UTMB, Sodexo Chiefs Link AI Deployment Failures to Internal Process Defects

UTMB与Sodexo负责人将AI落地失败归咎于内部流程缺陷 当企业试图自动化存在缺陷的内部流程，而非明确定义工作流并将工具直接与利润指标挂钩时，企业级人工智能项目便会走向瓦解。 专家小组强调，规模化应用生成式AI的主要障碍在于组织脱节与流程梳理不足，而非技术能力问题。为避免陷入仅提升一线局部便利却无法带来机构级ROI的陷阱，管理者必须将技术部署逆向锚定至损益杠杆，同时保持员工乐于采用的消费级软件易用性。与会嘉宾分享了从环境临床病历记录、转诊分诊到预测性供应链等运营突破，主张开展灵活且高度可迭代的试点项目。展望未来，演讲者指出当前正迎来关键转型，正从纯文本的后台自动化迈向高带宽多模态系统，以消除繁冗的行政官僚流程并重构直接的人际连接。 技术极少成为规模化应用 AI 的真正障碍，组织的系统性不一致与未清晰梳理的业务流程才是核心瓶颈，因为企业无法规模化其无法明确定义的事物。 (3:26) 仅针对一线便利性优化的概念验证（POC）常因无法证明管理层看重的投资回报率（ROI）而夭折，成功的落地必须建立从一线工作流触发点到损益表（P&L）杠杆的完整反向映射。 (7:32) 环境临床听写（Ambient scribes）与 AI 驱动的转诊智能分诊在医疗运营中成效显著，不仅有效遏制了医生的职业倦怠，还彻底改变了传统先到先得的排队积压风险。 (17:34) 企业级 AI 正迎来关键转折点，应用重心正在从后台排程与文档分析等局部效率优化，全面转向以多模态为支撑、实时面向人的人机交互体验升级。 (20:22) 面对快速演进的模型生态，企业必须摆脱对潜在风险的过度空谈，以快速部署与敏捷迭代取代旷日持久的静态试点，在快速响应用户反馈中积累信任。 (25:34)

Enterprise generative AI initiatives falter not from technical limits, but from organizational incoherence, misaligned incentives, and vague executive mandates. Cross-industry leaders argue that sustainable scaling requires backward-mapping AI projects to balance-sheet metrics, integrating consumer-grade ergonomics into existing workflows, and favoring rapid iteration over risk-averse perfectionism. Process Definition and the Frontline-to-P&L Disconnect (3:26) Peter McCaffrey (UTMB) contends that technology is rarely the true enterprise bottleneck, observing that organizations cannot scale or automate workflows they cannot clearly define. Early generative AI efforts frequently suffered from an evangelical phase where proofs of concept were optimized strictly for frontline convenience, such as clearing administrative work queues. David Wang (Tetrate) notes that executive 'AI fever' often dictates tool adoption without defining measurable business objectives, creating downstream technical friction and worker skepticism. When frontline time savings fail to translate into demonstrable institutional returns, leadership often writes off otherwise functional tools as business failures. To bridge this divide, Brian Carlson (Sodexo) emphasizes establishing clear financial projections, governance, and stage gates to decommission unviable experiments before they damage IT credibility. McCaffrey advocates a disciplined backward-mapping process: identify top profit-and-loss levers (such as yield or unit-cost reductions), isolate the specific operational outputs that influence them, map the required employee actions, and only then introduce targeted AI automation to remove workflow friction. Adoption Through Ergonomics and Workflow Integration (21:26) David Wang points out that ChatGPT permanently reshaped enterprise software expectations by demonstrating intuitive, consumer-grade ergonomics. If an internal AI deployment introduces clunky interfaces or friction, frontline staff will reject it regardless of executive mandates. Quinn Favret (Tavus) and Carlson stress that successful deployments must meet workers within their established environments and habits, warning that attempting to reform user behavior while simultaneously introducing new software almost guarantees failure. The panel underscores that generative AI delivers its strongest adoption by stripping away bureaucratic 'ceremony.' Modern medical workflows, for instance, force physicians to spend up to 95% of their attention navigating electronic health records and data collection rather than clinical decision-making. Ambient scribes succeeded because they directly mitigated clinician charting burnout without disrupting natural patient examinations. Similarly, Sodexo introduces computer vision and sensory inputs directly onto physical kitchen stations to track safety and shrinkage without burdening staff with traditional software inputs. Targeted Deployments Across Operations and Healthcare (14:38) Production use cases deliver verifiable returns when applied to concrete operational bottlenecks rather than speculative, diffuse rollouts. At Sodexo, algorithmic integrations optimize perishable supply chains for items like milk and produce to eliminate waste, automate complex workforce scheduling, and power its Product IQ platform to recommend cost-effective purchasing options for hospitality clients. Tetrate monitors developer velocity alongside model compute expenditures and deploys support systems that automate research so service representatives can focus on customer interactions. In healthcare delivery, UTMB replaced conventional first-in, first-out administrative queues with an AI-assisted referral triage model. When a patient receives a nephrology referral, the system checks chart history to assess clinical urgency, bumping high-risk cases to the front of the schedule rather than leaving them in a static four-to-six-week backlog. Favret notes that deployments are also expanding from back-office administrative automation toward interactive human experiences, including conversational roleplay coaching for corporate training, automated patient intake, and modern alternatives to rigid phone response menus. Multimodal Expansion and Overcoming Risk Paralysis (25:34) The panel identifies significant untapped enterprise value in multimodal models incorporating voice, video, and computer vision. In clinical diagnostics, AI models inspecting CT scans can identify microscopic pancreatic cancer before it is perceptible to human radiologists, while handheld ultrasound hardware turns raw imaging into point-of-care assessments. In financial services, multimodal vision extracts structured underwriting intelligence from legacy charts and unstructured graphs. Furthermore, large-scale public initiatives, such as Texas investing billions into dementia prevention, will lean on ambient voice biomarker analysis via consumer hardware. Capturing these operational gains requires overcoming institutional paralysis. McCaffrey criticizes healthcare leadership's tendency toward 'admiration of the risk,' arguing that regulatory hesitation ignores the severe risks already embedded in broken, fragmented status-quo operations. Panelists agree that organizations must replace static multi-month trials with fluid proofs of concept, maintaining trust by responding rapidly to frontline user feedback and prioritizing quick deployment velocity over the search for a theoretically perfect solution. While the panel provides clear frameworks for targeting business metrics and integrating ergonomics, organizations still face the operational friction of overhauling antiquated operational processes without overburdening workers. The enduring enterprise challenge lies in maintaining agile governance and employee trust when underlying foundational models evolve significantly faster than standard institutional planning cycles.

