Cummins and JPMC Executives Outline Strict AI Return Rules

康明斯与JPMC高管明确严格的AI回报准则 企业AI只有在组织指定业务财务主管负责绩效指标并将退役成本计入初始财务模型时，才能带来可衡量的财务回报。 与会嘉宾分享了AI的实际落地应用，涵盖从Cummins管理供应链关税影响，到Morgan Health优化医疗福利导引，以及Mural优化走向市场工作流。他们概述了用于评估自研与外购决策、计算总体拥有成本以及精简整合企业软件应用的实用框架标准。此外，讨论还强调了跨职能治理、财务主管对成效指标的权责归属，以及通过争取持怀疑态度的普通用户来扩大试点规模的策略。 康明斯（Cummins）利用图技术与本体构建企业数据翻译层，在 9 天内完成了针对数百万个零部件的台湾关税供应链冲击评估。 (3:04) 通过企业买家筛选的核心在于直击业务痛点并结合垂直领域数据，且 AI 带来的业务成效必须足以覆盖组织迁移与人员培训的变革成本。 (13:28) 评估自研还是采购（Build vs. Buy）以及五年的总拥有成本（TCO）时，解决方案必须无缝嵌入现有工作流，并提前计入模型迭代、厂商迁移和旧系统退役成本。 (23:04) 从试点转向规模化推广时，企业应将 KPI 看板所有权交由财务管控团队（Controllers）以明确业务收益，同时重点赋能普通用户并淘汰旧有系统，防止隐性成本侵蚀 ROI。 (34:44)

Capturing measurable return on investment from enterprise artificial intelligence requires solving urgent operational bottlenecks, enforcing strict financial accountability, and designing rollouts around average employees rather than technical enthusiasts. Anchoring Deployments in Workflow and Urgency (3:04) To generate tangible business value, artificial intelligence must target high-stakes operational friction rather than speculative enhancements. At Cummins, Jim Jacob observed that applying graph technologies and ontologies to a "burning platform"—sudden tariff changes affecting global supply chains—allowed the company to trace enterprise-wide impacts across millions of parts within nine days, circumventing costly, repetitive large language model token calls. Similarly, Preeti Vaidya noted that Morgan Health, which oversees care strategies for 160 million employer-insured individuals and has deployed nearly $300 million into health tech since 2021, focuses Retrieval-Augmented Generation (RAG) tools on complex benefits navigation and oncology support without disrupting existing clinical workflows. In go-to-market settings, Bill Dwoinen of Mural reported that AI helps sales teams rapidly synthesize 10-K filings and earnings calls into custom account plans. However, he warned that without active synthesis, tools generate uncurated "AI slop" that exposes internal performance deficits rather than improving client engagement. Evaluating Build-versus-Buy and Accelerated Obsolescence (23:04) Deciding whether to build or buy AI tools requires weighing internal opportunity costs, workflow fit, outcome transparency, and vendor reliability. Because model capabilities evolve in months rather than standard five-year total cost of ownership (TCO) cycles, Vaidya emphasized that organizations must treat model decommissioning, vendor migration, and switching costs as inevitable upfront financial obligations rather than hypothetical contingencies. Simultaneously, enterprise leaders must enforce software rationalization to prevent budget creep. Jacob warned that organizations frequently embrace the "romantic solution" of new AI offerings without performing the change management needed to retire legacy applications like PowerBI. Dwoinen reinforced this discipline from a revenue perspective, stating that evaluations conducted alongside corporate finance must explicitly ask what software tools—such as internal GPT experiments—can be eliminated before approving new vendor spend. Establishing Financial Ownership Through Business Controllers (34:44) Quantifying the economic impact of AI software often triggers organizational friction between technology builders and commercial units over who receives credit for productivity gains. Jacob noted that when an AI initiative produces substantial financial value, such as a $10 million benefit, corporate finance typically moves to capture the sum as enterprise contribution income, which can alienate operational business leaders who feel penalized. To resolve this dispute, Cummins assigned direct ownership of standardized AI performance dashboards to corporate business controllers rather than AI engineering teams. Vaidya noted that convening cross-functional committees of business heads, engineers, and finance leaders before launching pilots prevents technical teams from developing abandoned products and deters commercial leaders from pursuing technically infeasible tools. Overcoming Pilot Friction by Engaging Skeptics Transitioning AI from localized proof-of-concept tests to enterprise deployment represents the most challenging hurdle for organizations. Vaidya cautioned that initial pilot successes typically depend on "canary users" whose high AI literacy does not represent the median employee on a standard distribution curve. Full-scale adoption requires budgeting dedicated time and capital to upskill the broader workforce while deliberately pairing enthusiasts with skeptics to surface operational flaws. Dwoinen added that software procurement and governance have expanded from two stakeholders historically to as many as 13 across contemporary enterprises. Rather than piloting tools exclusively with power users, Dwoinen recommended recruiting disengaged staff and skeptics to locate genuine workflow barriers, while building IT governance capacity to handle complex integrations such as connecting external large language models directly to production CRM databases. The panel concluded that AI ROI depends less on model sophistication than on rigorous change management. Sustained gains require organizations to proactively retire legacy software, assign KPI ownership to financial controllers, and budget for the ongoing cost of upskilling average employees.

来自软件、医疗与制造业的企业高管指出，要实现真正的人工智能投资回报率（ROI），企业必须摒弃纯粹的技术炒作，将AI严格锚定于高紧迫性的核心业务工作流，并通过严密的财务核算、跨部门协同与包容怀疑者的变革管理来推动规模化落地。 聚焦业务紧迫痛点与真实工作流 (3:04) 企业部署AI若脱离具体业务场景，容易沦为无效的技术炫技。Cummins数字化与AI执行总监Jim Jacob强调，立项必须选择具有高紧迫性与高风险的“燃烧平台”问题。例如当突发关税政策威胁数百万零部件与代工伙伴的全球供应链时，Cummins并未依赖高昂且脆弱的大模型Token调用，而是结合图技术与本体构建翻译层，在9天内精准梳理出台湾进口零部件在整个企业的“长鞭效应”。他建议将管理海量电子表格的财务总监等领域专家与AI工程团队直接结对，以低成本的结构化本体解决特定业务痛点。 在以人为核心的工作流中，盲目引入AI反而会降低效率。Morgan Health执行总监Preeti Vaidya提醒，该机构虽管理着1.6亿由雇主投保的美国人医疗质量并顾问13家创企，但在急诊室为护士提供聊天机器人却会分散其分诊注意力；真正的价值在于利用检索增强生成（RAG）解析晦涩的福利条约，或通过旗下Embold和Thyme Care为专科与肿瘤患者提供导诊。Mural首席营收官Bill Dwoinen亦表示，销售团队虽可用AI提炼财报与10-K文件，但若缺乏业务语境就会产生泛滥的“AI垃圾（AI slop）”，工具更直接放大了员工的执行力差距。 自研采购五项法则与软件合理化 (23:04) 在技术选型与供应商评估上，盲目采购或全面自研都会损害回报率。Preeti Vaidya系统提出了自研与外购决策的五项评估法则：一是核心业务匹配度，确认底层问题是否必须由定制工具解决；二是机会成本，权衡由外部解决是否优于耗费内部资源；三是工作流整合度，确保方案无缝融入员工日常；四是结果透明度，审查技术底层机制；五是供应商伙伴信任度。她主张由技术、业务与财务团队组成的跨学科委员会共同制定技术路线，避免各自为政。 引进新系统的同时必须进行软件合理化与旧系统淘汰。Jacob指出，Cummins设立了精密性、架构适应性与经济可行性三大准则，但他警告许多企业往往沉溺于新AI工具的“浪漫解法”，却未做好变革管理去注销旧有的PowerBI等系统，造成软件总支出虚高。Dwoinen补充称，Mural在评估内外部工具（如自研的市场洞察工具SignalGPT与外部供应商）时，CFO的核心考量始终是“我们能够停用什么”，只有当新方案能明确取代旧系统或内部GPT配置时，替换软件的组织变革成本才是合理的。 总体拥有成本核算与主动退役规划 (24:42) 评估AI的真实成本需要打破传统软件预算认知。Jacob指出，如果一项AI功能的构建成本大幅超出业务问题本身的价值，高管支持就会迅速流失；反之，若工程团队能在特定预算内交付数倍于成本的商业回报，才能确立其在企业内部的信誉。Morgan Health自2021年以来已在医疗AI与科技领域部署近3亿美元，Vaidya基于此经验警告，单纯迎合销售话术或投资人期望而推行AI，不仅无法改善患者成效，还会直接消耗内部信任。 隐性支出与快速的技术换代要求企业重新定义总体拥有成本（TCO）。传统IT投资通常依据5年期的TCO模型核算，但Vaidya提出，AI技术以数月为周期的迭代节奏意味着现有模型的退役与技术转向不是假设，而是必然。构建团队常常将自研工具视为“心血结晶”而产生情感依附，但企业必须在初始财务模型中提前计入模型替换、数据迁移及供应商转换的硬性成本，避免被突发的重构支出吞噬ROI。 财务主导绩效度量与跨越试点鸿沟 (34:44) AI项目从小规模试点迈向生产环境时，价值度量权责往往引发部门纷争。为避免技术团队与业务单元在财务收益归属上发生摩擦，Cummins建立了标准化KPI仪表板，并明确将管理所有权交由业务财务主计长（Controllers）负责。Jacob解释，当AI创造1000万美元效益时，集团CFO倾向于将其作为贡献收益，容易让业务主管产生防备心理；让主计长从第一天起掌握绩效仪表板，能在不破坏损益表（P&L）约定的前提下确保财务与业务部门对价值产生共识。 此外，从试点向企业级规模化推广是实施中最艰巨的阶段。Vaidya提出在推广时需“让金丝雀呼吸”：试点成功往往依赖具有极高AI素养的早期狂热用户（“金丝雀用户”），但他们并不代表处于正态分布均值的中位数员工；企业必须为中位数员工的技能培训预留专项预算与时间，并主动将怀疑者与爱好者配对。Dwoinen也警示，现代软件采购决策主体已从过去的2人扩大到13个利益相关方，企业试点必须吸纳持怀疑态度的非活跃员工以暴露真实摩擦，并正视ChatGPT连接Salesforce等复杂集成对IT治理带来的严峻考验。 该讨论表明，实现AI商业回报的瓶颈不在于算法模型本身，而在于能否建立由财务主计长锚定真实ROI、由业务工作流定义痛点、并直面普通中位数员工培训与遗留系统退役的系统性治理框架。

