IBM Target $3 Billion Productivity Cut Via AI internal Deployment

IBM通过内部部署AI目标削减30亿美元生产力成本 IBM将675名呼叫中心员工转岗至AI相关职位，并通过将严格的中心辐射式治理架构与规模更小、适配特定领域的语言模型相结合，实现了45亿美元的生产力效益。 Madison Gooch 阐述了企业组织如何通过从分散式实验转向中心辐射式治理模式，弥合AI成效衡量与规模化应用之间的差距。她重点介绍了IBM内部带来45亿美元总生产力价值的“零号客户”（Client Zero）战略，详细说明了确立严格的ROI目标、利用领域专用模型以及开展结构化价值评估如何优化AI投资。她还结合人力资源成本削减及电信领域高性价比模型规模化落地的成功案例，具体展示了这些原则的应用。 尽管 85% 的 CEO 期望到 2027 年实现规模化 AI 的正向 ROI，但目前仅有 25% 的企业达到预期；衡量并交付可量化的业务价值仍是阻碍 AI 项目跨部门扩展的最大瓶颈。 (0:06) 企业 AI 实践表明纯去中心化治理难以带来高投资回报，成功的组织通常采用中心辐射型（hub-and-spoke）架构，在企业级统一标准与业务单元自主创新之间实现平衡。 (7:23) 作为自身技术的“零号客户”（Client Zero），IBM 内部推行 Watsonx 并在两年半内创造了 45 亿美元的生产力价值；其机制要求所有内部 AI 场景必须在两周内完成商业案例论证，并在一年内实现完整 ROI。 (9:01) 通过在人力资源领域上线 34 个以上的生成式 AI 场景，IBM 将净运营预算削减了 40%，员工体验评分提升 55 分，并将原呼叫中心 750 名员工中的 675 人成功转岗至新成立的数据科学与 AI 部门。 (14:08)

Madison Gooch of IBM argues that translating enterprise AI investment into quantifiable return requires moving beyond isolated pilots to a disciplined framework that standardizes value measurement, adopts hub-and-spoke governance, and matches model architecture to specific operational workloads. Scaling Barriers and Hub-and-Spoke Governance (0:06) According to IBM Institute for Business Value surveys of approximately 5,000 C-level executives, quantifying business value and managing governance risks remain the primary obstacles preventing corporate AI initiatives from scaling. Currently, 25% of CEOs explicitly report that their AI deployments fail to deliver anticipated value. While post-2022 experimentation favored decentralized adoption where individual business units built isolated tools, industry results show that purely decentralized models generate fragmented data silos and fail to produce sustainable returns. To overcome these barriers, organizations are shifting toward a hub-and-spoke operational model. Under this structure, a centralized core establishes architectural standards, safety governance, and baseline return on investment (ROI) metrics, while decentralized spokes retain the autonomy to execute specific use cases. IBM survey data indicates that 68% of companies with an 'AI-first' orientation maintain established data governance policies and manage early initiatives within structured innovation portfolios, even as only 27% of broader enterprises report having mature foundational AI capabilities. IBM's Internal Mandate and Value Taxonomy (9:01) To streamline its own operations, IBM adopted a 'Client Zero' strategy under CEO Arvind Krishna to test internal automation before offering it externally. Addressing IBM's 10% to 11% Selling, General, and Administrative (SG&A) expense ratio—which trailed top benchmarks of 2% to 3%—leadership mandated $3 billion in productivity savings over two years. The framework required every internal project to demonstrate full ROI and total cost of ownership in under one year. Under an enforced 'two-week rule,' initiatives unable to validate their business case within a fortnight were promptly dropped, while executive leadership held two-day quarterly reviews to track performance. Over two and a half years, the Client Zero approach yielded $4.5 billion in total productivity value. This total comprised approximately $3.0 billion in hard financial impact (direct ledger operating cost reductions and top-line revenue additions) and $1.5 billion in cost avoidance (calculated as dollarized employee time savings). IBM formalizes this evaluation across four pillars: pre-tax revenue generation, direct cost reduction, software and infrastructure cost avoidance, and employee productivity gains. HR Transformation and Skills Redeployment (14:08) IBM applied its value framework across human resources, deploying more than 34 generative AI use cases across talent development, compensation, benefits, and payroll. This rollout enabled a 40% reduction in the HR department's net operating budget alongside a 55-point increase in internal customer satisfaction scores. Complementary internal systems, such as the support application 'AskIT,' were deployed in under 100 days to address routine employee inquiries, targeting a 50% reduction in support volume against a baseline cost of roughly $18 per ticket. Rather than executing purely headcount-reduction strategies, IBM reorganized its operational staff. In one employee support call center of 750 workers, AI automation absorbed routine tasks and allowed the company to reassign 675 staff members into a newly created Department of Data Science and AI. These domain specialists were redirected to train, fine-tune, and monitor agentic workflows, translating operational human resources experience into automated recruitment and onboarding templates. Model Architecture Optimization and Governance Value (20:03) Enterprise ROI is directly constrained by model selection. Deploying massive, general-purpose large language models for high-volume, narrow workflows frequently causes severe cost overruns during inference. In one telecommunications partnership highlighted by Gooch, processing an initial 3 million customer call transcripts with a large model achieved 97% accuracy, but scaling the task across the full backlog of 330 million transcripts exhausted the allocated budget. Replacing the system with a smaller, domain-adapted model maintained the 97% accuracy rate while reducing infrastructure and inference costs by 10x. To align architecture with financial feasibility before deployment, IBM utilizes a four-phase assessment framework encompassing assessment scoping, discovery workshops, financial synthesis, and formal value case readouts. Gooch positions structured AI governance not as a restrictive compliance exercise, but as an active cost-avoidance driver that reduces risk and accelerates deployment velocity across both enterprise processes and public-facing analytical applications. While IBM demonstrates that disciplined metrics, smaller fine-tuned models, and centralized governance can convert experimental pilots into billions in ledger savings, the model depends on strict corporate enforcement—such as dropping unproven projects within two weeks and actively retraining displaced labor. The broader unresolved challenge is whether enterprises lacking IBM's extensive engineering infrastructure can replicate similar workforce reallocations and architectural cost cuts.

IBM 专家 Madison Gooch 指出，企业人工智能应用正面临从技术试验走向规模化投资回报（ROI）的严峻鸿沟；企业必须打破技术混淆与孤岛式试验，转向兼具中央治理与业务灵活性的轴辐式架构，并通过精细的四维价值量化与模型成本优化来实现可验证的财务回报。 衡量鸿沟与轴辐式治理转型 (0:06) 企业在评估人工智能投资回报时，常常错误地将传统机器学习、深度学习、大语言模型与自主智能体（Agent）混为一谈。单点应用的局部试验与跨业务部门的全面推广在投资规模和回报逻辑上存在本质差异。根据 IBM 商业价值研究院对约 5,000 名企业最高管理层的年度调研，虽然 ChatGPT 引发了普遍的试验热潮，但目前已有 25% 的 CEO 明确表示其 AI 项目未能达到预期价值，只有极少数项目能够真正进入全面投产阶段。缺乏可量化的业务价值衡量标准以及复杂的企业级治理风险，是阻碍项目扩展的两大核心障碍。 在组织架构层面，许多企业在 2022 至 2023 年间采取了完全去中心化的尝试，允许各业务部门独立构建解决方案，但这种模式最终被证明难以形成高投资回报。行业数据显示，68% 具备“AI 优先”思维的企业均建立了健全的数据治理政策，并将早期试验纳入结构化的创新投资组合。目前最具成效的路径是采用“轴辐式”（Hub-and-Spoke）治理模式：由中心枢纽统一制定技术标准、价值指标与风险治理规则，各分支部门则保有贴近一线场景的实施与创新自主权。 IBM零号客户机制与价值四支柱 (9:01) 为了在庞大组织中验证 AI 效能，拥有 28.6 万名员工的 IBM 将自身作为技术的“零号客户”，在向企业客户交付前率先内部部署 Watsonx 等技术。面对此前高达 10% 至 11% 的销售与管理费用率（SG&A），首席执行官 Arvind Krishna 提出了在两年内实现 30 亿美元生产力节约的目标。在实际执行中，该战略在两年半内累计创造了 45 亿美元的总生产力价值，其中包括约 30 亿美元的账面硬性影响（直接削减运营成本与增加新收入）以及约 15 亿美元的工时货币化成本规避。在此期间， Krishna 与跨职能执行委员会每个季度都会投入整整两天时间对各类用例进行严格复盘，并通过竞赛吸引了超过 20 万名员工参与基础创新。 IBM 将业务回报严格归入“硬性效益”与“软性效益”两大流向，并进一步细化为四项具体支柱：直接带来营收的税前收入、从账本上直接扣减的运营成本、通过合并重复系统或软件合同实现的成本规避，以及把员工从重复性劳动中解放出来的生产力节省。为了保障资本使用效率，IBM 实行严苛的准入规则：所有内部 AI 用例必须具备完整的总体拥有成本（TCO）核算，并在一年以内实现完全回本；若业务案例在立项两周内无法完成逻辑闭环，团队必须立即予以放弃并转向更高产值的方案。 人力资源场景落地与组织转型 (14:08) 在跨越人力资源、IT、销售、采购与供应链的 100 多个生成式及 Agentic 部署中，IBM 的人力资源部门展现了极高的落地成熟度，已有超过 34 个生成式 AI 用例处于生产环境。自动化转型打破了“降低成本必然牺牲员工体验”的传统假设：人力资源主管实现了 40% 的净运营预算削减，而内部员工满意度得分（CSAT）反而大幅跃升了 55 分以上。例如，通过在 100 天内从零构建的内部员工支持工具“AskIT”，企业针对单次约 18 美元的工单成本基准，设定了消除高达 50% 常规服务工单的目标，释放了巨额营运资金。 在人员组织配置上，IBM 对原有的 750 人规模员工支持呼叫中心进行了结构性重组。通过自动化工具接管日常问询，企业成功解放了其中的 675 名一线员工。IBM 未选择直接裁减人员，而是将这 675 名拥有深厚业务背景的领域专家成建制地转入新设立的“数据科学与 AI 部门”。这一调配使企业能够直接利用业务专家的实操经验来训练、微调模型并设计可重复的工作流，将一线业务感知转化为标准化的算法资产。 模型选型优化与端到端评估体系 (20:03) 技术架构层面的模型选型直接决定了企业级应用的推理成本边界与扩展上限。脱离场景盲目使用通用大模型容易导致推理性算力超支。在 IBM 协助某电信合作伙伴实施的客户关怀案例中，该企业最初调用大型 GPT 模型处理了 300 万条通话记录，准确率达到 97%，但在尝试将该方案扩展至积压的 3.3 亿条全量数据时，项目预算迅速耗尽。IBM 协助其换用经过私有领域数据微调的小型专用语言模型，在维持 97% 准确率不变的同时，将基础设施与推理整体成本压缩到了原来的十分之一。 为了加速企业客户的价值兑现，IBM 成立了专门的“AI 价值评估组织”，推行标准化的四阶段评估框架：第一步明确业务用例与价值驱动因素的对应范围；第二步通过调研访谈挖掘深层运营瓶颈；第三步进行财务综合建模与 ROI 测算；第四步向决策层汇报商业论证以统一跨部门共识。此外，IBM 将内部沉淀的招聘等端到端流程打包为预构建的智能体模板，并强调强有力的 AI 治理不仅关乎合规，更能直接规避安全与法律风险，从而加快模型和自主智能体的安全投产速度。 Madison Gooch 总结强调，AI 的可衡量回报并不局限于后台流程，也已在每年超 40 场大型现场体育与娱乐赛事的自动化数据洞察中得到证实。不过，从实验室小规模测试跨越到规模化投产，企业仍须直面数据治理欠账，并通过严格的周期考核与跨部门协作机制，破除传统技术项目回报滞后的惯性思维。

## Background 该视频由 Ai4 于 2025 年 8 月 27 日发布在 YouTube 平台上，由 IBM 美洲区 watsonx 副总裁 Madison Gooch 主讲。视频简介指出，尽管 85% 的 CEO 期望到 2027 年实现规模化 AI 的正向投资回报，但目前仅有 25% 达成预期；本次演讲深入探讨了企业衡量 AI ROI 的考量因素、常见障碍、IBM 内部的实践范式及加速实现价值回报的解决方案。 ## TL;DR 演讲深入剖析了当前企业在 AI 落地中“试点泛滥但生产落地稀少”的困境，指明价值量化模糊与治理缺失是阻碍 ROI 达标的核心瓶颈。Madison 重点分享了 IBM 践行“零号客户”（Client Zero）策略的实战成果：通过严苛的准入规则（两周定案、一年内见效）与轴辐治理架构，在内部落地超 100 个生成式与智能体 AI 用例，实现了 45 亿美元的生产力总回报，并成功推动 HR 与 IT 等核心业务的大幅降本增效。此外，她提出了涵盖硬性增收/降本与软性避费/工时节约的四大价值维度及标准四步评估法，并结合电信客户切换小语言模型（SLM）达成 10 倍降本与 UFC 赛事洞察等实例，强调底层架构选型与正向治理对规模化 ROI 的决定性作用。 ## Atmosphere & style 整场演讲风格干练、务实且极具科技巨头高管的说服力，通篇摒弃空洞的技术炒作，完全立足于硬核财务数据、组织架构演进与真实业务痛点展开。Madison 的台风从容自信、节奏紧凑，从开场直击行业 ROI 期望落差的冷峻剖析，到娓娓道来 IBM 内部“自啖狗粮”式的变革经验，情绪流转沉稳而笃定。整个分享过程充盈着强烈的实操指导性与商业理性，成功将前沿生成式 AI 的讨论从概念热潮拉回到资产负债表与业务成效的坚实地面。 ## Key takeaways - [00:06](https://www.youtube.com/watch?v=aLauVgmLDg0&t=6s) AI ROI 的评估需因地制宜：AI 涵盖从传统机器学习到大模型和智能体等技术光谱，局部嵌入式试点与企业级跨部门部署在价值评估逻辑上存在本质差异。 - [02:13](https://www.youtube.com/watch?v=aLauVgmLDg0&t=133s) 期望与现实存在明显鸿沟：仅有 25% 的 CEO 认为其 AI 项目达到了预期的商业价值回报，绝大多数企业面临“概念验证（PoC）泛滥但生产落地稀少”的困境。 - [03:26](https://www.youtube.com/watch?v=aLauVgmLDg0&t=206s) 阻碍 AI 规模化落地的两大核心挑战：一是缺乏对 AI 价值的清晰理解、量化与兑现能力，二是涵盖安全、合规、道德透明与风险缓释的 AI 治理体系尚未完善。 - [04:08](https://www.youtube.com/watch?v=aLauVgmLDg0&t=248s) AI 价值交付具有间接性与滞后性，企业必须统一组织对“硬性财务收益”与“生产力/工时节省”等指标的度量标准，以应对短期内快速见效的考核压力。 - [06:58](https://www.youtube.com/watch?v=aLauVgmLDg0&t=418s) 组织运营模式从去中心化走向“轴辐模式”（Hub and Spoke）：由中央核心赋能平台统一数据治理与度量标准，并向业务单元赋能，能带来远高于分散实验的 ROI。 - [09:01](https://www.youtube.com/watch?v=aLauVgmLDg0&t=541s) 推行“零号客户”（Client Zero）机制：IBM 坚持让内部研发的 AI 与自动化技术率先在自身 28.6 万名员工中全面部署，以优化过高的 SG&A（销售及行政开支）比例并打磨产品。 - [10:26](https://www.youtube.com/watch?v=aLauVgmLDg0&t=626s) IBM 两年半内实现 45 亿美元生产力回报：其中约 30 亿美元来自账面成本直接削减与新增业务收入，15 亿美元来自成本规避与工时金额化节约。 - [11:39](https://www.youtube.com/watch?v=aLauVgmLDg0&t=699s) 严苛的用例准入规则：IBM 要求立项商业案例必须在 2 周内完成论证且 1 年内兑现 ROI，并围绕“业务增收、账面减本、成本规避、工时节约”四大维度统一价值评估。 - [13:48](https://www.youtube.com/watch?v=aLauVgmLDg0&t=828s) 打破降本与体验不可兼得的矛盾：IBM HR 部门投产 34 个生成式 AI 用例，在实现运营预算降低 40% 的同时，员工体验评分逆势提升了 55 分以上。 - [15:28](https://www.youtube.com/watch?v=aLauVgmLDg0&t=928s) 将业务流沉淀为预构建智能体模板，并通过专门设立的“AI 价值评估组织”绘制专属价值图谱（Value Map），消除业务专家与技术架构人员之间的认知隔阂。 - [18:33](https://www.youtube.com/watch?v=aLauVgmLDg0&t=1113s) 业务人才转型与知识灌注：IBM 将原呼叫中心的 675 名接线员工转入数据科学与 AI 部门，把一线业务专家的智慧直接转化为企业级可复用的智能化流程资产。 - [20:23](https://www.youtube.com/watch?v=aLauVgmLDg0&t=1223s) 模型架构选型直接决定规模化成本：某电信客户在将客服摘要扩展至 3.3 亿通记录时，通过从通用大模型切换至定制小语言模型（SLM），在维持 97% 准确率的前提下实现了 10 倍的推理成本降幅。 - [21:54](https://www.youtube.com/watch?v=aLauVgmLDg0&t=1314s) AI 治理具备双重价值：治理不仅是合规与风控的防御手段，更能通过规范化流水线加速多模型与 Agent 的引入和部署，直接驱动业务营收与生产力增长。 ## Detailed walkthrough ### [00:06](https://www.youtube.com/watch?v=aLauVgmLDg0&t=6s) 演讲主题与核心要点概述 IBM 美洲区 watsonx 副总裁 Madison Gooch 在演讲开场时热烈地欢迎了现场观众，并明确了本次分享的主题：企业如何在纷繁复杂的 AI 项目中真正实现投资回报率（ROI）。在接下来的 20 分钟内，她为与会者梳理了本次演讲的三大核心要点：首先是分享行业洞察，揭示企业在部署 AI 时的常见模式及回报类型；其次也是最关键的部分，她将深入剖析 IBM 自身的实践经验，用具体案例展示 IBM 是如何在其业务中榨取 AI 价值的；最后，她将提供一套实操性极强的评估框架，指导企业如何将特定技术项目与明确的 ROI 进行精准绑定。 Madison 强调，当今的 AI 概念极其宽泛，涵盖了传统机器学习（ML）、深度学习、自然语言处理（NLP），以及新兴的基座模型（Foundation Models）、大语言模型（LLM）和 AI Agent。虽然这些技术都归于 AI 的大伞之下，但在计算 ROI 时，技术类型的选择与应用范围的广度将直接决定价值评估的方式。例如，一个仅在单一应用中嵌入 AI 的局部试点项目，其 ROI 算式和预期，与在整个企业跨部门大规模铺开 AI 解决方案有着天壤之别。因此，企业绝不能混为一谈，必须深入细节，切中要害地明确自己的计算逻辑。 - **演讲目标时间**：约 20 分钟 - **三大核心要点**：行业 ROI 部署模式洞察、IBM 内部实现 ROI 的实战案例、企业精准映射 AI ROI 的评估框架与指导建议 - **AI 涵盖的技术光谱**：机器学习（ML）、深度学习（DL）、自然语言处理（NLP）、基座模型（Foundation Models）、大语言模型（LLM）、AI Agent - **应用规模对比**：嵌入式 AI 应用的局部试点 vs 跨部门全企业级大规模部署 --- ### [02:13](https://www.youtube.com/watch?v=aLauVgmLDg0&t=133s) 当前市场现状：期望与现实的鸿沟 谈及目前的市场大环境，Madison 抛出了一个令人深思的现象：自 ChatGPT 问世以来，生成式 AI 在全球范围内掀起了空前的热潮与期待，但高管们的态度正逐渐从最初的“盲目兴奋”转向“理性审视”。如今，许多 C 级高管（C-suite）在经历了长期的探索与尝试后，开始严肃地追问：“我们做了这么多实验，到底换来了什么价值？如何才能让这些技术在业务中真正落地并带来收益？” 根据 IBM 商业价值研究院（Institute for Business Value, IBV）对全球约 5,000 名 C 级高管的深度调查，目前仅有 25% 的 CEO 认为其公司运行的 AI 实施项目达到了预期的回报价值。市场上绝大多数企业仍处于海量实验与试点（PoC）阶段，真正能够突破重围、成功投入生产环境（Production）并实现规模化价值的 AI 项目屈指可数。针对这种“试点泛滥但生产落地稀少”的困境，CEO 们在调查中一致指出了导致 ROI 难以达标的两大关键瓶颈。 ```mermaid flowchart LR accTitle: AI投资回报率期望与落地现状对比 subgraph S1["市场期待与尝试"] A["ChatGPT 掀起技术热潮"] --> B["海量 AI 概念验证与试点实验"] end subgraph S2["落地现实与困境"] C["仅 25% 的 CEO 认为 AI 达到预期 ROI"] D["极少数试点项目成功进入生产环境"] end B -. "落地瓶颈".-> D ``` - **调研机构与规模**：IBM 商业价值研究院（IBV）每年针对全球约 5,000 名 C 级高管进行专题调查 - **核心统计数据**：仅有 **25%** 的 CEO 表示其企业内部的 AI 举措带来了预期的商业价值 - **行业痛点**：概念验证（PoC）与试点项目极多，但成功转化为生产环境（Production）并规模化提效的项目比例极低 <ElicitationsGroup message="如果您想继续深入了解本演讲的后续内容，可以选择以下方向："> <Elicitation label="获取下半部分关于 IBM 内部 AI ROI 落地案例的详细总结" query="请继续总结 Turning AI ROI Expectations into Reality With IBM 演讲的下一部分，重点关注 IBM 自身的 AI 落地案例与框架。"/> </ElicitationsGroup> ### [03:26](https://www.youtube.com/watch?v=aLauVgmLDg0&t=206s) 高管调研的核心挑战：ROI 价值量化与治理 在对各大企业 C-level（高管层）的年度调研中，每年反馈出来的核心议题都呈现出惊人的一致性。阻碍企业实现 AI 规模化扩展的首要障碍有两个方面：第一是关于“价值”本身的认知障碍——企业不知该如何理解价值、如何计算价值，以及如何在整个组织架构中真正兑现价值；第二则是涵盖范围极广的“治理”（Governance）问题，包括如何在企业内部存在大量 AI 技术测试、实验和应用的情况下，做到风险缓释、安全合规、道德透明。 ```mermaid flowchart TD accTitle: AI规模化面临的两大核心挑战 subgraph CoreChallenges["企业AI规模化的两大核心阻碍"] A["价值实现与量化<br/>(ROI Value & Delivery)"] B["治理与风险控制<br/>(Governance & Safety)"] end subgraph GovernanceDetails["治理的关键维度"] B1["风险缓释 (Mitigate Risk)"] B2["安全可信 (Safe & Ethical)"] B3["全流程透明 (Transparency)"] end B --> B1 B --> B2 B --> B3 ``` - **阻碍规模化的第一大难题**：无法清晰理解、计算并在组织内交付 AI 带来的实际价值。 - **阻碍规模化的第二大难题**：涵盖安全性、道德准则、透明度及风险控制的全面 AI 治理。 --- ### [04:08](https://www.youtube.com/watch?v=aLauVgmLDg0&t=248s) 为何 AI 的 ROI 评估如此困难？ 传统商业教学中计算投资回报率往往非常简单明了——算出解决方案的总成本，预测何时能交付何种价值，明确项目上线后的核心考核指标即可。然而，AI 项目的价值交付通常具有“间接性”和“延迟性”，这使得传统的衡量方式频频失灵。企业内部不同业务部门对“成功”的定义大相径庭：有些部门坚持 AI 必须带来立竿见影的“硬性财务收益”（Hard Benefit），而另一些部门则认为带来生产力提升、成本规避（Cost Avoidance）或折算人员工时节省（Dollarized Time Savings）就已经足够。 与此同时，企业面临着越来越紧迫的“快速见效”压力。在当前的大环境下，“给我三年时间，我保证达到某种价值”的说辞已经不再被高管层接受。企业渴求以越来越快的节奏看到回报，因而会更加严苛地评估技术可行性、数据就绪度，并追问“在最短时间内能实现最大价值的切入点到底在哪里”。 此外，“数据孤岛与项目孤岛”是另一个巨大阻碍。许多团队对技术前景感到兴奋，但面对分散在各处的孤岛数据或尚未就绪的数据状态，往往陷入困惑：如果必须花大量时间先去治理和拉通底层数据，达成预期价值的时间周期就会被极度拉长。同时，规模化落地项目（Scaled Projects）与前期的 PoC/实验性项目（Experiments）在 ROI 的衡量维度上有着本质的不同。 - **衡量困难的根源**：AI 价值具有间接性与滞后性，难以用单一度量衡一概而论。 - **成功定义不统一**：业务部门在“硬性财务收益”与“生产力/工时节省”等软性指标之间缺乏共识。 - **快速见效的压力骤增**：多周期长期承诺不再受青睐，企业优先寻找技术可行、数据就绪且短期回报最高的场景。 - **数据就绪度瓶颈**：数据孤岛和治理滞后严重拖慢了从技术到价值转化的整体周期。 --- ### [06:28](https://www.youtube.com/watch?v=aLauVgmLDg0&t=388s) 企业 AI 基础能力与“AI 优先”思维的现状 针对企业在实现 ROI 方面的现状指标，最新调研数据展现了行业的分化与趋势：目前仅有 27% 的企业表示自身已经建立了相当成熟、稳健的 AI 基础能力（AI Foundational Capabilities）；然而在那些积极采纳“AI 优先（AI-first）”思维方式——即在面对任何业务问题时，都将 AI 技术作为首要考量和前沿解决手段的企业中，这一比例和推进速度正在发生明显变化。 - **基础能力建设现状**：仅有 27% 的组织认为自己拥有较为健全的 AI 基础底座能力。 - **思维模式转变**：68% 的企业正在积极引入或转向“AI 优先（AI-first）”的业务解决思路。 ### [06:58](https://www.youtube.com/watch?v=aLauVgmLDg0&t=418s) 组织架构与治理：从去中心化走向“轴辐模式”（Hub and Spoke） 在探索前沿 AI 落地时，领先组织展现出了非常明确的治理共识与战略定力。数据表明，68% 的高绩效企业在内部建立了扎实的数据治理体系与合规策略；同时，同样有 68% 的企业将当前开展的各种 AI 实验视为一个“创新组合”（Innovation Portfolio），系统性地在内部孵化这些创新，并清晰界定它们能为企业创造的具体商业价值。 回溯 2022 至 2023 年 ChatGPT 爆发初期，市场上涌现出大量单点式、偏消费级的 AI 工具。当时企业最核心的目标是“激发内部的好奇心社区”（spur a community of curiosity），鼓励全员动手实验新技术。这促使多数企业在初期采用了完全去中心化的模式，允许各个业务单元（BU）自由探索贴合自身业务的孤立用例。然而实践表明，单纯的去中心化模式所产生的投资回报率（ROI）和综合价值明显偏低。真正能拉开差距的是集中化（Centralized）或“轴辐模式”（Hub and Spoke）：由中央核心部门统一价值衡量标准、树立标杆实践，再将经验与基础设施赋能给各个业务分支，让业务部门在保留自主性的同时共享最佳实践。AI 的成功绝不仅仅取决于选几个用例写写算算，而是关乎人员、流程、治理与组织架构的深层演进。 ```mermaid flowchart TD accTitle: AI 组织运营模型对比 accDescription: 展示去中心化模式向轴辐模式演进并带来更高 ROI 的架构图 subgraph HubSpoke["轴辐模式（Hub & Spoke / 集中化赋能）- 高 ROI"] Hub["中央 AI 赋能中心 (Hub)<br/>• 统一数据治理与价值度量<br/>• 沉淀最佳实践与通用平台"] SpokeA["业务单元 A (Spoke)<br/>自主探索落地"] SpokeB["业务单元 B (Spoke)<br/>自主探索落地"] SpokeC["业务单元 C (Spoke)<br/>自主探索落地"] Hub --> SpokeA Hub --> SpokeB Hub --> SpokeC end subgraph Decentralized["早期去中心化模式 - 低 ROI"] BU1["孤立业务单元 1"] BU2["孤立业务单元 2"] BU3["孤立业务单元 3"] end ``` - **数据治理基准**：68% 的前沿实践机构拥有严密的数据治理策略，并将试验项目作为创新组合进行统一孵化与价值追踪。 - **模式优劣对比**：早期纯去中心化的自由实验虽然激活了组织好奇心，但在规模化交付 ROI 方面显著落后于轴辐模式。 - **四大关键支柱**：AI 落地不是单纯的技术工具采购，必须涵盖技术、人员（People）、流程（Process）与治理（Governance）。 --- ### [09:01](https://www.youtube.com/watch?v=aLauVgmLDg0&t=541s) IBM 内部实践：从削减 SG&A 到“实验室技术自身先用” 作为一家拥有 28.6 万名员工的百年科技巨头，IBM 历经了多次自我重塑。当 Arvind Krishna 接任 CEO 时，IBM 的领导层结构迎来了重大转变——与以往多位具有咨询背景的最高掌舵人不同，Krishna 是一位出身于研发与工程部门的纯正技术专家，同时具备敏锐的商业嗅觉。上任伊始，Krishna 审视了公司的资产负债表与运营结构，直指销售、一般及行政费用（SG&A）过高的问题。 在行业基准中，表现优秀的组织通常将 SG&A 支出控制在 5% 至 7%，顶尖组织甚至能降至 2% 至 3%，而当时的 IBM 这一比例高达 10% 到 11%。面对成本痛点，Krishna 提出了一套清晰直接的执行框架：IBM 实验室里研发的 AI 与自动化工具，必须首先在自身内部全面部署。通过将自身作为“第一客户”（Client Zero），IBM 借助自己开发的人工智能与自动化技术在内部实现大幅降本增效，并在把产品推向外部客户和合作伙伴之前，在自身 28.6 万人的真实复杂业务环境中完成全方位打磨与验证。 - **员工规模与背景**：IBM 拥有约 286,000 名员工，CEO Arvind Krishna 以研发与工程管理者的技术背景执掌公司。 - **SG&A 成本差距**：行业标杆企业 SG&A 占比为 5%–7%（极少数达到 2%–3%），IBM 当时高达 10%–11%，存在巨大的效率优化空间。 - **内部落地法则**："吃自己的狗粮"（Dogfooding），将自家 AI/自动化产品深度应用到企业内部运营以削减成本、提升生产力，验证成熟后再向全球客户推广。 ### [10:26](https://www.youtube.com/watch?v=aLauVgmLDg0&t=626s) IBM 的“零号客户”执行框架与 45 亿美元生产力回报 IBM 内部建立了所谓的“零号客户”（Client Zero）执行框架，在过去两年半的时间里，通过在内部全面部署 AI，已为公司自身带来了高达 45 亿美元的生产力价值回报。在这 45 亿美元的总收益中，约有 30 亿美元属于直接从账本中扣减的实际硬性成本节约，或是通过部署 AI 用例直接创造的新增业务收入；其余约 15 亿美元则属于成本规避（Cost Avoidance）或所谓的“工时金额化节约”（Dollarized Time Savings）。例如，如果员工 Alex 的时薪成本是 100 美元，AI 帮他节省了 100 个小时，企业面临的核心课题就是如何让 Alex 将被释放出来的这 100 个小时投入到其他高价值工作中去。 这一庞大体系的落地离不开高层领导力与全员自下而上创新的紧密结合。在治理层面，IBM 设立了由 CEO Arvind Krishna 亲自挂帅的指导委员会，Arvind 每季度都会抽出整整两天时间，极其深入地审查各业务部门部署的 AI 用例细节。与此同时，跨职能委员会负责广泛评估来自生态系统内部的各种构想。IBM 还举办了“watsonx 挑战赛”（watsonx Challenge），动员了超过 20 万名员工亲自动手使用自家技术构建 AI 能力，以此从基层充分汲取创意并注入执行框架中，全面推动组织生产力的提升。 - **45 亿美元生产力总回报构成**：30 亿美元为实际账面成本削减与直接创收，15 亿美元为成本规避与工时金额化收益。 - **高管深度参与**：CEO Arvind Krishna 每季度投入两天时间逐一深度复盘具体用例的细节与落地表现。 - **全员实践创新**：watsonx 挑战赛吸引了超过 20 万名员工参与实操构建，形成自下而上的创意征集机制。 ```mermaid flowchart TD accTitle: IBM Client Zero AI ROI Framework subgraph Governance["治理与创意驱动"] A["CEO 指导委员会<br/>(每季度深入复盘 2 天)"] --> C["跨职能评审委员会"] B["watsonx 挑战赛<br/>(20万+ 员工基层创新)"] --> C end subgraph Assessment["严格准入评估"] C --> D{"用例两周内定案<br/>且 1 年内见 ROI？"} D -- 否 --> E["果断剔除 / 转向其他项目"] D -- 是 --> F["投入部署与规模化"] end subgraph Value["价值交付 (45亿美元)"] F --> G["硬性收益 (~30亿美元)<br/>账面减本 / 业务增收"] F --> H["软性收益 (~15亿美元)<br/>成本规避 / 工时金额化"] end ``` --- ### [11:39](https://www.youtube.com/watch?v=aLauVgmLDg0&t=699s) 严苛的用例准入规则与 AI 价值定义的四大维度 在推进 AI 项目时，最关键的前期工作是在全公司统一对“价值”的衡量标准。Arvind Krishna 自上而下给团队下达了“两年内实现 30 亿美元价值”的硬性目标，这一巨大压力倒逼团队必须在价值定义上迅速达成共识。IBM 确立了极其严苛的业务用例立项原则：任何内部提交的 AI 商业计划书，必须明确总体拥有成本（TCO），且必须在一年以内实现投资回报（ROI）；如果一个构想无法在两周内完成清晰的商业案例论证，就会被毫不留情地立即剔除。这并不意味着该构想本身是个坏点子，而是在当时的目标约束下，它并不是当下达成业绩指标的最佳选择。 ``` ┌───────────────────────────────────────────────┐ │ IBM AI 价值评估的四大核心维度 │ └──────────────────────┬────────────────────────┘ │ ┌──────────────────────────┴──────────────────────────┐ ▼ ▼ ┌───────────────────────┐ ┌───────────────────────┐ │ 硬性收益 │ │ 软性收益 │ │ (Hard Benefits) │ │ (Soft Benefits) │ ├───────────────────────┤ ├───────────────────────┤ │ 1. 营业收入 / 税前利润│ │ 3. 成本规避 │ │ 通过 AI 与自动化 │ │ 治理影子 IT，整合 │ │ 直接带来业务增长 │ │ 冗余供应商与系统 │ │ │ │ │ │ 2. 账面成本直接削减 │ │ 4. 生产力与工时节约 │ │ 直接从财务账目中 │ │ 员工工时节约并重投 │ │ 剔除实际支出项目 │ │ 到更高价值业务中 │ └───────────────────────┘ └───────────────────────┘ ``` 根据 IBM 与众多企业的交流经验，企业级 AI 的价值评估体系通常落在两大“硬性收益”与两大“软性收益”构成的四大维度中： 1. **业务增收与税前利润增加（硬性收益）**：直接通过引入 AI 和自动化能力拉动业务增长、创造增量营收。 2. **账本成本削减（硬性收益）**：借助技术手段直接移除或大幅降低原有的运营开支，真实体现在财务账目的成本下降上。 3. **成本规避（软性收益）**：针对内部系统冗余和“影子 IT”进行清理。例如，当发现公司内部竟存在 37 种不同的仪表盘工具时，团队必须反思原因——这些往往是员工在特定场景下各自采购的产物，虽出于善意却无法在全组织规模化扩展。通过整合战略供应商合同，“榨干现有资产的每一滴价值”，从而优化运营架构并规避未来潜在的重复投入。 4. **生产力节约（软性收益）**：优化日常流程，将释放出的工时转化为等价的生产力资本并重新赋能核心业务。 ### [13:48](https://www.youtube.com/watch?v=aLauVgmLDg0&t=828s) IBM 内部的 AI 治理与规模化成果：降本与体验提升的双赢 为了真正衡量生产力节约以及把时间节约转化为具体的美元价值，IBM 建立了一套集中的跟踪体系与宏观的 AI 治理机制。这套机制不仅用于系统化管理用例目录（use case catalog），还在投产前后持续评估风险与交付价值。依托这套严密的评估体系，IBM 在内部全方位推进生成式与智能体 AI（Agentic AI）的落地，至今已在人力资源（HR）、IT、销售、营销、采购、财务和供应链等各大业务部门上线了超过 100 个生成式与智能体 AI 用例。 在这些业务板块中，IBM 的 HR 部门在 AI 应用的成熟度上表现尤为突出。目前 HR 领域已正式投产超过 34 个生成式 AI 用例，覆盖福利与薪酬、薪资核算、人才发展以及企业培训等核心环节。最令人瞩目的成效在于，它不仅帮助首席人力资源官（CHRO）削减了 40% 的净运营预算，还将员工体验评分（Employee Experience Score）逆势提升了 55 分以上。在传统企业管理认知中，大幅压缩预算与显著改善体验往往是不可兼得的矛盾指标，但借助深度集成的 AI 解决方案，IBM 实现了成本与满意度的双向突破。 在 IT 运营支持领域，IBM 同样展示了极高的敏捷交付速度。例如内部代号为 "AskIT" 的项目，团队从零开始构建仅耗时 100 天，上线后便直接分流并消减了 50% 的入站员工 IT 服务工单。考虑到在整体支持工作流中，每处理一张工单的平均成本约为 18 美元，这一系统为企业运营带来了极为直观且可量化的资金回报。 ```mermaid flowchart TD accTitle: IBM 内部 AI 治理与价值转化流程 subgraph 治理与评估架构 A["中央跟踪与 AI 治理机制"] --> B["用例目录管理与风险审查"] B --> C["全口径成本与价值量化模型"] end subgraph 业务落地与成效 C --> D["HR 业务场景 (34+ 投产用例)"] C --> E["IT 支撑场景 (AskIT 平台)"] D --> F["运营预算削减 40%<br/>员工体验分提升 55+ 分"] E --> G["100 天快速落地上线<br/>工单减少 50% ($18/单)"] end ``` - **全域覆盖**：在 HR、IT、销售、营销、采购、财务及供应链等环节落地超 100 个生成式及智能体 AI 用例。 - **HR 业务突破**：部署 34 个生成式 AI 用例，实现 CHRO 运营预算降低 40%，员工体验评分提升超过 55 分。 - **AskIT 落地成效**：仅用 100 天完成全流程开发，拦截消除 50% 的 IT 员工支持工单，按每单约 18 美元的综合工单成本直接变现投资回报。 --- ### [15:28](https://www.youtube.com/watch?v=aLauVgmLDg0&t=928s) 从内部实践到预构建智能体：AI 价值评估组织的赋能 衡量 AI 项目时，企业必须立足现状并通盘考量全口径成本（All-in Cost），这不仅包括直接的技术采购支出，更涵盖在特定用例上投入的人力成本与资源调配，最后再将这些要素精确映射至核心价值指标中。为了帮助外部客户避免重复走弯路，IBM 将内部验证成熟的场景沉淀并打包为即开即用的预构建智能体或智能体模板（Pre-built Agents / Agentic Templates）。这些工具绝非单一任务的简单脚本，而是针对整条端到端业务流设计的综合方案。 以人才招聘（Talent Acquisition）这一宏大业务链条为例，其中穿插了多项密集的 AI 任务：撰写职位描述（JD）、多渠道搜寻候选人、辅助面试执行以及自动化总结面试反馈等。IBM 将整套经过真实业务场景打磨的工作流预先固化到智能体中，把海量的训练成果与领域专业知识直接赋予企业客户。通过这种方式，客户无需再从头探索，能够以比 IBM 当初更快的速度兑现投资回报。 为帮助企业建立这种端到端的落地信心，Madison 在其团队内部专门组建了“AI 价值评估组织”（AI Value Assessment Organization）。该团队的核心职责是为客户定制专属的“价值图谱”（Value Map）与价值视角分析。在推进 AI 规模化转型的过程中，业务领域的专家与技术架构人员常常存在严重的认知鸿沟；而价值图谱的作用正是打破这种壁垒，让懂业务的人与懂技术的人在同一个语境下达成共识，清晰看到项目在降本增效上究竟能挖掘出多大的潜能。 - **全口径成本核算**：全面评估技术支出、人力劳动资源与各业务环节的实际消耗，精准对齐四大价值要素。 - **预构建智能体模板**：将成熟经验产品化，覆盖如人才招聘（JD 编写、人才搜寻、面试提问、结果摘要）等复杂完整的工作流。 - **AI 价值评估组织**：通过产出定制化的价值图谱（Value Map），消除领域专家与技术专业人员之间的隔阂，统一组织对 AI 潜在回报的预期。 ### [17:08](https://www.youtube.com/watch?v=aLauVgmLDg0&t=1028s) 明确业务价值驱动因素与动态价值评估框架 企业在特定业务领域（如人力资源 HR）应用 AI 时，首先需要清晰拆解价值驱动因素（Value Drivers），明确应用 AI 能够带来的具体收益描述，并建立标准的收益计算方式。对跨部门、跨用例的员工进行相关测算方法的技能培训至关重要。尽管将这些指标落实在纸面上看似简单，但在许多大型企业所面临的规模下，将这些框架清晰展示给各类利益相关者并达成组织内部的共识，是推动 AI 项目落地的前提。 在实践中，IBM 会与客户共同开展深度的“动态价值评估”（Dynamic Value Assessment）。以人才招聘与吸纳（Talent Acquisition）流程为例，评估框架会从企业的组织规模和年度新员工入职率入手，细化评估无论是外包招聘还是内部入职工作流中每个环节所耗费的时间和成本，最终测算出 AI 所能带来的具体生产力提升空间。 ```mermaid flowchart LR accTitle: 人力资源 AI 价值评估与量化路径 subgraph S1["1. 价值驱动与基准"] A["明确业务价值驱动因素"] --> B["获取基础数据 (组织规模 / 年招聘率)"] end subgraph S2["2. 成本与流程测算"] B --> C["拆解招聘/入职各环节耗时与成本"] end subgraph S3["3. 收益转化与落地"] C --> D["测算生产力收益与 ROI"] D --> E["组织与人才转型赋能"] end ``` --- ### [18:33](https://www.youtube.com/watch?v=aLauVgmLDg0&t=1113s) IBM 自身实践：呼叫中心转型与知识沉淀 谈到生产力提升与人员转型，IBM 分享了一个自身的标杆案例：在推出员工支持工具（Employee Support Tool）之前，IBM 有一个 750 人的内部 HR 呼叫中心。随着 AI 系统的上线，其中 675 名原本负责接听电话的员工不再需要从事常规接线工作。 IBM 并未直接裁撤这些人员，而是意识到这批员工才是真正精通 HR 业务流程、了解企业内部实际运转逻辑的核心人才。IBM 专门成立了一个数据科学与 AI 部门（Department of Data Science and AI），并将这 675 名业务专家吸纳其中。通过将他们的业务智慧与行业知识直接“灌注”（plumbing）到 AI 系统中，IBM 成功构建了可复用的企业级智能化流程，实现了从传统运营向高附加值 AI 建设的跨越。 - **人员结构优化**：750 名呼叫中心人员中，675 人成功转岗至数据科学与 AI 部门。 - **业务智慧复用**：将一线员工掌握的业务诀窍沉淀为企业级 AI 资产，打造可复用的流程体系。 --- ### [19:07](https://www.youtube.com/watch?v=aLauVgmLDg0&t=1147s) 标准化价值评估四步法与架构考量 为了在企业内部及与客户合作时系统化地量化 AI 价值，IBM 建立了标准化的四步价值评估流程，并建议大型企业在探索 AI 用例时全面推广这种全员参与的规范化机制： ```mermaid flowchart LR accTitle: IBM 四步价值评估流程 A["1. 评估范围界定<br/>(Assessment Scoping)"] --> B["2. 调研访谈与工作坊<br/>(Discovery & Workshops)"] B --> C["3. 综合归纳与价值建模<br/>(Synthesis & Value Modeling)"] C --> D["4. 价值案例成果汇报<br/>(Value Case Readout)"] ``` 1. **评估范围界定（Assessment Scoping）**：根据特定业务领域或用例梳理价值驱动因素与收益图谱。 2. **调研访谈与工作坊（Discovery Interviews and Workshops）**：深入业务一线，挖掘各业务部门未被察觉的痛点与真实需求。 3. **综合归纳与价值建模（Synthesis and Value Modeling）**：对收集到的业务数据进行量化分析与动态价值建模。 4. **价值案例成果汇报（Value Case Readout）**：向管理层与利益相关方输出清晰的投资回报业务用例。 最后，衡量 AI 的投资回报率不能仅仅局限于用例本身、人员规模或节省的工时，底层的技术选型与架构设计（Architectural Choices）同样直接决定了项目的长期 ROI 与运行成本。 ### [20:23](https://www.youtube.com/watch?v=aLauVgmLDg0&t=1223s) 电信客户实战案例：从小语言模型中获取 10 倍成本降幅与规模化 ROI Madison 指着展示图表中的粉色条形柱解释道，这些柱状图代表了 IBM 内部以及各行业客户在不同业务场景中，最初采用超大模型（Large Models）来解决问题的情况。她特别引导观众关注最右侧的一个真实电信运营商案例。该电信客户曾与 IBM 深入探讨过技术平台以及如何用自身专属数据微调和适配小语言模型（SLM），但最初仍选择在通用大模型上展开尝试。 ```mermaid flowchart TD accTitle: 电信客户转用小语言模型架构优化流程 subgraph ScaleChallenge ["规模化瓶颈 (GPT 大模型)"] A["处理 300 万通通话记录摘要"] --> B["准确率达 97%"] B --> C["目标：扩展至 3.3 亿通记录"] C --> D["推理与基础设施成本激增，预算严重超支"] end subgraph OptimizedSolution ["IBM 小模型优化方案"] D --> E["采用专属数据适配的小语言模型 (SLM)"] E --> F["维持 97% 相同准确率"] F --> G["推理成本直降 10 倍 (10x Cost Reduction)"] end ``` 大约一年前，该电信公司的负责人联系 Madison 坦言自己遇到了重大危机。他们开发了一个客户服务（Customer Care）场景的摘要应用，利用 GPT 模型成功对 300 万条客户通话记录进行了摘要，模型表现相当不错，准确率高达 97%。然而灾难在于，该公司全年的实际需求是处理 3.3 亿条通话记录，而仅仅在处理完这 300 万条时，整个项目的预算就已经被彻底耗尽。他们迫切需要一种在推理（Inference）和基础设施占用上成本显著更低的小模型，以便将规模真正放大到全公司预期的体量。IBM 随后介入协助，采用适配的小语言模型达成了完全相同的 97% 准确率，同时直接将推理成本削减了整整 10 倍（"reduced his cost by 10x"）。Madison 强调，早期做出的模型架构选择至关重要，它直接决定了业务场景能否走出单纯的实验并真正实现规模化 ROI。 - **试点规模与全量需求脱节**：仅处理了 300 万条通话记录预算即告告罄，而全量业务规模高达 3.3 亿条。 - **准确率与成本双赢**：转向专用小语言模型（SLM）后，在保持 97% 准确率不变的前提下，实现了 10 倍的成本降幅。 --- ### [21:54](https://www.youtube.com/watch?v=aLauVgmLDg0&t=1314s) AI 治理的双重价值与 UFC 体育娱乐落地闭幕 在演讲的收尾环节，Madison 纠正了业界关于 AI 治理（AI Governance）的普遍误区。大多数人往往将治理单纯视为规避风险与合规底线的防御性手段，认为它充其量能带来一些成本避免或缩减。但实际上，行之有效的 AI 治理能够直接赋能业务增长——它能为企业引入更多 AI 模型与 AI Agent 提供合规保障与流水线加速能力，从而大幅提升企业整体的生产力与营收。 Madison 随后分享了 IBM 与 UFC（终极格斗冠军赛）的合作案例作为佐证。UFC 利用双方联合打造的见解引擎（Insight Engine），能够以极快的速度自动生成深度赛事分析与洞察，并将其直接整合到每年举办的 40 多场全球直播赛事的内容流中。这证明即使在体育、媒体与娱乐等极具观赏性与趣味性的行业，AI 同样具备扎实可量化的业务价值。最后，Madison 热情邀请与会者前往 IBM 展台参观互动演示，与技术专家深入探讨如何在各自机构内部落地这些架构与价值评估方法。 - **AI 治理的真实红利**：不仅限于合规与风控，更能加速 AI Agent 及多样化模型的部署流程，驱动收入增长与生产力提升。 - **UFC 赛事洞察引擎**：为每年 40 多场大型现场直播赛事提供快速生成的高价值内容与数据洞察。 ## Notable quotes > [02:13](https://www.youtube.com/watch?v=aLauVgmLDg0&t=133s) 我们做了这么多实验，到底换来了什么价值？如何才能让这些技术在业务中真正落地并带来收益？ — Madison Gooch > [04:08](https://www.youtube.com/watch?v=aLauVgmLDg0&t=248s) 给我三年时间，我保证达到某种价值 — Madison Gooch > [06:58](https://www.youtube.com/watch?v=aLauVgmLDg0&t=418s) spur a community of curiosity — Madison Gooch > [09:01](https://www.youtube.com/watch?v=aLauVgmLDg0&t=541s) 吃自己的狗粮 — Madison Gooch > [11:39](https://www.youtube.com/watch?v=aLauVgmLDg0&t=699s) 两年内实现 30 亿美元价值 — Arvind Krishna > [11:39](https://www.youtube.com/watch?v=aLauVgmLDg0&t=699s) 榨干现有资产的每一滴价值 — Madison Gooch > [20:23](https://www.youtube.com/watch?v=aLauVgmLDg0&t=1223s) reduced his cost by 10x — Madison Gooch ## Mentioned Madison Gooch, Arvind Krishna, IBM, watsonx, IBM 商业价值研究院 (IBV), ChatGPT, AskIT, UFC, AI Value Assessment Organization ## Video description Madison Gooch, Vice President, watsonx, Americas, IBM Abstracts By 2027, 85% of CEOs expect investments in scaled AI initiatives to return a positive ROI, but currently, just 25% report that their programs have delivered as hoped. How do we close this gap and ensure responsible, productive and value-add AI is a part of your business? Join our session to hear: Factors to consider when measuring AI ROI, Common hurdles to achieving positive ROI on AI investments, How IBM looks at ROI for our AI implementations, Solutions to improve and optimize AI for delivering positive ROI. [Source](https://www.youtube.com/watch?v=aLauVgmLDg0)

[2:13] Madison Gooch: 我们做了这么多实验，到底换来了什么价值？如何才能让这些技术在业务中真正落地并带来收益？ [4:08] Madison Gooch: 给我三年时间，我保证达到某种价值 [6:58] Madison Gooch: spur a community of curiosity [9:01] Madison Gooch: 吃自己的狗粮 [11:39] Arvind Krishna: 两年内实现 30 亿美元价值 [11:39] Madison Gooch: 榨干现有资产的每一滴价值 [20:23] Madison Gooch: reduced his cost by 10x

