UnitedHealth, TIAA Chiefs Say AI Payoffs Depend on Process Overhauls

UnitedHealth与TIAA高管：AI回报取决于业务流程重构 只有在梳理运营工作流并在采用基础模型前彻底改造遗留数据层，生成式AI才能带来可衡量的企业价值。 在Alain Biem的主持下，来自UnitedHealth Group、TIAA、Cox Communications和Odin AI的企业AI负责人坦诚分享了在受监管的复杂环境中部署生成式与智能体AI的实践教训。与会嘉宾强调，应去除对AI“魔法”的盲目崇拜，将其转化为结构化的系统工程，并将数据本体和业务流程梳理置于单纯追求原始模型能力之上。他们详细介绍了从自动化财务差异分析、联络中心辅助到环境临床记录等实际落地案例，同时倡导采用模块化架构及“适度博采众长”的多供应商策略。归根结底，企业要取得真正成功，需要严格的时间审计、持续的可观测性，并将AI深度嵌入日常一线工作流程，而非仅止步于孤立的试点项目。 破除对AI“魔法”的虚幻预期，关键在于将底层业务流程系统性解构成模块化的多Agent工作流；实际复杂的企业级架构通常是由约10个可组合的子Agent协同工作，而非盲目堆砌数百个独立Agent。 (3:09) 在医疗与金融等强监管领域，负责任的AI治理需直接内嵌于部署代码中；以医疗场景为例，环境临床听音（Ambient Listening）与自动化预授权交换正大幅缓解医护人员倦怠，并化解约80%由行政流程造成的延误。 (12:39) “婚礼与婚姻”的隐喻揭示了AI落地现实：构建原型如筹办婚礼般迅速且令人兴奋，但长期与非确定性输出的系统共存则如漫长的婚姻，需要持续的组织变革管理并借助能力评估矩阵防范“金锤综合征”。 (20:16) 企业在多模型选择上应保持“适度开放”以规避供应商锁定；基础模型与外部框架皆为可替代的通用商品，企业专有的私域数据、业务流程、语义层与分类本体（Taxonomy）才是真正持久的核心资产。 (28:45)

Deploying generative AI across large enterprises requires shifting focus away from model hype and toward disciplined workflow re-engineering, internal data readiness, and ongoing operational change management. Demystifying AI into Modular Business Workflows (3:09) Enterprise clients frequently approach generative AI expecting effortless automation, but practical deployment requires breaking down business processes into modular, systematic agent architectures. Russell LaCour reported that Odin AI applied this approach at telecommunications company Airtel, where an 80-person finance team's month-long variance reporting across tens of millions of transactions was reduced to a few hours while exposing hidden revenue leakage. Eric Pace noted that while vendors often claim to deploy hundreds of agents, most robust production systems actually rely on roughly ten modular, composable sub-agents operating under a unified structure. Data Foundations and Modular Vendor Strategy (25:00) Successful AI deployment hinges on internal data architecture rather than external model providers. LaCour and Geeta Pyne stressed that organizations must structure internal standard operating procedures, documentation, active data pipelines, metadata, semantics, and ontologies before agentic systems can deliver sustainable value. While legacy structured database modernization remains a multi-year hurdle, Pace highlighted that generative models provide immediate returns on unstructured data, such as querying millions of customer service call transcripts or performing 100% audits across historical files rather than relying on small statistical samples. To balance model capabilities without vendor lock-in, the panel advised maintaining in-house engineering for proprietary integration while treating foundational model partnerships with calculated flexibility. Enterprise leaders should run parallel tests across providers like OpenAI, Anthropic, Google, and Microsoft, allocating capital across ecosystems because model capabilities fluctuate rapidly while foundational providers frequently pivot their business models. Risk Governance and Architecture in Regulated Sectors (9:27) In highly regulated environments like financial services and healthcare, AI risk must be managed systematically through platform controls and human oversight. Pyne explained that TIAA established a coded 'Generative AI Technology Gate as a Platform' to enforce corporate compliance and access controls directly within deployed code. Starting with low-risk tasks such as public data ingestion, document summarization, and data wrangling, TIAA's lean enterprise architecture team automated 80% to 90% of routine data quality analysis, allowing human architects to address the remaining edge cases. Incoming use cases are evaluated against an enterprise capability map and a 2x2 value grid to decide whether to custom-build differentiators or buy commodity tooling. In healthcare, UnitedHealth Group organized AI implementation across four operational verticals: Consumer, Care, Claims, and Engineering. Ambient clinical scribing tools have achieved rapid adoption to counter physician documentation burnout, while automated two-way data exchange targets prior authorization workflows, where an estimated 80% of delays stem from administrative and documentation bottlenecks across a $4.7 trillion national healthcare system. Change Management, Trust, and Success Metrics (20:16) Operationalizing generative AI presents challenges distinct from classical machine learning because generative outputs are non-deterministic and context-dependent. Alain Biem noted that automated summaries that satisfied credit risk analysts were rejected as useless by ESG analysts, demonstrating that traditional mathematical metrics like Area Under the Curve (AUC) cannot gauge generative success. Pyne emphasized that sustained user adoption relies on strict data trust; early hallucinations can destroy confidence, necessitating phased rollouts to targeted user cohorts alongside continuous observability stacks to track data drift. Addressing the operational reality where an estimated 4% of companies have extracted tangible value from agentic AI, Pace and LaCour advocated conducting thorough time audits to avoid spending engineering months automating low-impact tasks. In a representative five-week proof-of-concept timeline shared by Pace, two weeks were allocated to process mapping, two weeks to agent development, and one week to change management. At enterprise scale, UnitedHealth leadership framed ultimate success as expanding tooling beyond the ~40,000 technical staff to unburden the broader 400,000-person frontline workforce, holding initiatives to transparent financial pull-through and an ethical standard where leaders can speak about deployments with unreserved pride. Enterprise generative AI maturity requires moving past small pilot projects to unburden operational workers across entire organizations. However, sustained business value remains constrained by how effectively legacy enterprises master unglamorous data foundations, maintain trust amid subjective model outputs, and drive continuous cross-functional change management.

在 Ai4 的研讨会上，来自医疗、金融与电信行业的高管与技术专家指出，生成式人工智能（GenAI）的落地已从最初追求“魔法般”的预期，转向系统工程、流程解构与严格的治理框架。发言嘉宾强调，技术编码仅占部署工作的一小部分，企业成功的核心在于梳理内部数据资产、推行持续的变革管理，以及在摆脱单一供应商锁定的同时保持底层架构的自主掌控。 解构魔法：从流程梳理到模块化智能体 (3:09) 多位专家指出，许多企业对生成式 AI 的初始诉求往往脱离实际，寄希望于模型能不加干预地直接解决复杂问题。Odin AI 联合创始人兼首席技术官 Russell LaCour 表示，工程落地的关键在于祛除这种“魔法感”，把既有业务流程拆解为模块化、系统化的智能体工作流。若业务本身的流转逻辑未被清晰定义，AI 并不能凭空填补组织空白。以电信运营商 Airtel 为例，其以往的财务偏差分析需 80 名分析师耗时近一个月汇总上千万笔交易；通过多系统数据接口与多智能体编排，该流程缩短至数小时，并直接挖掘出潜在的收入流失漏洞。 Cox Communications 人工智能主管 Eric Pace 同样认为，不同规模的企业当前面临着相似的技术挑战，而行业常吹嘘的“部署上百个智能体”，实际上往往是由大约 10 个模块化子系统构成的复合体系。在软件开发与内容创作等场景中，LLM 正被用于需求梳理、代码编写与营销文案生成。Pace 提及一个反直觉的现象：在分析型任务中必须极力防范幻觉，但在创意设计场景下，适度的发散反而有助于拓展创意边界与压缩迭代周期。 治理与架构：从单点用例走向能力矩阵 (9:27) 在监管严格的金融与保险行业，直接推进复杂模型存在高度合规风险。TIAA 高级常务董事兼首席架构师 Geeta Pyne 介绍了“能力中心视角”（Capability-centric view）与“生成式 AI 技术门禁平台”的实践路径。该机构管理着约 1.4 万亿美元资产，面临海量法律与合规文档审查，其策略是先从公开数据及基模型切入，把企业风控政策与访问权限以代码形式固化在技术网关中，随后再扩展至私有资产，并始终保留“人在回路”（Human-in-the-loop）机制来攻克最后的业务校验。 为了避免各部门将所有问题都视作锤子下的“钉子”，Pyne 提倡借助 2x2 机会矩阵对用例进行客观评估，区分核心差异化能力与通用标准工具，从而明确自主开发、采购还是合作的决策树。与此同时，该团队利用生成式 AI 处理了 80% 至 90% 的底层数据清洗与质量分析，使精简的架构师团队得以专注于最后一公里的疑难治理。而在医疗垂直领域，UnitedHealth Group 的架构则涵盖客户、临床、理赔和工程四大板块。在每年支出达 4.7 万亿美元的美国医疗体系中，约 80% 的预先授权延误源于资料缺失等行政瓶颈；通过实时双向系统对接与环境录音（Ambient listening）转录，企业不仅缓解了医护人员的行政负担，也逐步化解了长期遗留的数据孤岛壁垒。 供应商与数据：保持适度多元并筑牢内部基底 (25:00) 面对 OpenAI、Anthropic、AWS、Google 和微软等基础模型与超大规模云厂商的快速更迭，小组探讨了避免厂商锁定的合作策略。与会嘉宾提出，在模型层面应当采取“适度多元”（Appropriately promiscuous）的策略，同时在多个相互竞争的生态中进行测试与资金分配，以规避押错技术路线的风险。由于底层模型提供商经常在水平平台、垂直应用及咨询服务之间调整重心，企业必须对自身的成熟度保持透明，并对核心资产保有独立掌控权。 这种外部灵活性必须建立在扎实的内部数据准备之上。Russell LaCour 强调，外部模型与框架瞬息万变，但企业专有的业务流程、标准作业程序（SOP）和制度文档才是不可替代的核心数据资产。Geeta Pyne 亦补充道，若缺乏对元数据层、语义本体和企业分类法的底层投入，智能体工作流将迅速遭遇效益瓶颈。此外，Eric Pace 提到，面对庞大结构化数据治理周期长的难题，生成式 AI 可在非结构化文本上快速发挥作用，例如通过调取全部通话记录提取客诉趋势，或对历史交易开展 100% 全量审计，彻底颠覆依赖人工抽样的传统审计模式。 度量与落地：超越传统指标并向全员渗透 (32:16) 在生产环境中衡量 GenAI 的成效不能仅依靠传统的数学指标。前 S&P Global 机器学习主管 Alain Biem 指出，与传统预测性 AI 依据 ROC 曲线或 AUC 进行客观评估不同，文本生成与摘要的质量高度依赖终端用户的业务场景；同一种摘要在信用风险分析师和 ESG 分析师眼中的评价可能完全相反。Geeta Pyne 强调，长期的用户采纳度与信任才是根本，早期的失真或幻觉会迅速瓦解用户信心，因此必须通过小规模试点验证数据准确性，并在架构中集成持续的可观测性（Observability）堆栈以监测系统漂移。 针对行业中仅有 4% 的企业从智能体 AI 中获取明显商业价值的现状，Eric Pace 和 Russell LaCour 认为团队应摒弃追求短期轰动效应的投机心态，通过工时审计挑选真正消耗数千工时的流程进行改造。在 Pace 分享的一个为期 5 周的概念验证（POC）案例中，业务流程评估占用 2 周，智能体工程实现占 2 周，变革管理占 1 周，表明技术开发仅占总体耗时的一小部分。此外，针对大型企业级转型，发言者提出了三项核心准则：通过经得起媒体和公众审视的“纽约时报测试”确保伦理合规；实现覆盖数千万至上亿用户的规模化效应；以及穿透组织层级，不仅赋能技术团队（约占 10%），更要真正减轻剩余 90% 一线运营人员的工作负荷，推动企业整体的流程重塑。 生成式 AI 在大型机构中的落地并非单纯的算法升级，而是一场依赖清晰流程梳理、严格数据治理与深度组织变革的系统工程。企业唯有在自主掌控核心数据资产的前提下保持技术架构的开放性，并建立以用户采纳与实际工时节约为导向的度量机制，才能跨越试点陷阱，实现可持续的规模化效益。

## Background 本内容由 Ai4 于 2025 年 8 月 27 日发布在 YouTube 平台上。视频记录了一场名为《行动中的生成式 AI》（Generative AI in Action）的圆桌讨论，汇聚了来自哥伦比亚大学、联合健康集团、TIAA、考克斯通信及 Odin AI 的高管与专家，重点探讨生成式 AI 与大模型在各行业中的实际落地案例、执行策略、治理机制以及规模化推广经验。 ## TL;DR 来自医疗、金融、电信及 AI 解决方案领域的领军人物深入剖析了生成式 AI 从“概念魔法”走向“工程系统”的真实路径。嘉宾们分享了财务差异分析自动化、创意营销、智能客服 Agent、合规网关构建及医疗环境聆听等具体场景，并指出企业落地 AI 时最大的阻力往往在于遗留系统集成与变革管理。讨论还确立了以“核心能力透镜”、小步试点（Cohort Testing）、全量审计及赋能基层非技术员工为核心的价值评估决策框架，强调企业唯有筑牢自身数据底座并保持透明务实的文化，才能实现 AI 的真正规模化落地。 ## Atmosphere & style 整场圆桌氛围务实、专业且充满机智幽默，嘉宾们用“婚礼与漫长婚姻”、“恰如其分的开放”等生动隐喻化解了复杂的技术治理难题。讨论既有拆解行业营销泡沫时的辛辣直率，也有面对医疗系统痛点时的深刻使命感与自省，真实再现了一线技术决策者在狂热技术浪潮下保持清醒、脚踏实地推进工程变革的严谨质感。 ## Key takeaways - [03:09](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=189s) 生成式 AI 的落地高度依赖行业、数据成熟度与组织文化，不存在一刀切的通用策略；首要任务是将客户的‘魔法幻想’拆解为确定性、可自动化的工程系统。 - [06:03](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=363s) 企业在 AI 落地初期面临的基础挑战高度相似，企业规模并非关键差异点；当前落地的重点场景包括容忍幻觉的营销创意、严格受控的联络中心以及研发全生命周期。 - [08:00](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=480s) 市场上宣称部署数百个 Agent 的说法大多是营销话术，企业底层通常只是构建了由十余个可组合子模块构成的复合系统。 - [09:27](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=567s) 强监管金融机构应优先利用 GenAI 处理法务与合规等‘枯燥苦差活’，通过建立策略网关代码化合规规则，并始终保留‘人类在回路中（HITL）’。 - [12:58](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=778s) 医疗领域的 GenAI 落地聚焦消费者端呼叫分流、医护端环境智能监听（Ambient Listening）、事前授权与理赔防欺诈、以及研发工程提效四大垂直支柱。 - [17:36](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=1056s) 许多 AI 突破的实质并非神秘魔法，而是利用大语言模型作为新界面，解决长期拖延的跨系统整合、业务流程梳理与底层数据治理工程。 - [20:16](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=1216s) 传统机器学习输出确定，而生成式 AI 充满不确定性；构建 AI 像短暂的婚礼，而真正的挑战在于漫长且艰难的变革管理与持续共存。 - [22:10](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=1330s) 企业架构应通过‘业务能力’与‘AI 擅长能力’的双重透镜筛选高价值用例，并建立明确的自研与外购（Build vs. Buy）决策树，避免将工程算力浪费在非核心商品化场景。 - [25:32](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=1532s) 面对底层模型大厂的频繁迭代，企业应采取‘恰如其分的开放’策略：保持极端透明与积极合作，同时进行多方押注以避免单一供应商锁定。 - [28:45](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=1725s) 基础模型会不断更迭，企业真正的核心护城河是自身的数据与业务流程；在深入 Agent 之前，必须先做好元数据、语义分类及非结构化数据的梳理。 - [32:15](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=1935s) 衡量生成式 AI 成败的核心指标正从传统算法数学指标（如 ROC/AUC）转向‘用户采纳度与信任’，需要通过小群验证与全生命周期可观测性来建立信心。 - [36:10](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=2170s) 在 AI 项目与 POC 实施中，编码与开发往往只占少部分时间，大部分精力必须投入在业务流程定义、痛点对齐与组织变革沟通上。 - [39:36](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=2376s) 大型企业成功推进 AI 依赖四大支柱：通过‘纽约时报测试’的真实效用、自下而上穿透的利润底线、数千万级的宏大应用尺度，以及走出 IT 象牙塔、赋能 90% 非技术一线员工。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=4s) 圆桌研讨开场与嘉宾背景介绍 主持人 Alain Biem 开场介绍了本场主题《行动中的生成式 AI》（Generative AI in Action）。从现场座无虚席的景象可以看出，生成式 AI 是当前技术与商业领域最炙手可热的话题。Alain 首先介绍了自己的经历：他曾任纽约人寿保险（New York Life Insurance Company）首席 AI 官、标普全球市场财智（S&P Global Market Intelligence）AI 负责人以及 Wayfair 机器学习与数据科学负责人；早年他在 IBM 担任科学家，参与 IBM Watson 平台的早期算法研发，目前还在哥伦比亚大学担任兼职教授，为研究生讲授机器学习与分析学。 随后，在场的几位跨行业专家依次进行了自我介绍： - **Dr. S. Mahin**：联合健康集团（UnitedHealth Group）首席 AI 转型与战略官。加入联合健康前，他曾在拜登政府负责医疗卫生领域的 AI、数据与网络安全事务，此前还曾担任麦肯锡管理咨询顾问，其本职背景是一名内科医生。 - **Geeta Pyne**：美国教师退休基金会（TIAA，Teachers Insurance and Annuity Association）高级董事总经理兼首席架构师。TIAA 拥有百年以上历史，管理着约 1.4 万亿美元的资产规模（AUM）。Geeta 拥有近三十年的数据架构与工程从业经验。 - **Eric Pace**：考克斯通信（Cox Communications）AI 负责人，负责美国南部广大区域的宽带与有线网络服务。他在数字化转型领域深耕多年，负责主导公司内部的 AI 落地。 - **Russell LaCour**：Odin AI 联合创始人兼首席技术官（CTO），专注于帮助大型企业部署面向内部流程与外部客户的 AI Agent，曾服务于 Airtel、软银（SoftBank）、IndusInd Bank 等大型跨国企业。Russell 在发言中对 Alain 致以敬意，调侃道当大多数人还在 30 年前对 AI 一无所知时，Alain 就已经在研究“能说话的计算机”了。 --- ### [03:09](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=189s) 从“魔法”到系统：服务提供商视角下的 AI 落地实践 主持人 Alain 指出，生成式 AI 的落地并没有放之四海皆准（one-size-fits-all）的通用策略，它极度依赖于所在行业、业务场景、数据成熟度以及企业内部的组织文化。作为 AI 解决方案提供商，Odin AI 的 CTO Russell 分享了客户在提出需求时的普遍心理与实际工程落地的巨大反差。 Russell 幽默地提到，客户在刚接触生成式 AI 时总是“期待魔法”，希望 AI 能无所不能甚至读懂人心；但在现实中，工程团队的任务是剥离这些不切实际的幻想，深入拆解日常业务流程，将复杂的“魔法”拆构为可自动化的确定性系统。以电信巨头 Airtel 的财务差异分析（financial variance reporting）为例，过去需要一个约 80 人的团队花费整整一个月，跨业务单元拉取和核对数千万笔交易数据，以找出季度之间和跨年度的财务变动。Odin AI 通过构建连接企业各系统的 AI Agent，不仅成功定位出潜在的收入流失（revenue leakage），还实现了自动生成分析报告，将原本需要数周的工作周期大幅缩减到了几个小时。 此外，Russell 强调基于大语言模型（LLM）的 AI Agent 在客服与知识密集型工作（如产品手册问答）中展现出了极大优势。由于 AI Agent 具备“无限的耐心”，它不仅能快速消化原本需要大量培训成本的知识库内容，还能承接海量的日常通用咨询，从而让人类员工解脱出来，专注于更具决策性和创造力的高价值任务。 ```mermaid flowchart LR accTitle: Airtel 财务差异分析流程优化对比 subgraph Legacy["传统人工流程（耗时约数周）"] A["80人专家团队"] --> B["核对各业务线数千万笔交易"] B --> C["季度/年度差异分析与报告"] end subgraph Agentic["AI Agent 自动化流程（耗时数小时）"] D["企业多源系统接入"] --> E["AI Agent 自动化分析与异常检测"] E --> F["识别收入流失 (Revenue Leakage)"] E --> G["自动生成财务分析报告"] end ``` --- ### [06:03](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=363s) 考克斯通信的内部落地：跨部门场景与 Agent 架构拆解 考克斯通信的 AI 负责人 Eric 接过话题，从大型企业内部实践的角度展开了深入剖析。Eric 首先指出了一个常见误区：虽然各行业的业务逻辑和企业文化千差万别，但**企业规模（公司大小）并不是决定 AI 落地难题的核心差异点**。无论是大企业还是小团队，大家在初期面对的基础挑战几乎是一模一样的。 Eric 戏称，业务部门提过来的很多 AI 需求，在优先级待办列表（backlog）里简直排在“把水变成葡萄酒”这类神迹之后。考克斯通信在推进生成式 AI 时，最初采取了“能力视角”（Capability View），试图寻找 LLM 最擅长什么；随后转为针对各业务部门的“职能视角”（Functional View）；目前则演进为开发具备公用事业属性的“企业级 Agent”（Enterprise Agents）。他们落地较快的重点场景包括： 1. **市场营销部门**：这是一个极其富饶的试验场（target-rich environment），AI 能大幅加速内容创意的规模化产出，缩减个性化文案的迭代周期。Eric 特别提到，在艺术和创意领域，模型的“幻觉”（Hallucination）反而是激发灵感的积极因素。 2. **联络中心（Contact Center）**：针对销售与客服等电话高频场景，结合护栏（Guardrails）与意图识别（Intents）的 LLM Agent 可以自主处理大量非辅助渠道的常规诉求，并在复杂情况下完成通话摘要，平滑转接给人类座席（虽然人类员工有时不愿意阅读 AI 生成的摘要，但这属于组织与人类协同问题）。 3. **软件开发生命周期（SDLC）**：从上游的用户故事（User Story）编写、代码开发一直覆盖到下游的自动化测试。 最后，Eric 揭穿了行业内常见的“宣传泡沫”。当听到某些企业声称自己“部署了 100 或 200 个 Agent”时，外界往往会产生焦虑感；但从底层架构来看，这纯粹是营销话术。实际上，企业通常只是构建了一个由 10 个可组合（Composable）的子模块构成的复合系统。“当他们说部署了 100 个 Agent 时，他们其实只是做了 10 个而已，千万别被这些话术给忽悠了。” - 考克斯通信正推动从传统职能孤岛式的工具应用，迈向模块化、可组合的企业级 Agent 体系。 - 营销创意场景对模型幻觉具备较高的容忍度与实用价值，而联络中心场景则高度依赖严格的护栏控制与意图路由机制。 ### [09:27](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=567s) 金融老牌巨头的 AI 落地策略：以合规网关与数据治理化解“繁重苦力活” 在谈到受严密监管的金融保险行业时，来自 TIAA 的首席架构师 Geeta Pyne 阐述了大型机构在部署生成式 AI 时的通用与特有路径。她打趣地回忆道，自己早在 30 年前技术极其艰难的时期就从事过神经网络与 AI 的研究，引得同台嘉宾幽默搭话道：“那我当时还在上小学呢！”针对拥有百年以上历史的 TIAA，Geeta 指出，公司内部积累了海量涉及法律、风险控制、合规和监管的文件，这些“没有人愿意去干的苦差事”（grunt work）恰恰是生成式 AI 的绝佳切入点。团队从文档理解、数据聚合与摘要提取起步，同时始终坚持保留“人类在回路中”（Human in the Loop）。 ```mermaid flowchart TD accTitle: TIAA 生成式 AI 落地与治理架构 accDescription: 展示从数据治理、策略网关到赋能业务与架构团队的完整落地流程 subgraph DataAndGovernance["数据治理与合规管控"] A["公共数据 / 内部数据治理"] --> B["访问控制与数据清洗"] B --> C["AI 政策与合规代码化<br/>(GenAI Tech Gate)"] end subgraph Applications["业务落地与人机协同"] C --> D["文档理解 / 摘要生成<br/>(法务·风控·合规 HITL)"] C --> E["财富顾问助手<br/>(Research Buddy)"] C --> F["架构与数据清洗<br/>(解决 80%-90% 基础工作)"] end ``` 在系统推进时，TIAA 极其强调数据阶梯与负责任 AI（Responsible AI）的先行。团队没有贸然尝试最复杂的场景，而是从公共数据与公开模型逐步过渡，搭建起名为“生成式 AI 技术网关”（Generative AI Technology Gate）的企业级平台，将合规与治理策略直接代码化嵌入到每一个使用场景中。在具体应用上，除了面向财富顾问的智能投研助手“Research Buddy”外，Geeta 作为首席架构师，最兴奋的应用是借助生成式 AI 协助其精简的架构师团队完成大量枯燥的数据清洗、数据理解与数据整理（Data Wrangling），“这能帮我直接产出 80% 到 90% 准确度的结果，而我们只需集中精力去攻关最后的一公里难题。” - **历史与包袱**：百年机构积累的大量法律合规文档，优先通过生成式 AI 自动化处理并结合人类把关。 - **策略平台化**：构建生成式 AI 技术网关，实现访问权限与负责任 AI 策略的自动化校验。 - **架构赋能**：利用模型处理数据整理与清洗，大幅减轻核心技术团队的底层工作量。 --- ### [12:58](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=778s) 医疗健康领域的四大矩阵：消费者、诊疗、理赔与工程效率的系统重塑 围绕医疗健康领域的 AI 应用，嘉宾将业务与技术能力映射为一个多维矩阵，并提炼为四大核心垂直领域：消费者（Consumer）、诊疗护理（Care）、理赔与行政（Claims）以及工程研发（Engineering）。在消费者端，无论是针对拥有数千万参保会员的联合健康保险（UnitedHealthcare），还是 Optum 旗下的药房（OptumRx）与医疗服务机构的患者，首要目标是借助 AI 升级数字化生态，从根本上减少用户被迫拨打客服电话的需求；而当电话接入时，AI 则通过实时摘要与跨系统数据连通，极大赋能一线客服，乃至直接实现部分通话的端到端自动化。 在诊疗与理赔层面，环境聆听（Ambient Listening）技术正在全行业以前所未有的速度普及。过去医疗系统推进电子病历（EMR）往往耗时十数年并斥资数十亿美元，而环境聆听正迅速变成行业基础设施，不仅为近年不堪重负的医生与护士极大减负，更为后续沉淀高质量病历数据提供了跳板。与此同时，嘉宾痛陈了美国医疗体系在理赔与行政上的巨大内耗——例如高达 80% 的事前授权（Prior Authorization）延误纯粹是因文档未同步等行政原因所致。通过在前后端实现数据透明、实时核验与双向系统互通，不仅能根除繁琐的行政延迟，更能有效遏制欺诈与浪费（Fraud and Waste）。 针对整个医疗生态，嘉宾情绪沉重且坚定地谈到了背后的使命感。他特别提及 2024 年 12 月 4 日至 5 日在社交平台 X（原 Twitter）上引发全网热议、因失去同事而令全公司彻底警醒的悲剧事件，深刻指出：“美国在医疗上每年支出高达 4.7 万亿美元，但我们实际换来的体验却远远不相称。”当前整个医疗体系面临深层困境，而技术团队、政府与生态各方必须通力合作，利用 AI 减轻体系负担，真正赋能患者与医生。 - **四大垂直支柱**： - **Consumer**：打通外部人群触点，优化数字生态以减少呼叫，赋能热线坐席实时摘要。 - **Care**：普及环境智能监听，解放医护录入病历的行政负担，沉淀临床 AI 数据底座。 - **Claims**：攻克事前授权等行政延误痛点，实现两端系统实时通信并防范理赔欺诈。 - **Engineering**：普及代码生成与研发辅助工具，提升全链路开发生产力。 - **产业反思**：面对 4.7 万亿美元的高额支出与系统割裂，AI 被寄予厚望成为推动医疗行政简化与诊疗平权的核心杠杆。 --- ### [17:36](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=1056s) 攻坚遗留系统与变革阻力：将 AI 的“魔法”还原为系统集成工程 主持人随后将焦点从场景原型引向落地实践中面临的严峻挑战：对于产品繁多、监管严苛且各部门存在不同亚文化的大型企业而言，变革管理（Change Management）与跨系统集成的阻力往往远超单纯的技术算法难题。 针对如何帮助不同企业客户在复杂的遗留系统（Legacy Systems）之上成功部署方案，嘉宾给出了极具洞察力的本质拆解：“这最终归结为‘什么是魔法’？一百年前如果你展示一部 iPhone，大家会觉得那是魔法；而在今天，它只是科学与技术。”他指出，当前的医疗健康体系依然高度依赖传统的电话录音和各自孤立的老旧系统，缺乏底层集成能力。在 ChatGPT 与铺天盖地的营销浪潮下，外界习惯于将任何连接医疗体系的尝试统统冠以“AI”的光环，但剥开这层神秘感，其本质依然是脚踏实地构建健全、规范的系统架构与数据互通管道。 - **文化与管理挑战**：企业内部的多部门壁垒与合规监管要求，使得 AI 落地必须兼顾交付与组织变革。 - **去神秘化认知**：许多所谓的 AI 突破，实质是在利用大语言模型这一新界面去解决长期拖延的跨系统整合与工程规范化问题。 ### [19:18](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=1158s) 去魅与工程化：将“魔法”还原为可落地的科学 Russell LaCour 继续深入探讨 AI 落地的根本逻辑。他指出，与客户合作时的首要任务是向他们阐明“科学实际上就是魔法，而魔法也是科学”的道理。AI 并不是凭空发挥奇迹的魔杖；团队必须一步一步地厘清数据到底存放在哪里、现有的业务流程是什么、整个链路中正在发生哪些具体环节。一旦把这些底层细节摸透，AI 所谓的“神秘魔法”就被破除了（dispels the magic）。当一切回归到清晰的科学时，围绕它构建系统、接入 AI 来执行任务就会变得非常顺理成章。 然而，如果数据依旧是一团未知的迷雾——比如医院想获取某些病历记录，却根本找不到检索路径，连业务流程都未曾定义——那么 AI 是不可能凭空推导并补全这种流程的。这本质上是一个严肃的“工程问题”，必须先由专业人员介入，把业务流程与数据链路完整绘制（map out）出来，AI 才能真正进场发挥价值。主持人 Alain Biem 随后总结道，这意味着必须顺应客户对事物的认知视角及其现有的遗留系统（legacy systems），以此为基础展开落地。 ```mermaid flowchart LR accTitle: 数据梳理与AI落地流程 subgraph S1["第一阶段：业务与数据梳理（工程问题）"] A["厘清数据位置与现状"] --> B["绘制业务流程与链路"] end subgraph S2["第二阶段：AI系统构建（科学落地）"] C["破除魔法迷思（去魅）"] --> D["构建工程系统并接入AI"] end B --> C ``` --- ### [20:16](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=1216s) 从传统 ML 到 GenAI：婚礼易办，漫长的“婚姻生活”难维系 Alain Biem 将话题抛给 Cox Communications 的 AI 负责人 Eric Pace：企业早在生成式 AI（GenAI）爆发前就拥有深厚的传统技术和基础设施积淀；当高管们（C-suite）纷纷要求“GenAI 来了，赶紧做点什么”时，内部团队该如何应对从传统 AI 向 GenAI 的过渡，并在现有架构与新技术之间取得平衡？ Eric Pace 直言，过去构建传统的经典机器学习（Plain old ML）模型时，团队其实早已预知输出的答案形态，只是等待工程师把模型训练出来以达到目标，因此围绕“拿到答案后该怎么做”的所有业务规划早已完备。但生成式 AI 则完全不同。为此，Eric 提出了一个引发全场共鸣的“婚姻隐喻”： > “构建 AI 就像你的婚礼——它棒极了、充满乐趣、转瞬即逝；可当婚礼结束时，你才恍然大悟：‘哦，我结婚了，现在我得跟我创造出来的这个东西朝夕相处了。’而且它完全不同，屋里多了一个时刻都在运转的其他存在，它并不总是按你的心意行事，你也不一定总能满足它的要求。” Eric 指出，传统 ML 的输出是确定的，而与生成式 AI 协作生活则充满了不确定性。因此，团队面临的最大挑战，是迫使业务伙伴在事前经历大量的论证沟通与“前后对比研究”（before-and-after study），让他们明白技术团队不可能只是把模型做出来、“往里一扔就拍拍屁股走人”。真正的成功依赖于极其漫长且艰难的**变革管理（Change Management）**流程。但现实往往是残酷的——“人们根本不想听你讲什么自身必须改变，他们只想用你的现成工具，只想挥舞现成的魔杖来解决问题”。 --- ### [22:10](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=1330s) 企业架构的解法：能力透镜、投资姿态与决策树 面对系统复杂混乱与业务部门盲目追逐热点的现象，TIAA 的高级董事总经理兼首席架构师 Geeta Pyne 提出了以架构为核心的破局方案。她指出，当下极易出现“手里拿着锤子，看什么都像钉子”的狂热现象，全公司都在涌现各种琐碎的用例。架构师的核心职责，就是通过清晰的框架为企业揭示底层逻辑、权衡取舍（Trade-offs）以及深远影响： 1. **双重透镜分析（Dual Lens）**： - 首先审视企业自身的核心业务能力全景（如 TIAA 的业务能力地图 Capability Map）； - 其次叠加“AI 究竟擅长什么”的 AI 透镜，结合两者的交集进行打分（Opportunity Score），在涌入的上千个需求中筛选出最高价值点。 2. **2x2 评估矩阵与核心分层（Core, Context or Commodity）**： - 利用清晰的 2x2 四象限网格，向业务方明确展示该能力在企业全局中的位置。 - 评估该场景是否属于企业的“差异化护城河”。鉴于工程技术人才极度稀缺，绝不能把宝贵的自研算力与工程资源浪费在非核心的商品化（commodity）模块上。 3. **投资姿态与决策闭环（Build vs. Buy Playbook）**： - 明确企业的投资定位：是要做到行业最佳（Best in class），还是仅保持市场平均水平（At market）？ - 建立结构化的决策树：到底该自研（Build）、外购（Buy）、合作（Partner），还是采购后进行二次扩展（Buy and extend）？ Geeta 强调，一旦这套“决策树与行动指南（Playbook）”与实际用例结合起来，架构师就无需再与业务人员争论某个用例的好坏，而是转变为“以原则为导向（Principle-driven）”的科学决策机制，引导整个组织做出明智选择。 --- ### [25:32](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=1532s) 供应商合作悖论：“恰如其分的开放”与快速洗牌的生态 Alain Biem 提出一个核心痛点：许多核心资产与底层大模型技术并不属于企业自身，而是掌握在 OpenAI、Anthropic、Google、AWS 或微软等大厂手中，企业该如何在内部挑战与外部依赖之间维持平衡？ 一位嘉宾风趣地跳出之前的婚姻隐喻，直言在 AI 的世界里与厂商合作反而需要保持“恰如其分的开放/多情（appropriately promiscuous）”： > “在 AI 的世界里，某一天厂商 A 表现更好，但可能到了下一天厂商 B 就反超了——就像上周刚刚发布了新模型一样。如果是在过去的传统合作中，最理想的模式是建立极深的信任并彻底绑定嵌入；但在当今的 AI 格局下，你必须在深耕与多方押注之间保持平衡。” 与会者探讨了当前企业普遍面临的博弈困境：如何能既吸引到 OpenAI、Anthropic、Google 等巨头顶尖工程师的深度支持，成为他们眼中的战略客户，同时又不在一棵树上吊死，保持敏锐的多方实验与适时押注？这要求企业像是在“与四家供应商同时展开深入对话，今天与这家走深一点，明天与另一家推进一步”。 为了在多方博弈中破局，嘉宾分享了两条核心准则： - **极致透明（Transparency）**：诚实地向供应商表明自身的技术现状、能做什么和不能做什么。即便因为坦白而可能暂时被排在对方医疗行业优先级列表的第二位（哪怕自身体量与潜在市场规模是第一大），也依然坚持真诚互信。 - **积极深耕（Aggressiveness）**：投入大量时间与资源去深度绑定核心合作，因为企业心里非常清楚——单打独斗绝对无法成功。 Alain Biem 补充指出，由于大模型技术迭代极快，生态格局变得越发复杂。过去企业只需选定一家云厂商（Google、Amazon 或 Microsoft）即可，而今天各大模型厂商既想通过生态锁定客户，又在频繁剧变自身的商业模式（例如 Anthropic 正在发力行业垂直领域，OpenAI 既涉足咨询又进攻垂直行业应用），这让下游企业很容易成为商业模式变动的被动承受者。与此同时，市场上也涌现出如 Odin AI 等新兴创新企业，试图在复杂的夹缝中寻找新的突破点。 ### [28:45](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=1725s) 掌控核心技术与“婚姻”隐喻：在拥抱生成式 AI 前先做好自身数据建设 关于企业内部技术自研与外部合作的关系，嘉宾们总结出一套基本法则：企业必须自建核心工程团队，牢牢控制核心技术并严格保护数据资产，同时借助外部生态伙伴来增强和扩展自身能力，并构建模块化、开放的架构以支持未来的持续迭代。 针对这一合作模式，现场引出了一段幽默的“婚姻隐喻”。嘉宾开玩笑称，在拉斯维加斯（Vegas）讨论结婚与约会格外应景，但在技术层面上，企业在“步入婚姻”之前必须先“做好自我提升”（work on yourself）。在生成式 AI 时代，“提升自我”的本质就是沉淀和治理好数据——基础模型提供商会不断更迭，技术框架也会层出不穷，唯独企业自身的数据才是最根本的护城河。业务流程、内部文档以及工作流规范本身都是数据，企业必须将这些资产整理成 AI 能够理解并执行的结构化或半结构化形式。 ```mermaid flowchart TD accTitle: AI 企业就绪与数据基础架构 subgraph S1["第一阶段：自我沉淀 (数据基建)"] D1["业务流程文档化"] D2["元数据、语义与分类体系 (Ontology & Taxonomy)"] D3["主动数据与非结构化资产梳理"] end subgraph S2["第二阶段：智能就绪 (Agentic AI)"] A1["多智能体/Agent 体系集成"] A2["灵活接入主流大模型 & 框架"] end subgraph S3["第三阶段：业务赋能"] B1["海量客户对话分析"] B2["全量审计与异常挖掘"] end S1 --> S2 --> S3 ``` - **底层数据就绪度**：企业若未在元数据（Metadata）、语义网（Semantics）、本体论（Ontology）和分类法（Taxonomy）等底层基础建设上投入，开发智能体（Agentic AI）时将很快遭遇边际效益递减。 - **拥抱非结构化数据**：将传统结构化数据全面整理成 AI 可用状态往往任务艰巨。如果企业想走快速见效的捷径，应当利用生成式 AI 处理非结构化数据，发挥其“化无形为有形”（make the invisible visible）的大规模计算优势。 - **高价值落地场景示例**： - **客户通话记录全量分析**：企业通常积累了数以百万计的客户对话录音/文本（Transcripts），人工根本无法逐一阅读，而生成式 AI 能够轻松对全量文本进行跨维度提问与洞察挖掘。 - **全量合规审计**：传统审计流程依赖人工对少量样本进行抽样并推断异常；生成式 AI 则能一键审核全量文件，不仅彻底消除抽样盲区，还颠覆了以往审计师按部就班的工作模式。 --- ### [32:15](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=1935s) 重新定义成功指标：从传统算法性能到用户采纳与信任 在讨论企业落地成效时，主持人指出传统 AI 时代通常以 ROC 曲线、AUC 等清晰的数学性能指标衡量成败，而生成式 AI 的效果则带有较强的主观性和模糊性。即便是最基础的文本摘要（Summarization）功能，在不同受众眼中的价值也截然不同——例如在标普（S&P），同一套摘要算法提供给信用风险分析师和 ESG 分析师，得到的评价往往大相径庭。因此，如何定义成功成为了设计系统时的首要问题。 从架构师的视角来看，衡量的首要核心指标是“用户采纳度”（Adoption）。无论是合规治理 Agent 还是辅助财富顾问做投研的智能助手，若用户不愿意使用，系统就毫无价值。采纳失败往往归结于用户对数据和输出结果缺乏“信任与信心”（Trust and Confidence）。为了建立这种信任，企业在架构设计和推进策略上必须遵循循序渐进的路径，并搭配全生命周期的可观测性（Observability）。 - **小步试点与小群验证（Cohort Testing）**：先从特定业务用户群小范围试水并快速迭代，验证信任闭环后再全面推广上线。 - **内建全链路可观测性（Observability）**：AI 系统不可“一放了之”（fire and forget），必须建立端到端的持续监控与反馈循环，确保输出质量长期稳定。 - **拒绝“中彩票”心态**：有行业调查显示仅约 4% 的企业在 Agentic AI 中看到了显著价值。企业管理者不应指望一上来就部署能让 CFO 或整个 C-suite 震惊的巨型项目，而应优先普及组织层面的 AI 素养，逐步拓展到后台职能自动化，避免“用昂贵方案解决错误问题”。 --- ### [36:10](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=2170s) 业务价值矩阵与 POC 落地节奏 针对如何精准锁定业务场景，嘉宾们提出了一套清晰的时间与价值评估框架。许多业务部门往往误以为 AI 应当优先解决自身每周最烦琐的报告任务，但如果该报告每周仅耗费 1 小时，花费三个月为其定制 AI 系统只会招致高层的质疑。 | 场景类型 | 组织耗时规模 | 业务优先级 | 投资回报（ROI）评估 | |:--- |:--- |:--- |:--- | | **低频琐碎任务**（如某部门周报） | 全组织每周 ~10 小时 | 低 | 不建议投入长周期定制开发，难以体现商业价值 | | **高耗时核心流程**（如全量工单/审计分析） | 全组织每月数千小时 | 高 | 优先推进，规模化提效成果显著 | 为确保项目聚焦于真正创造生产力价值的环节，嘉宾以一个概念验证（POC）项目的时间线为例，说明了“业务理解与变革管理”所占的真实比重： ```mermaid flowchart LR accTitle: POC 实施阶段时间分配 subgraph Time["5周 POC 时间线"] P1["前2周<br/>业务流程梳理"] P2["中间2周<br/>Agent 开发与测试"] P3["最后1周<br/>变革管理与推进"] end P1 --> P2 --> P3 ``` 在整个 5 周的周期中，真正编写代码和构建 Agent 只花了 2 周，其余整整 3 周时间全部投入在业务流程定义与变革沟通中，这表明在企业落地 AI 的过程中，理解业务痛点与推动组织变革才是决定成败的关键。 ### [37:42](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=2262s) 从原型冲刺到规模化：交付业务目标的核心聚焦 在谈到项目落地的执行节奏时，讲者回顾了团队如何在极度聚焦的状态下迅速取得突破。当时团队完全摈弃了与核心产出无关的旁枝末节，所有的沟通和精力都纯粹扑在两件事上——“这个产品究竟要解决什么问题”以及“它最终需要呈现为什么形态”。在这样高密度的同频协作下，团队仅用了短短两周时间就彻底完成了阶段性开发。 当然，讲者也坦承，把一个两周内跑通的成果真正推向全规模（full scale）部署、逐步推广至其他业务团队，后续依然有大量艰巨的工作要做。但他着重强调，如果在项目初期没有把时间真正花在业务实质的深入探讨与对齐上，缺乏这种全局视角，团队从一开始就会偏离靶心、错失良机。 --- ### [37:56](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=2276s) 领导层变革下的坦诚文化：老牌掌舵人回归与AI治理新范式 主持人随后切入了一个关键而现实的难题：在大型复合型企业中，业务线盘根错节、利益相关者诉求各异，“成功”往往是一个高度模糊的概念。在如此复杂的企业生态中，团队究竟是依赖某种固定的评估工具来向各方证明“大功告成”，还是有着更为深层的衡量体系？ 面对这一提问，讲者坦率地分享了公司近期经历的高管层剧变与内部心路历程。他略带调侃地提到，如果大家关注股市或社交媒体，就会知道公司最近“显然经历了一些波折”。就在同年五月，公司的前任功勋首席执行官、年届73岁的董事会主席史蒂夫·赫姆斯利（Steve Hemsley）重新出山执掌帅印。讲者幽默地将这一幕比作迪士尼的鲍勃·艾格（Bob Iger）或星巴克的霍华德·舒尔茨（Howard Schultz）重返老东家的经典剧情。讲者坦白承认，当初正是离任的前任CEO将自己招入公司，因此当73岁的老将赫姆斯利重返岗位时，他内心的第一反应其实是心头一紧，暗想“这下可要遭殃了（damn, this is going to suck）”。 然而，事实却完全出乎意料，迎来了惊人的反转。这位资深领导者展现出了极具现代感的务实与开明风格。在刚刚结束的两天董事会会议上，赫姆斯利向董事会、外部投资者以及全公司40万名员工清晰传达了极其坚决的透明度原则：“各位，这就是我们正在做的事情，这是我们的落地路径，哪些已经生效，哪些还在碰壁。”他严禁报喜不报忧的隐瞒文化，直接向团队喊话：“别对我藏着掖着，必须实事求是（Don't hide. Be real with me）。”正是在赫姆斯利的大力推动下，团队开始自上而下重新确立整套企业级AI的成功定义。 --- ### [39:36](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=2376s) 衡量企业级AI成功的四大支柱：效用、底线、规模与全员赋能 在全新的治理氛围下，讲者与管理层共同梳理出评判AI项目是否成功的四个核心标准。首先是**切实的实用价值**：AI必须实打实地解决现实痛点，真正帮到客户支持倡导者、陷入困境的业务线、一线的医生或是就诊的患者。讲者提出了一个极其严苛但直观的红线法则——“《纽约时报》测试”：“如果你无法带着自豪感在《纽约时报》记者面前公开谈论这项AI应用，那从一开始就干脆别做。” ```mermaid flowchart TD accTitle: 企业级AI成功的四大衡量支柱 subgraph S1["价值立足点"] A["实际效用<br/>纽约时报测试"] --> B["商业底线追踪<br/>平衡股东与社会使命"] end subgraph S2["规模与影响力"] B --> C["行业与组织启发<br/>数千万至上亿级用户规模"] C --> D["穿透基层全员赋能<br/>从4万技术人员延伸至40万全员"] end ``` 第二项支柱是**底线价值追踪与企业使命的平衡**。医疗企业的使命具有双重属性：一方面要切实造福大众的健康福祉，另一方面也必须在商业体系中为股东创造利润，二者缺一不可。很多管理层往往在“将AI仅视作无需精细核算价值的赋能工具”与“极度严密地追踪ROI”之间摇摆不定。团队最终达成共识：绝不成为短期指标的傀儡，但也绝不含糊其辞，必须建立起透明的穿透机制，自下而上清晰追踪AI对每一条具体业务线利润底线的真实拉动。 第三与第四项支柱则关乎**规模化与全员穿透力**： - **行业启发与宏大尺度**：如果一个AI应用只是局限在某个孤立环节的小打小闹（onesie），根本无法激发组织与行业的变革雄心；唯有当它真正触达2000万乃至2亿人群时，才具备震撼行业的影响力。 - **走出IT象牙塔，深度赋能基层**：在公司约40万名员工的庞大体量中，技术全职员工（FTE）约为3万人，另有约1万名技术外包合同工，技术人员仅占整体员工规模的10%左右。讲者尖锐地指出，如果AI变革仅仅停留在让这4万名技术人员减负提效，而没有改变剩下36万名非技术员工的工作体验，那么这项技术变革就只是困在了工程研发层，根本未曾触及公司的核心脉搏。真正的成功，在于深入到任何一个业务分支的四级基层岗位，能真真切切看到一线普通员工因为手中的技术工具而过得更轻松、工作更高效，并且激发基层员工从自身视角重塑业务流程，进而撬动整个企业自下而上的深刻重塑。 --- ### [42:07](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=2527s) 主持人总结与圆桌收官致谢 主持人对这一评估框架表示高度认同。他总结道，AI的价值核心终究来源于业务端利益相关者的共创；只有将业务伙伴视为共同定义价值的同盟，再借助量化指标与持续的采纳推广，才能形成长久良性的闭环。即便在某些场景下时间节约或间接创收的归因相对模糊，只要能够确保持续适应外部变革、敏捷前行，AI就能展现出深远的生命力。 随着全场讨论渐入尾声，在主持人温馨的主持提示下，台下响起了热烈的掌声，本次聚焦生成式AI落地案例与实践真知的圆桌讨论圆满落下帷幕。 ## Notable quotes > [06:03](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=363s) 当他们说部署了 100 个 Agent 时，他们其实只是做了 10 个而已，千万别被这些话术给忽悠了。 — Eric Pace > [09:27](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=567s) 这能帮我直接产出 80% 到 90% 准确度的结果，而我们只需集中精力去攻关最后的一公里难题。 — Geeta Pyne > [12:58](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=778s) 美国在医疗上每年支出高达 4.7 万亿美元，但我们实际换来的体验却远远不相称。 — Dr. S. Mahin > [17:36](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=1056s) 这最终归结为‘什么是魔法’？一百年前如果你展示一部 iPhone，大家会觉得那是魔法；而在今天，它只是科学与技术。 — Dr. S. Mahin > [20:16](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=1216s) 构建 AI 就像你的婚礼——它棒极了、充满乐趣、转瞬即逝；可当婚礼结束时，你才恍然大悟：‘哦，我结婚了，现在我得跟我创造出来的这个东西朝夕相处了。’而且它完全不同，屋里多了一个时刻都在运转的其他存在，它并不总是按你的心意行事，你也不一定总能满足它的要求。 — Eric Pace > [20:16](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=1216s) 人们根本不想听你讲什么自身必须改变，他们只想用你的现成工具，只想挥舞现成的魔杖来解决问题 — Eric Pace > [25:32](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=1532s) 在 AI 的世界里，某一天厂商 A 表现更好，但可能到了下一天厂商 B 就反超了——就像上周刚刚发布了新模型一样。如果是在过去的传统合作中，最理想的模式是建立极深的信任并彻底绑定嵌入；但在当今的 AI 格局下，你必须在深耕与多方押注之间保持平衡。 — Dr. S. Mahin > [37:56](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=2276s) 各位，这就是我们正在做的事情，这是我们的落地路径，哪些已经生效，哪些还在碰壁。 — Steve Hemsley > [37:56](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=2276s) 别对我藏着掖着，必须实事求是（Don't hide. Be real with me）。 — Steve Hemsley > [39:36](https://www.youtube.com/watch?v=_WYXTUmjBbA&t=2376s) 如果你无法带着自豪感在《纽约时报》记者面前公开谈论这项AI应用，那从一开始就干脆别做。 — Dr. S. Mahin ## Mentioned Alain Biem, Dr. S. Mahin, Geeta Pyne, Eric Pace, Russell LaCour, Steve Hemsley, Bob Iger, Howard Schultz, Ai4, Columbia University, New York Life Insurance Company, S&P Global Market Intelligence, Wayfair, IBM, IBM Watson, UnitedHealth Group, UnitedHealthcare, Optum, OptumRx, TIAA, Cox Communications, Odin AI, Airtel, SoftBank, IndusInd Bank, McKinsey & Company, OpenAI, Anthropic, Google, AWS, Microsoft, The New York Times, X (Twitter), Disney, Starbucks, ChatGPT, iPhone ## Video description Alain Biem, Adjunct Professor, Columbia University; Former Chief AI Officer, New York Life, S&P Global Market Intelligence, Columbia University Russell LaCour, Chief Technical Officer, Odin AI Eric Pace, Head of AI, Cox Communications Geeta Pyne, Senior Managing Director and Chief Architect, TIAA Abstracts This session will showcase some of the most successful use cases of generative AI such as LLMs and multimodal AI across various industries. Panelists will explore what made these implementations effective, from the initial identification of opportunities to the execution and scaling of AI solutions. Learn about the specific industries, challenges, and strategies that led to these breakthrough successes, and take away actionable insights on how to replicate this success in your own organization. Whether it's in content creation, customer service, or operational efficiency, discover why these case studies stand out and how you can apply their lessons. [Source](https://www.youtube.com/watch?v=_WYXTUmjBbA)

[6:03] Eric Pace: 当他们说部署了 100 个 Agent 时，他们其实只是做了 10 个而已，千万别被这些话术给忽悠了。 [9:27] Geeta Pyne: 这能帮我直接产出 80% 到 90% 准确度的结果，而我们只需集中精力去攻关最后的一公里难题。 [12:58] Dr. S. Mahin: 美国在医疗上每年支出高达 4.7 万亿美元，但我们实际换来的体验却远远不相称。 [17:36] Dr. S. Mahin: 这最终归结为‘什么是魔法’？一百年前如果你展示一部 iPhone，大家会觉得那是魔法；而在今天，它只是科学与技术。 [20:16] Eric Pace: 构建 AI 就像你的婚礼——它棒极了、充满乐趣、转瞬即逝；可当婚礼结束时，你才恍然大悟：‘哦，我结婚了，现在我得跟我创造出来的这个东西朝夕相处了。’而且它完全不同，屋里多了一个时刻都在运转的其他存在，它并不总是按你的心意行事，你也不一定总能满足它的要求。 [20:16] Eric Pace: 人们根本不想听你讲什么自身必须改变，他们只想用你的现成工具，只想挥舞现成的魔杖来解决问题 [25:32] Dr. S. Mahin: 在 AI 的世界里，某一天厂商 A 表现更好，但可能到了下一天厂商 B 就反超了——就像上周刚刚发布了新模型一样。如果是在过去的传统合作中，最理想的模式是建立极深的信任并彻底绑定嵌入；但在当今的 AI 格局下，你必须在深耕与多方押注之间保持平衡。 [37:56] Steve Hemsley: 各位，这就是我们正在做的事情，这是我们的落地路径，哪些已经生效，哪些还在碰壁。 [37:56] Steve Hemsley: 别对我藏着掖着，必须实事求是（Don't hide. Be real with me）。 [39:36] Dr. S. Mahin: 如果你无法带着自豪感在《纽约时报》记者面前公开谈论这项AI应用，那从一开始就干脆别做。

[0:04] Russell LaCour: talking computers 30 years ago when none of us even knew what AI was. [6:03] Eric Pace: right behind turning water into wine in our backlog. [6:03] Eric Pace: hallucination in a creative sense can be a good thing. [6:03] Eric Pace: When they've built a hundred agents, they've really built 10—don't let them lie to you. [9:27] Geeta Pyne: when things were really, really hard [9:27] I was in elementary school. [19:18] Russell LaCour: explaining that science is actually magic and magic is science. [20:16] Eric Pace: A lot of people in this room are married. Y'all probably spend a lot of time planning your wedding. And you all probably spend zero time planning your marriage, right? Building AI is like your wedding. It's great. It's fun. It's going to go quick. And then when it's over, you realize, 'Oh, I'm married and now I got to live with this thing that I've built.' [25:00] When it comes to partnership in this AI world, I think it's actually important to be appropriately promiscuous... One day person A is better and the next day person B... last week we had GPT-5, right? [28:45] Russell LaCour: work on yourself [30:44] Eric Pace: make the invisible visible [30:44] Eric Pace: you can watch the auditors sit there—if you're an auditor, I'm sorry—you can watch the auditors sit there and pick their nose. They don't know what to do with themselves when we're done. [32:16] Alain Biem: in the eye of the beholder.