## Background 该内容由 Ai4 于 2026 年 8 月 21 日发布在 YouTube 平台上。视频记录了一场关于企业级 AI 投资回报率的高管圆桌论坛，汇集了来自 PSP Growth、康明斯（Cummins）、摩根健康（Morgan Health）及 Mural 的业务与技术领袖，重点探讨如何打破 AI 噱头泡沫、精准识别具备财务可行性的落地场景并实现可衡量的商业价值。 ## TL;DR 论坛围绕企业如何从 AI 投资中榨取真实 ROI 展开了深度拆解。三位行业高管分别从 GTM 销售武装、复杂医保福利导航以及百年工业巨头利用知识图谱应对关税危机的实战经验出发，阐述了以业务工作流为核心而非盲目追逐技术的落地原则。讨论进一步剖析了技术选型标准、自研与外购权衡、总拥有成本（TCO）中的应用精简与退役成本，以及如何通过业务财务总监把关指标来化解内部利益博弈。最后，嘉宾们强调在从 PoC 试点跨越到企业级规模化时，必须让关键支持者喘息，主动吸纳怀疑论者的真实反馈并筑牢底层治理体系。 ## Atmosphere & style 整场对话洋溢着干练、务实且不讲空话的工业界决策者气场，彻底剥离了市场上泛滥的 AI 营销泡沫与浮夸修辞。四位嘉宾互动默契、直奔痛点，通过“翻雪山上学”、“AI 泔水”、急诊室添乱以及 CFO 拷问等极其生动形象的譬喻，将原本枯燥的企业架构与财务核算讲解得妙趣横生。讨论在冷静剖析技术局限、代币成本陷阱与组织阻力的同时，展现出直面复杂真实世界的建设性力量，给听众带来一种极其清醒、接地气且充满实操启发的沉浸体验。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=4s) AI 落地容易陷入高期望与低回报的落差，企业唯有将技术可行性与真实的商业价值紧密结合，才能榨取真正的 ROI。 - [03:04](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=184s) 销售团队应警惕 AI 生成的空洞内容（AI Slop），AI 工具不仅是情报武装手段，更是一面能迅速区分员工能力与态度的“试金石”。 - [05:53](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=353s) 在医疗福利等复杂信息场景中，利用 RAG 消化庞杂条款并结合对话式 AI，能有效实现大规模信息解析与优质就医导流。 - [11:59](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=719s) 工业企业无需强行将分散数据集中迁移，通过知识图谱与业务本体论（Ontology）实现就地智能互联，能快速应对关税等突发供应链危机。 - [13:28](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=808s) 企业买家对 AI 供应商有着严格筛选标准：方案必须紧密绑定战略核心、基于专有背景提供行动洞察，且收益必须大幅超越变革成本。 - [17:07](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=1027s) AI 方案要获得高层支持需瞄准企业的“燃烧平台”（紧迫危机），同时警惕 LLM 无节制调用带来的昂贵 Token 推理成本。 - [20:16](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=1216s) 医疗等专业场景必须坚持“工作流优先”，拒绝为了技术而技术破坏核心流程；自研与外购需基于核心壁垒与机会成本审慎决策。 - [24:42](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=1482s) AI 时代的总体拥有成本（TCO）核心在于“下线与转向成本”，企业从部署初始就必须为工具的高速迭代和方案退役做好准备。 - [28:03](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=1683s) 引入 AI 新工具时必须坚决淘汰旧有 Legacy 资产（如冗余 BI 系统），避免系统叠加导致 IT 总体成本严重失控。 - [32:04](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=1924s) 面对 CFO 审查预算时的“用了它能少用什么”拷问，AI 方案的价值在于能否切实减少对存量软件与冗余人力流程的依赖。 - [35:20](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=2120s) 通过让各业务线财务总监（Controller）把控 AI 价值看板，既能客观核定收益，又能保护业务部门的 P&L 利益，避免内部财务博弈。 - [38:41](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=2321s) 技术、业务与财务三方自始至终同席决策是落地根基，规模化推广时必须正视中位数普通用户的技能重塑（Upskilling）时间与成本。 - [41:47](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=2507s) 概念验证与试点不应只依赖极少数狂热先锋（金丝雀用户），主动拉入抵触工具的“怀疑者”才能暴露真实盲点并完善治理。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=4s) 论坛开场与嘉宾背景介绍 小组讨论由 PSP Growth 董事总经理兼科技主管 Momei Qu 主持。PSP Growth 是由前美国商务部长（奥巴马政府时期）、普利兹克家族核心成员 Penny Pritzker 创立的成长型投资机构。Momei Qu 谈到，普利兹克家族有着深厚的传统产业根基（如凯悦酒店集团、Marmon Group 等），因此 PSP Growth 在探索前沿 AI 技术的现实落地时，恰恰处于“技术可行性（what is possible）”与“商业价值（what is valuable）”的交汇点。当下行业普遍存在对 AI 的高期望与商业回报不及预期的落差，本次对话的核心就在于探讨企业如何从 AI 投资中榨取真实的 ROI。 随后，三位来自不同行业背景的嘉宾依次进行了自我介绍： - **Bill Dwoinen**：视觉协作平台 Mural 的首席营收官（Chief Revenue Officer, CRO），掌管公司的 Go-to-Market（GTM）体系与客户组织，致力于通过视觉协作促进跨部门协同与行动落地。 - **Preeti Vaidya**：摩根大通（JPMorgan Chase）旗下 Morgan Health 的数据与 AI 执行总监。Morgan Health 专注于为全美通过雇主获取医保的 1.6 亿人群改善医疗的公平性、可负担性与医疗质量；Preeti 既统管内部 AI 用例落地，也兼任旗下 13 家医疗科技投资组合公司的 AI 战略顾问。 - **Jim Jacob**：百年工业制造巨头康明斯（Cummins Inc.）数字化、高级分析与人工智能执行总监。作为全球动力与发电设备领导者（不仅驱动重型卡车，还在为当下激增的数据中心提供关键能源支持），Jim 带领团队在传统制造业环境中通过数据与 AI 创造了数以亿美元计的商业回报。 --- ### [03:04](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=184s) 落地实战（一）：GTM 团队的情报武装与“试金石”效应 讨论切入具体而微的真实用例。Mural 的 Bill Dwoinen 从销售与客户团队的实际场景出发，强调 AI 的核心必须是“驱动商业结果（outcomes）”。过去，一名优秀的销售代表在拜访客户前需要耗费大量时间翻阅繁杂的财报电话会议实录（Earnings calls）和 10-K 年报，如同“当年顶着风雪上下学都要走上坡路”一般笨拙且低效；而今天，借助 AI，团队在初次见面前就能迅速形成高度成熟、经过深度研究的客户洞察（Point of view），并在全组织范围内建立统一的标准打法。 然而，表面的便捷也伴随着尖锐的落地痛点。Bill 犀利地指出了三个具体细节： 1. **警惕“AI 泔水（AI Slop）”**：信息量的爆炸导致销售容易拿着 AI 生成的废话或满纸空谈（Word vomit）直接堆砌给客户。真正的难点在于如何过滤杂音，把海量情报转化为可落地的行动方案（Actionable alignment）。 2. **底层数据跨系统连接的阻碍**：缺乏真实业务系统和工具的有效串联，生成的洞察便无法立足。 3. **团队分化与偷懒者的“试金石”**：AI 是一面照妖镜。它能迅速区分出三类人——努力从“优秀跨越到卓越（Good to great）”的骨干、面对信息洪流感到无所适从需要培训的普通员工，以及试图利用 AI 抄近道、糊弄差事的偷懒者。 --- ### [05:53](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=353s) 落地实战（二）：破译复杂雇主医保福利体系 Morgan Health 的 Preeti Vaidya 带来了医疗健康领域极具社会意义与情感重量的真实痛点：雇主担保医疗保险（Employer-Sponsored Insurance）体系下的“就医与福利导航（Benefits navigation）”。当患者身体出现异样症状时，往往伴随着极度的恐惧、焦虑与迷茫，根本分不清自己应该去紧急救护（Urgent Care）、抽血检验还是直奔急诊室（ER）。对于替重病亲属操持就医的家属（Caregivers）而言，缺乏专业健康素养（Health literacy）更让漫长的报销政策和转诊网络如同天书。 ```mermaid flowchart TD accTitle: 医疗福利与就医导航 AI 工作流 subgraph S1["患者端痛点"] P1["出现突发/未知症状"] --> P2["焦虑、恐惧且缺乏健康素养"] end subgraph S2["AI 驱动的信息解析与导航"] P2 --> RAG["RAG 检索增强系统<br/>（消化复杂医保福利条款）"] RAG --> A1["明确报销范畴/拒保项目"] RAG --> A2["匹配对口医疗设施/等级"] end subgraph S3["精准优质医疗机构对接"] A1 & A2 --> BOT["对话式医疗 AI 导航"] BOT --> C1["Embold / Quantum Health<br/>（推荐高评分医生/专科）"] BOT --> C2["Thyme Care<br/>（复杂肿瘤学诊疗与陪护）"] end ``` Preeti 指出，没有哪个单一个体能够吃透如此庞杂的条款，这正是 AI 展现超凡价值的领域： - **RAG（检索增强生成）赋能福利速查**：将企业复杂的保险福利合同导入 RAG 系统，患者只需简单提问，就能立刻获知哪些费用可报销、哪些被剔除、该去哪个网点找哪类专科医生。 - **对话式 AI 对接优质诊疗**：获取信息后，更关键的是找到真正负责、高水准的医疗专家。其投资组合中的 Embold（现并入 Quantum Health）打造了对话式 AI 工具，直接引导患者寻访高评分专科名医；另一家 portfolio 公司 Thyme Care 则专门深耕肿瘤学（Oncology），在患者面临确诊的毁灭性打击时提供全流程导医。这种“大规模信息消化 + 关键决策指引”的能力，具有跨行业的普适借鉴意义。 --- ### [08:33](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=513s) 落地实战（三）：百年工业巨头的“数字孪生”与关税风暴 康明斯（Cummins）的 Jim Jacob 采用生动的“谷歌地图（Google Maps）”类比开篇。他向台下提问有多少人开发技术、又有多少人用谷歌地图——早期谷歌耗费大量人力由学生测绘各地道路与门牌，随后将这些支离破碎的数据集打通整合成统一界面；如今人们早已告别将纸质地图摊在汽车引擎盖上规划路线的时代，只要车前出现拥堵，系统就会毫秒级计算并自动重定向。现代大语言模型（LLM）的角色正是如此：它们为底层冷冰冰的机器数据提供了一套拟人化的“数字孪生（Digital twin）”交互翻译层。 这一逻辑在百年制造业巨头康明斯中发挥着不可替代的作用： - **历史包袱与数据割裂**：康明斯历经百年发展、海量业务迭代与连续并购，企业内部积攒了极其割裂的异构数据系统。AI 团队的核心任务被精辟提炼为六个字：“连接、解读、交付（Connect, interpret, provide）”。 - **商业成果与实战大考**：团队近年来已落地数百个项目，累计创造了数亿美元的净收益。 - **关税（Tariff）危机化解战**：当国际关税政策突然变更时，制造业面临噩梦般的供应链混乱。康明斯作为生产大型原始设备（ODM/OEM）的厂商，涉及数以百万计的零部件、遍布全球的数百万贸易伙伴、数以万计的终端客户以及错综复杂的生产基地。如何在短时间内摸清每一个部件在不同产地、不同税率下的成本暴露敞口并迅速调整采购策略，成为 AI 连接与解析数据能力的严峻考验。 ### [11:59](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=719s) 利用知识图谱与本体论应对关税挑战的三大核心原则 Jim Jacob 继续深入阐述了康明斯（Cummins）如何通过 AI 与图技术（Graph Technology）打破分散的数据壁垒。在应对错综复杂的关税政策变动时，企业无需强行将所有底层数据迁移到一个庞大的集中式数据库中，而是利用本体论（Ontologies）深度连接分散在各处的异构数据集。通过这套连接架构，团队在关税新规出台后仅耗时 9 天就完成了全企业范围的财务与业务影响分析。例如，当发现某个来自中国台湾的特定零部件受到关税波及时，系统能够顺藤摸瓜，直接追踪并呈现该零部件在下游引发的完整“长鞭效应”（Bullwhip Effect），帮助企业在集团层面实现主动应对与成本追回。 基于这一实战经验，Jim 总结了在企业内落地 AI 技术的三个核心原则： - **拒绝数据硬集中，专注智能互联**：不要试图将所有数据搬到一个地方，而是运用 AI 与知识图谱技术就地连接异构数据源。 - **业务场景化语义解读**：将深层数据与分析结果翻译成采购、规划、财务、法务和人力资源等业务部门能直接理解的语言。 - **让终端用户拥有“导航地图”**：把最终工具交到一线业务人员手中，如同给他们一张“谷歌地图”，让他们自主在实际工作流中探索与决策。 ```mermaid flowchart TD accTitle: 基于图技术与本体论的关税供应链分析架构 subgraph S1["底层数据层 (Disparate Data)"] D1["采购系统数据"] D2["供应链物流数据"] D3["财务与海关关税数据"] end subgraph S2["图谱与语义层 (AI & Ontology)"] G1["知识图谱引擎 Graph Tech"] O1["业务本体模型 Ontology"] end subgraph S3["业务洞察与决策 (Downstream Functions)"] B1["供应链长鞭效应分析"] B2["财务/关税精准追回"] B3["跨部门业务协同 (采购/法务/HR)"] end D1 --> G1 D2 --> G1 D3 --> G1 G1 --> O1 O1 --> B1 O1 --> B2 O1 --> B3 ``` --- ### [13:28](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=808s) 企业买家与 CRO 的技术筛选漏斗：拒绝空洞 Demo 与信息过载 主持人 Momei Qu 针对台下众多初创企业创始人与 AI 开发者抛出一个极为现实的问题：当前市场上充斥着各种 AI 概念与产品，大型企业的决策者究竟依据什么标准来筛选供应商？初创团队如何才能跨过门槛拿到面谈机会，而不是在初筛阶段就被当作“噱头营销”淘汰出局？ 作为 Mural 的首席营收官（CRO），Bill Dwoinen 率先分享了采购决策方的痛点。他感叹如今的销售拓客仿佛丧失了基本的艺术，自己邮箱里每天塞满了毫无针对性的冷启动邮件，千篇一律地询问“想看个 Demo 吗”或是堆砌着“应用合理化（App Rationalization）”、“降本增效”等泛滥的行业黑话。面对这种嘈杂环境，Bill 明确提出了他用于判断项目真实价值的三重硬性过滤标准： 1. **是否紧密绑定企业的核心战略目标**：AI 方案必须能回答最核心的问题——它能否拉近企业与客户的距离？能否直接驱动预期的业务成果？能否改善员工体验？ 2. **是否植根于企业专有背景并真正让团队变强**：仅仅提供通用信息或炫酷 Demo 毫无意义。Bill 举例提到，许多演示在展示财务分析时只会告诉你“公司现金流上升了”，但这种浮于表面的信息根本无法帮助销售或业务团队建立独到的业务视角（Point of View）。AI 必须深度融入企业的价值支柱（Value Pillars）和方法论中，过滤无用噪音，交付具备行动指引的高价值洞察。 3. **变革成本（Cost of Change）与实际收益的权衡**：推动数百名员工迁移到新工具的隐形成本极其高昂。新产品绝不能仅仅因为“更便宜”就被采用，它必须在体验与业务产出上展现出无可替代的突破性优势。 --- ### [17:07](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=1027s) 抓住企业“燃烧的平台”：业务专家介入与大模型代币成本陷阱 Jim Jacob 补充指出，技术方案要获得大企业管理层的全力支持，必须精准定位到企业的“燃烧平台”（Burning Platform）——即那些如果不解决就会引发重大损失或能够立刻产生立竿见影财务价值的紧迫危机。以关税追踪为例，因为涉及直接的资金挽回与合规风险，项目能够直接获得公司最高层高管的背书与资源倾斜。 与此同时，Jim 强调了领域专家（SME）在 AI 落地中的决定性地位。技术提供方在业务部门眼中往往只是“新面孔”，真正的破局点在于找到那些深陷在两千张 Excel 表格里焦头烂额的财务分析师或内控主管，将他们拥有的专业业务逻辑转化为领域本体（Ontology）。此外，Jim 还向在场团队发出了关于大语言模型（LLM）调用成本的警告： - **警惕过高的大模型调用与 Token 成本**：将自然语言翻译为机器语言虽然方便，但“便利性是有昂贵代价的”。如果日常业务运维持续无节制地依赖昂贵的大模型推理，代币（Token）成本将迅速吞噬项目的投资回报率。 - **关注投入产出比的数学方程**：在企业财务模型中，如果解决方案本身的构建与运行成本过高，业务部门绝不会买单。只有当技术成本控制在线性增长（$\bar{x}$），而交付的业务价值呈现指数级爆发（$x^2$ 甚至 $x^3$）时，AI 团队才能真正赢得业务伙伴的热烈拥抱。 --- ### [20:16](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=1216s) 医疗场景下的工作流优先原则与自研或采购抉择 Morgan Health（摩根大通旗下医疗健康部门）数据与 AI 执行总监 Preeti Vaidya 从投资与业务落地双重视角剖析了 AI 评估逻辑。她透露，自 2021 年以来，团队已在医疗健康领域投资近 3 亿美元。如今几乎每个向投资人汇报的项目都会言必称 AI，但其中充斥着大量“为了 AI 而 AI”的伪需求——往往只是因为销售团队要求加个卖点，或是为了迎合投资机构的融资门槛。 Preeti 直言，这种为了技术而技术的做法正在严重侵蚀利益相关者与用户的信任。她以急诊室（ER）场景为例进行了生动的说明： - 在分诊和抢救争分夺秒的急诊室里，护士最需要的是陪伴和观察患者；如果此时强行推行一个生成式 AI 聊天机器人（GenAI Chatbot），迫使护士分心与对话工具互动，不仅无法提效，反而会拉长患者排队分诊的时间，破坏整个核心医疗流程。 - **工作流优先于技术亮点**：真正的解决方案必须从实际业务工作流出发，搞清楚要解决的具体痛点是什么，再看 AI 应当如何无缝嵌入，而不是拿着像“谷歌地图”一样耀眼的新技术到处硬套问题。 在探讨企业决策者面临的“自研还是采购”（Build vs. Buy）难题时，Preeti 提出了评判商业回报率的核心考量框架： - **业务核心度定位**：该问题是否属于企业最核心的差异化竞争壁垒？ - **工具专用性评估**：该痛点是可以通过市场上的通用型工具解决，还是必须依赖高度定制化的专业方案？ - **内部研发的机会成本**：如果选择组织内部团队自研，所消耗的工程资源、时间周期与放弃其他关键业务的机会成本到底有多大？ ### [23:36](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=1416s) 评估 AI 方案与供应商的五大核心维度 在探讨 AI 项目评估时，评估框架最终被提炼为五个关键判断维度。首先，如果内部开发，必须能合理解释自研成本的必要性；其次是即便是外购（Buy），该方案能否无缝集成到现有工作流中；第四是方案产出的结果是否值得信赖，企业是否能够“掀开引擎盖查看内部原理”（see under the hood）。 而最关键的第五点则是企业是否能信任该供应商成为长期赋能 AI 的合作伙伴。面对 AI 技术的快速迭代，企业必须确保所选供应商和构建的方案能够长期共同演进，在业务领域内持续集成与创新。只有通过跨部门利益相关者、财务专家等共同组成的决策委员会（Council）通力合作，才能真正释放业务价值。 ```mermaid flowchart TD accTitle: AI方案评估的五大核心维度 subgraph S1["成本与集成"] A["1. 自研成本合理性"] --> B["2. 现有工作流集成能力"] end subgraph S2["信任与长期价值"] C["3. 结果可信度与黑盒透明化"] --> D["4. 供应商长期战略伙伴关系"] D --> E["5. 跨部门委员会协同释放价值"] end B --> C ``` - **五大考量要点**：自研成本合理性、现有流程集成度、算法产出透明度与信任度、供应商长期演进能力、跨职能委员会协同决策。 --- ### [24:42](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=1482s) 总体拥有成本（TCO）：AI 时代不仅是计算，更是“下线与转向成本” 从收益端转向成本端时，总拥有成本（Total Cost of Ownership, TCO）的计算在 AI 浪潮中迎来了根本性改变。虽然行业对传统 5 年周期的 TCO 测算并不陌生，且经历了过去的云计算转型，但 AI 的不同之处在于极快的创新速度。企业必须将“退役或转向其他方案的成本”（cost of decommissioning or pivoting）置于首要位置。 作为构建者（Builder），工程师通常会将项目视为心血，希望它能长青不衰。