By 2027, 85% of CEOs expect investments in scaled AI initiatives to return a positive ROI, but currently, just 25% report that their programs have delivered as hoped. [0:06] Madison Gooch: 25% of CEOs explicitly report that their AI initiatives are not delivering the level of value they had originally hoped for. [5:46] Madison Gooch: 27% of organizations report already having robust foundational AI capabilities established. [5:46] Madison Gooch: 68% of companies adopting an "AI-first mentality" have solid data governance practices and policies in place. [9:01] Madison Gooch: $4.5 billion total productivity value generated internally across two and a half years. [9:01] Madison Gooch: ~$3.0 billion in hard impact, consisting of direct ledger cost reductions and new revenue generated through deployed AI use cases. [9:01] Madison Gooch: ~$1.5 billion in cost avoidance, calculated as "dollarized time savings" [11:34] Madison Gooch: Every proposed internal AI use case was required to show a complete Total Cost of Ownership (TCO) and demonstrate a full ROI in less than one year. [14:08] Madison Gooch: The Chief Human Resources Officer (CHRO) achieved a 40% net budget reduction by automating core HR functions. [14:08] Madison Gooch: Despite deep operational spend cuts, employee satisfaction scores surged by over 55 points. [17:10] Madison Gooch: By deploying AI tools, they automated routine inquiry handling and freed up 675 of those employees. [20:03] Madison Gooch: The smaller architecture matched the 97% accuracy mark while cutting overall infrastructure and inference costs by 10x.

[0:06] So nice to be with you guys today. Um [0:06] So nice to be with you guys today. Um looking forward to diving in here a [0:07] looking forward to diving in here a [0:07] looking forward to diving in here a little bit for the next 20 minutes or so [0:10] little bit for the next 20 minutes or so [0:10] little bit for the next 20 minutes or so into a topic all about how we achieve uh [0:13] into a topic all about how we achieve uh [0:14] into a topic all about how we achieve uh an ROI right from all of the AI projects [0:16] an ROI right from all of the AI projects [0:16] an ROI right from all of the AI projects that many of you guys I'm certain are [0:18] that many of you guys I'm certain are [0:18] that many of you guys I'm certain are experimenting with and looking at within [0:19] experimenting with and looking at within [0:19] experimenting with and looking at within your organization today. So, three main [0:21] your organization today. So, three main [0:21] your organization today. So, three main takeaways um you know looking for you [0:23] takeaways um you know looking for you [0:23] takeaways um you know looking for you guys to have just we're going to share a [0:25] guys to have just we're going to share a [0:25] guys to have just we're going to share a couple of different industry insights [0:26] couple of different industry insights [0:26] couple of different industry insights and what we're actually seeing is [0:28] and what we're actually seeing is [0:28] and what we're actually seeing is patterns of deployment and types of ROI [0:30] patterns of deployment and types of ROI [0:30] patterns of deployment and types of ROI that are being achieved um throughout [0:33] that are being achieved um throughout [0:33] that are being achieved um throughout different clients within the industry. [0:35] different clients within the industry. [0:35] different clients within the industry. I'm going to spend most of the time on [0:37] I'm going to spend most of the time on [0:37] I'm going to spend most of the time on kind of segment two and three though and [0:39] kind of segment two and three though and [0:40] kind of segment two and three though and giving you guys some practical examples [0:41] giving you guys some practical examples [0:41] giving you guys some practical examples of how IBM has been able to achieve ROI [0:44] of how IBM has been able to achieve ROI [0:44] of how IBM has been able to achieve ROI and also how we are working with [0:46] and also how we are working with [0:46] and also how we are working with enterprises to really map what ROI from [0:49] enterprises to really map what ROI from [0:49] enterprises to really map what ROI from very specific projects and technologies [0:51] very specific projects and technologies [0:51] very specific projects and technologies look like and some of the key uh [0:53] look like and some of the key uh [0:53] look like and some of the key uh elements that I would encourage you to [0:54] elements that I would encourage you to [0:54] elements that I would encourage you to think about as you are mapping ROI [0:57] think about as you are mapping ROI [0:57] think about as you are mapping ROI inside of your own organization. [1:00] inside of your own organization. [1:00] inside of your own organization. So to get started, I I probably don't [1:02] So to get started, I I probably don't [1:02] So to get started, I I probably don't need to explain to anybody sitting in [1:04] need to explain to anybody sitting in [1:04] need to explain to anybody sitting in this room what AI is, but I do want to [1:06] this room what AI is, but I do want to [1:06] this room what AI is, but I do want to make a point. AI is a vast concept [1:09] make a point. AI is a vast concept [1:09] make a point. AI is a vast concept today. People could be talking about [1:10] today. People could be talking about [1:10] today. People could be talking about traditional technologies that are [1:12] traditional technologies that are [1:12] traditional technologies that are related to machine learning, deep [1:13] related to machine learning, deep [1:13] related to machine learning, deep learning, natural language processing. [1:15] learning, natural language processing. [1:15] learning, natural language processing. They could be talking about foundation [1:16] They could be talking about foundation [1:16] They could be talking about foundation models and LLMs. They could be talking [1:18] models and LLMs. They could be talking [1:18] models and LLMs. They could be talking about AI agents. And while all of those [1:20] about AI agents. And while all of those [1:20] about AI agents. And while all of those are encompassed under the umbrella, [1:23] are encompassed under the umbrella, [1:23] are encompassed under the umbrella, right, of core AI, it actually matters [1:26] right, of core AI, it actually matters [1:26] right, of core AI, it actually matters when we start to evaluate the ROI and [1:28] when we start to evaluate the ROI and [1:28] when we start to evaluate the ROI and the value that we get from this [1:30] the value that we get from this [1:30] the value that we get from this technology because the scope of how you [1:32] technology because the scope of how you [1:32] technology because the scope of how you use AI could sometimes change and be [1:35] use AI could sometimes change and be [1:35] use AI could sometimes change and be dependent upon how you measure the value [1:37] dependent upon how you measure the value [1:37] dependent upon how you measure the value of it. And we're going to talk a little [1:39] of it. And we're going to talk a little [1:39] of it. And we're going to talk a little bit about what that means and kind of a [1:41] bit about what that means and kind of a [1:41] bit about what that means and kind of a broader architecture framework for [1:43] broader architecture framework for [1:43] broader architecture framework for measuring value today. Okay. So, just as [1:47] measuring value today. Okay. So, just as [1:47] measuring value today. Okay. So, just as an example, as you see here below, [1:48] an example, as you see here below, [1:48] an example, as you see here below, right? Um, somebody that's running a [1:50] right? Um, somebody that's running a [1:50] right? Um, somebody that's running a pilot using an app that has embedded AI [1:53] pilot using an app that has embedded AI [1:53] pilot using an app that has embedded AI is going to have a very different ROI [1:55] is going to have a very different ROI [1:55] is going to have a very different ROI calculation and set of expectations than [1:58] calculation and set of expectations than [1:58] calculation and set of expectations than if you're rolling out a wide swath of [2:00] if you're rolling out a wide swath of [2:00] if you're rolling out a wide swath of these technologies uh across your [2:02] these technologies uh across your [2:02] these technologies uh across your enterprise in a cross departmental type [2:05] enterprise in a cross departmental type [2:05] enterprise in a cross departmental type way. Okay. So, we have to get down to [2:07] way. Okay. So, we have to get down to [2:07] way. Okay. So, we have to get down to the brass tax of how we actually [2:08] the brass tax of how we actually [2:08] the brass tax of how we actually calculate and understand those things [2:11] calculate and understand those things [2:11] calculate and understand those things inside of our organization. [2:13] inside of our organization. [2:13] inside of our organization. So what's happening in the market today? [2:15] So what's happening in the market today? [2:15] So what's happening in the market today? Right? 25% of CEOs say that many of [2:19] Right? 25% of CEOs say that many of [2:19] Right? 25% of CEOs say that many of these AI initiatives that they are [2:21] these AI initiatives that they are [2:21] these AI initiatives that they are running inside of their business are not [2:23] running inside of their business are not [2:23] running inside of their business are not eliciting the type of value that they [2:25] eliciting the type of value that they [2:25] eliciting the type of value that they hoped. I think all of us experienced [2:27] hoped. I think all of us experienced [2:27] hoped. I think all of us experienced right when the I call it the advent of [2:29] right when the I call it the advent of [2:29] right when the I call it the advent of chat GPT came out. There was so much [2:31] chat GPT came out. There was so much [2:31] chat GPT came out. There was so much promise and so much excitement about the [2:33] promise and so much excitement about the [2:33] promise and so much excitement about the potential of these technologies [2:35] potential of these technologies [2:35] potential of these technologies enterprisewide. [2:37] enterprisewide. [2:37] enterprisewide. And while I still think that there's a [2:38] And while I still think that there's a [2:38] And while I still think that there's a lot of excitement there, I think that [2:40] lot of excitement there, I think that [2:40] lot of excitement there, I think that there's many sea level executives that [2:41] there's many sea level executives that [2:41] there's many sea level executives that are now looking and saying,"All right, [2:43] are now looking and saying,"All right, [2:43] are now looking and saying,"All right, it's been a while. We've done some [2:44] it's been a while. We've done some [2:44] it's been a while. We've done some experimentation. We've gotten familiar [2:46] experimentation. We've gotten familiar [2:46] experimentation. We've gotten familiar with the technologies that are out [2:48] with the technologies that are out [2:48] with the technologies that are out there. Where is the value and how do we [2:50] there. Where is the value and how do we [2:50] there. Where is the value and how do we deliver it?" Right? Um how do we [2:52] deliver it?" Right? Um how do we [2:52] deliver it?" Right? Um how do we actually make this come to life inside [2:54] actually make this come to life inside [2:54] actually make this come to life inside of our organization? And so we still see [2:57] of our organization? And so we still see [2:57] of our organization? And so we still see a vast amount of experimentation [3:00] a vast amount of experimentation [3:00] a vast amount of experimentation happening today, but we see a much much [3:03] happening today, but we see a much much [3:03] happening today, but we see a much much smaller portion of those experiments [3:05] smaller portion of those experiments [3:05] smaller portion of those experiments actually making it into production. And [3:08] actually making it into production. And [3:08] actually making it into production. And really there are two main key drivers [3:10] really there are two main key drivers [3:10] really there are two main key drivers that CEOs would tell you. And we survey [3:13] that CEOs would tell you. And we survey [3:13] that CEOs would tell you. And we survey uh something called the Institute for [3:14] uh something called the Institute for [3:14] uh something called the Institute for Business Value at IBM. We survey about [3:16] Business Value at IBM. We survey about [3:16] Business Value at IBM. We survey about 5,000 different sea level executives [3:18] 5,000 different sea level executives [3:18] 5,000 different sea level executives every single year and different reports [3:20] every single year and different reports [3:20] every single year and different reports across generative AI. And there are two [3:23] across generative AI. And there are two [3:23] across generative AI. And there are two incredibly consistent themes yeartoyear [3:26] incredibly consistent themes yeartoyear [3:26] incredibly consistent themes yeartoyear that continue to come out from these sea [3:28] that continue to come out from these sea [3:28] that continue to come out from these sea levels that we survey. Number one is [3:31] levels that we