在Ai4组织的企业生成式AI研讨会上，来自医疗、企业服务与技术基础设施领域的专家指出，AI规模化落地的核心瓶颈在于组织流程的不连贯与业务价值脱节，而非技术缺陷；企业必须将一线工具人机工学与顶层损益表指标紧密对齐，并通过敏捷迭代与多模态交互实现真正的业务重塑。 流程定义与损益表对齐决定落地成败 (3:26) UTMB首席数字与AI官Peter McCaffrey指出，技术极少是真正的瓶颈，企业无法自动化和扩展自身无法清晰定义的工作流；若缺乏端到端映射的标准流程，盲目部署AI只会放大既有的组织缺陷。Tetrate产品负责人David Wang补充道，管理层常因患上缺乏明确商业目标的“AI发热症”，导致技术决策沦为瓶颈并引发基层疑虑；与此同时，ChatGPT将用户的人机工学预期提升至消费级标准，若企业AI工具交互体验欠佳，即便战略目标再明确也会遭到员工排斥。 Sodexo高管Brian Carlson与McCaffrey一致认为，早期AI试点过度偏向一线员工的“局部便利”（例如快速清理待办任务），但如果省下的工时未转化为可衡量的商业产出，高管层便会将其定性为财务失败。Carlson介绍Sodexo通过“Sodexo Spark”框架将能力分类以驱动内生需求，并设立严格的项目退出机制（stage gates）以维护IT信誉。McCaffrey则提出严谨的逆向映射路径：必须首先锁定直接影响的损益表（P&L）杠杆（如产能提升或单位成本削减），倒推具体业务产出和员工操作，进而准确定位AI替代摩擦的切入点，形成清晰的价值链路。 剔除官僚琐事与深入核心运营场景 (17:34) 多位专家指出，生成式AI的核心价值在于剔除各行业的“行政仪式感”（ceremony），使专业人员回归本职。McCaffrey介绍，过去医生受制于数字化系统的繁琐操作，高达95%的心智用于数据检索等后勤琐务；UTMB引入环境临床记录工具（ambient scribes）让医生摆脱病历输入负担，有效遏制了医护人员离职与倦怠问题。针对医疗资源挤兑，UTMB将传统的“先进先出”转诊队列改造为AI智能分诊系统，在肾病转诊中直接对比患者病历评估危急程度，避免重症患者面临长达4至6周的延误危险。 在非临床与服务领域，Sodexo利用Product IQ推荐算法协助集团采购客户寻找低成本高品质方案，并部署生鲜易腐品预测性订购算法以减少浪费，同时运用计算机视觉监控厨房安全。Tavus联合创始人Quinn Favret强调，企业应借鉴Anthropic的策略“在用户既有习惯中相遇”，将AI无缝嵌入既有物理工位与工具中；若试图在部署技术的同时强行改变员工工作流，往往极易招致抵制与失败。 迈向多模态分析与高触感交互体验 (35:00) 企业AI部署正在跨越纯文本大模型的范畴，向语音、视频与计算机视觉深度演进。Favret观察到，企业正从单纯的后台效率优化转向交互体验升级，例如在企业培训中引入对话式AI模拟实时角色扮演，针对语调与肢体姿势提供即时辅导，并逐步取代刻板传统的电话交互式语音应答（IVR）系统。 在专业诊断与产业端，多模态正释放巨大红利。McCaffrey列举指出，AI分析腹部CT影像能在人眼可见病灶之前检出早期胰腺癌微观标志物；结合Butterfly iQ等便携式手持超声设备，AI能将像素直接转化为床旁诊断参考，使医生转向自动化诊断结果的裁决者。在公共健康方面，德州正规划利用消费级设备捕获环境语音生物标志物以推进痴呆症早期筛查。此外，实时音视频辅导已在技术支持场景中减少了50%的现场上门服务，并将疗效提升30%至40%；金融与保险机构则依托多模态视觉从历史图表中提取非结构化数据，用于精准核保与精算分析。 克服风险瘫痪并确立敏捷迭代机制 (25:34) 针对高度监管环境中的保守倾向，McCaffrey批评了医疗行业对合规风险的“孤芳自赏”（admiration of the risk），直言维持现状本身就包含着医疗供给不足的巨大临床风险；企业必须拿出研发抗癌化疗方案的审慎与紧迫感，探索安全可行的执行路径而非因噎废食。Favret强调，鉴于基础模型每周都在迭代演进，长达数月的静态概念验证（POC）已不合时宜，企业必须推行高频流动的POC，并快速响应员工反馈以维系信任基础。 Wang与Carlson将当前对AI的焦虑与当年企业争论互联网或公有云技术的情形相提并论，指出停滞不前即是在被动落后。与会专家在闭幕总结中一致建议：企业应秉持“速度优于完美”的原则，挑选具备高迭代能力的团队和供应商进行单点突破，及早确立清晰量化的衡量基准，通过在真实业务流中持续试错来建立长期技术壁垒。 本场讨论反映出企业AI布局正从盲目的概念验证转向对业务实质回报的深度审视。未来规模化应用的关键，在于管理者能否在复杂多变的合规要求下打破部门壁垒、建立跨越人机工学与损益指标的敏捷治理体系；而在高度去中心化的传统服务与一线现场，如何低成本且持续稳定地维护多模态工具的长期可用性，仍是企业需要应对的现实考量。

## Background 该视频由 YouTube 频道 Ai4 于 2026 年 8 月 22 日发布。视频记录了一场关于在大型企业中规模化部署生成式 AI 的圆桌研讨，聚焦于如何跨越概念验证（POC）走向全组织级应用，探讨跨部门协同、高管对齐、组织文化与战略框架等核心议题。 ## TL;DR 来自医疗、企业服务、前沿 AI 实验室及基础设施领域的专家共同指出，阻碍企业 AI 规模化的根本瓶颈在于组织流程的割裂与未被清晰定义的业务痛点，而非技术本身。讨论深入剖析了从自下而上的“局部便利”转向自上而下“可衡量业务价值”的必要性，并展示了在医疗智能分诊、供应链优化、研发效能提升与多模态人机交互等场景下的真实落地成果。专家们一致建议，企业应走出盲目罗列风险的停滞思维，摒弃对“绝对最优解”的执念，通过极速迭代、精准量化指标与端到端流程重构，将员工从繁琐杂务中解放出来以释放真正价值。 ## Atmosphere & style 整场圆桌讨论务实而犀利，嘉宾们摒弃了空洞的技术布道与狂热口号，直击企业推进 AI 时面临的真实阵痛与管理短板。各方观点交织碰撞、节奏紧凑，在揭示“薛定谔的部署”与医疗系统繁文缛节等现实困境时既幽默辛辣又发人深省。随着讨论从批判痛点转向多模态赋能与实战建言，现场洋溢着工程理性与行动导向的紧迫感，给观众带来一场高度专业、接地气且充满实战洞见的思维激荡。 ## Key takeaways - [03:26](https://www.youtube.com/watch?v=g1gW5GWNss8&t=206s) 企业 AI 规模化落地的核心障碍不是技术本身，而是组织流程的不连贯性与缺乏明确定义的业务标准，导致自动化与扩展无从谈起。 - [07:32](https://www.youtube.com/watch?v=g1gW5GWNss8&t=452s) 摆脱 AI 布道时代的投资回报陷阱：自下而上关注的一线局部便利往往与管理层关注的底线收益脱节，必须将技术动作直接对齐核心业务财务指标。 - [07:32](https://www.youtube.com/watch?v=g1gW5GWNss8&t=452s) 索迪斯提出 Spark 体验框架，将 AI 能力拆解为客户、消费者与员工三大体验支柱，通过直观呈现技术收益将落地推行从被动推销转为主动拉动需求。 - [10:48](https://www.youtube.com/watch?v=g1gW5GWNss8&t=648s) 企业级 AI 必须以端到端全流程视角建立治理机制，避免局部单点用例的孤立优化对上下游协作产生负面外部性。 - [12:14](https://www.youtube.com/watch?v=g1gW5GWNss8&t=734s) 组织信任的存续取决于反馈与敏捷迭代的响应速度，面对高频更迭的技术和潜在故障，快速响应与持续优化是防止信任瓦解的关键晴雨表。 - [14:38](https://www.youtube.com/watch?v=g1gW5GWNss8&t=878s) 研发效能与客户支持场景已从凭感觉评估走向精准量化，企业正严格核算功能上线周期、算力消耗并剥离无差别的繁重前置杂务。 - [16:29](https://www.youtube.com/watch?v=g1gW5GWNss8&t=989s) 医疗场景中环境语音听写（Ambient Dictation）大幅减轻医生职业倦怠，而基于 AI 的转诊智能分诊能够打破传统先到先得队列，动态置顶急危重症患者。 - [20:36](https://www.youtube.com/watch?v=g1gW5GWNss8&t=1236s) 生成式 AI 正迎来范式转移：应用重心从单纯的后台降本与效率优化，大步迈向面向人际互动（Human-Facing）的多模态与端到端体验重塑。 - [21:12](https://www.youtube.com/watch?v=g1gW5GWNss8&t=1272s) 众多 POC 无法规模化源于忽视消费级的人机工效学体验，以及缺乏阶段门禁（Stage-gating）止损机制与从 P&L 财务杠杆逆向推导的清晰链路。 - [26:56](https://www.youtube.com/watch?v=g1gW5GWNss8&t=1616s) 在医疗等高风险强监管领域，过度审视风险而维持现状同样是巨大风险，需吸取当年电子病历仅增加机械点击流杂务的数字化教训。 - [30:30](https://www.youtube.com/watch?v=g1gW5GWNss8&t=1830s) 治理逻辑应从消极罗列出错风险转向建立科学度量与监控体系，与其被动修补低效的手动流程，不如底层重构原生自动化工作流。 - [32:07](https://www.youtube.com/watch?v=g1gW5GWNss8&t=1927s) 深入现场一线赋能需在用户现有工作流中相遇，通过“剥离工作仪式感”减少繁文缛节与案头录入，让员工聚焦于核心决策与客户服务。 - [35:00](https://www.youtube.com/watch?v=g1gW5GWNss8&t=2100s) 多模态能力爆发释放非文本数据红利，不仅在医疗影像中展现超人类视觉诊断潜力，还在企业端大幅减少售后现场上门并盘活历史资产。 - [40:33](https://www.youtube.com/watch?v=g1gW5GWNss8&t=2433s) 落地起步阶段速度胜过完美，企业应优先绑定具备高敏捷迭代能力的团队，避免因过度苛求完美而拖慢落地转化。 - [41:14](https://www.youtube.com/watch?v=g1gW5GWNss8&t=2474s) 敏捷迭代的核心在于聚焦单点并务实地就可行方案达成共识，而非死磕无法预知所谓的最优解，同时应尽早确立精准的量化衡量指标。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=g1gW5GWNss8&t=4s) 圆桌讨论开场与嘉宾阵容介绍 TechCrunch 特约撰稿人、《Signals》AI 与创新通讯主理人 Kate Park 担任本次研讨的主持人。本次讨论的主题聚焦于“如何在大型企业中规模化落地生成式 AI（Generative AI）”，从概念验证（POC）走向全组织级部署。四位来自医疗、跨国企业服务、前沿 AI 实验室以及基础设施领域的行业专家受邀进行跨维度探讨： - **Peter McCaffrey**：德克萨斯大学医学分校（University of Texas Medical Branch，简称 UTMB）首席数字与 AI 官（CDAO），病理学家出身。UTMB 是一个集科研、教学和临床诊疗于一体的学术医疗系统，作为休斯顿地区广大患者的关键医疗安全网，其核心目标是提供可规模化、可靠且高质量的普惠医疗。 - **Brian Carlson**：索迪斯北美区（Sodexo North America）首席技术、数据、数字化与创新官，全面掌管传统 IT、AI、数字化及创新业务。索迪斯主要为医院（患者营养与零售餐饮）、企业办公场所、大学校园和体育场馆提供餐饮与综合设施管理服务。 - **Quinn Favret**：Tavus 联合创始人兼 COO。Tavus 是一家位于旧金山的 AI 研究实验室，自 2021 年（“在 AI 真正爆火之前”）便深耕该领域，已累计融资约 7500 万美元。他们专注于构建驱动实时交互的面部 AI 智能体（Face-to-face AI agents，内部代号为 "PALs"），让智能体具备逼真的外观、视觉与理解能力并可部署至各行业企业中。 - **David Wang**：Tetrate 产品负责人，负责构建帮助企业观测、管理和治理（Govern）AI 流量的基础设施，使企业清晰洞察组织内部 AI 的使用现状与优化空间，并将技术数据转化为实际业务收益。 --- ### [03:26](https://www.youtube.com/watch?v=g1gW5GWNss8&t=206s) 企业 AI 规模化落地的瓶颈：技术不是核心障碍，组织流程不连贯才是根因 针对主持人提出的核心问题——“阻碍 AI 在大型组织中落地的最大挑战究竟是技术本身，还是组织与人员的就绪度”，嘉宾们一致认为技术从来不是真正的瓶颈所在。 Peter McCaffrey 尖锐地指出，即便管理层不去主动推进技术部署，AI 技术也会不可避免地“渗透（Infiltrate）”进组织；但单纯拥有无限强大的技术根本无法奇迹般地解决复杂的医疗系统难题。医疗乃至大型企业面临的本质问题是“组织不连贯性（Organizational Incoherence）”——即究竟由谁在做什么、出于什么目的而做，内部往往缺乏统一认知。AI 并未制造这种断裂，而是像一面照妖镜将其彻底曝光：“你无法对连自己都无法明确定义的事物进行规模化与自动化（You can't scale and automate what you can't define to begin with）。”如果连“新确诊心力衰竭的患者如何完成建档与患者教育”这种标准化流程内部都缺乏共识，AI 就无从谈起。 Brian Carlson 紧接着补充，在面向成千上万名员工推广时，必须具备明确的问题、标准化的流程、合格的人员以及清晰的推进计划，以确保落地的“一致性与可预测性”。 ```mermaid flowchart TD accTitle: Enterprise AI Deployment Dilemma subgraph Bottlenecks["核心瓶颈"] A["业务流程未定义 / 组织不连贯"] --> B["AI 无法规模化与自动化"] end subgraph Approaches["企业落地路径"] C["自下而上（Bottom-Up）"] --> E["个人生产力工具赋能<br/>易产生局部便利与ROI脱节"] D["自上而下（Top-Down）"] --> F["聚焦特定高价值业务问题<br/>（如销售线索转化、培训客服）"] end Bottlenecks -.