然而在当前时代，系统退役“不再是一个‘如果’（if）的问题，而是一个‘何时’（when）的问题”。企业在初始部署时就必须预判几个月内工具迭代的现实，将转向新方案或更换供应商的成本纳入考量，避免追求一步到位的所谓“完美方案”，而是保持方案随时间演变的能力。 --- ### [28:03](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=1683s) 技术选型三原则与“沉没成本”陷阱：不可忽视的应用精简与自研重塑 在技术与架构落地层面，企业可以依据三大标准进行筛选：第一是技术必须具备足够的先进性（Sophisticated）；第二是该技术能够无缝融入企业的组织架构与系统底座；第三则是经济可行性。 然而许多企业在引入 AI 时掉入了成本叠加的陷阱：企业引进了昂贵的 AI 新技术，却往往“舍不得停用原有的 PowerBI 或旧分析系统”，导致新技术直接叠加在旧 IT 架构之上，引发 TCO 的剧烈膨胀。真正的变革管理不能只停留在对“浪漫的新方案”的迷恋上，而必须强制剥离遗留系统的成本。 ```mermaid flowchart LR accTitle: AI 引入过程中的技术选型与系统精简路径 subgraph Selection["选型三原则"] T1["技术先进性"] T2["架构契合度"] T3["经济可行性"] end subgraph Rationalization["避免成本叠加"] R1["引入 AI 新方案"] --> R2["坚决剔除 Legacy 资产\n(如冗余 BI 软件)"] R2 --> R3["实现净效率与 ROI 提升"] end Selection --> Rationalization ``` 同时，“自研还是外购”（Build vs. Buy）的逻辑也发生了剧烈重塑。当前几乎所有供应商都宣称能解决问题，但外购往往面临双方利益冲突（供应商想多赚钱，企业想省钱）甚至形成“寄生”关系的风险。随着大语言模型（LLM）能够协助解决机器与代码层面的复杂问题，企业现在更有可能借助稳固的基础平台，在内部构建本体（Ontology）、解决具体问题并牢牢掌控核心 IP 与差异化能力。 - **三大选型标准**：技术先进性、企业架构适应性、财务经济可行性。 - **变革管理核心**：引入 AI 提升效率的同时，必须主动淘汰旧有分析工具，否则系统叠加会导致 IT 成本失控。 - **自研新优势**：借助 LLM 的辅助，企业无需庞大团队即可自建并沉淀核心 IP，避免外部伙伴的利益博弈。 --- ### [32:04](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=1924s) 倒置的采纳路径与 CFO 的灵魂拷问：“用了它能减少什么？” 在具体企业实践中，AI 工具的落地展现出一种完全“反向”（backwards）的特征。传统企业软件采购有着严密的场景定义、商业案例叙述和 ROI 衡量指标；而这次 AI 往往是“先发放账号，让全员去用”。从实际用量前 25 名的超级用户来看，使用场景极其分散且跨职能——从 GTM 体系、客户成功（CSM）到高管助理（EA），每个人都在用其辅助决策或提速。这使得统一衡量 ROI 变得异常复杂，管理层不得不反向对这批深度用户展开调研，排查哪些是单人应用（single player），哪些是协同场景（multiplayer）。 在与 CFO 沟通 AI 预算时，最核心的考量往往被精炼为一句话：“**如果我们用了这个，能少用什么？**”（"what could we use less of?"）。如果一款号称能提供全面市场信号分析的外部解决方案，并不能替代团队内部自研的“Signal GPT”或减少对其他软件的依赖，那么它就无法通过应用精简（App Rationalization）的审查。 - **反向落地特征**：先铺设 AI 牌照，再反向梳理高频用户的实际应用场景与业务价值。 - **CFO 审查逻辑**：AI 工具的真正 TCO 并非仅仅看其采购价格，而是看其是否能直接减少对其他存量软件和人力流程的依赖。 --- ### [34:44](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=2084s) 从概念验证（PoC）试点走向生产级落地 在激烈的 AI 供应商竞争中，即便某个方案凭借优势突破重围并成功完成了试点（Pilot），如何将受控环境下的试点真正转化为面向全公司的生产级（Production）系统，依然是一道巨大的门槛。试点通常运行在相对封闭、受限的环境中，而一旦走向大规模生产部署，面对系统集成、数据管道、成本控制与业务稳定性，整个评估逻辑和衡量标准都将面临全新的考验。 ### [35:20](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=2120s) 从单点实验到规模化标杆：化解财务与业务的利益博弈 当一个 AI 实验在小规模试点中取得正向成果后，究竟是什么让它能够真正嵌入业务流程，升级为全公司津津乐道的标杆案例？康明斯（Cummins）的 Jim Jacob 指出，衡量技术的价值就像评估“一台笔记本电脑究竟值多少钱”一样，是技术人员必须向业务管理者自证效益的经典难题。技术投资的核心诉求本应是提高生产力和效率，但技术团队单方面宣称“我创造了多少价值”往往缺乏公信力，业务端会认为技术只是方案的一部分，脏活累活全是业务线完成的。 为了打破这种信任僵局，康明斯在早期就明确了“究竟由谁来发起 ROI 质询”，并建立了一套指标极其严格的量化看板。关键在于，这套看板的管理权并没有留在技术部门，而是交给了各业务部门的财务总监（Controller）。这种机制巧妙规避了极具破坏性的企业内部财务博弈：如果技术部门为某个业务线节省或创造了 1000 万美元的效益，并在企业层面大肆宣扬，公司首席财务官（CFO）就会理直气壮地要求业务负责人上缴这 1000 万美元作为利润贡献（Contribution Income），直接影响财报计划与业绩承诺。这往往会导致业务负责人震怒，责怪技术团队向财务团队“告密”。 通过让业务部门的主计长参与指标把关与价值裁定，财务人员既能确认技术的真实产出，又愿意持续出资支持后续的解决方案。Jim Jacob 强调，企业必须保护业务线极其敏感的损益表（P&L）约定，只有在尊重现有财务框架的前提下推进，AI 项目才能从试验阶段稳健走向长期嵌入。 ```mermaid flowchart TD accTitle: 康明斯 AI 价值核算与利益协同机制 subgraph S1["1. 业务与技术协同"] A["业务部门提出痛点"] --> B["技术团队开发 AI 方案"] end subgraph S2["2. 价值核算与管控"] B --> C["业务财务总监 (Controller) 统一管理看板与 KPI"] C --> D["客观核定业务价值与生产力收益"] end subgraph S3["3. 利益平衡与持续投资"] D --> E["保护业务线 P&L 承诺<br/>避免 CFO 直接索要利润分成"] E --> F["财务与业务共同出资支持 AI 规模化推广"] end ``` - **权属下放**：价值看板由业务财务总监（Controller）直接把控，而非技术开发人员自卖自夸。 - **利益保护**：避免在 10-K 或 8-K 财务计划制定中意外刺破业务负责人的既有利益底线，确保业务端愿意主动拥抱技术。 --- ### [38:41](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=2321s) 跨部门协同与规模化法则：“让金丝雀喘口气” 摩根健康（Morgan Health, JPMC）的 Preeti Vaidya 补充指出，无论是探索初期还是准备规模化推广，最忌讳在孤岛中进行“点对点”的零散沟通。技术、财务和业务三大核心利益相关者必须自始至终同处一室：技术人员闭门造车只会做出无人问津的工具；业务人员单独决策则容易提出脱离实际、技术上根本不可行的空想方案。只有三方坐在一起，共同对可行性、预期价值及落地成本达成共识，项目才有坚实的根基。 ``` ┌──────────────────────────────────────────────┐ │ 正态分布 (用户群体) │ │ │ │ ▲ │ │ / \ │ │ / \ │ │ / \ │ │ / \ │ │ / \ │ │ / 中位数 \ │ │ ────────────/ 普通用户 \───────────── │ │ 【怀疑者群体】 【金丝雀先锋】 │ │ 需重点拉入概念验证 需适时休整喘息 │ └──────────────────────────────────────────────┘ ``` 对于绝大多数技术主管而言，从试点（Pilot）跨越到企业级规模化（Scaling）是整个周期中最艰难的一关。Preeti 提出了一个生动的观点——"让金丝雀喘口气"（Let the canary breathe）。试点之所以能够成功，往往是因为团队挑选了一群极具热情、AI 素养极高、乐于尝鲜的“金丝雀用户”（Canary Users）担任先锋。然而，这些先锋绝不代表企业正态分布中的“中位数普通用户”。 当项目从验证概念转向全员普及时，必须清醒地认识到：普通员工在技能培训、心智重塑和能力升级上都需要投入大量的真实时间和资金成本，这绝非随手带过的附带工作。Preeti 建议，在大规模推广阶段应主动将“怀疑论者”与“AI 狂热者”进行配对，同时吸纳两者的视角，才能克服企业级采用过程中的各种摩擦力。 - **三方同席机制**：立项之初就让技术、业务、财务三方同处一室，对可行性与成本边界达成硬性共识。 - **正视普及成本**：从概念验证走向企业级 AI 原生（AI-native），需要对中位数用户提供系统性的技能重塑（Upskilling）与时间预算。 - **阵营互补**：在落地推广中将持怀疑态度的员工与早期拥护者搭配，兼听反馈以扫清落地阻碍。 --- ### [41:47](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=2507s) 聚焦核心问题、重用怀疑者与强化企业级治理 Mural 的首席营收官 Bill Dwoinen 呼应了前面的观点，并分享了一个关键行业数据：几年前软件与技术的采购决策通常只需要 2 个人参与，而如今涉及的利益相关方已经激增到了 13 个人。出现这种现象，正是因为集中化专断决策与业务部门野蛮采购之间存在双重风险——特别是在销售团队中，业务人员往往“遇到任何细小需求都想买一个新工具”，造成工具泛滥与资源浪费。 在推行试点项目时，Bill 强调绝对不能只拉拢最喜欢尝鲜的核心超级用户（Power Users），反而要把团队里“最抗拒、目前根本不用工具的 10 个人”拉进试点。超级用户只会给出顺从的赞美，而只有那些拒绝使用的怀疑者，才能提供最真实的阻力反馈，揭示系统设计中的真正盲点，从而帮助团队找到让所有人顺利跟进的路径。 最后，Bill 提醒大家不可忽视日益复杂的底层治理要求。例如在 Go-to-Market 团队中，仅仅是打通 ChatGPT 与 Salesforce 数据源这一个集成动作，对于一个编制精干的 IT 团队来说就是一项极其繁重的工程，涉及权限配置、合规管控与实时监控等多重挑战。企业推行 AI 的所有讨论，最终必须回归最初的原点：“我们究竟在试图解决什么问题？它如何让我们变得更好、更快？我们又该如何带领全员一同踏上这段旅程？” - **决策链路演变**：企业级软件决策主体从 2 人扩展至 13 人，治理、配置与监控的复杂性已达到全新层级。 - **逆向试点策略**：主动引入对 AI 持抵触态度的群体，挖掘未被采用的根本原因，取代单一依赖高频用户的失真反馈。 - **底层价值回归**：明确业务靶心，在严密的企业治理与基础设施支撑下稳步拉动全员采纳。 ## Notable quotes > [03:04](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=184s) 当年顶着风雪上下学都要走上坡路 — Bill Dwoinen > [08:33](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=513s) 连接、解读、交付 — Jim Jacob > [17:07](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=1027s) 便利性是有昂贵代价的 — Jim Jacob > [03:38](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=218s) 抓取并展示信息与真正提供可付诸行动的对齐方案之间存在巨大差距 — Bill Dwoinen > [14:54](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=894s) 它能够拉近我们与客户的距离吗？能够改善员工体验吗？ — Bill Dwoinen > [17:07](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=1027s) 寻找那些焦头烂额、被两千张 Excel 表格淹没的人 — Jim Jacob > [24:42](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=1482s) 退役或转向不再是一个‘如果’的问题，而是一个‘何时’的问题 — Preeti Vaidya > [32:04](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=1924s) 如果我们用了这个，能少用什么？ — Bill Dwoinen > [35:20](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=2120s) 究竟由谁来发起 ROI 质询 — Jim Jacob > [38:41](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=2321s) 让金丝雀喘口气 — Preeti Vaidya > [41:47](https://www.youtube.com/watch?v=SJU-FKSYEA4&t=2507s) 我们究竟在试图解决什么问题？它如何让我们变得更好、更快？ — Bill Dwoinen ## Mentioned Momei Qu, Bill Dwoinen, Preeti Vaidya, Jim Jacob, PSP Growth, Mural, Morgan Health, JPMorgan Chase, 摩根大通, Cummins Inc., 康明斯, Penny Pritzker, 凯悦酒店集团, Marmon Group, Embold, Quantum Health, Thyme Care, RAG, LLM, PowerBI, Signal GPT, ChatGPT, Salesforce ## Video description Jim Jacob - Executive Director - Digitalization, Advanced Analytics and Artificial Intelligence, Cummins Inc Momei Qu - Managing Director, PSP Growth Preeti Vaidya - Executive Director, Data & AI, Morgan Health, JPMC Bill Dwoinen - Chief Revenue Officer, Mural Too many AI initiatives deliver buzz but not business results. This panel brings together executives and practitioners who have cracked the code on achieving real ROI with AI. Panelists will explore how to identify financially sound use cases, build business cases, measure success rigorously, and iterate effectively. Expect actionable strategies for aligning technical possibilities with financial imperatives and unlocking value that actually moves the needle. Leave with a clear idea of how to pick the right use case for your business! [Source](https://www.youtube.com/watch?v=SJU-FKSYEA4)

[3:04] Bill Dwoinen: 当年顶着风雪上下学都要走上坡路 [8:33] Jim Jacob: 连接、解读、交付 [17:07] Jim Jacob: 便利性是有昂贵代价的 [3:38] Bill Dwoinen: 抓取并展示信息与真正提供可付诸行动的对齐方案之间存在巨大差距 [14:54] Bill Dwoinen: 它能够拉近我们与客户的距离吗？能够改善员工体验吗？ [17:07] Jim Jacob: 寻找那些焦头烂额、被两千张 Excel 表格淹没的人 [24:42] Preeti Vaidya: 退役或转向不再是一个‘如果’的问题，而是一个‘何时’的问题 [32:04] Bill Dwoinen: 如果我们用了这个，能少用什么？ [35:20] Jim Jacob: 究竟由谁来发起 ROI 质询 [38:41] Preeti Vaidya: 让金丝雀喘口气 [41:47] Bill Dwoinen: 我们究竟在试图解决什么问题？它如何让我们变得更好、更快？

[19:35] Jim Jacob: If you cost just xbar and you deliver x² squared, x cubed, you're going to get really a big hug. [28:03] Jim Jacob: You didn't do the change management falling in love with the new romantic solution that came out.