[0:04] Uh the session today is about Gen AI in [0:06] action I think and I think judging by [0:09] the entrance here this is a very [0:10] exciting topic right now. So my name is [0:13] Alan Ben and I'll be the moderator for [0:15] this panel. Um so let me just quickly [0:18] introduce a bit of myself. So um um I'm [0:22] the former chief AI officer for New York [0:24] Life Insurance Company. Um and before [0:26] that I was a chief AI science is the [0:29] head of AI for S SNP global market [0:31] intelligence divisions. Also was at some [0:34] point head of machine learning and data [0:36] science for wfair.com. Um I start my [0:39] career at IBM as a quote unquote a [0:42] scientist nerd working on some of the [0:44] algorithm on the IBM Watson platforms [0:48] and also working very closely with IBM [0:50] customers. Um I also teach at Colombia. [0:53] I am I teach machine learning and [0:55] analytics to grad students at Columbia [0:57] University. So um thanks for being here [0:59] and I will ask some of my fellow [1:00] panelists to introduce themselves. We [1:03] can start from the other side from S. [1:06] >> Hello everybody. Uh say Mahin I am the [1:09] chief AI transformation and strategy [1:11] officer at United Health Group. Um, I'm [1:14] priority united. I was uh overseeing AI [1:17] and data and cyber for healthcare in the [1:20] Biden administration and before that [1:22] management consultant at Mackenzie and I [1:24] am an internal medicine doc. Great to be [1:26] here. Thank you. [1:27] >> Hello. Good afternoon. Gita Pine. I'm [1:29] the senior managing director and chief [1:31] architect for TIAA. [1:34] Um, many of you may not know, so I'll [1:35] spell it. Teachers Insurance and Annity [1:38] Association. We are a 100 plus years old [1:40] company. We manage about 1.4 4 trillion [1:43] asset under management. I've been with [1:45] the company for about two years, but I [1:46] have been I feel like I've been working [1:48] ever since I was born. So like three [1:50] decades on and about like I've been [1:52] working across several industries mostly [1:55] around data architecture engineering and [1:59] very pleased to be here looking forward [2:00] to the conversation. [2:02] >> Awesome. Hello, I'm Eric Pace. I'm the [2:04] head of AI at Cox Communications. We're [2:06] a service provider for internet [2:08] connectivity and cable services uh all [2:10] along the southern belt of the United [2:12] States. Uh I've been in this role for [2:15] about 8 months, maybe a year and a half [2:17] depending on how you want to count the [2:18] days. Uh and I've got a long digital [2:20] transformation journey behind that. [2:22] >> Awesome. Hi everybody. My name is [2:24] Russell Vort, CTO and co-founder of a [2:27] company called Odin AI. So we specialize [2:30] in working with large enterprises to [2:32] deploy AI agents for internal processes [2:35] and external processes. We've worked [2:38] with companies like Airtel, Soft Bank, [2:40] indis bank and we primarily help in [2:45] identifying roles where automation and [2:48] AI can actually make a difference. Uh [2:51] excited to be on the panel today humbled [2:53] that you know I'm going to give our [2:55] moderator a little bit more credit than [2:57] he gave himself. This man was talking [2:59] about talking computers 30 years ago [3:02] when none of us even knew what AI was. [3:04] He was studying it. So very honored to [3:06] be with this panel. Thank you everyone. [3:09] >> Glad to be here. Thank you for being [3:10] that. So uh gen AI in action use cases. [3:14] So we're very very happy that we having [3:16] a very good representation of many [3:18] industry in this stage. What as you best [3:20] know gen AI depend on the industry, [3:22] depend on the use case, depend on your [3:24] data, depend on the culture in your [3:26] company. So there is no one solution fit [3:28] all kind of strategy right now. They [3:30] always need to adapt any templiz [3:32] solution you may have into particular [3:34] story. So I'm going to ask is my [3:36] panelist to kind of provide some kind of [3:38] examp use case they have right now. So [3:41] pro I start with you because you are [3:43] basically working for client. You are [3:45] the AI startups coland you deliver [3:48] solution for customer. [3:49] >> Could you just walk us a few example [3:51] about what kind of ask do you have from [3:54] customer application you working for [3:55] them? Yeah, absolutely. So, everything [3:58] always starts off with magic. Everybody [4:00] wants magic. They want the AI to do [4:02] anything and everything, read their [4:04] mind. And the real world is [4:08] what is the business process that's [4:10] happening on a day-to-day basis where we [4:12] can insert a system that can understand [4:15] what you're trying to do and automate [4:17] it, take apart the complexity of the [4:20] magic and actually make it a system. So [4:24] for Airtel, what we did was their [4:26] financial variance reporting was a team [4:28] of about 80 people over the course of a [4:30] month getting together all of their [4:32] business units and tens of millions of [4:35] transactions and actually figuring out [4:36] what changed from this quarter to last [4:39] quarter to the last few years. And this [4:41] process we one connected all the systems [4:45] to an AI agent and actually were able to [4:48] find revenue leakage. We were able to [4:50] generate reports and we were able to do [4:52] it within a few hours rather than a few [4:55] weeks. So we've worked everything from [4:58] huge financial cases with companies like [5:00] Airtel to even customer support cases [5:03] which I know Eric can speak a lot about [5:05] where you're actually doing for instance [5:08] product manual questions. you're doing [5:10] this kind of knowledge work that you [5:12] have so many people that you're training [5:14] on, but really AI agents are very [5:16] capable with the power of LLMs to answer [5:20] common questions and they're much more [5:23] capable because they're infinitely [5:25] patient and so it alleviates and gives [5:28] people the ability to now focus on real [5:30] tasks to focus on making decisions and [5:33] allows in my opinion for companies to [5:36] grow and be more creative. So, but I'll [5:39] let Eric speak more to that because he [5:42] has been front and center for [5:43] >> and this is a great segue because Eric [5:45] now from your inside view you're working [5:47] for a telecom company. So, you are [5:49] basically dealing with what happened [5:50] inside the company at the corporate [5:52] level. So, could you not take a segue [5:54] from him being outside and you inside [5:56] tell us more about at least into the [5:58] telecommunication industry what you guys [6:00] are popping up into uh into cost [6:02] communication? [6:03] >> Yeah, sure. I think the um the one thing [6:05] that I'll add to what you were saying [6:07] earlier, it it it a lot of it is [6:09] industry specific. A lot of it is based [6:11] on your culture, etc. But the one thing [6:12] that I have found that is not [6:14] necessarily a differentiator yet is the [6:16] size of your company. It doesn't matter. [6:18] I've met companies that are very small [6:20] and very large and we talk to a lot of [6:21] people and all of the companies of all [6:23] sizes are all solving the same types of [6:25] problems. You start to get into [6:26] tailoring of how you solve those [6:28] problems once you get into some more [6:30] industry specificity or have to deal [6:31] with your culture, etc. Um but but I do [6:34] agree I do agree with Russell the most [6:36] of the most of the time we tell people [6:38] that that the things that they're asking [6:40] for are right behind turning water into [6:42] wine in our backlog. [6:44] >> Um but we are focused in a lot of [6:45] different places across the enterprise. [6:47] We went from a bit of a capability view [6:49] uh trying to figure out what types of [6:51] things the LLM's or AI or generative AI [6:53] was best at and then we shifted to more [6:55] of a functional view and I feel like [6:56] we're actually shifting a little bit [6:57] back towards what capabilities we want [6:59] to deploy more utility based or [7:01] enterprise agents against. Uh but some [7:03] of the use cases that we've got uh in [7:06] flight range from things in the [7:07] marketing departments because it's a [7:09] target-rich environment relative to [7:11] getting content creative um into place [7:14] at scale and at speed and reduce [7:16] iteration cycles relative to you know [7:18] things like personalized messaging uh [7:20] etc. You know a lot a lot of times what [7:22] we are I find I I try to get people to [7:25] understand that hallucination in a [7:26] creative sense can be a good thing. Um [7:28] we are working in our contact centers a [7:30] lot. The industry requires us to be on [7:32] the phones a lot with our customers both [7:33] for sales and service. And we do drive a [7:35] lot of unassisted channel support [7:37] relative to what we want our customers [7:39] to experience. Uh and a lot of times [7:41] what we have found is that as we as as [7:43] we uh delve deeper into what enterprise [7:46] agents or agents are are capable of [7:47] specifically with LLM backing if you [7:50] give them some guardrails and you give [7:51] them some intents to handle whether [7:53] they're from a sales or service [7:54] perspective uh they do they do quite a [7:56] good job of talking to customers through [7:58] at least you know a bulk of the issues [8:00] that they have or some of the things [8:01] that they need. uh they can't solve all [8:03] of the most complex problems yet, but [8:05] they also do a pretty good job of [8:06] summarizing a lot of those interactions [8:07] and then doing handoffs to our human [8:10] workforce when necessary. Um sometimes [8:12] we do have a human problem when the [8:14] humans don't want to digest or consume [8:16] some of the summaries that we provide [8:18] for them, but I don't think that's [8:19] necessarily different in the AI space. [8:22] Um [8:22] >> that's a different panel. [8:23] >> It's a different panel, right? Exactly. [8:25] That's right. Uh and and then we've got [8:27] like I said there's a there's there's a [8:29] long list of cross functional uh um use [8:32] cases that we've got whether it's it's [8:35] pushing deeper into the SDLC with what [8:37] we're doing in it from development up [8:39] through user story creation and testing. [8:41] Uh we're deploying agents all over the [8:43] place I think or sometimes is what it [8:44] feels. But I will dispel a little bit of [8:46] what I told some some folks in a room [8:48] like this earlier. Uh it it it can be [8:51] daunting when you hear you know someone [8:52] up on stage or someone you know giving a [8:54] talk that's like we deployed 100 agents [8:56] we deployed 200 agents and everyone's [8:58] like 200 we've only deployed one what [9:00] the hell are you talking about a lot of [9:01] times the marketing gets in front of