-> Approaches ``` Quinn Favret 从服务跨国 500 强到街边小店的广泛客户实践出发，将企业的落地路径归纳为两种截然不同的模式： - **自下而上（Bottom-Up）模式**：给团队提供通用生产力工具，以提升创造力、效率并优化决策，但面临全员培训和管理分散的挑战。 - **自上而下（Top-Down）模式**：企业带着极其具体的硬指标痛点而来（例如“提高官网销售线索转化率”或“培训员工以更好处理客服支持”），采用针对性产品在隔离场景中快速攻坚。两种路线没有绝对对错，但多数企业正在向特定场景明确的 Top-Down 方案倾斜。 David Wang 幽默地将当前的 AI 部署状态比喻为“薛定谔的部署（Schrödinger's problem）”——在任何给定时刻，AI 在一家组织里都同时处于“既部署了又没部署”的叠加态。只要不去追问，它就悄然存在于各个角落；一旦追问，大家就会对“什么是真正的部署”争论不休。他强调，如果没有建立起企业级的“信任（Trust）”，AI 就会永远停留在薛定谔状态。 --- ### [07:32](https://www.youtube.com/watch?v=g1gW5GWNss8&t=452s) 从“局部自嗨”到“高管对齐”：摆脱 AI 布道时代的投资回报陷阱 在深入探讨妨碍 AI 项目实质推进的阻力时，嘉宾们剖析了从高管层认知模糊到一线员工激励机制错位的多重断层。 David Wang 强调了领导层方向明确的重要性。许多管理层陷入了盲目的“AI 狂热（AI fever）”，只是一味高呼“我们需要更多 AI”，却无法给出具体的商业目标。这种模糊性会直接引发全组织的迷茫，导致技术选型失去优先级指导，使每项技术都沦为绊脚石；最终推向市场的技术不仅无法被业务人员消化吸收，甚至会引发员工对 AI 的最坏猜想与抵触心理。 针对员工认知与落地阻力，Brian Carlson 分享了索迪斯的应对举措。索迪斯将技术攻关重点转向变革管理与沟通，推出了名为 **“Sodexo Spark”** 的全新框架体系，将 AI 能力归纳并拆解为三大核心体验支柱： - **客户体验（Client Experience）** - **消费者体验（Consumer Experience）** - **员工体验（Employee Experience）** 通过这一框架，员工能够直观理解“这项技术对自己有什么好处（What's in it for them）”，进而变被动推行为主动拉动需求（Pull the demand）。 Peter McCaffrey 随后总结了企业正在经历的关键范式转变：过去很长一段时间的 AI 推进处于“高度布道化时代（Highly evangelical era）”，即为了证明 AI 能创造价值，团队深入到所谓的“一线（Frontline）”去询问“什么能让你感受到价值”。这种做法虽能快速促成大量 POC（概念验证），但一线员工的价值诉求往往局限于“本地便利性（Local convenience）”（例如只想清空眼前的待办任务工作流），这种诉求虽合理却与管理层（C-suite）关注的底线收益脱节。一线员工感到省时省力，并不等于企业获得了可衡量的 ROI，除非节省的带宽能切实转化为其他产出。如果优化方向无法对齐那些真正需要捍卫预算投资的副总裁（VP）级别高管，项目在基层看似成功，在全企业层面上依然会走向失败。 ### [10:48](https://www.youtube.com/watch?v=g1gW5GWNss8&t=648s) 端到端优化、治理平衡与团队信任的双重挑战 在企业级 AI 战略的讨论中，Brian Carlson 率先从大型企业运营的视角切入，指出不能仅仅孤立地看待局部的单一微观用例（niche use cases）。在复杂的企业环境中，如果某一个部门获得了局部胜利，却给上下游流程的其他环节带来了损失，这种单点突破就毫无意义。因此，企业真正需要的是对“端到端全流程（end-to-end process）”进行全面审视，并建立一种平衡的治理机制，确保 AI 部署能够为整个组织带来全局收益，这也是目前企业面临的核心挑战。 ```mermaid flowchart TD accTitle: 孤立用例与端到端协同治理对比 subgraph Siloed["局部孤立视角 (容易导致部门失衡)"] A["单点微观用例优化"] --> B["某一部门获益 / 效率提升"] B -. 负外部性.-> C["上下游协作受阻 / 整体失衡"] end subgraph EndToEnd["端到端协同视角 (全局共赢)"] D["端到端全流程审视"] --> E["平衡的治理体系"] E --> F["全链路体验优化与跨部门共赢"] end ``` Quinn Favret 随后将讨论引向了“信任（Trust）”的维度。他强调，无论 AI 的愿景多么宏大，归根结底都必须回答一个具体问题：员工在日常工作中是否真正相信 AI 能够稳定、可靠地完成既定任务？要建立这种信任，企业既需要推出经过深度测试、深度微调且切实可用的工具，又必须在“现实可用性”与“可能性的艺术（the art of the possible）”之间取得平衡。当前模型与技术产品日新月异，新工具层出不穷，企业管理者面临的现实难题是：如何在紧跟行业最前沿创新步伐的同时，确保团队对现有部署的工具保持坚定的信任。 --- ### [12:14](https://www.youtube.com/watch?v=g1gW5GWNss8&t=734s) 一线脱节、组织协同与通过敏捷迭代构建信任 Peter McCaffrey 进一步指出了企业在自上而下与自下而上推进 AI 时普遍暴露的深层矛盾——一线员工与高管层之间的认知脱节。想要真正实现可量化的投资回报率（ROI），整个组织必须对“价值是如何被创造出来的”达成连贯一致的理解，明确一线指标如何直接关联并驱动核心业务指标。Peter 直言：“这实际上是一个管理问题，而不是 AI 技术问题。” Brian Carlson 对此表示认同，并补充道，像 Sodexo 这样极其依赖一线人力的企业，其核心目标正是通过将员工从繁琐的后台事务中解放出来，把省下来的宝贵时间与精力全部重新投入到直接服务客户、消费者和宾客的一线互动中。 David Wang 接着分析称，组织的内在协同性（organizational coherence）是所有企业梦寐以求却极难达成的目标，这在 AI 时代不仅没有消失，反而因为 AI 带来的极高推进速度而被进一步放大——企业跑得越快，自身绊倒的地方就暴露得越明显。关于如何建立信任，David 提出： - **反馈响应速度是信任的关键晴雨表**：当业务利益相关方提出诉求时，企业必须通过快速迭代证明 AI 具备应对这些需求的能力；如果无法快速响应，信任就会迅速瓦解。 - **迭代速度决定信任存续**：传统企业以往从未面临过如此高频的迭代要求，一旦跟不上这一速度曲线，员工就会对技术失去信心。 - **跌倒后的应对策略**：Quinn 总结道，高速奔跑中必然会遭遇挫折或系统故障，防止信任崩塌的唯一方式就是“倾听第一、迅速响应并持续迭代优化”。 --- ### [14:38](https://www.youtube.com/watch?v=g1gW5GWNss8&t=878s) 落地实效：量化研发效能与解放客户支持流程 主持人 Kate Park 将话题转向当前正在切实产生业务价值的实际用例，询问各位嘉宾在自身企业或客户场景中看到哪些落地成果。Tetrate.io 的 David Wang 首先代表高科技与金融服务技术领域分享了经验，指出企业高管最关注的两大业务成果是**研发人员生产力（开发创新成本）**与**客户体验**： - **从“感觉（Vibes）”走向精准量化**：软件工程的范式已被 AI 彻底改变。领先的企业不再仅凭“感觉开发变快了”来做评估，而是精确量化功能交付速度与创收产品上线周期的提升，并将其与消耗的 AI 算力及工具成本进行严格对标。 - **剥离无差别的繁重事务（Undifferentiated Heavy Lifting）**：在客户服务与准备环节中，大量的调研、背景准备和基础劳动被 AI 取代，使团队能够将核心精力集中在“真正照顾客户”本身，而非“为照顾客户做耗时准备”。 --- ### [16:29](https://www.youtube.com/watch?v=g1gW5GWNss8&t=989s) 供应链、采购优化与医疗场景中的 AI 实战价值 Sodexo 的 Brian Carlson 介绍了 AI 贯穿其全业务价值链的多个成熟应用。Sodexo 推出了一款名为 **Product IQ** 的集采赋能产品，为酒店、餐厅等客户提供智能采购建议，实时推荐性价比更高、配送更准时的替代产品，并通过功能完备的独立采购门户优化整体支出。在内部供应链端，AI 被用于预测牛奶、生菜等生鲜产品的精准采购量，有效防止过度采购或供货不足。此外，Sodexo 还在劳动力排班优化以及面向消费者的移动端个性化推荐引擎中深度集成了 AI。 ```mermaid flowchart LR accTitle: 医疗转诊智能分诊对比流程 subgraph Traditional["传统排队模式"] A["患者转诊接入"] --> B["先到先得 FIFO 队列"] B --> C["非临床准入人员分配"] C --> D["病情延误风险 (滞留4-6周)"] end subgraph AISmart["AI 智能分诊模式"] E["患者转诊接入"] --> F["AI 深度解析病历与病情"] F --> G["动态优先级排队"] G --> H["紧急重症即时优先就诊"] end ``` 随后，UTMB 的 Peter McCaffrey 详细拆解了医疗健康领域的两大核心落地成果： - **全局分析与效能评估**：利用 AI 自动评估业务成效与转诊率变化，避免团队被繁重庞杂的常规数据汇报工作压垮，提高决策纪律性。 - **环境语音听写（Ambient Dictation / AI 医生助手）**：医生在问诊时只需将手机放在桌上与患者面对面交谈，AI 会自动完成病历文档书写并预先开立医嘱。这不仅保留了医患沟通的人文温度，更大幅降低了高昂且紧缺的医生资源的职业倦怠（physician burnout）。 - **转诊队列智能优先级重排**：以肾脏科（nephrology）转诊为例，传统模式依赖行政准入人员按“先到先得（FIFO）”被动排期，容易导致重症患者在长达 4 至 6 周的等待期内病情恶化。引入 AI 后，系统会自动审查患者病历与转诊单，智能识别危急程度并动态置顶高危病例，确保关键医疗资源第一时间覆盖最急需的患者。 --- ### [19:48](https://www.youtube.com/watch?v=g1gW5GWNss8&t=1188s) 从后台降本走向面向人际交互的 AI 新范式 面对医疗场景的案例，Quinn Favret 产生强烈共鸣，指出医疗领域的现状印证了一个普遍规律：“你解决的问题越多、效率提升得越高，涌现出的新挑战和新机遇也会成倍增长。” Quinn 提到，过去数年行业内的 AI 落地主要集中在后台提效、自动化、流程调度、智能分析，以及像 Granola 这样的个人会议转录记录工具上。而当前整个行业正在迎来一个重要拐点：**AI 部署正从单纯的后台降本（Back Office），快速迈向面向人际互动（Human-Facing）的全新交互场景**。例如在企业培训与组织发展（L&D）领域，AI 已经被深度用于高度逼真的角色扮演（Roleplay）和沉浸式业务模拟，实现真正拟真、双向互动的复杂沟通训练。 ### [20:36](https://www.youtube.com/watch?v=g1gW5GWNss8&t=1236s) 生成式 AI 的范式转移：从效率优化到体验重塑 在人机交互与多模态能力的演进中，AI 不仅能理解文本，还能实时捕捉肢体姿势、面部动作以及语调变化并提供即时反馈。这种深度交互正在迅速渗透至医疗问诊接诊、患者入院指引、术前术后护理，乃至求职面试等关键场景，成为企业建立首要品牌印象的前沿阵地。当前企业部署 AI 正在经历一场深刻的范式转换：在接下来的发展周期中，AI 的应用重心将从单纯追求“效率提升与成本压缩”，大步迈向“端到端体验优化”，全面替代过往那些僵化笨拙、体验倒退的传统交互式语音应答（IVR）系统与机械化流程。 ```mermaid flowchart TD accTitle: AI 应用演进与落地成功要素 subgraph Evolution["AI 演变趋势"] direction LR E1["传统阶段：纯效率与降本<br/>（替代人工、死板 IVR）"] --> E2["新兴阶段：体验深度优化<br/>（多模态感知、重塑交互）"] end subgraph ScaleFactors["规模化落地的核心支柱"] direction TB S1["人机工效体验 (Ergonomics)<br/>对齐消费级产品体验"] S2["明确业务价值 (P&L Levers)<br/>从财务杠杆逆向推导动作"] S3["阶段门禁与赞助 (Stage-gating)<br/>及早止损以捍卫团队信誉"] S4["敏捷流动态势 (Fluid POC)<br/>紧跟底层模型快速迭代"] end Evolution --> ScaleFactors ``` --- ### [21:12](https://www.youtube.com/watch?v=g1gW5GWNss8&t=1272s) 为什么许多 AI POC 无法规模化落地？人机工效与业务价值闭环 针对主持人关于“为何众多前景光明的 AI 概念验证（POC）最终折戟沉沙、无法走向企业级规模化”的提问，技术提供商与企业实践者给出了多维度的诊断。从技术界面的角度来看，决定规模化成败的关键变量往往在于“人机工效学”（Ergonomics）与最终交互体验。过去业界普遍默认企业级软件的交互体验极为糟糕，但以 ChatGPT 为代表的生成式 AI 本质上是消费级形态的产品，其核心词汇在于自然的“聊天”（Chat），这为大众设立了极高的交互预期。如果在企业内部推广一套 AI 工具时忽略了这种直观、顺畅的工效设计，再多的管理层对齐也难以挽救糟糕的落地效果。 从业务推进与组织治理的视角来看，缺乏严谨的立项与退出机制是导致试点搁浅的另一大主因。一个成功的项目必须在启动前就明确界定问题陈述与衡量“成功”的具体量化标准，这通常是一套严格的财务核算，并需要得到销售、财务或一线业务团队等对应业务负责人（Executive Sponsorship）的全力背书。更为关键的是建立“阶段门禁”（Stage-gating）机制：如果一个方案在纸面上都无法跑通商业逻辑，就应果断止损放弃，而不是硬着头皮推进注定失败的技术探索，否则技术团队在业务部门眼中的信誉将彻底丧失。 在医疗等复杂组织内部，阻碍规模化的深层矛盾在于“价值创造与员工日常意愿之间的摩擦”。