[0:04] Welcome. I'm really excited to kick off our panel  today around how you derive real ROI from AI, [0:10] which I know has been on a lot of people's  minds. Uh so, first introductions. My name [0:15] is Mome Q. I will be your moderator today. I am a  managing director and head of tech at PSP Growth, [0:22] uh which is an investment firm backed by Penny  Pritsker. We um my group specifically we look [0:28] at the most cutting edge AI technologies and  how it can be applied in the real world. Um [0:33] and the interesting angle that we sit on is our  company PSP u being backed by the Prrisker family [0:40] who was the founding family behind Hyatt Hotels.  Marman Group Penny herself was former secretary [0:45] of commerce under uh President Obama. And so a lot  of our DNA is actually rooted in more traditional [0:51] industries. So we sit right at the intersection  of what we think what we call what is possible [0:57] and and what is valuable. And today I'm really  excited to continue on that thread of conversation [1:03] uh with a really great group of panelists up  here on stage. So I will now pass it to them [1:08] to introduce themselves. We're really excited to  be here. Yep. Hi everyone. Thank you very much [1:13] for coming today. Uh Bill Dwinan, chief revenue  officer at Mural. I am responsible for our go-to [1:19] market and customer organizations and mural really  is at the center of driving alignment and action [1:24] through visual collaboration and excited  for the conversation today. Hi everyone, [1:30] I'm Prrii. I'm an executive director of data and  AI at Morgan Health. Morgan Health is a part of [1:37] JP Morgan Chase and we are focused on improving  equity, affordability and quality of health care [1:45] for 160 million Americans who get their health  insurance from their employers. In my role, [1:52] I'm focused on both our internal AI use cases  and also advise our portfolio companies. We [1:58] have about 13 of them on their AI strategy. So  think of like complex problem space and a lot [2:05] of real world impact with an eye towards ethics.  Excited for the conversation today and glad to be [2:11] here. Thank you Priy. Uh hi everybody. My name is  Jim Jacob. I'm from Cumins. Uh you probably would [2:19] have heard of the company maybe not but if you  think about a big truck then Cummins comes in. [2:24] But in addition to that, if you're thinking  about data centers because I think some of [2:28] you might be thinking about them recently, then  comes come in too. So we are a power company [2:34] uh one of the largest um in the world and I lead  AI and analytics for the company. I'm here with my [2:40] brilliant team. They're the brainiacs. I am they  just propped me up here. was looking forward to [2:47] uh enjoying the discussion sharing some of the  knowledge that we have come through with uh trying [2:52] to figure out ROI and I think we've we've been  successful in doing that for a few years. So love [2:58] to see how the model can be replicated further.  Thank you. Uh terrific. So I think a good place [3:04] to start um I always find it better to orient  um our minds into something really concrete. [3:10] Um so maybe the first question to kick off is can  you give an example of a reall life use case? Um [3:17] uh give us all the gritty details. What  are something some of the tangible impact [3:22] um that this AI was supposed to that AI was able  to deliver in this particular use case? Uh and [3:27] what benefits are you seeing? Anybody can can  jump in and start. I'll we can go left to right, [3:32] right to left. Fantasy football draft. Perfect.  [laughter] So I I would say for us so for me with [3:38] AI it's all about outcomes. So I mentioned I lead  our go to marketing organization and our customer [3:42] organization. So how do we leverage AI to help  drive outcomes across our teams and what I love [3:50] one of the use cases I would say I love more than  anything is how it's helping us just get closer [3:55] to our customers and prospects is today you can  just show up with a point of view. you can show up [4:01] um so much more educated and I will say I'm going  to date myself a bit like the uh you know I used [4:07] to have to walk to school uphill both ways is  as a salesp person you used to have to jump on [4:11] earnings calls or look through 10ks or all these  different things to get research and now you can [4:17] do so much to not only have a consistent point  of view to bring to your customer it also drives [4:24] consistency across the organization and so I will  say it's one of the favorite use cases that we [4:30] have and you see it it you see it show up with  how our teams show up with customers. Now the [4:36] gritty details are a few things and I think it  it comes down to number one what I see with AI [4:44] across my organization is people showing up with  more information but how do you actually turn that [4:50] into action and alignment and so we have to not  just show up to customers with AI slop or word [4:56] vomit. How do you actually turn it into something  that is actionable? And so it's not, you know, [5:00] the challenge is we're showing up almost with  too much information. Second thing is just how [5:05] do you connect to all the different systems and  tools to really have a legitimate point of view. [5:11] Then the third thing I think it reveals what  I do think that AI shows is like it shows who [5:18] wants to go from good to great. Like who wants to  go from being a you know good account team member [5:23] to a great account team member. It also shows  who's a bit overwhelmed and who are the folks [5:28] that need help using AI. Then it also shows  who's trying to take shortcuts. And I'm not [5:34] saying it kind of sniffs out the lazy people,  but in a way it kind of does. And so, you know, [5:38] I think I'm I'm really proud of that use case.  you see it showing up consistently across our [5:43] organization and it also I think also provides a  great experience for our customers and prospects [5:53] for us at Morgan Health uh similar learnings  but all of our use cases have to go back to [5:59] our mission which is the employer sponsored  insurance lens so in that vein the healthcare [6:05] benefits navigation problem is a really really  complex problem and I think it's one of the most [6:10] consequential as well. Here's what the problem  actually looks like. A patient similar to your [6:17] focus of like the customer first, right? A patient  starts with symptoms and then from there they need [6:24] to figure out where they need to go, whether this  is an urgent care problem. Do they need labs? Do [6:30] they need to go to the emergency department?  And all this is happening when they're scared. [6:36] They're confused. they still have those symptoms  happening to them. This is especially difficult [6:43] for caregivers because they're navigating all  of this for a beloved family member or somebody [6:48] that they're responsible for. And this is a big  problem especially when people don't have like the [6:55] health literacy that we expect our healthcare  system expects. And this is where AI can be [7:02] really really helpful. It can be helpful in two  ways. One is this information processing problem, [7:09] right? That no one human can process information  like AI can. There's so much information on the [7:16] internet holding all that information together  is a really good thing that AI can help us with. [7:22] So helping people understand their benefits  landscape. It's a simple retrieval augmented [7:28] generation system where you feed it the benefits  documents and a patient can get quick information [7:34] about what is covered, what is not covered, which  facility they should go to, which specialist they [7:40] can go to. All of that important piece, those  pieces of information that they need can be at [7:45] their fingertips. And then once they have that  information, and this is even more critical, [7:51] right? They need to know which is a good quality  provider, a specialist, a doctor that they need to [7:57] go to. And one of our portfolio companies,  Embold, it's now part of Quantum Health, [8:03] has actually built a chatbot that helps or like  a conversational AI tool that helps uh patients [8:09] navigate and find really good quality providers.  Another one of our portfolio companies, Timeare, [8:16] has done this for oncology. So you can imagine  getting an oncology diagnosis is really really [8:22] life-changing. So there's a lot of potential  for AI that we're seeing and some of this like [8:28] healthcare specific but that idea of like where  AI can be really helpful I think translates like [8:33] across industry. Thank you Bill and Priy that was  good. I'm going to tell you something probably [8:39] very similar but I'd like to break it down a  little bit so that I u leave maybe one two or [8:45] three things for you to think about um because I  think the entire objective of the discussion is to [8:51] give you a few thinking points uh as you go back.  