survey. Number one is [3:31] levels that we survey. Number one is that achieving this type of value, how [3:33] that achieving this type of value, how [3:33] that achieving this type of value, how to understand it, how to calculate it, [3:35] to understand it, how to calculate it, [3:35] to understand it, how to calculate it, and how to deliver it among the [3:37] and how to deliver it among the [3:38] and how to deliver it among the organization is really the number one [3:39] organization is really the number one [3:39] organization is really the number one impediment to not scaling. And number [3:42] impediment to not scaling. And number [3:42] impediment to not scaling. And number two, there's lots of things wrapped up [3:44] two, there's lots of things wrapped up [3:44] two, there's lots of things wrapped up in a word that I would like to call [3:46] in a word that I would like to call [3:46] in a word that I would like to call governance. Right? So, how to mitigate [3:48] governance. Right? So, how to mitigate [3:48] governance. Right? So, how to mitigate risk, how to do this safely, how to do [3:50] risk, how to do this safely, how to do [3:50] risk, how to do this safely, how to do it ethically, how to do it transparently [3:53] it ethically, how to do it transparently [3:53] it ethically, how to do it transparently across the business when you know there [3:55] across the business when you know there [3:55] across the business when you know there are a vast amount of AI technologies [3:58] are a vast amount of AI technologies [3:58] are a vast amount of AI technologies likely being tested, experimented with, [4:00] likely being tested, experimented with, [4:00] likely being tested, experimented with, and used inside of an environment. Okay, [4:03] and used inside of an environment. Okay, [4:03] and used inside of an environment. Okay, those are the two key elements, ROI and [4:05] those are the two key elements, ROI and [4:05] those are the two key elements, ROI and governance. We'll talk about both today. [4:08] governance. We'll talk about both today. [4:08] governance. We'll talk about both today. So, why is measuring this concept of ROI [4:12] So, why is measuring this concept of ROI [4:12] So, why is measuring this concept of ROI so hard at times, right? Weren't we all [4:14] so hard at times, right? Weren't we all [4:14] so hard at times, right? Weren't we all probably taught in school or somewhere [4:16] probably taught in school or somewhere [4:16] probably taught in school or somewhere in uh you know our growings inside of an [4:19] in uh you know our growings inside of an [4:19] in uh you know our growings inside of an organization how to calculate a [4:21] organization how to calculate a [4:21] organization how to calculate a simplistic ROI like what's the total [4:23] simplistic ROI like what's the total [4:23] simplistic ROI like what's the total cost of one of these solutions? Uh what [4:25] cost of one of these solutions? Uh what [4:25] cost of one of these solutions? Uh what value are we going to deliver and when [4:27] value are we going to deliver and when [4:27] value are we going to deliver and when what's kind of the key objective that [4:29] what's kind of the key objective that [4:29] what's kind of the key objective that we're looking to um you know give back [4:31] we're looking to um you know give back [4:31] we're looking to um you know give back to the organization with one of these [4:33] to the organization with one of these [4:33] to the organization with one of these projects being deployed. I think we have [4:35] projects being deployed. I think we have [4:35] projects being deployed. I think we have to think about sometimes AI value is [4:37] to think about sometimes AI value is [4:37] to think about sometimes AI value is often indirect or delayed okay for a [4:40] often indirect or delayed okay for a [4:40] often indirect or delayed okay for a variety of different reasons. um [4:43] variety of different reasons. um [4:43] variety of different reasons. um business units within your organization [4:45] business units within your organization [4:45] business units within your organization may define success differently. People [4:47] may define success differently. People [4:48] may define success differently. People in your organization may have an idea [4:49] in your organization may have an idea [4:50] in your organization may have an idea that you know an AI experiment or an AI [4:52] that you know an AI experiment or an AI [4:52] that you know an AI experiment or an AI use case must deliver hard benefit while [4:54] use case must deliver hard benefit while [4:54] use case must deliver hard benefit while others may say things like productivity [4:57] others may say things like productivity [4:57] others may say things like productivity and cost avoidance or dollarized time [4:59] and cost avoidance or dollarized time [4:59] and cost avoidance or dollarized time savings uh of individuals in your [5:01] savings uh of individuals in your [5:01] savings uh of individuals in your organization may be good enough. So not [5:03] organization may be good enough. So not [5:03] organization may be good enough. So not having that clear definition of what [5:06] having that clear definition of what [5:06] having that clear definition of what value looks like or what that key [5:07] value looks like or what that key [5:07] value looks like or what that key definition of success is is challenging. [5:10] definition of success is is challenging. [5:10] definition of success is is challenging. um pressure to show ROI and value at a [5:13] um pressure to show ROI and value at a [5:13] um pressure to show ROI and value at a very early rate and pace uh is [5:16] very early rate and pace uh is [5:16] very early rate and pace uh is increasing. Okay, we see this increasing [5:18] increasing. Okay, we see this increasing [5:18] increasing. Okay, we see this increasing in enterprises. It's no longer good [5:20] in enterprises. It's no longer good [5:20] in enterprises. It's no longer good enough. If you say it will take me three [5:21] enough. If you say it will take me three [5:22] enough. If you say it will take me three years, but I promise I will achieve X [5:24] years, but I promise I will achieve X [5:24] years, but I promise I will achieve X level of value. People are looking for [5:26] level of value. People are looking for [5:26] level of value. People are looking for what can they do faster and faster and [5:28] what can they do faster and faster and [5:28] what can they do faster and faster and faster. They're looking at hey what is [5:30] faster. They're looking at hey what is [5:30] faster. They're looking at hey what is feasible from a technical perspective, [5:33] feasible from a technical perspective, [5:33] feasible from a technical perspective, from a data readiness perspective, and [5:35] from a data readiness perspective, and [5:35] from a data readiness perspective, and where can I achieve the highest value in [5:37] where can I achieve the highest value in [5:37] where can I achieve the highest value in the shortest amount of time right? So [5:39] the shortest amount of time right? So [5:39] the shortest amount of time right? So that pressure continues to rapidly [5:41] that pressure continues to rapidly [5:41] that pressure continues to rapidly increase. Um we think about data and [5:44] increase. Um we think about data and [5:44] increase. Um we think about data and project silos. This is kind of one uh [5:46] project silos. This is kind of one uh [5:46] project silos. This is kind of one uh very large impediment that I typically [5:48] very large impediment that I typically [5:48] very large impediment that I typically will see is people get really excited [5:50] will see is people get really excited [5:50] will see is people get really excited about these projects and the potential [5:51] about these projects and the potential [5:51] about these projects and the potential of this technology, but they say if my [5:53] of this technology, but they say if my [5:54] of this technology, but they say if my data is not in a state where it's truly [5:55] data is not in a state where it's truly [5:56] data is not in a state where it's truly ready or it's not understood or it sits [5:58] ready or it's not understood or it sits [5:58] ready or it's not understood or it sits in a whole bunch of different silos, can [6:00] in a whole bunch of different silos, can [6:00] in a whole bunch of different silos, can I really achieve this value? and and [6:02] I really achieve this value? and and [6:02] I really achieve this value? and and what's the elongated time period that's [6:04] what's the elongated time period that's [6:04] what's the elongated time period that's act that it's going to take me to fix [6:06] act that it's going to take me to fix [6:06] act that it's going to take me to fix that part of the equation in order to [6:09] that part of the equation in order to [6:09] that part of the equation in order to reach that level of value. And then [6:11] reach that level of value. And then [6:11] reach that level of value. And then you'll hear me talk about an example [6:13] you'll hear me talk about an example [6:13] you'll hear me talk about an example today of you know a scaled project and [6:17] today of you know a scaled project and [6:17] today of you know a scaled project and that ROI is looked at very differently [6:19] that ROI is looked at very differently [6:20] that ROI is looked at very differently than the experiments than we talk about. [6:22] than the experiments than we talk about. [6:22] than the experiments than we talk about. And this is uh this is an important [6:24] And this is uh this is an important [6:24] And this is uh this is an important notion that you'll hear me give kind of [6:25] notion that you'll hear me give kind of [6:25] notion that you'll hear me give kind of a real client example of. So if we think [6:28] a real client example of. So if we think [6:28] a real client example of. So if we think about some of the indicators of [6:30] about some of the indicators of [6:30] about some of the indicators of achieving ROI right let's talk uh about [6:34] achieving ROI right let's talk uh about [6:34] achieving ROI right let's talk uh about kind of what organizations are telling [6:36] kind of what organizations are telling [6:36] kind of what organizations are telling us today right 27% of organizations [6:39] us today right 27% of organizations [6:39] us today right 27% of organizations report that they already have fairly [6:41] report that they already have fairly [6:41] report that they already have fairly robust AI foundational capabilities in [6:44] robust AI foundational capabilities in [6:44] robust AI foundational capabilities in place 68% okay who are adopting what we [6:49] place 68% okay who are adopting what we [6:49] place 68% okay who are adopting what we call an AI first mentality meaning they [6:51] call an AI first mentality meaning they [6:51] call an AI first mentality meaning they would solve a problem thinking about AI [6:53] would solve a problem thinking about AI [6:54] would solve a problem thinking about AI technology being a very first and [6:55] technology being a very first and [6:55] technology being a very first and forward approach 68% of those entities [6:58] forward approach 68% of those entities [6:58] forward approach 68% of those entities have fairly solid data governance [7:01] have fairly solid data governance [7:01] have fairly solid data governance practices and policies inside of their [7:03] practices and policies inside of their [7:03] practices and policies inside of their organization. Okay. And 68% of these [7:06] organization. Okay. And 68% of these [7:06] organization. Okay. And 68% of these organizations also think about many of [7:08] organizations also think about many of [7:08] organizations also think about many of these experiments that are being run as [7:11] these experiments that are being run as [7:11] these experiments that are being run as their innovation portfolio. How do they [7:13] their innovation portfolio. How do they [7:13] their innovation portfolio. How do they incubate this innovation internally? [7:16] incubate this innovation internally? [7:16] incubate this innovation internally? Okay. And then how do we draw the line [7:18] Okay. And then how do we draw the line [7:18] Okay. And then how do we draw the line to what type of value that creates for [7:20] to what type of value that creates for [7:20] to what type of value that creates for an organization? Now when you know again [7:23] an organization? Now when you know again [7:24] an organization? Now when you know again if you follow me back to kind of the [7:25] if you follow me back to kind of the [7:25] if you follow me back to kind of the advent of chat GBT and what I'll call [7:27] advent of chat GBT and what I'll call [7:27] advent of chat GBT and what I'll call all of the singlethreaded relatively [7:29] all of the singlethreaded relatively [7:29] all of the singlethreaded relatively consumerfocused tools that are in the [7:32] consumerfocused tools that are in the [7:32] consumerfocused tools that are in the market today around AI. Um we want to [7:36] market today around AI. Um we want to [7:36] market today around AI. Um we want to actually spur a community of curiosity [7:38] actually spur a community of curiosity [7:38] actually spur a community of curiosity inside of our businesses and [7:40] inside of our businesses and [7:40] inside of our businesses and enterprises. You want people to be [7:42] enterprises. You want people to be [7:42] enterprises. You want people to be