像环境智能听诊（Ambient AI）这类工具之所以能迅速铺开，是因为它直击临床医生的职业倦怠痛点，员工有自驱力去使用；然而，许多具有巨大财务价值的后台任务（例如捕捉漏报计费）往往意味着新的增量工作量。要打破这一瓶颈，业务线负责人必须清晰展示技术动作与财务价值之间的直接视线（Line of sight to value）： - 明确核心财务杠杆：业务主管必须指出前三大损益（P&L）抓手是来自于产能提升还是成本削减。 - 逆向因果推演：从最终目标反向倒推——产生这一财务结果需要哪些具体动作？触发该动作的前置步骤是什么？一路倒推直至找到 AI 能够切入并接管的底层工作流节点。 - 确立推行动机：唯有让全体参与者深刻认同“做这件麻烦事是因为它能直接创造可衡量的业务价值”，项目才不至于因员工个人偏好的分歧而中途偏航。 --- ### [25:34](https://www.youtube.com/watch?v=g1gW5GWNss8&t=1534s) POC 的流变：在快速迭代的 AI 前沿中保持动态敏捷 从解决方案构建方的视角来看，能够在企业级部署中跑赢同行数月乃至数年的企业，普遍具备极快的响应速度：他们清晰定义成功指标，并在落地过程中维持极高的迭代与创新频率。当前的 AI 落地陷入了一种“薛定谔悖论”（Schrödinger's paradox）——企业往往既无法断言 AI 已经完全落地，也无法断言其尚未部署。 POC 固然为管理层提供了结构化的试错沙盒，但随着底层基础模型和工具链以周、月为单位高频更迭，传统的静态 POC 正在演化为一种动态流转的持续验证流程。企业必须学会在多个维度上同步推进与试错，推动技术边界不断向前。POC 的形式不会消亡，但其在企业内部的定义、评估方式与战略地位已经发生了根本性重塑。 --- ### [26:56](https://www.youtube.com/watch?v=g1gW5GWNss8&t=1616s) 医疗与高风险领域的 AI 抉择：坐视不理是最大的风险 在强监管、高风险的医疗行业中，如何划定“AI 自主决策”与“AI 辅助人类决策”的界线是一个核心命题。很多医疗机构常常陷入对潜在风险的过度审视（admiration of the risk），却选择性忽视了一个残酷事实：在当前医疗体系本就极不完备的现实下，“坐视不理、维持现状”（Not doing anything）同样是巨大的风险。合规与安全性固然是必须跨越的红线，但面对技术演进，机构应当具备如同寻找攻克癌症的新型化疗药物一般的紧迫感，以严密的工程逻辑去探索正确的实现路径，而不是任由组织文化停留在罗列风险清单并以此为由裹足不前。 回顾本世纪初医疗数字化的历史教训，当年业界天真地推行“有意义使用”（Meaningful Use）法规，试图通过全面推行电子病历（EMR）解决医疗问题，结果却是在陈旧落后的体系上硬套了一套数字化外壳，最终给医护人员留下了如同“PDF 文件海洋”般繁琐割裂的点击流界面。 - 临床决策被严重挤压：医生本应将核心精力放在诊断与照护决策上，现实中却有 95% 的脑力被前置的杂务、信息检索和系统打卡消耗殆尽，真正留给患者决策的时间仅剩可怜的一小部分。 - 医疗体验的历史倒退：对比二三十年前，患者进门有熟悉的接待员无缝流转信息给护士，护士测完体征转交医生，医生能在极短时间内提供精准决策；而如今的远程医疗或线上系统，却逼迫患者重复填写四遍相同的表单，医生深知患者只是机械地勾选“否”，导致医患双方陷入无效沟通的恶性循环。 - AI 的破局使命：将医护人员从荒谬的数字化后勤杂务中解救出来，即使只能将临床决策时间的占比恢复到一半，也是对当代医疗体系的一次重大拯救。 ### [30:30](https://www.youtube.com/watch?v=g1gW5GWNss8&t=1830s) 流程重塑与治理逻辑：从“罗列风险”转向“度量与监控” 在探讨医疗及企业服务中低效流程的弊端时，嘉宾们指出，许多传统流程因为拼凑了低劣的软硬件而出现体验倒退——例如一项原本高效的工作，却因为反复机械地核对信息而耗费极高价值专家的 20 分钟时间，最终效果依然不尽如人意。面对这种痛点，企业面临两种根本不同的路径选择：“自动化现有的手动流程”或者“重新构建全新的原生自动化流程”。嘉宾强调，在当今 AI 技术演进的大潮中，“原地踏步就等于落后，因为其他人都在全力向前推进”。 针对医疗等强合规行业对 AI 风险的顾虑，讨论指出，与其不断罗列 AI 可能会出错的各种场景而止步不前，不如聚焦于“如何把事情做对”。这种技术变革与当年的数字化转型、互联网普及乃至公有云的引入如出一辙——每一次技术浪潮初期，行业都会经历摸索阶段，但最终都会建立起成熟的衡量与治理体系。AI 的渗透是不可阻挡的必然趋势，企业应当将核心精力放在如何建立科学的评估指标与监控机制，确保其安全、受控地创造价值。 ```mermaid flowchart LR accTitle: AI Adoption Governance Cycle subgraph Mindset["认知转变"] A["传统思维: 罗列风险 / 拒绝变革"] --> B["现代思维: 构建指标 / 动态治理"] end subgraph Execution["落地路径"] C["被动修补: 自动化低效手动流程"] D["底层重构: 创建原生自动化工作流"] end B --> D ``` --- ### [32:07](https://www.youtube.com/watch?v=g1gW5GWNss8&t=1927s) 深入一线业务场景：以“剥离繁文缛节”赋能传统与现场员工 当话题转向 Sodexo 这样深耕于后厨、物业设施和现场一线运营的非传统科技环境时，Brian Carlson 阐述了 AI 在基层员工中的实际应用场景。Sodexo 在各个驻点拥有数以千计的一线员工，他们日常需要进行菜单规划、厨余浪费追踪与物资管理。随着多模态能力的成熟，AI 在健康安全管理以及库存损耗防范（Shrinkage Prevention）方面展现出了巨大的应用潜力，成为未来数年内持续深耕的重点用例。 Quinn Favret 补充道，向一线工人推广技术的关键在于“在用户所在的地方与他们相遇（Meet people where they are）”。他以 Anthropic 为例指出，成功的技术部署不是强迫用户颠覆现有的作业习惯去适应全新工具，而是把 AI 深度嵌入到现有的工作流与环境中，尽量减少单个人员需要做出的改变。David Wang 对此深表赞同，并提出了“剥离工作仪式感（Stripping away the ceremony）”的概念： - **餐饮服务业**：剥离繁杂的数据录入仪式，让员工将更多时间投入到理解和真正服务客户上； - **医疗卫生领域**：减少大量冗长病历书写等案头工作，让医生专注于核心的医疗决策与患者救治； - **软件工程领域**：减少繁琐的方案审批流程，让工程师专注于高价值的实际问题解决。 --- ### [35:00](https://www.youtube.com/watch?v=g1gW5GWNss8&t=2100s) 超越文本的多模态爆发：语音、影像与视频释放的数据新红利 当前行业对生成式 AI 的讨论大多局限于文本，但多模态（语音、图像、视频）正在开启全新的价值空间。Peter McCaffrey 从病理学与医学影像的角度指出，AI 不仅能自动化人类已有的识别能力（如辅助确认结节），更具备超越人类肉眼极限的全新诊断能力。例如在腹部 CT 扫描中，早期胰腺癌的蛛丝马迹在人眼看来极难察觉，但 AI 算法能够敏锐捕获，这相当于创造了一种全新的床旁诊断范式。随着 Butterfly iQ 等便携式超声设备的普及，前沿数据的采集变得极其轻便，医生的角色正在从传统的“靠手触摸肿块”转变为“对 AI 风险预警进行综合裁决”。此外，得克萨斯州投入数十亿美元用于失智症预防，这类筛查未来将不再局限于医院，而是通过 Siri 等日常语音交互，利用更丰富的多模态生物标志物实现早期被动监测。 ```mermaid flowchart TD accTitle: Multimodal AI Value Across Industries subgraph Healthcare["医疗健康领域"] H1["微型硬件普及 (如 Butterfly iQ)"] --> H2["AI 超人类视觉 (早期胰腺癌/病灶捕获)"] H3["日常语音/交互数据"] --> H4["常态化生物标志物监测 (失智症预防)"] end subgraph Enterprise["企业与服务领域"] E1["视频/面部情绪分析 (教学辅导/客户支持)"] --> E2["减少 50% 现场上门 / 提升 30-40% 疗效"] E3["非文本历史资产 (金融图表/保险保单影像)"] --> E4["数字化第二阶段 (精细核保与客户洞察)"] end ``` Quinn Favret 从构建对话式音频与视频模型的角度进一步阐述了富媒体数据的独特价值。与纯文本相比，视频和音频能够传递极高维度的信息： - **情绪与认知状态感知**：在远程辅导中，系统可以通过面部微表情与语调判断学生是因微积分难题而感到沮丧，还是跃跃欲试希望挑战更难的问题； - **交互式远程排障**：在客户支持场景下，用户只需将摄像头对准故障打印机，多模态 AI 便能实时指导其完成维修。这类部署已在实际场景中帮助客户减少了 50% 的现场上门服务，并将干预疗效提升了 30% 至 40%。 David Wang 补充指出，在金融服务与保险行业，多模态 AI 标志着“数字化转型的第二阶段（Phase 2 of Digital）”。许多核心资产（如历史财务走势图、精算数据图表、扫描保单）本质上都是图像形式，多模态模型能够深度解析这些非文本资产，帮助金融机构更精准地制定保险策略并形成核心竞争壁垒。 --- ### [39:56](https://www.youtube.com/watch?v=g1gW5GWNss8&t=2396s) 专家建言：AI 落地起步阶段的核心法则 在圆桌讨论进入尾声之际，主持人 Kate Park 邀请各位嘉宾为刚踏上 AI 征程的企业提供避坑指南与供应商建议。Quinn Favret 率先总结道，在技术瞬息万变的当下，企业面临的各种建议可能极其复杂，但最核心的一条法则是：“速度胜过一切（Speed matters more than anything else）”。无论企业当前处于技术探索的哪一个阶段，快速验证、快速迭代才是拉开竞争差距的关键。 ### [40:33](https://www.youtube.com/watch?v=g1gW5GWNss8&t=2433s) 速度与完美的博弈：通过敏捷迭代与找对伙伴实现落地 在探讨企业如何跨越 AI 采纳曲线时，嘉宾们深入分析了技术落地过程中速度与质量之间的核心张力。无论一家企业当前处于技术创新的最前沿，还是暂时处于落后状态，真正决定其能否从 AI 中获取实际收益的唯一要素，就是看他们能以多快的速度将技术投入实际应用并落地部署。然而在实践中，“速度”最大的天敌往往就是对“完美”的过度执念。如何在快速推进与追求完美之间找到最佳平衡点，被证明是企业最终能否取得最佳业务成果的关键驱动力。 随后，现场嘉宾将这一观点进一步升华，提出企业在推动项目时，必须将“可迭代性”（iterability）置于所有其他品质之上。特别是当团队作为开路先锋率先探索新技术时，切忌去选择那些态度消极、满腹怀疑的合作团队；相反，必须挑选那些具备敏捷迭代能力、愿意同频共振的伙伴并肩前行，唯有这样才能在快速试错中建立正向循环。 ```mermaid flowchart TD accTitle: 速度、迭代与落地的平衡循环 subgraph Mindset["战略思维"] A["拒绝完美主义陷阱"] --> B["追求落地与采用速度"] end subgraph Execution["执行落地"] B --> C["筛选具备高迭代能力的团队"] C --> D["快速试错与敏捷验证"] end D --> E["达成最优业务成果"] E -.->|持续反馈| A ``` - **速度决定收益**：企业从 AI 获益的实际能力，直接取决于从评估到应用落地的速度有多快。 - **摒弃完美主义陷阱**：过分苛求完美往往会扼杀速度，必须在推进速度和技术精细度之间实现动态平衡。 - **挑选高迭代潜力的团队**：起步阶段不要浪费时间试图说服持怀疑态度的队伍，而要优先与能够持续、敏捷参与迭代的伙伴深度绑定。 --- ### [41:14](https://www.youtube.com/watch?v=g1gW5GWNss8&t=2474s) 聚焦明确定义与量化指标：达成共识而非死磕“最优” 紧接着，其他嘉宾从方法论层面对快速迭代进行了更具体的拆解。所谓的“快速学习”（learn fast），本质上就是一种积极正向的“快速失败”（fail fast）。企业在推行 AI 时切忌“病急乱投医”、像漫无目的地“往墙上扔意面”（throw spaghetti at the wall）那样盲目试探。真正的捷径反而是保持极度聚焦，一次只专注攻克一件事，脚踏实地将方案推进到底，同时必须清晰定义“达到良好状态到底意味着什么”（what good looks like）。 进一步的探讨指出，在所有能成功快速迭代的团队中，存在一个极其微妙却决定性的认知差异：团队需要达成的共识，是“就某个方案达成一致”，而不是“非要找到那个绝对最优的解”。在技术演进飞速的当下，根本没有人能预知何为绝对的“最优”；如果非要死磕完美答案，只会陷入无休止的内耗与停滞。此外，团队必须从定性评估全面转向定量衡量——明确“良好”的量化指标，远比口头定性地描述更具指导意义。及早确立这种量化指标和务实的共识机制，能帮企业在后续扩展阶段避免付出巨大的沉没成本与试错阵痛。最后，主持人在掌声中致谢各位嘉宾，圆满结束了整场圆桌讨论。 - **快速学习与聚焦单点**：拒绝广撒网式的盲目试错，一次只做一件事并坚持做到底，把失败转化为快速吸收经验的通道。 - **求共识，不求所谓的最优解**：“同意某一个方案”（agree on something）远比“必须就最优解达成一致”（agree on the best thing）高效得多，因为在未知领域没人能预先定义真正的最优。 - **量化指标先于定性描述**：将评估标准从单纯的定性（qualification）转向精准的量化（quantification），及早消除未来可能出现的业务隐患与阵痛。 ## Notable quotes > [03:26](https://www.youtube.com/watch?v=g1gW5GWNss8&t=206s) 你无法对连自己都无法明确定义的事物进行规模化与自动化（You can't scale and automate what you can't define to begin with）。 — Peter McCaffrey > [03:26](https://www.youtube.com/watch?v=g1gW5GWNss8&t=206s) 薛定谔的部署（Schrödinger's problem）——在任何给定时刻，AI 在一家组织里都同时处于“既部署了又没部署”的叠加态。 — David Wang > [12:14](https://www.youtube.com/watch?v=g1gW5GWNss8&t=734s) 这实际上是一个管理问题，而不是 AI 技术问题。 — Peter McCaffrey > [14:38](https://www.youtube.com/watch?v=g1gW5GWNss8&t=878s) 将核心精力集中在“真正照顾客户”本身，而非“为照顾客户做耗时准备”。 — David Wang > [19:48](https://www.youtube.com/watch?v=g1gW5GWNss8&t=1188s) AI 部署正从单纯的后台降本（Back Office），快速迈向面向人际互动（Human-Facing）的全新交互场景。 — Quinn Favret > [26:56](https://www.youtube.com/watch?v=g1gW5GWNss8&t=1616s) 坐视不理、维持现状（Not doing anything）同样是巨大的风险。 — Peter McCaffrey > [30:30](https://www.youtube.com/watch?v=g1gW5GWNss8&t=1830s) 原地踏步就等于落后，因为其他人都在全力向前推进。 — Peter McCaffrey > [32:07](https://www.youtube.com/watch?v=g1gW5GWNss8&t=1927s) 在用户所在的地方与他们相遇（Meet people where they are）。 — Quinn Favret > [39:56](https://www.youtube.com/watch?v=g1gW5GWNss8&t=2396s) 速度胜过一切（Speed matters more than anything else）。 — Quinn Favret > [41:14](https://www.youtube.com/watch?v=g1gW5GWNss8&t=2474s) “同意某一个方案”（agree on something）远比“必须就最优解达成一致”（agree on the best thing）高效得多，因为在未知领域没人能预先定义真正的最优。 — David Wang ## Mentioned Kate Park, TechCrunch, Signals, Peter McCaffrey, 德克萨斯大学医学分校 (University of Texas Medical Branch / UTMB), Brian Carlson, 索迪斯北美区 (Sodexo North America), Quinn Favret, Tavus, PALs, David Wang, Tetrate, Sodexo Spark, Product IQ, Granola, ChatGPT, Anthropic, Butterfly iQ, Siri ## Video description Peter McCaffrey - Chief Digital and AI Officer, The University of Texas Medical Branch Kate Park - Editorial Contributor, TechCrunch Brian Carlson - NorAm Chief Tech, Data, Digital & Innovation Officer, Sodexo Quinn Favret - Co-Founder and COO, Tavus David Wang - Head of Product, Tetrate.io Scaling generative AI such as LLMs and multimodal AI from POC to enterprise-wide deployment requires not just technical readiness, but strategic vision. This session will focus on the broader organizational challenges and opportunities of integrating AI at scale. Panelists will discuss theories, best practices, and innovative approaches to embedding AI systems across business units to drive value. Explore how companies are navigating cross-functional alignment, ensuring leadership buy-in, and fostering a culture that supports widespread AI adoption. Learn about the strategic frameworks that can maximize the impact of generative AI on a large scale. [Source](https://www.youtube.com/watch?v=g1gW5GWNss8)

[3:26] Peter McCaffrey: 你无法对连自己都无法明确定义的事物进行规模化与自动化（You can't scale and automate what you can't define to begin with）。 [3:26] David Wang: 薛定谔的部署（Schrödinger's problem）——在任何给定时刻，AI 在一家组织里都同时处于“既部署了又没部署”的叠加态。 [12:14] Peter McCaffrey: 这实际上是一个管理问题，而不是 AI 技术问题。 [14:38] David Wang: 将核心精力集中在“真正照顾客户”本身，而非“为照顾客户做耗时准备”。 [19:48] Quinn Favret: AI 部署正从单纯的后台降本（Back Office），快速迈向面向人际互动（Human-Facing）的全新交互场景。 [26:56] Peter McCaffrey: 坐视不理、维持现状（Not doing anything）同样是巨大的风险。 [30:30] Peter McCaffrey: 原地踏步就等于落后，因为其他人都在全力向前推进。 [32:07] Quinn Favret: 在用户所在的地方与他们相遇（Meet people where they are）。 [39:56] Quinn Favret: 速度胜过一切（Speed matters more than anything else）。 [41:14] David Wang: “同意某一个方案”（agree on something）远比“必须就最优解达成一致”（agree on the best thing）高效得多，因为在未知领域没人能预先定义真正的最优。

[3:26] Peter McCaffrey: You can't scale and automate what you can't define to begin with. [7:32] Brian Carlson: what's in it for them [13:10] Peter McCaffrey: If you can't show them that their needs... could accommodate their feedback, you will erode their trust. Eventually, they will turn against you. [21:26] Peter McCaffrey: line of sight to value [25:34] Peter McCaffrey: admiration of the risk [28:37] Peter McCaffrey: sea of PDF files [28:37] Brian Carlson: you can automate a manual process or you can create an automated process [32:07] Quinn Favret: meet people where they are [33:57] ceremony [39:56] throwing spaghetti at the wall

[0:04] Good morning everyone. Thanks for joining today.  Uh for those who don't know me, my name is Kate [0:09] Park. I'm a contributing writer for Techrunch  and I also run my own newsletter, Signals, [0:16] um focused on AI and innovation  across Asia and beyond. Okay. So [0:22] um I'm very excited today uh to we are  going to talk about scaling AI scaling [0:27] generative AI across organizations. Uh I'll  be your moderator for the next 45 minutes and [0:34] uh we will we have a great uh great panelists  here today each approaching this challenging [0:40] from a different uh different perspective.  So let me briefly introduce our panelists. [0:50] Okay. Peter McAffrey, chief digital and AI officer  at the University of Texas Medical Branch. Brian [0:57] Carson, North America chief technology data  digital and innovation officer at Sodexo. Queen [1:03] Fabret, co-founder and COO of Tabos and David  Dang, head of product at Tetrate. I'd love for [1:10] each of you to briefly introduce yourself uh and  tell us about what your organization does. Yeah, [1:18] let's all start. Hey everyone, I'm Peter McAffrey.  So, uh, I'm a pathologist by training. My main [1:23] job at the University of Texas Medical Branch  is I head our our digital and AI efforts. I'm [1:27] our digital and AI officer. Um, you know, we're  an academic health care system. We do research, [1:32] teaching, patient care. Ultimately though, we  serve as a safety net for many patients in the [1:36] Houston area and is to provide scalable, reliable,  highquality care to everyone. Thank you. Uh, [1:44] hello, I'm Brian Carlson. I head the North  American tech teams for Sedexo North America [1:50] which incorporates your standard IT stuff, your  AI, your digital, your innovation, those types [1:57] of things. We do what we call food services and  facility management which uh we sell to hospitals [2:02] for patient nutrition and retail, corporate  offices, college campuses, sports and stadiums, [2:08] things like that. Uh and we do a combination of  food service as well as facilities management. [2:13] Thank you. Oh, yeah. Go ahead. Awesome. Hey, good  morning everyone. My name is Quinn Favre. I'm the [2:18] co-founder of TAVIS. Tavis is a AI research  lab based in San Francisco. Specifically, [2:23] a lot of what we're building are models to  power real-time interaction. Right? So, these [2:27] face-to-face AI agents, we call them PALS, that  you can actually talk to and they look like you, [2:32] they see like you, they understand like you, and  you can deploy