So just for me to get a gander real quick, how [8:57] many of you are technology developing, you know,  solution developing kind of people, right? How [9:07] much how many of you use Google maps? [laughter]  Right? I mean a majority of us use Google maps, [9:16] right? That's going to be the theme of what I'm  going to talk to you about today. that is all of [9:22] us have a way to interact with a machine because  of Google maps. Um when Google started Google maps [9:32] started you had a bunch of students going around  mapping all these roads and writing them out. [9:38] Google purchased all of them and that was the  creation of different disperate data sets that [9:46] Google then connected together to represent it to  everybody in a very simple way. So gone are the [9:52] days I remember traveling in the US and you you  you put a map on the bonnet um or the hood and [10:01] you you map out your direction and you go. I mean  today you get in the car and if you have a traffic [10:09] jam it'll say redirecting making it very optimal  for you. In a simple sense that's exactly where we [10:17] are today. We have very sophisticated LMS Google  maps that create a translation for people like [10:27] everybody to be able to interact with machine  data in a very sophisticated way. So in the [10:33] most simplest of terms, Google created a digital  twin of that information so that we can interact [10:41] better. Bring that down to corporate America.  I'm in a company called Cummins. Like I told you, [10:47] it's a very complex company. It's a company that  is over 100 years old. So you can imagine the [10:52] multitude of disruptions, changes, evolution of  the company with acquisitions and all that changed [11:01] the fabric of data right across and I think many  of you have the same problem which is desperate [11:07] data systems data sets are everywhere. Bill talked  about connecting our entire mission in Cumins at [11:15] least the team that we are in is basically three  things. We connect, we interpret and we provide [11:22] it to the folks like Google map users inside  the company. Um there are many many projects, [11:28] hundreds of projects that we've done and generated  over hundreds of millions of dollars in the [11:32] company. One recent one is in all of our minds is  tariffs. When tariff came in, it was a nightmare. [11:41] Imagine the number of components for a company  that manufactures big ODMs, original devices and [11:47] OEM products. You got millions and millions. We  got millions and millions of uh trading partners [11:53] and then millions of thousands and thousands of  customers bringing all and and multiple locations [11:59] that have a variety of tariff implications because  of the connect technology that we use and that is [12:08] being able to use AI use graph technology and  connect deep into disparate data sets represent [12:16] the data through ontologies. We were then able  to present that data to the finance community, [12:22] business community in nine days after it came.  So if there was one component coming from Taiwan, [12:31] we were able to say that component had an effect  downstream, we could see the entire bullwhip [12:39] effect of that component right through. So we were  able to manage tariffs enterprisewide. But that [12:45] was the pieces that we did. So the three things  I'd like to leave with you as you think about [12:50] it is using AI sophistication to connect and you  don't need to bring everything together. That's [12:57] the wrong thing to do. You connect into disparate  data sets where they are using graph technology [13:03] using fin sophistication with AI interpret that  data to the functions sourcing planning finance [13:11] legal HR and then let use Google maps to figure  out what they do. So that means you're bringing [13:19] the technology right through the through the  people who are trying to use it to apply it [13:23] into their into their workspace. That's what I'd  learned. Yeah, really great insights. Thanks for [13:28] all those examples. Um I do think they're very  tangible and really great use cases. Um I'm sure [13:33] one question that the audience would love to  know given the show of hands of how many people [13:37] are builders and some of them pro some of you guys  probably are building earlier stage companies are [13:43] potentially selling to large enterprises. I'm sure  a lot of these entrepreneurs or CEOs would love to [13:48] be as embedded in a you in a such a use case that  you described. um what would you say uh is either [13:55] your filter as you look at because there's there's  a lot of AI technologies out there right now. So [14:00] um what would it take to one I guess get a  meeting with you with one of you and move [14:06] through the organization and two what is your  filter in terms of this is something real that [14:11] I want to learn more about versus this feels like  hype and I probably won't take a second meeting. [14:17] I I'll go first again. [laughter] Okay. So, uh,  the other day I I kind of So, I mean, I mean, I, [14:24] uh, I I will say, you know, a few days ago, I had  one of those moments of like, whatever happened [14:28] to the art of prospecting because so much that we  get, you know, when I get reached out to you is, [14:35] do you want a demo? Do you learn like like the all  the things you shouldn't do? And so, my filters, [14:40] and I said this the other day, I like gosh, even  for a sophisticated buyer, it's confusing because [14:45] everyone's saying the same thing. everyone  saying optimization, app rationalization, [14:50] all all the buzzwords. And so for me, it comes  down to two things. Number one, does this connect [14:56] to our priorities? So if I said, does it help us  get closer to our customers? Does it help us with [15:02] the outcomes that we're driving? Does it does it  help drive employee experience? That's number one. [15:08] Then number two is does it make us better? And  I'll give you an example. The other day we were [15:12] going through this great demo. It was showing all  this really cool stuff and the first question that [15:17] I asked it is that is any of your intelligence  grounded in us. So we invested all this time to [15:24] ground what I was describing to you in our value  pillars in our um in our methods in our the way we [15:31] think about transformation um to filter out the  things that aren't relevant for our teams. And [15:36] so when they go through, let's say, an earnings  statement, great, you know, cash is up. That's [15:41] good to know. But there's things when you're like,  "Yeah, that's not going to help me build a point [15:45] of view." And so that to me is the biggest thing  is like what we don't need is more information [15:51] because the challenge today is how do we actually  take that and drive alignment and drive action. [15:56] So those are the two main main filters for me is  like does it tie to the two or three things we're [16:02] driving? And as a CRO, it's outcomes, right?  It's so black and white for me. Like does it [16:08] help us with the outcomes that we need across  the organization? And then number two, does [16:13] it actually make our teams better? And what I'm  seeing is so much will give you more intelligence, [16:20] but it's not grounded in the things it needs to be  grounded in for relevancy. And it's going to just [16:25] be more information that I think will um create  more confusion. And the last thing is the cost [16:31] of change. And I think you all know that. I'm sure  you know that as well. and we're even seeing it in [16:35] some of our motions with customers and prospects  is that hey, if we're going to move um you know [16:42] 400 people over to something, it can't just be  cheaper. You know, it actually has to be better [16:48] and impactful. And so you are a bit more mindful  of what is that experience for the cost to change. [16:53] So in the spirit of three things, it's does it tie  to what we're focused on? Does it drive outcomes? [16:59] Does it actually make us better? or does it just  give us more information to take no action on? [17:07] Oh yeah, it's true. Actually, I actually like  what Bill said. I I think connecting the way [17:12] you connected it is fantastic. I would bring a  couple other components into it. Um I think it has [17:19] to be a burning platform, right? It it could be a  burning platform for um inventory planning, let's [17:27] say, or something that doesn't see tangible value  for the company immediately. But if it's a burning [17:33] platform that actually sees value instantly, like  tariff in the case that we talked about because [17:38] you can claw back now and all of that, it gives  you far more oomph for the project to be a success [17:46] because now you have the most senior executives in  the company right behind it. So this is critical. [17:51] It might feel trivial but it is critical to be  selfish on the projects that you select that has [17:58] a real burning platform. That's the first thing.  