the [9:03] message and this coming from a guy who [9:05] ran was in marketing for a while I think [9:08] the way you have to wrap your heads [9:09] around that is is you have many levels [9:12] of agents that can comprise the system [9:13] and so when someone talks about one [9:15] agent I think what they're really [9:16] talking about are the 10 agents that [9:18] they built in a very you composable way [9:20] to make up whatever that agent is doing [9:22] and so when they've built a hundred [9:23] agents they've really built 10 don't let [9:25] them lie to you [9:27] >> thank you sir so we moving that into you [9:29] know you get and actually you said you [9:31] guys are working into regulated industry [9:33] you know insurance and then healthcare [9:35] so tell us about in at least what kind [9:38] of use case you guys are working on taa [9:40] >> yeah absolutely so I mean a lot of what [9:43] what Eric talked about I think that [9:46] applies pretty much everywhere right if [9:48] you look at regardless of size a type of [9:50] business you probably you have something [9:52] to for your engineers to be become [9:54] productive or workforces you have things [9:57] to run your operations so IT operations [9:59] or business operations then you have [10:01] your customer support your sales [10:03] marketing so I think all those areas are [10:06] pretty much generic but what I would say [10:08] where we have seen are really great [10:11] value starting with generative AI before [10:13] even getting aic and by the way I did [10:16] you know neural networks and AI 30 years [10:17] back when things were really really [10:20] hard. So we can we can debate and on our [10:22] legacy. [10:23] >> I was in elementary school. [10:26] >> So uh so um what I was going like so for [10:29] say for the way we have seen is we are a [10:32] 100 plus years old company and this has [10:35] we have a tons of documentorientated [10:38] legal risk compliance regulatory stuff [10:41] that we don't like. I mean we have to be [10:43] compliant. You have to be but no who [10:45] wants to do the grunt work. So a lot of [10:47] that are perfect for so we started with [10:51] lot of you know document you know [10:53] understanding aggregation summarization [10:56] and but obviously with human in the loop [10:59] so legal risk compliance or even [11:01] customer service agent but the key is we [11:05] also looked at you know what kind of [11:07] data you need so that you are going with [11:10] increasing complexity not starting with [11:13] the most complex [11:15] >> defining our responsibility AI policies [11:17] although it is not a policy conversation [11:19] but I think it is important you cannot [11:21] talk about generative AI and agentic [11:22] without that right so and what kind of [11:24] data is needed so am I starting with [11:26] only public data and public models and [11:30] from there so get your arms around [11:32] making sure to your point Eric's point [11:34] around whether you start with one or 100 [11:36] or 200 what kind of data you are okay to [11:39] use do you have the right access control [11:42] is this data clean so I think putting [11:44] the whole thing vertical around has been [11:46] our so we built our PL you know platform [11:49] what we called generative AI you know [11:51] technology gate as a platform that's [11:53] where basically policies are coded at [11:55] every use cases so so use cases are [11:58] everywhere you know we have you know for [12:00] our wealth advisor right you know we [12:03] call research body because you can do a [12:05] tons of research right so we're doing [12:07] but I think the my most you know I get [12:09] excited about you know stuff that again [12:12] we don't want to do there is a lot of [12:14] data quality data data cleansing, data [12:17] you know understanding, data wrangling I [12:19] am using a lot because I run I as a [12:22] chief architect I have a very skinny [12:24] small team for architects but I get to [12:27] support like whole you know entire [12:29] business so I am using for a lot of that [12:31] kind of a work which is helping me to [12:33] get 80% 90% accurate results and then I [12:37] can only solve the last mile problem. [12:39] Okay, that makes sense. And then say I [12:42] know as you know you know United Health [12:44] you guys you know your role you need to [12:46] take care of the transformation [12:47] basically the change management aspect [12:48] on top of the delivery aspect. So this [12:50] provides some extra challenge. So before [12:52] we go into the challenge, but just get [12:54] us some of the prototype, you know, [12:55] application you guys are cooking under [12:57] the hood. [12:58] >> And it's so similar to what you've [13:00] already heard, obviously just applied to [13:02] a different industry and what all of us [13:04] are basically [13:07] living in is a matrix, right? with one [13:11] access you have what are the [13:13] applications relevant to my industry or [13:15] the things like engineering is not [13:16] relevant to my industry but it's just [13:18] you know table stakes things that need [13:20] to get done and done better through AI [13:21] that can be done better and then on the [13:23] other you have what can AI uniquely do [13:25] whether that's summarization or voice or [13:28] whatever it might be right the obviously [13:30] the more generative thinking now so uh [13:32] and and it's and then you just say for [13:34] each of these okay for a box so in in my [13:37] kind of simple brain or united simple [13:39] brain We have kind of four verticals [13:42] right it's consumer claim care and [13:47] engineering essentially consumer is [13:51] every interaction you have with people [13:53] that are not employed by you essentially [13:55] so that is uh members of United [13:58] Healthcare which are insurance carrying [14:00] members that is patients uh of optimum [14:04] whether that's optimum the pharmacy [14:05] optimum rx or optimum the uh care [14:08] delivery organization that is, you know, [14:11] um just that that whole universe of [14:13] people outside, how do you use AI to [14:16] interact with them better? And that kind [14:17] of part of that is your whole digital [14:19] ecosystem using AI to make that [14:21] infinitely better so that you could [14:23] actually avoid the call, right? And [14:25] actually give them a much deeper, [14:27] richer, personalized, empowering [14:29] experience. Um the other part of that [14:32] then is when you get into the call, what [14:34] can you do to make that call as [14:35] efficient, as effective as possible? [14:37] both in empowering the advocates [14:39] answering the call with summarization [14:41] with better information with you know uh [14:44] connected data sources and the like as [14:46] well as you know obviating the need for [14:49] a human on the call versus the call [14:51] itself. Um that's the consumer piece. On [14:55] the care piece, you know, the big thing [14:57] you've all heard about in healthcare [14:59] right now where there's been histo for [15:01] our industry where tech adoption is [15:03] usually like 17 years and requires [15:05] high-tech and billions of dollars to get [15:06] EMRs going. Ambient listening has taken [15:08] off like wildfire. But ambient is really [15:12] going to be a utility in our industry. [15:14] It's going to be the thing. Yes, it [15:16] saves you some time. Yes, it eases some [15:18] administrative burden. and most [15:19] importantly uh makes doctors lives [15:22] doctors who've been decimated with [15:24] demand in the last 5 years especially [15:25] since co makes their lives better and [15:28] and nurs's lives better as well but uh [15:30] even long term the big opportunity is [15:33] what does that do for getting more [15:34] information about the patient documented [15:36] so that you can apply AI in different [15:38] ways both administrative ways as well as [15:40] clinical and care ways and then the the [15:42] claims the engineering I'm not going to [15:44] repeat it's same stuff for all of us [15:45] right productivity etc etc um the the [15:48] claim names piece. [15:50] So much of what stinks about healthcare [15:53] now is what an administrative pain it is [15:56] to get anything done. You know, 80% of [15:58] the time that there's an issue with a [16:01] prior authorization, for instance, [16:02] obviously a big topic in the last year. [16:04] Um, that's because of some [16:06] administrative delay. I don't have this [16:08] documented. I don't have that. I don't [16:09] have this. All of that can be eliminated [16:12] by managing a claim more effectively, [16:14] doing it real time, capturing all the [16:16] information, and having the two-way [16:18] communication between systems that you [16:20] know, humans don't have to create these [16:23] delays because they're no longer [16:24] involved in that particular process, [16:25] right? So, you you you take it on the [16:27] front end from a having transparency on [16:30] what you got to pay to scheduling to [16:32] this and that all the way to the back [16:33] end of making sure that you have high [16:35] fidelity on payment and that there isn't [16:37] fraud and waste and the like. and [16:39] everything in between. There's a lot of [16:40] administrative simplification uh and [16:42] imprative [16:44] improvement happening. And the last [16:45] thing I'll say on this is at the end of [16:47] the day, the vision here, I think we all [16:50] agree and if you don't believe me, go [16:52] look at X or Twitter on December 4th and [16:55] December 5th of 2024. That's when we [16:57] really really woke up to it as a company [16:59] when we lost one of our colleagues [17:01] honestly. But I think we all agree [17:03] something is fundamentally broken with [17:06] healthcare in America today given what [17:08] we spend 4.7 trillion. Uh what we get is [17:12] just not there. So the the fundamental [17:15] hypothesis is we can use this technology [17:18] and our best intentions and our best [17:20] efforts and partnering with the [17:22] government and working as an ecosystem [17:23] to ultimately empower the entire system, [17:27] empower patients, empower doctors uh to [17:30] be better and unbburden the system along [17:32] the way. So that is amazing because you [17:36] discuss basically some of the use case [17:37] but you also have put the segue into the [17:40] challenges basically and then in the [17:42] case of United Health you guys are [17:43] basically a big company you're having [17:45] many