them across different enterprises. [2:36] So, we've been building since 2021 or so. so long  since before AI was cool and um since we had big [2:42] conferences like this and uh since then we've  raised about $75 million or so to pursue this. [2:46] So it's been quite the quite the fun ride. Thank  you. Hello everyone. I'm uh David Wang. I look [2:52] after product management for for Tetrate. Um what  Tetrate does is we help you to understand your [2:58] AI traffic, help you to understand and govern  it. And that means for the people who are using [3:04] AI in your organization um the enterprise can  understand how AI is being used and what could [3:10] be changed about it. And so my job basically is  to build the infrastructure that allows this to [3:15] happen and increasingly help our customers use the  information from our infrastructure for business [3:21] gains. All right. Thank you everyone. Let's  jump right in. So we are going to talk about [3:26] let's start with where we are today. Uh where  are we right now? Everyone is talking about [3:31] AI transformation. Uh but how difficult is it  actually to roll out AI across large organization. [3:38] So is it is the biggest challenge the technology  itself or getting people and organizations ready [3:44] to actually adopt it? So who want to go first?  Peter, you want to go first? Yeah. Yeah. So, I'll [3:51] tell you I um there's definitely technology needed  to enable that, but what we always talk about at [3:57] UTMB is it's not really the bottleneck. Like even  if we don't deploy tech, tech tech will infiltrate [4:05] the organization anyway. Not that that's our plan,  but that will happen. And so the the real question [4:11] is so what? If you have infinite technology,  that's not going to magically fix healthcare. [4:16] I think the problem that we have is as a system  there's just a lot of organizational incoherence [4:21] in what's being done by whom for what purpose  and AI has revealed that didn't create that but [4:26] it's really revealed that that you can't scale and  automate what you can't define to begin with and [4:32] so if no one can say yes this is the process  that everyone mutually understands about how [4:38] a patient with a new diagnosis of heart failure  gets on board and educated then that's the root [4:43] cause and to me that's really the thing that  impacts acts any deployment whether it's AI or [4:48] not AI and that's the root issue we have because  we're just a complex org with lots of different [4:53] uh sub teams and processes that we're trying to  wrangle I build on that a little bit you know to [5:00] me it's you got to have a problem to solve you  got to have a process you got to have people [5:05] and you got to have a plan to actually drive it  out because you're probably like us where you're [5:10] trying to scale across thousands and thousands of  people and it has to be consistent it has to be [5:15] predictable and it has to be a use case that's  clearly defined. Yeah, it's it's been really [5:21] interesting from our purview too. I I completely  agree with what you all said. We work with a [5:25] really diverse set of clients, right? So, everyone  from Fortune 500s to Fortune 10s to enterprises in [5:30] different countries to a mom and pop shop, right?  And we've seen two different styles of approaches [5:35] when it comes to really implementing AI across the  enterprise. One of which is, I think, what we're [5:39] starting to mention up here, which is a bottom-up  approach, right? How do we give our team members, [5:43] our groups, our our people the right tools to help  them be more creative, productive, efficient, um, [5:49] and ultimately make better decisions? And so that  has its own distinct and unique challenges. The [5:53] other side of the coin and what we're also seeing  is a really big trend towards top- down enterprise [5:58] deployments, right? Where companies are coming in  and they say, "Hey, I have this specific problem, [6:02] right? I have a problem with converting sales  leads on my website. I have a problem with [6:07] training and and educating my team members so they  can handle support requests better. And they're [6:11] coming in and they're looking for solutions or  specific products or areas to really drive that [6:17] down. And it's much more about hey, how do we do  a great roll out um in a isolated manner versus [6:22] how do we train everyone to use this in different  ways? And there's definitely not a right answer [6:26] between either of them. But it is interesting to  see some enterprises definitely gravitate towards [6:30] one end or the other, some fall in between. And  um I I think that within itself is something [6:34] worth calling out as well that it's not just  giving people the tools but it could actually [6:37] be rolling out specific solutions as well. Yeah.  So just completely agree with this uh infiltration [6:44] terminology and I think the thing interesting  I see is that um most people won't even be able [6:50] to tell you what what it looks like when AI is  deployed truly. So well I think uh tongue and [6:56] cheek I think AI deployment and AI transformation  is kind of like a Schroingers problem. At any [7:01] given moment uh AI is both rolled out and not  rolled out in your organization and you won't know [7:06] until you ask the question and then people start  to debate over the definition of rolled out. So [7:11] um I would say overall AI is already here. It's  everywhere but then is it deployed? That depends [7:16] on your definition of what deployed is and a lot  of times the bottleneck is uh for some people [7:23] to call it deployed it has to be trusted and uh  without trust it is in the shorter state. Okay. [7:32] The next question is very similar but what are  the biggest things that prevent AI deployment [7:38] uh from moving forward? Is it like a technical  limitation or security concerns or simply getting [7:44] people aligned? So yeah I can yeah I I'll get  started with that. So it's like what I see is [7:52] um yes to all the like every every aspect  there's bottleneck but in different sequence [7:59] I think there's a huge leadership component  in terms of even defining the goal of an AI [8:05] roll out the business objective um and I think  that's often fuzzy because um sometimes leaders [8:10] just get AI fever