I think the second thing is you've got to get [18:02] the subject matter experts of those problems where  the problem resides you in in the tariff scenario. [18:10] You've got to get the finance guy who's actually  trying to deal with it with 2,000 spreadsheets [18:15] trying to figure it out. you've got to get him  or her uh working with you to be able to resolve [18:21] that problem. That's when you see the difference  because the p the business leader or the function [18:27] leader who wants this resolved you're a fresh  face if you're a technological producer but [18:34] the person the controller or the finance person  who is helping the person giving them the right [18:38] credible information that's the first that's the  first person. So you get that individual connected [18:44] with the um with the technology, give him a Google  maps and you are ready to roll. I'm sorry to use [18:51] trivialize it, but the fact that we use LLMs  today, very costly LLMs, by the way, I'd say, [18:58] you know, and use that to translate our natural  language into machine language and come back out. [19:06] It's very convenient. And what what do we know  about convenience? it costs, right? And so we [19:12] know that tokens are going to keep coming up. But  if you don't want to do that and you want to be, [19:17] you know, economical for your company and resolve  it, you know, faster way, then I think the best [19:23] thing is to get the subject matter expert, create  an ontology, do the job. You might use some LLM, [19:30] you know, conversions and translation at the  point, but you're not going to be running with it, [19:35] you know, on a daily basis using LLM. And that's  that's something to keep in mind because you [19:40] can we have we have you know many companies are  probably seeing this now but we are also looking [19:46] at this um the cost equation if the function  of x means the result of a solution is going [19:54] to going to be way more expensive to build than  the solution you're building your business folks [20:01] are not going to be behind you right if you cost  just xbar and you deliver x² squared x cubed, [20:09] you're going to get really a big hug. That's  the way it it it'll work for us. So pretty. [20:16] Yeah. Uh this it's a really packed question.  Thank you, Mo. [laughter] And I'm sure Yeah, [20:22] there's more. Okay. Uh and um I think perhaps like  the the reason why all of us are in this room. Um, [20:31] so maybe I'll take a perhaps a different lens  because at Vorggan Health, um, we also have an [20:37] investments team and we are investing in companies  that again are like kind of aligned to our [20:41] mission. Since 2021, we have invested nearly $300  million in companies that are moving the needle in [20:48] healthcare. And I kind of advise on companies that  have like a technology component, an AI component, [20:55] a data component. And as you all can imagine,  there are a lot of companies now that have an [21:01] AI component. And a lot of that is perhaps AI just  for the sake of AI because the sales team said so. [21:11] Uh the investor is expecting it and you perhaps  not even like reach the pitch if there is no AI. [21:17] So that is the reality that we're in today. Going  back to like what it takes to get a call back. [21:24] um it is the it's the it's definitely not AI for  the sake of AI, right? I think we've all seen [21:31] that come across our desks and that is problematic  because that erodess trust if it's just AI for the [21:38] sake of AI. So that stakeholder trust, that user  trust that you need to build and it is problematic [21:44] because it might not necessarily lead to better  outcomes for your customers, your patients, [21:49] whatever it is that you're trying to drive. So  that's something that like we definitely keep an [21:54] eye on making sure that it's really driving value.  So what's not a good solution, right? If everybody [22:00] in the room is excited about AI as a technology  and not the problem that is being solved, [22:08] that's a problem. You really need to start with  the workflow and the solution that you're building [22:15] and how AI kind of integrates into it instead  of wow I have this shiny new AI capability to [22:21] borrow that Google maps analogy I have Google  maps now now what can I do with Google Maps is [22:27] not going to get us very far right it's what are  my core problems what are the things that we need [22:33] to solve going back to that healthcare example a  nurse in the ER is not going to be very happy with [22:40] the Genaii chatbot because that takes her away  from her patients. You as a patient also don't [22:46] want a nurse who's like having conversations  with a conversational AI tool while you're [22:51] probably needing attention and getting waiting for  triage and all of the things that are critical in [22:56] in that setting. So knowing your workflow is  very critical. Flipping the lens a little bit, [23:04] build versus buy, right? that that decision what  is that ROI statement there so there's like what [23:11] it boils down to is really like five rules of  the road or five five big rules right one is [23:19] is this core to our business problem what is this  bringing us is this something that can be solved [23:24] with a generalized tool or do we need something  specialized and then once you have that answer [23:31] what is what is the opportunity cost of building  it in inhouse right If there's a solution that [23:36] can solve it reasonably well, you must justify  the cost of building it internally. Next, is [23:43] it going to be even if you're buying, right? Can  this solution be integrated into our workflows? [23:49] If you're buying, the fourth point is can we trust  the outcomes? Can can we see under the hood? And [23:57] then fifth point which is the most important point  can we trust this vendor to be a long-term partner [24:06] in enabling AI for us with the rapid pace of AI  innovation. We need to make sure that the vendors, [24:15] the solutions that we're building can be partners  in enabling long-term AI solutions for us, [24:23] integrating and innovating in the space and then  also making sure that there's real value unlocked [24:30] through that council of like stakeholders, finance  people all in the same room. That's I think what [24:38] unlocks the value. So that's those are the five  things that it boils down to for us. Yeah, it [24:42] makes a lot of sense. Um, these are great examples  and hopefully u you guys all found that helpful [24:47] too. And I think most of what we talked about  until now touches on the benefit side, right, [24:54] of what AI um can drive. And I think the other  part of the equation, right, is the cost and that [25:00] goes into the ROI. Um, I already heard the word  token set. So, I think you guys are dying to go [25:04] in that direction anyway. Uh, so let's let's talk  about that. And I think one um question that I've [25:09] heard pop up more recently than than before is how  do you measure the total cost of ownership for AI? [25:16] Uh I think we went into AI being there's this is  so much more efficient. This is faster and and in [25:21] some ways that's true but there's a lot of hidden  cost too in terms of when you think about compute [25:26] which is rising significantly when you think about  maintenance when you think about everything else [25:30] that goes into making AI work. Um I guess first  how do you even measure um what the actual cost [25:35] is? Maybe PT, I'll start with you. Sure. Well,  um [laughter] um so I think um the one thing that [25:47] um we weren't doing before in our I think so  AI technology is like any other technology and [25:56] everybody is I think familiar in this room.  Uh we are no strangers to the total cost of [26:00] ownership calculations. five year horizon however  like whatever is your technology horizon you do [26:07] the total cost of ownership I think one thing  with AI that is different there and I don't see [26:14] AI technology adoption very different from like  the waves that came before the cloud journeys and [26:20] we've we've been here before right the thing  that's different with AI is the rapid pace of [26:25] innovation and one thing that needs to come to  the forefront in those calculations is the cost [26:33] of decommissioning or pivoting to a different  solution. That can be hard to imagine at the [26:41] forefront of a project, right? As a builder,  you're inclined to you've like it's your it's [26:47] your brainchild. You've spent so much time on it.  