curious about the potential of these [7:44] curious about the potential of these [7:44] curious about the potential of these technologies and I would argue you want [7:45] technologies and I would argue you want [7:46] technologies and I would argue you want them to be hands-on okay in terms of [7:48] them to be hands-on okay in terms of [7:48] them to be hands-on okay in terms of their experimentation. And so we saw I [7:50] their experimentation. And so we saw I [7:50] their experimentation. And so we saw I would say back in the 2022 2023 time [7:53] would say back in the 2022 2023 time [7:53] would say back in the 2022 2023 time period lots of businesses and [7:55] period lots of businesses and [7:55] period lots of businesses and enterprises that adopted what was more [7:57] enterprises that adopted what was more [7:57] enterprises that adopted what was more of a decentralized approach, right? So [8:00] of a decentralized approach, right? So [8:00] of a decentralized approach, right? So how did business units kind of go off [8:02] how did business units kind of go off [8:02] how did business units kind of go off and experiment and things that made [8:04] and experiment and things that made [8:04] and experiment and things that made sense to them, solve use cases that made [8:06] sense to them, solve use cases that made [8:06] sense to them, solve use cases that made sense to their business? And what you [8:07] sense to their business? And what you [8:07] sense to their business? And what you actually will see on this chart here is [8:09] actually will see on this chart here is [8:09] actually will see on this chart here is a case for more of a centralized or more [8:13] a case for more of a centralized or more [8:13] a case for more of a centralized or more so a hub and spoke type model. Having [8:16] so a hub and spoke type model. Having [8:16] so a hub and spoke type model. Having some function of centralization to [8:18] some function of centralization to [8:18] some function of centralization to really bring all of this together to [8:20] really bring all of this together to [8:20] really bring all of this together to align what a value definition is to [8:23] align what a value definition is to [8:23] align what a value definition is to align what good looks like in the [8:25] align what good looks like in the [8:25] align what good looks like in the organization and propagate those [8:27] organization and propagate those [8:27] organization and propagate those learnings into kind of this spoke model [8:30] learnings into kind of this spoke model [8:30] learnings into kind of this spoke model where other people can take the best of [8:32] where other people can take the best of [8:32] where other people can take the best of what has happened within that [8:33] what has happened within that [8:33] what has happened within that centralized function but still have a [8:35] centralized function but still have a [8:35] centralized function but still have a level of autonomy of what they are able [8:37] level of autonomy of what they are able [8:37] level of autonomy of what they are able to do on their own. We now know that a [8:40] to do on their own. We now know that a [8:40] to do on their own. We now know that a decentralized approach, as you'll see [8:42] decentralized approach, as you'll see [8:42] decentralized approach, as you'll see here, does not produce actually as much [8:45] here, does not produce actually as much [8:45] here, does not produce actually as much ROI or value as these other models. And [8:48] ROI or value as these other models. And [8:48] ROI or value as these other models. And so, this isn't just about technology. [8:49] so, this isn't just about technology. [8:50] so, this isn't just about technology. It's not just about putting pencil to [8:51] It's not just about putting pencil to [8:51] It's not just about putting pencil to paper on a use case that you select. [8:54] paper on a use case that you select. [8:54] paper on a use case that you select. This is also about people and process [8:56] This is also about people and process [8:56] This is also about people and process and governance and how this evolves [8:58] and governance and how this evolves [8:58] and governance and how this evolves throughout your organization. [9:01] throughout your organization. [9:01] throughout your organization. So, I want to tell you a little bit [9:02] So, I want to tell you a little bit [9:02] So, I want to tell you a little bit about how IBM does this internally. So, [9:04] about how IBM does this internally. So, [9:04] about how IBM does this internally. So, IBM is a 286,000 person company. I would [9:07] IBM is a 286,000 person company. I would [9:08] IBM is a 286,000 person company. I would argue that we've reinvented ourselves uh [9:10] argue that we've reinvented ourselves uh [9:10] argue that we've reinvented ourselves uh several times. If I, you know, take you [9:12] several times. If I, you know, take you [9:12] several times. If I, you know, take you on a short journey with me here, we when [9:15] on a short journey with me here, we when [9:15] on a short journey with me here, we when Arvin Krishna came in and became our CEO [9:17] Arvin Krishna came in and became our CEO [9:17] Arvin Krishna came in and became our CEO several years ago, he was a researcher [9:20] several years ago, he was a researcher [9:20] several years ago, he was a researcher and he was in charge of our engineering [9:22] and he was in charge of our engineering [9:22] and he was in charge of our engineering division. That was what his trade was. [9:23] division. That was what his trade was. [9:23] division. That was what his trade was. And I would argue as a technologist, he [9:25] And I would argue as a technologist, he [9:25] And I would argue as a technologist, he also had a fairly keen eye uh of [9:29] also had a fairly keen eye uh of [9:29] also had a fairly keen eye uh of business acumen. Okay? and he came into [9:31] business acumen. Okay? and he came into [9:31] business acumen. Okay? and he came into IBM at the time. This was a very [9:33] IBM at the time. This was a very [9:33] IBM at the time. This was a very different CEO choice for us, by the way, [9:35] different CEO choice for us, by the way, [9:35] different CEO choice for us, by the way, than what we had had in years past. We [9:37] than what we had had in years past. We [9:37] than what we had had in years past. We had traditionally had people that had [9:39] had traditionally had people that had [9:39] had traditionally had people that had come from more of a consultative uh type [9:41] come from more of a consultative uh type [9:41] come from more of a consultative uh type state. And when Arvin came in, he said, [9:44] state. And when Arvin came in, he said, [9:44] state. And when Arvin came in, he said, you know, um looking at kind of our [9:46] you know, um looking at kind of our [9:46] you know, um looking at kind of our balance sheet here, we're probably not [9:48] balance sheet here, we're probably not [9:48] balance sheet here, we're probably not in the best position of what IBM's [9:49] in the best position of what IBM's [9:49] in the best position of what IBM's spending on things like SGNA. Like [9:52] spending on things like SGNA. Like [9:52] spending on things like SGNA. Like commendable organizations spend in the 5 [9:54] commendable organizations spend in the 5 [9:54] commendable organizations spend in the 5 to 7% range. Spectacular organizations [9:56] to 7% range. Spectacular organizations [9:56] to 7% range. Spectacular organizations spend in the two to three. we're [9:58] spend in the two to three. we're [9:58] spend in the two to three. we're probably at like 10 to 11%. So we said [10:01] probably at like 10 to 11%. So we said [10:01] probably at like 10 to 11%. So we said this shouldn't be a novel idea but like [10:02] this shouldn't be a novel idea but like [10:02] this shouldn't be a novel idea but like we're going to set up an execution [10:04] we're going to set up an execution [10:04] we're going to set up an execution framework and we're going to use the [10:05] framework and we're going to use the [10:05] framework and we're going to use the very things that we create inside of our [10:07] very things that we create inside of our [10:07] very things that we create inside of our labs i.e. AI and automation to deliver a [10:11] labs i.e. AI and automation to deliver a [10:11] labs i.e. AI and automation to deliver a robust amount of productivity inside of [10:13] robust amount of productivity inside of [10:13] robust amount of productivity inside of our own business. We're going to take [10:15] our own business. We're going to take [10:16] our own business. We're going to take cost out through doing it and we're [10:17] cost out through doing it and we're [10:17] cost out through doing it and we're going to use every tool that we create [10:19] going to use every tool that we create [10:19] going to use every tool that we create before we propagate those out to other [10:21] before we propagate those out to other [10:21] before we propagate those out to other clients and partners for use. Okay. So [10:24] clients and partners for use. Okay. So [10:24] clients and partners for use. Okay. So we set up what you see here kind of we [10:26] we set up what you see here kind of we [10:26] we set up what you see here kind of we call this our client zero execution [10:28] call this our client zero execution [10:28] call this our client zero execution framework and IBM has now delivered $4.5 [10:31] framework and IBM has now delivered $4.5 [10:31] framework and IBM has now delivered $4.5 billion dollars of productivity back to [10:33] billion dollars of productivity back to [10:33] billion dollars of productivity back to our organization across the last two and [10:35] our organization across the last two and [10:35] our organization across the last two and a half years. About three billion of [10:38] a half years. About three billion of [10:38] a half years. About three billion of that is within the realm of actual cost [10:41] that is within the realm of actual cost [10:41] that is within the realm of actual cost taken out of the ledger or revenue that [10:43] taken out of the ledger or revenue that [10:43] taken out of the ledger or revenue that has been generated through the AI use [10:45] has been generated through the AI use [10:45] has been generated through the AI use cases that we've deployed. About a [10:47] cases that we've deployed. About a [10:47] cases that we've deployed. About a billion and a half dollars of that is in [10:49] billion and a half dollars of that is in [10:49] billion and a half dollars of that is in the realm of cost avoidance or what we [10:51] the realm of cost avoidance or what we [10:51] the realm of cost avoidance or what we would call dollarized time savings. So, [10:53] would call dollarized time savings. So, [10:53] would call dollarized time savings. So, if Alex cost me a hundred bucks uh an [10:56] if Alex cost me a hundred bucks uh an [10:56] if Alex cost me a hundred bucks uh an hour and I save him a h 100red hours, [10:57] hour and I save him a h 100red hours, [10:58] hour and I save him a h 100red hours, how do I get Alex to do something else [10:59] how do I get Alex to do something else [10:59] how do I get Alex to do something else with the hundred hours that I have given [11:01] with the hundred hours that I have given [11:01] with the hundred hours that I have given back to him? And you'll see on the lefth [11:04] back to him? And you'll see on the lefth [11:04] back to him? And you'll see on the lefth hand side, we had a steering committee [11:06] hand side, we had a steering committee [11:06] hand side, we had a steering committee that looked at this. This was Arvin [11:08] that looked at this. This was Arvin [11:08] that looked at this. This was Arvin himself and our CEO spending about two [11:10] himself and our CEO spending about two [11:10] himself and our CEO spending about two days every quarter getting very deep [11:13] days every quarter getting very deep [11:13] days every quarter getting very deep within the detail of the AI use cases [11:15] within the detail of the AI use cases [11:15] within the detail of the AI use cases that we were deploying a crossf [11:17] that we were deploying a crossf [11:17] that we were deploying a crossf functional committee to look at ideas [11:18] functional committee to look at ideas [11:18] functional committee to look at ideas that were being served up inside of the [11:20] that were being served up inside of the [11:20] that were being served up inside of the IBM ecosystem. We had a Watson X [11:23] IBM ecosystem. We had a Watson X [11:23] IBM ecosystem. We had a Watson X challenge where we got over 200,000 [11:25] challenge where we got over 200,000 [11:25] challenge where we got over 200,000 people involved in building hands-on [11:27] people involved in building hands-on [11:27] people involved in building hands-on capabilities with our technology and we [11:29] capabilities with our technology and we [11:30] capabilities with our technology and we solicited ideas from the grassroots of [11:32] solicited ideas from the grassroots of [11:32] solicited ideas from the grassroots of IBM to bring into this framework to [11:34] IBM to bring into this framework to [11:34] IBM to bring into this framework to figure out how we could become a more [11:37] figure out how we could become a more [11:37] figure out how we could become a more productive organization. Okay. So you [11:39] productive organization. Okay. So you [11:39] productive organization. Okay. So you see a little bit of the execution [11:40] see