product having many department [17:47] you're having regulation you having the [17:48] fed behind you but you also have the [17:50] culture and you have different [17:52] subculture in the company and that [17:54] create a lot of challenge that need to [17:55] be faced on top of that beyond simply [17:58] the technical challenge that that that [18:00] you faces. So I want to talk more about [18:02] those challenges. I want to dig more [18:03] about about that. Um and I want to [18:06] divide the challenge view from inside [18:08] from the core private point of view [18:10] basically. Um because some of the [18:12] challenge you face are basically uh [18:14] thing that we have to do inside before [18:16] we expose them to customer. You talk [18:17] about that a bit more and so I want to [18:20] start from the customer point of view. [18:23] Chris you know um I start with you. You [18:25] are working for different client. They [18:28] have their own system, their own legacy [18:30] system. You need to make things work for [18:32] them. How do you do that? What's the [18:33] challenge you're facing for doing this? [18:35] >> Great question. And to speak really to [18:37] say and [18:40] it comes down to what is magic? [18:42] A 100 years ago, if you showed someone [18:44] an iPhone, it's magic. Today, it's [18:47] science. [18:49] You look at it and it makes sense. It's [18:50] technology. [18:52] >> And so, the health care system operates [18:54] on legacy phone calls. They're calling [18:58] records. They don't necessarily have [19:00] integrations and systems. And so today, [19:03] because of AI, because of large language [19:05] models, because of Chad GPT and really [19:07] good marketing, we now call integrating [19:11] healthcare as AI. [19:12] >> But really, it's building proper [19:14] systems. And so with every client that [19:18] we work with, the first thing is [19:20] explaining that science is actually [19:23] magic and magic is science. We have to [19:26] go step by step and understand where the [19:29] data lies, what are your processes, what [19:31] are these different things that are [19:33] happening along the way. And once we do [19:36] that, it dispels the magic. And once it [19:39] becomes science, it becomes very easy to [19:42] build a system around it, to put an AI [19:44] in there to do it. But as long as it is [19:47] unknown or those data records that the [19:50] hospital is trying to get, there's not a [19:52] way to find them. There's not a process [19:54] defined. AI won't be able to figure out [19:57] that process. It's an engineering it's [20:00] an engineering problem and job that [20:04] requires someone to go in and map it out [20:06] and then AI can come and actually make a [20:09] difference. [20:10] >> So you need to adapt to the customer [20:11] viewpoint of things and the legacy [20:13] system and then do that. [20:14] >> Exactly. [20:15] >> Now [20:16] um let's say you are inside the company [20:19] and I've gone back to you Eric on that. [20:21] You guys have been doing things for a [20:22] long time even before you have geni you [20:24] have your infrastructure you have a lot [20:25] of thing going right now then you having [20:28] you know the se coming to you saying [20:29] geni is here do something about that you [20:31] need to deal with transforming from AI [20:33] to genai make that in actions deal with [20:36] what is there before you move to what is [20:38] new how do you dealing with that [20:41] >> yeah I think when when when [20:43] we were when we're building you know [20:45] plain old machine learning models right [20:46] >> the difference is we we knew what the [20:48] answer was and we were just waiting for [20:50] someone to build the model to there, [20:51] right? And so all of the planning around [20:53] what we were going to do with the answer [20:54] was already done. [20:55] >> To Russell's point, and I the metaphor [20:57] that I like to use, and I'm sure, you [20:58] know, there's a lot of people in the [21:00] room, we spend a lot of time uh talking [21:02] about this in particular. And and I [21:04] think much of the the chagrin of our [21:05] business partners, we we end up forcing [21:08] them through a lot of conversations to [21:09] do a lot of the before and after study. [21:11] Like I'm I'm sure I'm not going to make [21:13] you raise your hands, but like a lot of [21:14] people in this room are married. Y'all [21:16] probably spend a lot of time planning [21:17] your wedding. And you all probably spend [21:19] zero time planning your marriage, right? [21:22] Yeah. AI is the same thing. Building AI [21:24] is like your wedding. It's great. It's [21:26] fun. It's going to go quick. And then [21:27] when it's over, you realize, oh [21:29] I'm married and now I got to live with [21:31] this thing that I've built, right? And [21:33] it's different. And there's someone else [21:35] here that's doing something all the time [21:37] and they don't always do what I want [21:38] them to do or I don't always do what [21:40] they want me to do. And you have to [21:42] really understand what the difference is [21:43] when you live and work with these things [21:44] that you've now built, not knowing like [21:46] a machine learning model what it's going [21:48] to produce. [21:49] >> So a lot of the the challenge that we [21:51] face is getting people to understand [21:52] that we can't just build this thing for [21:54] you, drop it in, and walk away. You have [21:56] to go through a long process of change [21:57] management in order to find success in [22:00] what we've built for you. Right? So it [22:02] it's it's you know, people don't want to [22:03] hear that they've got to change. They [22:04] just want to use what you've got. They [22:05] want to use the wand. [22:06] >> Okay. Want you to solve the problem. [22:07] That's right. That way. [22:10] So, anything you want to add to do that? [22:12] I know that with your view for [22:14] architect, you need to be well [22:16] connected, well designed before you [22:18] move. It's not always the case, right? [22:20] You find things that are quote unquote [22:22] messy or you need to build something [22:24] that already cannot be unbuilt right [22:26] now. So, what are those challenges in [22:28] your viewpoint for this? Yeah, I think u [22:31] I'll go back to I think earlier Eric you [22:33] would talked about a [22:35] capabilitycententric capability centric [22:37] view versus maybe functions and use [22:40] cases right what I'm finding is and [22:43] which is kind of expected is many a time [22:46] it is very easy to get excited about hey [22:49] I have a hammer let me everything starts [22:51] looking like a nail right so we got this [22:53] thing everybody is now coming with the [22:55] use cases I think as an architect being [22:58] able able to show the vis like you know [23:01] if you do this is there already somebody [23:05] has solved that problem or not because [23:07] you don't have the visibility or even if [23:09] you do this you might be you know [23:12] creating just like you know yeah you got [23:14] married but whatever you did the wedding [23:16] planning but you didn't think about the [23:18] marriage kind of a thing so if you build [23:19] something here are the trade-offs and [23:21] implications right also the decisioning [23:24] around is it a how should we build it do [23:26] we build it or buy it or partner it or [23:29] extend it, buy it and extend it. Right? [23:31] So there is so many decisioning. So what [23:33] I think architecture helps is a number [23:36] of creating that lens around what do you [23:39] do as a company and then I put a lens [23:42] around what is AI good for. So can I [23:44] have an AI lens on say case in point TIS [23:47] capability map so you can narrow down [23:50] when thousand things are coming in which [23:52] is the highest like what is your [23:54] opportunity score where is the highest [23:56] value right nothing like extraordinary [23:59] it's just simple ordinary 2x2 grid but [24:02] you show in a way that people can see [24:05] where in the business we are talking and [24:07] then also having the criteria around is [24:10] it really your differentiating factor do [24:13] you want to put because talent is very [24:15] scarce. Do you want to throw your [24:17] engineering talent for building those [24:19] things right? So differentiating that [24:21] stratification of like you know your [24:23] core context or commodity and entire [24:26] your investment posture like do you want [24:28] to be the best in class? Do you want to [24:30] be you know just at market and then [24:32] having the decision. So to me this kind [24:35] of almost like a decision tree and a [24:37] playbook. If you can complement that [24:40] with the use cases it just helps to [24:43] simplify the you are not fighting with [24:45] people then that this is a bad use case [24:47] or this you become a very principled [24:50] driven way of guiding the organization [24:53] to make the right choices. [24:55] >> Okay. Thank you so much for that. And I [24:57] think that's actually re's equation is [25:00] you need to make those choices you're [25:01] saying right. You need to decide where [25:02] to put where to where [25:04] >> but sometimes the technology the asset [25:07] you have in your hand does not belong to [25:10] you they may belong to open air it may [25:12] belong to cloud it may belong to [25:13] anthropy it may belong to Amazon it [25:15] belong to somebody else [25:17] >> how do you deal with that the fact that [25:19] you need to basically you have the [25:20] challenge that you face internally you [25:22] also have to make sure that you're [25:23] dealing with the right partnership and [25:25] stuff say I don't know if you want to [25:26] say a few word on that as you see into [25:29] right now whatever you you know how you [25:31] That's so [25:32] >> I'm going to I'm going to break from the [25:33] marriage analogy a bit and say when it [25:37] comes to partnership in this AI world, I [25:41] think it's actually important to be [25:43] appropriately promiscuous. [25:45] >> And and and the reason I say that is, [25:50] you know, the potential partner one day [25:54] person A is better and the next day [25:58] person, you know, last week we had GBT5, [26:00] right? I I agree with this for AI. for [26:03] AI for AI and that's this is AI for not [26:06] marriage for all right so no but [26:09] seriously it it's it's it's it's [26:11] actually quite hard because historically [26:14] when you think about