You've invested it. You want it to be evergreen. [26:52] I'm a builder first. I've been there, right? And  so, but with AI, with the rapid pace of innovation [26:59] that we're kind of the era that we're living in,  this decommissioning is not an if situation. It's [27:06] a when question now. And so, we really need to  start I mean my my two cents here are really [27:12] need to start thinking about what is what will  it take to pivot to a different solution? What [27:18] will it cost to move to a different vendor or  build a solution internally? I think the answer [27:26] ultimately is not like when today when we're  deploying solutions it's not to get a perfect [27:31] AI solution out there because everything is not  aligned right like not all your data is in the [27:36] same place or not the tool the sophistication  of the tool is also like not where it will be [27:42] in like a matter of months. So keeping that into  account, having a realist view into what it takes [27:50] to pivot from the solution or innovate with the  solution should be baked in so that there are no [27:55] surprises and you're thinking of your solution  evolving with time as well in that way. Well, [28:03] I can go next. Yeah, perfect. Priy, I think you uh  you explained that well. A couple of things that I [28:09] might add into this is um we use three criterias.  I love threes. Uh three criterias. Number one, [28:20] it has to um it has to be the most sophisticated  solution that you can get. Why? Because we are in [28:27] a capital market and you can select the best  thing you want, right? Uh economics becoming [28:32] a different problem. But number one is you  figure out the best technology for yourself. [28:37] Number two is it should be you should be able to  adapt that technology in your company fabric and [28:43] architecture and structure might might sound  trivial but it's something very important to [28:48] consider. And number three is economics.  If you have these three things worked out, [28:53] you now have the first of the equation taken care  of which is you're going to pick out a technology [28:59] that will work for you that is adaptable and  that's economically viable for you. That's the [29:03] first thing. But now comes the piece that u that  pretty also talked about which is you will realize [29:11] this you're going into AI you're getting into the  journey you've invested heavily into it you're [29:16] seeing some value into it now comes the part about  bottomline dollars for the total cost of ownership [29:22] guess what you didn't let go PowerBI I'm sorry  I didn't mean it like that but you didn't let go [29:28] the analytics user technologies that you're using  before you are having a new technology adding on [29:38] to the technology landscape that you already had.  So, no wonder you're going to have a massive cost, [29:43] right? No wonder you're going to have additional  pieces that you're considering because you didn't [29:49] do the change management the way you're supposed  to do it. You did the change management falling [29:54] in love with the new romantic solution that came  out, right? So, now so, so you got to think about [30:00] this really seriously. Once you figure out  what's your best ROI uh solution to work on, [30:07] the next thing is for you to get the technology  or the partner to build it on and then after that [30:11] you've got to figure out how are you going to  get away from legacy cost of that solution you [30:17] have to the new one because by definition if  you're using AI you're going to improve value [30:22] you're going to improve efficiency you're going to  improve uh productivity by definition if you're in [30:27] if you're investing in technology the outcome of  that for a corporation is that you're going to get [30:32] more value out of it just by definition. So you've  got to figure that one out. But I think there's [30:37] another key point uh that was mentioned too that  is do you build internally or buy outside today [30:47] it's chaotic because every vendor who calls me  tells me that they can solve the problem fine of [30:54] course they that's how they open the door I get  some of them could probably do that but the more [31:00] important part is could you build an ontology  internally uh connection scenario internally [31:08] solve the problem internally and own the IP and  the differentiation of that internally versus go [31:14] and get now a third person who is going to be a  part of this journey that you don't know whether [31:21] it's going to be parasitic for your journey it's  you don't know whether it's going to reduce the um [31:26] you know the ROI value proposition going forward  because by definition right you have another [31:32] partner they want to make money you want to save  money and that that doesn't work very fluidly. [31:38] So build versus buy is in a very different point  today. Why? Because you can have translated LLMs [31:46] helping you resolve machine problems. So you  might not need the whole team, but you might [31:51] need a very solid platform for you to be able to  figure that out and then bring it into production [31:56] and then move economically um um with a much  different footprint. They're too hard to follow [32:04] and I think when I reached out to you I said I can  help you solve the problem right by the way just [32:08] so you know. Um so I'll I'll um I want to keep it  quick because there's so many great questions and [32:12] we have about 12 minutes left. So I I think for  me bringing it back to ROI there's your point [32:17] about the speed of innovation and AI we we did  it backwards right usually when you buy software [32:24] you go through an evaluation right you have the  use cases scoped out you have the business value [32:28] narrative you have everything prepared so you  know what you're measuring with us it was one [32:34] day we have AI licenses go and the diversity of  use cases aren't so easy to measure and when I I [32:40] just actually looked at our top 25 users and the  diversity from functions is wild. I love seeing I [32:48] mean I think I'm number three in our organization  and if you look across GTM systems or CSMS or [32:56] um our EAS are using it. It's just the use cases  it's not as straightforward as looking at ROI on [33:03] workday or looking at ROI on a tool where you kind  of scope it the right way. And so the ROI part is [33:11] uh challenging and we're doing it the other way  around. So now we're trying to go through it and [33:14] say, "Okay, let me go talk to these 25 people.  What are you doing? How is it helping you make [33:19] decisions? How's it helping you move faster?  And is it single player or multiplayer?" And [33:24] so we're investigating there. And the last thing  I'll say is the biggest difference that I've seen [33:28] when I approach we have a great CFO and I would  say that if he was in the room as well. And what [33:34] I have seen when we look at different solutions  is the first question he will ask is what could [33:40] we use less of? And so the whole idea of cost of  ownership is beyond just the what you're looking [33:46] at. The example I gave earlier about the tool or  the the you know endtoend system that was talking [33:52] about signals and all these different things I can  share with you. The first question was well would [33:57] that take your team out of would you stop doing  the we call it signal GPT? We created a GPT that [34:04] runs all the research, all the cool things to uh  make us show up a bit more intelligent. And the [34:09] answer was I don't think so. And so it wasn't  necessarily the cost of ownership of that tool [34:14] or that solution. It was it would not reduce our  dependency on something else because everyone's [34:21] going through app rationalization and not in a  way that's just to save money. It's like if we're [34:26] using this, how do we stop using that? And from  a GTM standpoint, everyone sounds the same. And [34:32] so that's the biggest thing I I've seen show up is  not just that solution itself, it's what would you [34:37] use less of if you're using more of this. Yeah.  