a little bit of the execution [11:40] see a little bit of the execution framework on the side but measuring [11:42] framework on the side but measuring [11:42] framework on the side but measuring value was one of the key things that we [11:44] value was one of the key things that we [11:44] value was one of the key things that we did up front. We said everybody might [11:47] did up front. We said everybody might [11:47] did up front. We said everybody might not fully agree at the onset, but we [11:50] not fully agree at the onset, but we [11:50] not fully agree at the onset, but we have to align what IBM's value [11:52] have to align what IBM's value [11:52] have to align what IBM's value definition is for this exercise because [11:54] definition is for this exercise because [11:54] definition is for this exercise because again, we had a big target, right? We [11:56] again, we had a big target, right? We [11:56] again, we had a big target, right? We had to deliver $3 billion within a [11:59] had to deliver $3 billion within a [11:59] had to deliver $3 billion within a two-year period. That was the target [12:01] two-year period. That was the target [12:01] two-year period. That was the target that Arvin gave from the top down. And I [12:04] that Arvin gave from the top down. And I [12:04] that Arvin gave from the top down. And I would argue that really pushed us [12:05] would argue that really pushed us [12:05] would argue that really pushed us forward into aligning this definition [12:07] forward into aligning this definition [12:07] forward into aligning this definition that you see here on the bottom of the [12:08] that you see here on the bottom of the [12:08] that you see here on the bottom of the chart, which was an AI use case must [12:11] chart, which was an AI use case must [12:11] chart, which was an AI use case must produce for IBM one of four benefits. [12:15] produce for IBM one of four benefits. [12:15] produce for IBM one of four benefits. And we must be able to achieve the ROI. [12:19] And we must be able to achieve the ROI. [12:19] And we must be able to achieve the ROI. Any business case that was submitted [12:20] Any business case that was submitted [12:20] Any business case that was submitted internally, we had to understand the [12:22] internally, we had to understand the [12:22] internally, we had to understand the total cost of ownership for this [12:23] total cost of ownership for this [12:23] total cost of ownership for this solution. And the ROI had to be [12:25] solution. And the ROI had to be [12:25] solution. And the ROI had to be generated in less than a one-year [12:27] generated in less than a one-year [12:27] generated in less than a one-year period. And if a business case like that [12:29] period. And if a business case like that [12:29] period. And if a business case like that for an idea could not be submitted and [12:31] for an idea could not be submitted and [12:31] for an idea could not be submitted and solidified within a two-eek period, we [12:33] solidified within a two-eek period, we [12:33] solidified within a two-eek period, we removed it. We said move on to something [12:35] removed it. We said move on to something [12:36] removed it. We said move on to something else. Might not have been a bad idea, [12:38] else. Might not have been a bad idea, [12:38] else. Might not have been a bad idea, but it wasn't the best idea for us at [12:40] but it wasn't the best idea for us at [12:40] but it wasn't the best idea for us at the time to actually achieve what we had [12:43] the time to actually achieve what we had [12:43] the time to actually achieve what we had been given as as a target. And so our [12:45] been given as as a target. And so our [12:46] been given as as a target. And so our aligned value definition, and I would [12:47] aligned value definition, and I would [12:47] aligned value definition, and I would say that most enterprises that I talk [12:49] say that most enterprises that I talk [12:49] say that most enterprises that I talk to, right, they they span somewhere [12:52] to, right, they they span somewhere [12:52] to, right, they they span somewhere within this realm of two hard benefits, [12:55] within this realm of two hard benefits, [12:55] within this realm of two hard benefits, right? And this is what we would [12:57] right? And this is what we would [12:57] right? And this is what we would consider to be pre-tax income or revenue [12:59] consider to be pre-tax income or revenue [12:59] consider to be pre-tax income or revenue generated based upon AI and automation [13:02] generated based upon AI and automation [13:02] generated based upon AI and automation being used or cost actually taken out of [13:05] being used or cost actually taken out of [13:05] being used or cost actually taken out of the ledger. Okay? Where were able to [13:07] the ledger. Okay? Where were able to [13:07] the ledger. Okay? Where were able to remove some cost element of what we were [13:10] remove some cost element of what we were [13:10] remove some cost element of what we were doing because of the use of this [13:11] doing because of the use of this [13:11] doing because of the use of this technology. On the soft benefit side, we [13:15] technology. On the soft benefit side, we [13:15] technology. On the soft benefit side, we looked at cost avoidance. If we had 37 [13:17] looked at cost avoidance. If we had 37 [13:17] looked at cost avoidance. If we had 37 different dashboarding tools, we had to [13:19] different dashboarding tools, we had to [13:19] different dashboarding tools, we had to pause and look at oursel in the mirror [13:21] pause and look at oursel in the mirror [13:21] pause and look at oursel in the mirror and say why? like that probably happened [13:23] and say why? like that probably happened [13:23] and say why? like that probably happened because of a bunch of shadow creation of [13:25] because of a bunch of shadow creation of [13:25] because of a bunch of shadow creation of people needing a tool at a time and a [13:27] people needing a tool at a time and a [13:27] people needing a tool at a time and a place and they went and acquired it and [13:29] place and they went and acquired it and [13:29] place and they went and acquired it and it was you know it was all with kind of [13:31] it was you know it was all with kind of [13:31] it was you know it was all with kind of good measure at the time but it's not [13:33] good measure at the time but it's not [13:33] good measure at the time but it's not something that was scalable across our [13:35] something that was scalable across our [13:35] something that was scalable across our organization. So, how did we use [13:37] organization. So, how did we use [13:37] organization. So, how did we use strategic vendors, strategic contracts [13:39] strategic vendors, strategic contracts [13:39] strategic vendors, strategic contracts that we had to squeeze more juice out of [13:40] that we had to squeeze more juice out of [13:40] that we had to squeeze more juice out of that lemon and ultimately consolidate [13:43] that lemon and ultimately consolidate [13:43] that lemon and ultimately consolidate some of what we were doing from an [13:44] some of what we were doing from an [13:44] some of what we were doing from an operating perspective? And then of [13:46] operating perspective? And then of [13:46] operating perspective? And then of course productivity savings, right? Or [13:48] course productivity savings, right? Or [13:48] course productivity savings, right? Or those dollarized time savings, okay, of [13:51] those dollarized time savings, okay, of [13:51] those dollarized time savings, okay, of what we were looking at. This is how we [13:53] what we were looking at. This is how we [13:53] what we were looking at. This is how we aligned value. We set up a central [13:55] aligned value. We set up a central [13:55] aligned value. We set up a central tracking and a broader kind of [13:57] tracking and a broader kind of [13:57] tracking and a broader kind of overarching AI governance mechanism to [14:00] overarching AI governance mechanism to [14:00] overarching AI governance mechanism to look at the catalog of use cases to [14:02] look at the catalog of use cases to [14:02] look at the catalog of use cases to assess risk but to also assess value in [14:06] assess risk but to also assess value in [14:06] assess risk but to also assess value in terms of what we were able to deliver. [14:08] terms of what we were able to deliver. [14:08] terms of what we were able to deliver. And if you look at just a few areas of [14:11] And if you look at just a few areas of [14:11] And if you look at just a few areas of the KPIs and the value of what IBM has [14:13] the KPIs and the value of what IBM has [14:13] the KPIs and the value of what IBM has accomplished aligned to these [14:15] accomplished aligned to these [14:15] accomplished aligned to these definitions, I'll call out just a couple [14:17] definitions, I'll call out just a couple [14:17] definitions, I'll call out just a couple for you here. Right, there are many. [14:18] for you here. Right, there are many. [14:18] for you here. Right, there are many. We've deployed over a hundred generative [14:21] We've deployed over a hundred generative [14:21] We've deployed over a hundred generative and agentic AI use cases across the IBM [14:23] and agentic AI use cases across the IBM [14:23] and agentic AI use cases across the IBM framework in HR, IT, sales, marketing, [14:27] framework in HR, IT, sales, marketing, [14:27] framework in HR, IT, sales, marketing, procurement, finance, supply chain, you [14:29] procurement, finance, supply chain, you [14:29] procurement, finance, supply chain, you name it. Um the use cases are there. But [14:32] name it. Um the use cases are there. But [14:32] name it. Um the use cases are there. But if you look at something like HR where [14:34] if you look at something like HR where [14:34] if you look at something like HR where we are quite mature, I would argue uh in [14:38] we are quite mature, I would argue uh in [14:38] we are quite mature, I would argue uh in terms of the AI that has been [14:39] terms of the AI that has been [14:39] terms of the AI that has been implemented there, over 34 generative AI [14:42] implemented there, over 34 generative AI [14:42] implemented there, over 34 generative AI use cases deployed today. a 40% net [14:45] use cases deployed today. a 40% net [14:45] use cases deployed today. a 40% net operating reduction in the budget for [14:47] operating reduction in the budget for [14:47] operating reduction in the budget for our CHRO. Okay, this is across things [14:49] our CHRO. Okay, this is across things [14:49] our CHRO. Okay, this is across things like benefits and compensation, payroll, [14:52] like benefits and compensation, payroll, [14:52] like benefits and compensation, payroll, talent, development, learning, things of [14:54] talent, development, learning, things of [14:54] talent, development, learning, things of that nature. But at the same time, able [14:57] that nature. But at the same time, able [14:57] that nature. But at the same time, able to reduce the operating budget and [14:58] to reduce the operating budget and [14:58] to reduce the operating budget and improve the employee experience score by [15:01] improve the employee experience score by [15:01] improve the employee experience score by over 55 points. Typically, those types [15:03] over 55 points. Typically, those types [15:03] over 55 points. Typically, those types of value metrics don't correlate, right? [15:05] of value metrics don't correlate, right? [15:05] of value metrics don't correlate, right? A reduction in spend, doing more with [15:08] A reduction in spend, doing more with [15:08] A reduction in spend, doing more with less, but improving the experience at [15:09] less, but improving the experience at [15:09] less, but improving the experience at the same time. You'll see speed to [15:11] the same time. You'll see speed to [15:11] the same time. You'll see speed to market with things like ask it. 100 days [15:14] market with things like ask it. 100 days [15:14] market with things like ask it. 100 days to build a solution from scratch to [15:16] to build a solution from scratch to [15:16] to build a solution from scratch to remove 50% of the incoming employee [15:18] remove 50% of the incoming employee [15:18] remove 50% of the incoming employee support tickets. That has value for us [15:21] support tickets. That has value for us [15:21] support tickets. That has value for us when we think about every support ticket [15:22] when we think about every support ticket [15:22] when we think about every support ticket costing us about $18 across the [15:25] costing us about $18 across the [15:25] costing us about $18 across the workflow. Okay, so just a couple of [15:28] workflow. Okay, so just a couple of [15:28] workflow. Okay, so just a couple of examples here. I shared a little bit [15:30] examples here. I shared a little bit [15:30] examples here. I shared a little bit about these benefits and this is just [15:32] about these benefits and this is just [15:32] about these benefits and this is just another view of how we start to think [15:34] another view of how we start to think [15:34] another view of how we start to think about these things. We look at our [15:35] about these things. We look at our [15:36] about these things. We look at our current state. We look at the potential [15:37] current state. We look at the potential [15:37] current state. We look at the potential for adoption and savings and we look at [15:39] for adoption and savings and we look at [15:39] for adoption and savings and