actually present, if you can add  this reuse ratio of what you're doing with your [20:52] you know, with your AI agents that you're  building, just put that one number. I'm not [20:56] asking you to reorg yourself, I'm not asking you  to blow up your organization. I'm just trying to [21:01] give you some incremental things that you can  actually look at without reorging, what can you [21:06] do to kind of get the executive attention and be  able to show the value that you and your team is [21:11] actually building on, right? So, reuse ratio,  if you can actually add that to your scorecard [21:16] and be able to share that with your CFO because  the reuse ratio shows the value that's being [21:21] generated to the organization and justification of  the cost that they you spent on. And then it also [21:27] tells the adoption of what is happening within  like it actually gives you those three different [21:31] values immediately. So, that actually will start  resonating with which is what you need to have as [21:37] your best friend to make sure that the investment  keeps kind of flowing as you go through. [21:43] Okay. Um real quick. So, these are the four  numbers like in the value, adoption, reliability, [21:48] and velocity, you can actually put in on your  CFO report quickly as part of the scorecard. Um [21:55] and I think they should If you I know you guys are  taking pictures, but I'm hoping that this will be [21:59] available in in in the site as well. So, the five  moves that I talked about was just does not need [22:08] reorganization at all, right? So, is make an  inventory of your AI assets, right? Calculate [22:14] the reuse ratio, pick the three enterprise  outcomes that you can actually show value on, [22:19] convert into two of them with at least take two  which can actually get you through the outputs, [22:25] and then publish one kill criteria, right? So,  remember the stage gates that we talked about? It [22:30] should not be just a stage gate, but it can also  figure out a way to kill if it's not adding value. [22:35] So, that way you can use the stage gate to kill  the projects that is absolutely unnecessary for [22:40] you. Okay. And then the last slide, I'm sorry,  I'm running fast because I'm already past my [22:46] time. I'm trying to make sure I catch up onto  it. So, the enterprises that win the next 2 [22:50] years will not be the ones that build the most  AI, but they'll be the ones that actually build [22:55] it only once and reuse as much as possible.  So, that's my punchline for today. Thank you.