and they say we just need more  AI without defining what it's for and that puts [8:15] the organization in a state of confusion and if  you start off with that confusion And of course [8:19] the technology route becomes much harder because  um there's no path to determine to prioritize or [8:27] guide deep deployment and then every tech becomes  a bottleneck. And then finally like even if you [8:33] do get some tech out there people won't be able  to absorb it and even understand what it's for [8:37] and therefore sometimes they just make the worst  assumption about AI and you won't get adoption. [8:42] So that's that's what I see. Yeah. I'm going to  I'm going to go back to my people part of the [8:47] process. And uh you know, we've been putting a lot  of effort into communication, change management, [8:53] explaining things. We've even recently kind of  rebranded or readjusted our messaging to what [8:59] we're calling now Sedexo Spark, which is how we  group our capabilities into client experience, [9:05] consumer experience, and employee experience.  And we're hoping that's going to help people [9:09] kind of understand in what's in it for them and  where they can go to find the latest and greatest [9:15] tools and capabilities for them. So we've kind of  gotten through some of the other more technical [9:20] components of things now and a lot of it's really  focused on how do we get people excited, engaged [9:26] and have them pull the demand through the process.  I think one thing I would say is really to echo [9:32] one of the points you made. I think and we as an  organization have veered more towards these top- [9:37] down models because I think one of the challenges  we've observed there was an era many of us are [9:43] still in it of AI deployment that was I would  call highly evangelical meaning the whole idea [9:48] was to convince you that it's possible to have  value and the way to do that was to go throughout [9:52] the enterprise to the quote unquote front line  whatever that means and then ask them what makes [9:57] you feel value and this is does effectively get  PC's but the problem is that the incentives and [10:04] value at the front line are really not the same  as in the seauite. They're almost always valid, [10:10] but they're about local convenience. I want my  work queue to be done and the and this is we're [10:17] trying to mature out of this to a degree because  even if you achieve that, it doesn't have an ROI [10:23] in of itself unless it's put bandwidth somewhere  else or whatever. And this is an organizational [10:29] capability. And so I think the big challenge  is you can you you will realize that the thing [10:35] you are tuning to and optimizing for is often not  actually relevant to the VP person who's going to [10:42] defend the investment and this uh can be a success  at the front line but will be a failure kind of [10:48] at the seauite level and cautioning against that  I think is an important lesson in all that right [10:54] you want to add something can I add something on  top of his because we're both in enterprise right [10:59] To me, it's that endto-end process and how you  optimize the endto-end experience for everyone [11:04] in the loop. And I to me, that's what you were  saying. It's just I word it a little differently, [11:07] but it's really important not to look at  little use cases, niches, because it wins [11:11] for one and someone else loses. And how do you  have a balanced governance that ensures that [11:16] it's a win for the organization, right? And  that's that's the real challenge right now, [11:20] I think. Yes. And ultimately to to the first point  too, all of this comes back to trust, right? Now, [11:26] do do you trust in your day-to-day life things  like that? Th those are maybe pie in the sky [11:29] questions, but do you trust it to reliably execute  or deliver the job that you have deployed it to [11:34] do, right? And so how do you make your team trust  AI, right? Ultimately becomes that question. And [11:42] um having these really deep, really fine-tuned  deployments or these tools that you know work, [11:47] you've tested, you can give to them is something  that's incredibly important. But on the flip side, [11:50] you also need to balance that with the art of  the possible, right? models are moving so fast, [11:55] products are moving so fast, companies are popping  up every day. And so how do you balance both and I [12:00] mean this is a literal question by the way. Um  how do you balance both trust right with your [12:04] teams rolling out these proper proper tools and  getting them to believe in them and see them do [12:08] the job with also innovating as fast as the market  is moving and keeping up with the innovation. [12:14] And I wonder um curious if you you guys see  this is like uh as you try to navigate this [12:19] top down bottom up you often reveal that sometimes  there is a true disconnect between the frontline [12:26] workers and the executives. I mean what you're  describing to be able to create that ROI implies [12:32] like a coherent understanding throughout the whole  organization about how value is created. How does [12:36] a frontline job relate a frontline metric relate  to a business metric? And that's actually a hard [12:41] problem. That is not an AI problem. That is a that  is a management problem. I think to build on that [12:47] a little bit is we look at it right we want to put  our employees especially frontline in front of the [12:52] consumer the guest the client more like we're  starved for that level of attention and time. [12:58] So whatever we do to free them up or give them  something more efficiently from a back office [13:03] or a point of view we're completely redeploying  to engage with consumers and clients and guests. [13:10] Yeah. Yeah, I guess a comment I would make, you  know, several lots of good points raised. I think [13:13] the organizational coherence is probably like the  ultimate goal of all enterprises to achieve and [13:19] is so hard to do and it was a problem before AI,  it's problem after AI. It's often exacerbated by [13:24] AI because the now that you can move quickly, it's  revealing where you trip over yourself uh and how [13:30] that is challenging. To your point about building  trust, I think absolutely correct and one of the [13:34] things I've