we look at the cost and this is again guys an [15:41] the cost and this is again guys an [15:41] the cost and this is again guys an all-in cost. What kind of resource is it [15:43] all-in cost. What kind of resource is it [15:43] all-in cost. What kind of resource is it going to take you to achieve this [15:45] going to take you to achieve this [15:45] going to take you to achieve this technology cost the labor that's [15:47] technology cost the labor that's [15:47] technology cost the labor that's associated uh uh across the actual use [15:50] associated uh uh across the actual use [15:50] associated uh uh across the actual use case that you choose and how does it map [15:52] case that you choose and how does it map [15:52] case that you choose and how does it map into these four value elements. [15:55] into these four value elements. [15:55] into these four value elements. Now I want to share just while we're on [15:57] Now I want to share just while we're on [15:57] Now I want to share just while we're on the subject of a use case like HR. Now, [16:00] the subject of a use case like HR. Now, [16:00] the subject of a use case like HR. Now, something that IBM has done is we've [16:02] something that IBM has done is we've [16:02] something that IBM has done is we've taken many of those use cases inside of [16:04] taken many of those use cases inside of [16:04] taken many of those use cases inside of our own environment and we've actually [16:05] our own environment and we've actually [16:05] our own environment and we've actually plumbed those into pre-built agents or [16:08] plumbed those into pre-built agents or [16:08] plumbed those into pre-built agents or agentic templates. Now, these are not [16:10] agentic templates. Now, these are not [16:10] agentic templates. Now, these are not like single threaded function things. [16:11] like single threaded function things. [16:11] like single threaded function things. These are wide things like talent [16:13] These are wide things like talent [16:13] These are wide things like talent acquisition. And if we were to break [16:15] acquisition. And if we were to break [16:15] acquisition. And if we were to break down a workflow like talent acquisition, [16:17] down a workflow like talent acquisition, [16:17] down a workflow like talent acquisition, there's all kinds of AI tasks that are [16:18] there's all kinds of AI tasks that are [16:18] there's all kinds of AI tasks that are wrapped up in that. It could be writing [16:20] wrapped up in that. It could be writing [16:20] wrapped up in that. It could be writing a job description. It could be sourcing [16:22] a job description. It could be sourcing [16:22] a job description. It could be sourcing candidates. It could be interviewing [16:23] candidates. It could be interviewing [16:23] candidates. It could be interviewing interviewing and summarizing what we get [16:25] interviewing and summarizing what we get [16:25] interviewing and summarizing what we get from a candidate. The workflow is fairly [16:27] from a candidate. The workflow is fairly [16:27] from a candidate. The workflow is fairly vast. So, we plumbed these into [16:29] vast. So, we plumbed these into [16:29] vast. So, we plumbed these into pre-built agents for other clients to be [16:31] pre-built agents for other clients to be [16:31] pre-built agents for other clients to be able to use. And our hope is actually [16:32] able to use. And our hope is actually [16:32] able to use. And our hope is actually that you achieve a faster ROI than what [16:34] that you achieve a faster ROI than what [16:34] that you achieve a faster ROI than what IBM did, right? We've done a fair bit of [16:37] IBM did, right? We've done a fair bit of [16:37] IBM did, right? We've done a fair bit of training. There's a fair bit of domain [16:38] training. There's a fair bit of domain [16:38] training. There's a fair bit of domain understanding that's in there to [16:40] understanding that's in there to [16:40] understanding that's in there to accelerate the value that you get. But [16:42] accelerate the value that you get. But [16:42] accelerate the value that you get. But what's really important when I partner [16:44] what's really important when I partner [16:44] what's really important when I partner with a client is to look at what I call [16:45] with a client is to look at what I call [16:46] with a client is to look at what I call a value map like this. And I've set up [16:48] a value map like this. And I've set up [16:48] a value map like this. And I've set up an organization inside of my team at IBM [16:51] an organization inside of my team at IBM [16:51] an organization inside of my team at IBM that we call it an AI value assessment [16:54] that we call it an AI value assessment [16:54] that we call it an AI value assessment organization. And we produce these value [16:56] organization. And we produce these value [16:56] organization. And we produce these value perspectives. Okay, this is really [16:58] perspectives. Okay, this is really [16:58] perspectives. Okay, this is really important to aligning domain experts and [17:00] important to aligning domain experts and [17:00] important to aligning domain experts and technical professionals inside of your [17:02] technical professionals inside of your [17:02] technical professionals inside of your organization to get everybody on the [17:04] organization to get everybody on the [17:04] organization to get everybody on the same page about what you have the [17:05] same page about what you have the [17:06] same page about what you have the potential to achieve. Okay, so we lay [17:08] potential to achieve. Okay, so we lay [17:08] potential to achieve. Okay, so we lay these types of things out. What are the [17:10] these types of things out. What are the [17:10] these types of things out. What are the value drivers, okay, across different [17:12] value drivers, okay, across different [17:12] value drivers, okay, across different functions of HR? What are some of the [17:14] functions of HR? What are some of the [17:14] functions of HR? What are some of the benefit desri descriptions that you [17:16] benefit desri descriptions that you [17:16] benefit desri descriptions that you should expect if you're applying AI in a [17:18] should expect if you're applying AI in a [17:18] should expect if you're applying AI in a domain like HR? And then simply how do [17:21] domain like HR? And then simply how do [17:21] domain like HR? And then simply how do you calculate these things? Because it's [17:23] you calculate these things? Because it's [17:23] you calculate these things? Because it's actually important to skill people up on [17:25] actually important to skill people up on [17:25] actually important to skill people up on how you do that. across different [17:27] how you do that. across different [17:27] how you do that. across different domains and across different use cases [17:29] domains and across different use cases [17:29] domains and across different use cases inside of your organization. So this [17:31] inside of your organization. So this [17:31] inside of your organization. So this might look simple to some, but I promise [17:34] might look simple to some, but I promise [17:34] might look simple to some, but I promise getting the alignment inside of an [17:35] getting the alignment inside of an [17:36] getting the alignment inside of an enterprise at the scale of which many of [17:38] enterprise at the scale of which many of [17:38] enterprise at the scale of which many of you guys probably tend to operate, [17:40] you guys probably tend to operate, [17:40] you guys probably tend to operate, putting this on paper, putting it in [17:42] putting this on paper, putting it in [17:42] putting this on paper, putting it in front of a wide variety of different [17:43] front of a wide variety of different [17:43] front of a wide variety of different stakeholders and coming to a conclusion [17:46] stakeholders and coming to a conclusion [17:46] stakeholders and coming to a conclusion is important. [17:48] is important. [17:48] is important. We literally then do put pen to paper, [17:50] We literally then do put pen to paper, [17:50] We literally then do put pen to paper, right? We are doing these dynamic value [17:52] right? We are doing these dynamic value [17:52] right? We are doing these dynamic value assessments with our clients and I'm [17:54] assessments with our clients and I'm [17:54] assessments with our clients and I'm just showing you one that references the [17:56] just showing you one that references the [17:56] just showing you one that references the talent recruitment and talent [17:57] talent recruitment and talent [17:57] talent recruitment and talent acquisition process here. We would start [18:00] acquisition process here. We would start [18:00] acquisition process here. We would start here and we would go much deeper, right? [18:01] here and we would go much deeper, right? [18:01] here and we would go much deeper, right? But you'll see a framework. We look at [18:03] But you'll see a framework. We look at [18:03] But you'll see a framework. We look at things like the size of organiz the [18:05] things like the size of organiz the [18:05] things like the size of organiz the organization. Typically an HR [18:07] organization. Typically an HR [18:07] organization. Typically an HR organization is going to know what is [18:09] organization is going to know what is [18:09] organization is going to know what is their new hire rate on an annual basis, [18:12] their new hire rate on an annual basis, [18:12] their new hire rate on an annual basis, right? Where we can begin to estimate [18:13] right? Where we can begin to estimate [18:13] right? Where we can begin to estimate some of these elements. uh we look at [18:15] some of these elements. uh we look at [18:15] some of these elements. uh we look at the process cost for onboarding whether [18:18] the process cost for onboarding whether [18:18] the process cost for onboarding whether you're doing that from an outsource [18:19] you're doing that from an outsource [18:19] you're doing that from an outsource recruitment perspective or you're [18:21] recruitment perspective or you're [18:21] recruitment perspective or you're looking across that workflow that I [18:23] looking across that workflow that I [18:23] looking across that workflow that I described and how much time and how much [18:25] described and how much time and how much [18:25] described and how much time and how much cost is associated with each one of [18:27] cost is associated with each one of [18:28] cost is associated with each one of those elements. ultimately getting down [18:30] those elements. ultimately getting down [18:30] those elements. ultimately getting down to what can our productivity gain be and [18:33] to what can our productivity gain be and [18:33] to what can our productivity gain be and in IBM's case when we launched like an [18:35] in IBM's case when we launched like an [18:35] in IBM's case when we launched like an employee support tool as an example we [18:37] employee support tool as an example we [18:37] employee support tool as an example we had 750 people sitting in a call center [18:39] had 750 people sitting in a call center [18:39] had 750 people sitting in a call center we took 675 of those individuals who we [18:43] we took 675 of those individuals who we [18:43] we took 675 of those individuals who we no longer needed to staff to pick up the [18:44] no longer needed to staff to pick up the [18:44] no longer needed to staff to pick up the phone they were the people that actually [18:46] phone they were the people that actually [18:46] phone they were the people that actually knew how to get work done in HR and we [18:49] knew how to get work done in HR and we [18:49] knew how to get work done in HR and we created a department of data science and [18:51] created a department of data science and [18:51] created a department of data science and AI for those individuals to be a part of [18:54] AI for those individuals to be a part of [18:54] AI for those individuals to be a part of to contribute to things like this and [18:56] to contribute to things like this and [18:56] to contribute to things like this and actually plumbing their intelligence [18:58] actually plumbing their intelligence [18:58] actually plumbing their intelligence into AI so that we can make better use [19:01] into AI so that we can make better use [19:01] into AI so that we can make better use of it and create repeatable processes [19:03] of it and create repeatable processes [19:03] of it and create repeatable processes inside of our own framework. So to give [19:07] inside of our own framework. So to give [19:07] inside of our own framework. So to give you a little bit of an idea of how we do [19:08] you a little bit of an idea of how we do [19:08] you a little bit of an idea of how we do this with our clients today and this is [19:10] this with our clients today and this is [19:10] this with our clients today and this is also the process that we follow [19:12] also the process that we follow [19:12] also the process that we follow internally at IBM for assessing value, [19:15] internally at IBM for assessing value, [19:15] internally at IBM for assessing value, we literally look at an assessment [19:17] we literally look at an assessment [19:17] we literally look at an assessment scoping, right? So we kind of go through [19:19] scoping, right? So we kind of go through [19:19] scoping, right? So we kind of go through many of those value drivers or a benefit [19:21] many of those value drivers or a benefit [19:21] many of those value drivers or a benefit map like what