the best [26:16] partnerships it's all about building [26:18] trust going deep deep and deeper and [26:20] really turning it into this like really [26:23] you know wedded and embedded uh kind of [26:25] operation in this AI landscape you kind [26:29] of have to somehow figure out how to do [26:31] both and I think we're all wrestling [26:33] with this which is how do we get the top [26:37] engineers from open if you're me the [26:39] question is how do I get anthropics top [26:41] engineers open AI's top engineers Google [26:44] you know all these guys Microsoft etc [26:46] how do I get their best people and how [26:48] do I become the the one they want while [26:53] making sure that I'm not only keeping [26:55] tabs but actually appropriately [26:58] experimenting and then ultimately making [26:59] bets uh you on on the right ones at the [27:02] right time. And to to sort of do that [27:05] dance where I'm having a voice [27:07] conversation with four vendors and then [27:10] going a little deeper with one and then [27:11] a little deeper with another and then [27:13] that is very hard to do and I actually [27:15] people have figured it out. I'd love to [27:17] hear it. What I would say is the thing [27:19] that has I think to a certain degree [27:21] worked for us I think two things. One is [27:25] we are transparent like we are honest [27:27] about where we are and what we can do [27:29] and what we can't do. And if that means [27:31] we're going to be number two on your [27:33] priority list as a healthcare player, [27:35] even though we might be one in size and [27:36] theoretical TAM for you as a company, so [27:39] be it, right? But we are transparent. [27:41] That's one. Um, but two, we're also [27:43] aggressive. Like we are deep in time and [27:46] invested in these partnerships because [27:48] we know with absolute certainty that we [27:50] cannot do it alone. [27:51] >> No, that does make sense. And I think to [27:53] add to what you're saying, um, I think [27:56] it's compounded by the fact that the [27:58] technology moves so quickly. [28:00] >> Yeah. So even those vendor that we talk [28:02] connect today I mean in the past we used [28:03] to just go to one of the cloud vendor [28:06] Google Amazon uh you know Microsoft and [28:09] then having the system and then do but [28:11] today with geni on the loop and the and [28:14] all of them trying to create an [28:15] ecosystem where you lock you in but at [28:17] the same time they changing the business [28:19] model right so anthrop is going to the [28:21] vertical right now open air try to do a [28:23] bit of consulting while they're going to [28:24] do also the verticals right now they're [28:26] changing the business model and we are [28:28] the victim for that in term of [28:29] developing doing some of those things [28:30] right um and then we have some [28:32] innovations you know the Odin AI of the [28:34] world coming here and trying to you know [28:37] make something to you know to work so [28:38] yes I think I I understand in my [28:40] personal experience I did I did what [28:42] you're doing basically I will always [28:43] make sure that we have control about a [28:45] core technology that we develop right [28:48] something that you need to have in house [28:49] so for that you need to have the [28:50] engineer in house so that you control [28:52] the core technology but then you use the [28:55] partnerships to augment your core [28:57] technology with some partnerships So [28:59] that is the basic formula that's [29:01] >> control the technology and protect the [29:02] data. [29:03] >> You do both at the same time. [29:05] >> Exactly. [29:05] >> You have to [29:06] >> and then make sure you have a modular [29:08] open architecture that you can build [29:09] upon and that change just into term of [29:12] some of the success list that we can do [29:13] from here. [29:13] >> Can I add something [29:15] >> because I like the marriage analogy. I [29:16] think it's good. I like how you [29:18] deconstructed [29:19] >> lose it. No. [29:20] >> So when it comes to AI, I would it's [29:23] like marriage and being promiscuous I [29:25] think a little bit. But the core is [29:28] before you get married, you have to work [29:30] on yourself, right? In order to in order [29:32] to get married, you need to make sure [29:34] that you're ready. [29:35] >> Gone. [29:36] >> I love that we're having this [29:37] conversation in Vegas. [29:38] >> Get it done in like 10 minutes. [29:40] >> So, and hear me out. This [29:41] >> go to the gym and get get it done, guys. [29:44] >> So, the reason is data. Data is working [29:47] on yourself. And model providers will [29:49] come and go. Frameworks will come and [29:51] go. [29:52] >> Y [29:52] >> data is the most important thing in the [29:55] age of AI. When I talk about processes [29:57] in business, that's just data. Have you [30:00] created a process around what you do in [30:03] your business? Do you have [30:04] documentation? Do you have a it's just [30:05] data and it's in a form that AI can then [30:08] take action on it? [30:09] >> Absolutely. And I'll just add like yeah, [30:10] I mean you just touch the card, right? [30:12] You know, like the what is the data [30:15] readiness for agentic, right? We don't [30:17] talk enough about and and and the use [30:19] cases can come and go, we can change, [30:21] right? But how are we doing at the data [30:23] like you know whether it is active data [30:25] or metadata and the semantic and the [30:27] ontology and the taxonomy without that [30:29] we all can talk because we understand [30:32] each other the right where is who is [30:34] investing that so investing on [30:36] foundations of you know those before you [30:39] go too far I think will be otherwise you [30:41] will see the diminishing return very [30:43] quickly [30:44] >> so so watch this so take take two [30:46] examples right so take the case studies [30:48] and put them into action based on what [30:49] you're saying it's it it becomes is a [30:51] daunting task to go get all of your data [30:53] when it's structured to a place where [30:55] your AI can start to use it, right? [30:57] That's where I like to say go back to a [30:59] machine learning process and do it the [31:01] old the old way. If you want to go find [31:04] an easy path, easy path through data and [31:07] you want to apply generative AI to it, [31:09] go down an unstructured data path, [31:11] right? So think about what generative AI [31:13] can do for you to make the invisible [31:15] visible because ultimately Gen AI can [31:17] operate at a scale that none of you [31:19] humans can, right? So, I'll give you two [31:22] examples of of some case studies that [31:23] are easy for you to take back to your [31:26] companies. We have we have millions and [31:28] millions of of transcripts, right? We [31:30] talk to our customers all the time. All [31:31] of you talk to your customers all the [31:33] time. You can't have people sit around [31:34] and read all those transcripts, but AI [31:36] can. So, go go throw all of your [31:39] transcripts into a model and start [31:41] asking it questions or go back to, you [31:44] know, some audit function. If if all of [31:45] you work for a company of moderate size, [31:48] you have an audit process. Most of the [31:50] time you have your auditors come in, [31:52] they look at a sample set of your data, [31:54] they make some assumptions and then [31:55] whatever they find in terms of [31:56] anomalies, they go extrapolate and they [31:59] draw a lot of decisions based on what [32:01] those anomalies are. Right now AI can do [32:03] the whole audit process for you, look at [32:05] every single file instead of just the [32:07] sample set that the auditors looked at. [32:08] Then you can watch the auditors sit [32:10] there, if you're an auditor, I'm sorry, [32:12] you can watch the auditors sit there and [32:13] pick their nose. They don't know what to [32:14] do with themselves when we're done. [32:15] >> Yeah. Yeah. Now, so now [32:19] we all good. We hear about the [32:20] challenges. I want to talk about the [32:22] success. I want to talk about what makes [32:24] something successful, [32:26] right? So, and we we know the [32:28] difference. We know that in the past [32:29] traditional AI, you have your [32:30] performance metric, you have the AR [32:32] curve, the AU, you know, the [32:34] performance. But today AI is bit fuzzy [32:36] in nature, right? What is success? Uh [32:38] you guys are talk about summarization [32:40] for example as one of the core thing [32:41] that you guys are using in corporate, [32:43] right? we talk about that gal we talk [32:45] about that and also you um you talk [32:47] about that a simple summarization is [32:50] definitely in the eye of the beholder if [32:52] you have to summarize something for I [32:55] useful for S&P if you when my team was [32:57] developing summarizations we give that [32:59] to let's say you know a um a credit a [33:02] risk analyst and we give that to [33:04] somebody who does EG different view this [33:06] is good this is bad same algorithms [33:09] right so something as simple as [33:11] summization is [33:14] It's very hard to quantify success [33:16] because the success depend on the usage [33:19] >> and the uses have many you know have [33:21] many profile of users. So let's talk [33:23] about that what are the condition how do [33:25] we define success [33:26] >> in the case right now. So so get outside [33:29] with you here. So you are the architect [33:32] you want to make things successful [33:33] before they design it. [33:35] >> What will be the ideal how will be [33:37] success for you in that context? [33:39] >> Yeah to be I mean there are various [33:40] facets of success right? Um but one of [33:43] the most important thing I feel like is [33:45] adoption. [33:46] >> Mhm. [33:46] >> So whether like you talked about I mean [33:49] I can roll out you know a regulatory [33:51] governness like you know agent right or [33:56] the even you know giving this advisor [33:59] wealth advisor when we roll out some [34:01] agent to help them do the research body [34:03] right. I think I always look for [34:05] actually h how the adaption is there and [34:08] if those adaptions are not happening [34:11] then what are the root causes it there [34:13] are several factors and many a time if [34:15] if they are like not like once they if [34:18] they get a is it's very easy to lose the [34:21] confidence that if you don't have trust [34:24] on the data once if you lose