And maybe taking on the theme of measurement for [34:44] a little bit too. What else goes into how you  measure success for AI? Um and we're I'm taking [34:50] this maybe a step further is uh let's say um  you know in the land of all this AI competition, [34:57] someone does break through and now they have a  pilot. Um now the the issue now translates to [35:03] how do we become a pilot into something in more  production and there's I think right a new set [35:08] of thought process and equation that goes into  that because pilots tend to be in more confined [35:14] environment and there's a lot more that goes  into whether that's successful even for companies [35:20] um that might be able to show a positive number  on a kind of on a smaller scale. So um what makes [35:27] you know something from an experiment maybe a  successful experiment uh to something that becomes [35:32] embedded maybe going going back to the first  question becomes a poster child that you guys [35:36] would talk about as the first use case. Well I'll  go first then um real quickly um you know a laptop [35:45] how do you measure its value? It's been an age-old  problem for technology folks who make technology [35:54] their livelihood to convince the business guys  that um look you've got to invest on this and [36:00] you're going to get a benefit out of it. goes back  to the point I made if you invest in technology [36:04] in the company the result should be productivity  improvement efficiency given right so what we did [36:11] uh at Cumins is um very very early on we asked the  question of who's asking this question first thing [36:20] uh in most cases it's a business person it's the  finance person it's the you know it's the business [36:25] side leader who asked the question technology  builders don't really look at it in terms of [36:31] well I created this much of value and even if you  say that they won't believe you because they'll be [36:36] like yeah sure you were one part of the solution  and somebody else did all the work right so we [36:43] created this dashboard I mean straightforward  dashboard to be able to um uh measure very well [36:50] with very strong KPIs but the ownership of the  dashboards was the most important it went into [36:55] the controller of the businesses so if a a problem  comes up comes from the business. The controller [37:03] is involved. Our team is involved. The controller  manages what goes in there. Because here's what [37:08] you want to avoid. You want to avoid the situation  where the company CFO and the business leader has [37:15] an agreement on a number that the business leader  is going to contribute to the company. That's how [37:20] financial plans are made. Your 10ks, 8Ks,  etc. If you resolve for a problem and you [37:27] generate a $10 million of benefit for the business  leader, it is right for the CFO to come and say, [37:34] challenge the business person, say, "You got  $10 million more from corporate services. I [37:38] want that $10 million contribution income."  Right? It's fair. It's a fair discussion. [37:43] The business leader is not happy about this  because the business leader is now with you [37:48] saying why on earth did you go and you know expose  that information to the finance community. So all [37:55] that ruckus happens in finance because that's  where the beans are counted. So we thought the [38:02] best thing to do was to go to the controller let  him resolve the problem and he or she is able to [38:07] tell us look you are creating value. We are very  blessed with the controller we have um and we get [38:15] the financers to be able to invest in what we're  trying to solve for because of that. And they're [38:21] seeing tremendous value coming in. But you've got  to stay away from the debate where your value is [38:28] going to kind of disrupt the P&L agreements and  gen you know that agreement uh is a very delicate [38:35] one holy grail stuff. So I think if you stay  inside those frameworks you you're you're good [38:41] to go. Yeah, I agree. Um I think one of the things  that we have done I I loved how you described this [38:48] in terms of like that whole landscape of like  point conversations happening in silos. One of [38:55] the things that we've done that's worked really  well when scaling or even you know this happens [39:01] even before you're piloting your solutions is  making sure that technology finance and business [39:10] stakeholders are in a room when this conversation  in the same room when this conversation is [39:15] happening. Technology making decisions without  business stakeholders leads to a tool that nobody [39:21] uses. business stakeholders making decisions  leads to tools that are unrealistic, infeasible, [39:29] things that technology will probably never agree  to if they were in the room. So having all of [39:33] these people in the same room, these stakeholders  agreeing on what's feasible, what's going to drive [39:40] value, and how much it costs from the get-go is  super important. Point conversations don't help [39:48] though. Um it's it's really that like coming  together of the stakeholders that is I think [39:53] my two cents that I would advise. The other  piece of this as a technologist since most of [39:59] us in this room are technologists perhaps leaders  bringing these solutions to our teams moving the [40:05] needle forward for our organization. for you.  Acknowledging that going from pilot to scaling [40:13] is the most difficult stage of taking any solution  bringing it to the whole enterprise. Acknowledging [40:21] the risks early and often is super critical.  And then the second piece of advice is your [40:29] pilot worked because as technologists you got a a  set of canary users. These canary users were very [40:36] motivated. They had a level of AI literacy that  led them to be the Canary users that were carrying [40:42] your pilot forward. When you're going from pilot  to scaling, you need to let the canary breathe. [40:50] So, your canary needs a break. Your canary is not  your model user. When going from pilot to scaling, [40:57] you need to get in more people in the room  that represent your, if you think of a normal [41:02] distribution, your median user. The types  of skills that this user needs, training, [41:08] upskilling, that is a cost. Both time and  financial implications come out of that. That is [41:16] not an afterthought. That when going from pilot to  scaling is very different. You're not proving the [41:20] concept. you're bringing the whole team together  in using a solution, scaling it up so that you [41:27] become AI native. So let your canary breathe  is my advice to you. Um and the more you can [41:35] pair skeptics with enthusiasts will help because  both of those perspectives are really valid in [41:42] getting a solution to be widely adopted at the  enterprise scale. I'll I'll bring us home and [41:48] I let the canary breathe. I love that one. And uh  so what I'll say is is kind of more of the same, [41:52] right? And and there was a data point a few years  ago that showed that decisions around software [41:57] and technology went from two people involved to  13 people involved because there's a danger on [42:03] either side of it, right? If it's or centralized  decisions without the business, the business is [42:08] like what the heck then especially in sales  orgs, we're guilty of this. If we need a tool [42:12] for everything, like that's dangerous, too. And  so what's that middle ground of bringing in all [42:17] the right people? And it goes back to what I said  a few questions ago is it has to start with what [42:23] are we solving for when I use AI or look at data I  always start with a point of view of I'm trying to [42:31] figure out X or I'm trying to solve for Y and it's  the same thing with when you approach a pilot of [42:36] what are we truly trying to solve for. I think the  biggest thing, if there's one thing that I hope [42:40] you take away from this part of it, one thing that  we're trying to do different is not pull in power [42:47] users to pilots. We want to pull in the skeptics.  Yeah. I want to pull in the 10 people not using [42:54] things today to figure out why. So, of course, if  you have your top 10 super users of tool X or you [43:01] go through and say, "Hey, who's using AI the most?  Let's let's get their point of view." I want the [43:05] people not using it. I want the people that have  that point of view because that's when you really [43:10] get the feedback that you need to bring people  along. And I think what I'll end on is that same [43:14] concept of it all starts with what are we trying  to solve for? How do we make sure we not check all [43:22] the boxes? But now more than ever, the governance  of these tools is next level. you know, turning on [43:28] the uh for us in the go to market organization,  turning on the ability to uh through chat GBT [43:35] source salesforce, that was a lot for a pretty  small, you know, uh small but mighty IT team. And [43:41] so we have to be mindful of these sorts of things.  And that's where I've seen organizations even slow [43:46] down. It's like the way that you govern, the way  that you provision, the way that you monitor has [43:50] changed uh significantly. And so it always comes  back to what are we trying to solve for? How does [43:56] it make it better? How does it make us better  and faster and how do we bring everyone along [43:59] with us for the journey? What a great place to end  on. We have oh and five seconds early. Um thank [44:05] you everyone for coming and please give another  round of applause to our panelists. [applause]