you just saw aligned to a [19:24] map like what you just saw aligned to a [19:24] map like what you just saw aligned to a specific use case or aligned to a [19:26] specific use case or aligned to a [19:26] specific use case or aligned to a specific domain. We then will partner [19:28] specific domain. We then will partner [19:28] specific domain. We then will partner with clients and we'll do things like [19:30] with clients and we'll do things like [19:30] with clients and we'll do things like discovery interviews and workshops. We [19:32] discovery interviews and workshops. We [19:32] discovery interviews and workshops. We might find out a thing or two that you [19:34] might find out a thing or two that you [19:34] might find out a thing or two that you might not know, right, about other [19:35] might not know, right, about other [19:35] might not know, right, about other domains or people inside of your [19:37] domains or people inside of your [19:37] domains or people inside of your business and the problems that they're [19:39] business and the problems that they're [19:39] business and the problems that they're trying to solve. We then will do the [19:41] trying to solve. We then will do the [19:41] trying to solve. We then will do the synthesis and the value modeling. You [19:43] synthesis and the value modeling. You [19:43] synthesis and the value modeling. You just saw a very simple example of that. [19:45] just saw a very simple example of that. [19:45] just saw a very simple example of that. And then we do an overarching kind of [19:47] And then we do an overarching kind of [19:47] And then we do an overarching kind of value case readout. And I would suggest [19:50] value case readout. And I would suggest [19:50] value case readout. And I would suggest that for the use cases you're looking to [19:52] that for the use cases you're looking to [19:52] that for the use cases you're looking to explore in your own enterprises, [19:54] explore in your own enterprises, [19:54] explore in your own enterprises, adopting a process like this that is [19:56] adopting a process like this that is [19:56] adopting a process like this that is known throughout your organization, [19:58] known throughout your organization, [19:58] known throughout your organization, okay? And that people are participatory [20:00] okay? And that people are participatory [20:00] okay? And that people are participatory in is is quite important. [20:03] in is is quite important. [20:03] in is is quite important. I will also tell you though that [20:05] I will also tell you though that [20:05] I will also tell you though that technology and these kind of [20:06] technology and these kind of [20:06] technology and these kind of architectural choices matter in terms of [20:09] architectural choices matter in terms of [20:09] architectural choices matter in terms of ROI as well. Sometimes we think about it [20:11] ROI as well. Sometimes we think about it [20:12] ROI as well. Sometimes we think about it that it's just about the use case or [20:14] that it's just about the use case or [20:14] that it's just about the use case or it's just about the amount of people [20:16] it's just about the amount of people [20:16] it's just about the amount of people that we have or the amount of time that [20:18] that we have or the amount of time that [20:18] that we have or the amount of time that we're spending on something. But I want [20:20] we're spending on something. But I want [20:20] we're spending on something. But I want to give you kind of a use case here. [20:21] to give you kind of a use case here. [20:22] to give you kind of a use case here. What you're seeing on this graphic is [20:23] What you're seeing on this graphic is [20:23] What you're seeing on this graphic is something by which in these pink bars [20:26] something by which in these pink bars [20:26] something by which in these pink bars there were very large models that were [20:28] there were very large models that were [20:28] there were very large models that were being used to solve for a use case [20:30] being used to solve for a use case [20:30] being used to solve for a use case across a wide variety of of different um [20:33] across a wide variety of of different um [20:33] across a wide variety of of different um of different clients and different use [20:35] of different clients and different use [20:35] of different clients and different use cases internally at IBM. But I'll call [20:37] cases internally at IBM. But I'll call [20:37] cases internally at IBM. But I'll call your attention to the one that is uh [20:39] your attention to the one that is uh [20:39] your attention to the one that is uh over here to your foremost right side. [20:42] over here to your foremost right side. [20:42] over here to your foremost right side. This was a telecommunications company [20:43] This was a telecommunications company [20:43] This was a telecommunications company that I worked with and we had had [20:46] that I worked with and we had had [20:46] that I worked with and we had had several conversations about IBM's [20:47] several conversations about IBM's [20:48] several conversations about IBM's technology platforms, small language [20:49] technology platforms, small language [20:50] technology platforms, small language models, what it would mean to adapt a [20:51] models, what it would mean to adapt a [20:51] models, what it would mean to adapt a small language model with their data to [20:53] small language model with their data to [20:53] small language model with their data to solve for a use case. And remember at [20:55] solve for a use case. And remember at [20:55] solve for a use case. And remember at the very beginning I told you scalable [20:57] the very beginning I told you scalable [20:57] the very beginning I told you scalable ROI is very different than just an [20:59] ROI is very different than just an [20:59] ROI is very different than just an experiment in and of itself and what [21:00] experiment in and of itself and what [21:00] experiment in and of itself and what value can be delivered from it. This [21:03] value can be delivered from it. This [21:03] value can be delivered from it. This individual called me, this has probably [21:04] individual called me, this has probably [21:04] individual called me, this has probably been about 12 months ago, and said,"I [21:06] been about 12 months ago, and said,"I [21:06] been about 12 months ago, and said,"I think I have a problem. I have a [21:08] think I have a problem. I have a [21:08] think I have a problem. I have a customer care use case by which I've [21:10] customer care use case by which I've [21:10] customer care use case by which I've been able to do three million different [21:11] been able to do three million different [21:11] been able to do three million different call transcripts and summarize them uh [21:14] call transcripts and summarize them uh [21:14] call transcripts and summarize them uh across uh the GPT model, right?" And he [21:17] across uh the GPT model, right?" And he [21:17] across uh the GPT model, right?" And he said, "It's performing relatively well. [21:18] said, "It's performing relatively well. [21:18] said, "It's performing relatively well. I'm getting about 97% accuracy. Uh but I [21:21] I'm getting about 97% accuracy. Uh but I [21:21] I'm getting about 97% accuracy. Uh but I kind of have a problem. I've got 330 [21:24] kind of have a problem. I've got 330 [21:24] kind of have a problem. I've got 330 million transcripts to do overall and [21:26] million transcripts to do overall and [21:26] million transcripts to do overall and I've blown the budget for the project." [21:29] I've blown the budget for the project." [21:29] I've blown the budget for the project." He needed a much smaller model that [21:31] He needed a much smaller model that [21:31] He needed a much smaller model that would hit him much differently from an [21:32] would hit him much differently from an [21:32] would hit him much differently from an inference and an infrastructure [21:33] inference and an infrastructure [21:34] inference and an infrastructure perspective in order to scale to the [21:36] perspective in order to scale to the [21:36] perspective in order to scale to the aggregate of what his company was [21:39] aggregate of what his company was [21:39] aggregate of what his company was expecting, right? And so we did that. We [21:41] expecting, right? And so we did that. We [21:41] expecting, right? And so we did that. We got to the exact same level of accuracy [21:44] got to the exact same level of accuracy [21:44] got to the exact same level of accuracy but with a much smaller model and [21:45] but with a much smaller model and [21:45] but with a much smaller model and reduced his cost by 10x. So think about [21:48] reduced his cost by 10x. So think about [21:48] reduced his cost by 10x. So think about these types of choices, right? The [21:50] these types of choices, right? The [21:50] these types of choices, right? The architectural choices that you make [21:52] architectural choices that you make [21:52] architectural choices that you make inside of these cases, they matter. [21:54] inside of these cases, they matter. [21:54] inside of these cases, they matter. Also, [21:55] Also, [21:55] Also, I would also just tell you in closing, [21:57] I would also just tell you in closing, [21:57] I would also just tell you in closing, right from an AI governance perspective, [21:59] right from an AI governance perspective, [21:59] right from an AI governance perspective, most people think that this is [22:00] most people think that this is [22:00] most people think that this is fundamental just to mitigating risk in [22:02] fundamental just to mitigating risk in [22:02] fundamental just to mitigating risk in their business, there is some cost [22:04] their business, there is some cost [22:04] their business, there is some cost avoidance and reduction benefits, but [22:06] avoidance and reduction benefits, but [22:06] avoidance and reduction benefits, but doing AI governance right will also get [22:09] doing AI governance right will also get [22:09] doing AI governance right will also get you uh a lot of revenue and productivity [22:11] you uh a lot of revenue and productivity [22:11] you uh a lot of revenue and productivity gain inside of your business when we [22:13] gain inside of your business when we [22:13] gain inside of your business when we think about using more AI models, using [22:15] think about using more AI models, using [22:15] think about using more AI models, using more AI agents, and accelerating the [22:18] more AI agents, and accelerating the [22:18] more AI agents, and accelerating the development of these use cases across [22:20] development of these use cases across [22:20] development of these use cases across the board. Okay, so just in closing, um [22:23] the board. Okay, so just in closing, um [22:23] the board. Okay, so just in closing, um we've got pretty cool use cases just [22:25] we've got pretty cool use cases just [22:25] we've got pretty cool use cases just like the UFC. Uh if any of you guys did [22:26] like the UFC. Uh if any of you guys did [22:26] like the UFC. Uh if any of you guys did know that IBM was partnering with the [22:28] know that IBM was partnering with the [22:28] know that IBM was partnering with the UFC, they even have a value case to the [22:30] UFC, they even have a value case to the [22:30] UFC, they even have a value case to the insight engine that they've built in [22:32] insight engine that they've built in [22:32] insight engine that they've built in terms of equipping themselves to be able [22:34] terms of equipping themselves to be able [22:34] terms of equipping themselves to be able to generate insights much more quickly [22:36] to generate insights much more quickly [22:36] to generate insights much more quickly and building this into the content [22:38] and building this into the content [22:38] and building this into the content profile of 40 plus live events that they [22:40] profile of 40 plus live events that they [22:40] profile of 40 plus live events that they have every single year. So even fun [22:42] have every single year. So even fun [22:42] have every single year. So even fun things like this that we see out in the [22:44] things like this that we see out in the [22:44] things like this that we see out in the open in the media with sports and [22:46] open in the media with sports and [22:46] open in the media with sports and entertainment have AI value associated [22:49] entertainment have AI value associated [22:49] entertainment have AI value associated with them as well. I'd encourage you to [22:51] with them as well. I'd encourage you to [22:51] with them as well. I'd encourage you to stop by the IBM booth if you want to [22:53] stop by the IBM booth if you want to [22:53] stop by the IBM booth if you want to chat more. There's demos of much of this [22:55] chat more. There's demos of much of this [22:55] chat more. There's demos of much of this technology in action and plenty of [22:57] technology in action and plenty of [22:57] technology in action and plenty of experts that can talk to you just about [22:59] experts that can talk to you just about [22:59] experts that can talk to you just about what we did in terms of the value [23:00] what we did in terms of the value [23:00] what we did in terms of the value perspectives and how you go about [23:02] perspectives and how you go about [23:02] perspectives and how you go about achieving this in your own organization. [23:04] achieving this in your own organization. [23:04] achieving this in your own organization. Have a great rest of the conference.