your trust [34:26] so to me success depends on have we [34:29] built the trust and confidence in that [34:32] in the data in the results that we are [34:34] bringing in and which means when we are [34:36] rolling it out from adoption have we [34:38] have a cohort like starting small and [34:41] then starting to then before you go and [34:43] scale right to me those are the kind of [34:45] things like you start how I deploy [34:48] before I know that I'm deploying it you [34:51] know across the and then having this [34:53] continuous feedback loop right [34:55] >> observability is a critical aspect of [34:58] everything I we do and especially in [35:00] with AI so that you got to today [35:02] something is working doesn't mean you [35:04] just forget and fire and forget and go [35:06] something else right having that [35:08] observability built in as a part of your [35:10] stack [35:10] >> and having that you know a line of sight [35:13] non-stop is my my to me that's success [35:16] >> success so adoption right now and you [35:18] can [35:19] >> no like I'm just I I love what you just [35:20] said like an hour ago across the hall [35:22] this lady was talking about how 4% of [35:24] companies have found noticeable value [35:26] from agentic AI and so you either have [35:29] to sit up here and say success is a [35:31] lottery ticket which if you're chasing [35:32] Hey, like you have to you have to pay [35:34] you have to play to win. Or you can just [35:36] look at it as, hey, this whole thing is [35:38] very new right now. And if you're [35:39] getting it right, people are using it [35:41] and you're growing with it. Don't don't [35:42] don't go look for some CFO head turning [35:45] moment where you're going to deploy [35:46] something at scale that drives the [35:48] seauite nuts. It's just it's not going [35:50] to happen today, right? You're going to [35:52] have to start small, build up, and find [35:54] use cases where you can drive literacy [35:55] and adoption across the entire company [35:58] and then start to figure out how to [35:59] scale the back office functions and [36:01] everything that goes with it. and and [36:02] then solving the wrong problem with the [36:04] this is is the worst thing you can do. I [36:07] mean that's why solving the right [36:08] problem in the right way is the most [36:10] important thing. Right. [36:10] >> I see. I see. She had mentioned a grid [36:13] earlier which so anytime we go into an [36:15] enterprise the first thing that we're [36:17] looking at is if you're looking at a [36:20] business process how much time are you [36:22] actually spending because a lot of [36:23] people come with this idea that like oh [36:25] AI is going to save me from doing this [36:28] report every week. Well that report [36:30] every week takes an hour. So if I spend [36:32] three months building out a system for [36:34] you to save an hour, you're going to be [36:36] mad at me and your seuite's going to be [36:38] upset because we didn't actually do [36:40] anything worthwhile. So actually [36:42] defining these things where it's like if [36:43] I'm able to put a system in for this [36:46] process, how much time is it actually [36:48] going to save you? How much value is it [36:50] actually going to drive? And what you'll [36:51] find is in most business processes, [36:54] especially as companies grow larger, [36:55] it's very easily definable. you know [36:58] this reporting structure is taking 10 [37:01] hours a week across the org versus this [37:04] task is taking thousands of hours a [37:06] month across the org. Okay, let's start [37:08] here. [37:09] >> Yeah. [37:09] >> So adoptions and you're saying [37:10] productivity that's why you basically [37:12] referring to right now and then you're [37:14] saying continuous monitoring and [37:15] continuous change over times because [37:17] things change over time right now. [37:19] >> I'll give I'll give like a like a a [37:20] reference to what Russell was just [37:22] talking about. Granted, this is for a PC [37:24] only, right? So just wrap your head [37:25] around like we were looking to introduce [37:27] a capability. Uh we wanted to we wanted [37:30] to build something limited to test and [37:32] figure out if it was going to work or [37:33] not. [37:34] >> Spent two weeks talking about the [37:35] business process. We spent two weeks [37:37] building the agent. And then we spent a [37:39] week talking about change. [37:40] >> And so three weeks total not doing [37:42] anything other than talking about what [37:43] this thing was going to do and what it [37:44] needed to look like. And it was two [37:46] weeks to get the thing done. Now granted [37:47] we've got more work to do to get it to, [37:49] you know, full scale and you know how to [37:50] roll it out to other teams. But if [37:52] you're not spending time talking about [37:53] it, same perspective you're going to [37:55] miss. [37:56] >> I see. Say given your complex nature on [37:59] the business I know success is maybe a [38:01] bit fuzzy in a sense because it depend [38:03] on many kind of business stakeholder you [38:05] had different type of industry right [38:06] now. Do you have a a um a tool that you [38:11] use to go to your system saying look [38:13] done you guys be happy or something like [38:15] this? How do you find success in that [38:17] scenario? Yeah, look, um, [38:21] our company's obviously I I don't know [38:23] how much you follow the stock market or [38:26] Twitter, uh, but our company's obviously [38:28] seen better days and we our old CEO is [38:31] back as our new CEO and he took over in [38:35] I think it was May, uh, Steve Hemsley [38:37] and he was the board chair. So he's it's [38:40] like the Bob Iger, Howard, like it's [38:42] that move, right? I'm going to come back [38:43] and be the CEO. And um he's 73 years [38:47] old. And I have to admit, and the last [38:50] CEO is kind of one of the guys that [38:52] brought me to the company. And when he [38:54] joined, I have to admit I was like, [38:56] damn, this is going to suck. And it [39:00] turns out it's kind of been the [39:02] opposite. Uh he wants to take our board [39:05] meeting was just the last two days. He [39:06] he wants to take to the board. He wants [39:09] to take to to investors. He wants to [39:11] take across all 400,000 employees. Hey [39:14] guys, this is what we're doing. This is [39:16] how we're doing it. This is what's [39:18] working. This is what's not working. [39:19] He's all about transparency. He's all [39:22] about continu. He's like, don't hide [39:24] Be real with me. Right? So, we [39:28] together with Steve are defining success [39:31] in AI for the company. And I think it's [39:34] a few things, right? [39:36] First of all, um it has to actually be [39:40] doing useful stuff, right? Is it [39:42] actually helping people? Is it helping [39:45] the advocate? Is it helping the [39:47] struggling business? Is it helping the [39:49] doctor? Is it helping the patient? If [39:51] it's not helpful, maybe don't do it, [39:53] right? And and really the rule is if you [39:56] can't talk about it with a New York [39:57] Times reporter, don't do it. And talk [39:59] about it with pride, right? So that's [40:02] thing one. Is it actually helpful? two, [40:04] yeah, we're we have a two-part mission. [40:06] One is to actually help the, you know, [40:09] the American people with healthcare and [40:11] the other is to make shareholders money, [40:13] right? And that's capitalism. It's a [40:16] two-part mission. And we believe in that [40:19] mission. We believe you can do both. And [40:21] the the hard the the the debate that I [40:24] think a lot of us have had, you know, [40:25] when you're with management teams is [40:27] like, okay, are we gonna [40:30] see this as something that's an enabler [40:32] and not track the value or are we going [40:34] to intensely track the value? And I [40:37] think [40:38] >> we have fallen on the side of you know [40:39] what, [40:40] >> we're not going to be beholden to the [40:41] value, but let's track the value. Let's [40:44] be super transparent about actual pull [40:46] through. how is that actually impacting [40:48] bottom line you know at a line level and [40:51] then roll that up. Uh so that's number [40:53] two. Number three is are you inspiring [40:57] the organization? Are you inspiring the [40:59] industry? So that's gets down to [41:01] adoption and scale because you're not [41:03] going to be inspiring the organization [41:04] if you have a onesie that doesn't [41:06] actually matter to anybody but if it [41:07] matters to 20 million people or 200 [41:10] million people that might inspire, [41:12] right? So, and then I think you know the [41:15] the the last thing as a true measure of [41:18] success and you got at this a little bit [41:20] is you know if it's stopping at tech we [41:23] have roughly 400,000 people roughly [41:27] I think it's 30,000 tech FTEES and [41:30] another 10,000 contractors. So maybe 10% [41:32] of the workforce is tech give or take [41:34] those numbers can be slightly off. Um, [41:36] but if all you're doing is transforming [41:39] 40,000 people and not 400,000 people and [41:42] empowering and unbburdening 40,000 and [41:45] not 400,000, then you've just stopped at [41:47] the engineering level. You you haven't [41:48] actually touched your company. And so uh [41:51] making sure that you can go four levels [41:53] down into any business and see that that [41:56] person's life is be becoming better and [41:58] more productive and more effective [41:59] because of the technology at their [42:00] fingertips and that they are learning [42:02] how to reimagine processes at their [42:03] level such that we can do it at higher [42:05] levels. That's I think ultimately [42:07] success [42:07] >> that makes sense. Basically value come [42:10] from the you know the um the stakeholder [42:13] like you say. So the business will be a [42:14] partner for you to defining those value. [42:17] Then you quantify that with certain [42:18] adoptions that you go forward right now [42:21] and then you could also link to maybe [42:22] them having using it to save times maybe [42:26] to get some revenue which is a bit fuzzy [42:27] or doing that continuously to make sure [42:29] that they adapt you know for the change [42:31] they're moving forward. So I know we run [42:32] out of times but I just want to again [42:35] give a round of applause to our [42:36] panelists today and [42:38] >> yes give a round of applause for our [42:39] lovely panel members.