observed that is the like soft signal  of trust is responsiveness to feedback. Not all [13:39] feedback should be responded to, granted, but when  you have stakeholders who have needs, it if you [13:45] can't show them that their needs, allegedly this  hyperiterable technology could accommodate their [13:50] feedback, you will erode their trust. Eventually,  they will turn against you. And so, it's got to [13:55] be expressed through the iteration speed. But  then that's the thing that as an enterprise [13:59] is the hardest for us because you've never had  to iterate as fast as what you can do now. So, [14:05] it's a different game. And if you fall off the  curve doing that, people will lose trust in the [14:10] technology. So it's a it's an acute challenge of  the moment. I think it's a fundamental juiposition [14:16] that's really interesting, right? Because you're  moving faster than ever. So you have to run. And [14:19] when you're running, you're going to trip. And  so when you do trip, inevitably we'll we'll all [14:23] trip at some point with a deployment or a tool.  How do you not let that impact trust? And and I [14:27] think exactly what you're saying. You move fast.  You show responsiveness to it. You iterate on the [14:32] feedback. You listen first and foremost, and you  continue to run fast. All right, great. Thank you. [14:38] Thank you for sharing your insight. And let's uh  talk about what's actually working today. So, uh I [14:45] would like to ask what is AI actually doing today?  Whether inside your own organization or with your [14:50] customers, uh what are some examples where AI is  delivering real value? So, some kind of use cases [14:56] you can share, anything you would like to share  about your customers examples. Yeah, I can I can [15:03] start with that. So uh I'm I'm in the high-tech  business and our customers tend to be financial [15:09] services technology organizations. So one of  the biggest uh the two sort of business outcome [15:15] concerns folks are usually concerned with in an  innovation organization is developer productivity. [15:20] So your innovation the cost of innovation and uh  the second is customer experience. I'll speak to [15:25] both quickly. in terms of um innovation  cost of innovation fundamentally software [15:31] engineering has been changed by AI and what's  being successful is people actually u measuring [15:39] the impact of increasing velocity of generating  features and revenue generating products and be [15:45] able to correlate that with the cost of AI used.  So that's something that I'm seeing starting to [15:51] happen. So it's not just the vibe that things  are moving faster but a quantification of what [15:56] is faster and what is the cost of doing so  that translates to better experience and [16:01] more revenue. On the flip side in terms of um uh  cost reduction I think like you know a lot of the [16:08] uh toil around research and preparation and  uh kind of like undifferiated heavy lifting [16:16] now is being uh replaced with AI. So I see that,  you know, for a lot of our customers internally, [16:22] more time is spent um taking care of our customers  versus preparing to take care of customers. Yeah. [16:29] So we're we've got AI projects running in just  about every part of our our value chain process. [16:35] So I'll go through maybe a couple that I think  are the most relevant for the group. You know, [16:40] one of our businesses is a a buying group type  of power where we provide opportunity for people [16:46] to buy products like hotels, uh, restaurants,  stuff like that. And we have a product called [16:51] Product IQ, which actually helps recommend if  you buy this product, it's cheaper, it's better, [16:56] it can get there on time, things like that. So,  we actually have a fully functional website where [17:01] clients can purchase that product and use it to  optimize their spend. Um, we're doing something [17:07] similar on our supply chain side to make sure we  buy the right products at the right times and get [17:12] them there. And when it's fresh things like milk  and lettuce and stuff like that, you really don't [17:17] want to over buy or under buy. Um, so we have  a couple of cases like that. We have several [17:23] things running in terms of labor scheduling, labor  optimization. Uh and then in our consumer apps, [17:29] our mobile apps, we have some recommendation  engines and things like that that allow us to [17:34] kind of recommend stuff to consumers if they may  like it or not like it. Yeah, I think uh you know [17:41] for us there there's many AI use cases deployed.  I think the most pervasively universally useful [17:47] thing has actually been AI as an analytics tool to  your point like being able to say is it effective [17:54] and how do I analyze whether the referral rate  change for this is like really laborious the [17:58] more you do the more you have to do that and it's  easy to get suffocated by just the avalanche of [18:03] now reporting on the thing you're doing. It's been  really helpful to be able to kind of go faster at [18:08] like analysis of outcomes and the and the size of  an opportunity just be more disciplined about it. [18:14] And then some of the things where you know I've  I've seen AI really positive affect the enterprise [18:19] ambient dictation is probably the obvious one in  healthcare. So the idea you put the phone on the [18:23] table and you talk to face to face and I'm not in  the medical record but my documentation is done [18:27] my orders are teed up like it preserves the sort  of human bond of the of the appointment. It also [18:33] materially prevents physician burnout. Those are  very hard resources to replace. They're expensive. [18:38] They're complex, you know, and it's it you need  to preserve that workforce and keep the sort of [18:43] exodus of burnt out employees, especially postco  there's been a lot of workforce burnout. The other [18:47] one I think is interesting is healthcare is in  this interesting spot where the jobs to be done [18:52] seems to be growing exponentially because the more  you can diagnose things that are wrong with you, [18:56] the more labor there is to like adjudicate that  and what happens with it. And so you already have [19:01] a strained workforce. So how does AI help you  manage these things? So one of the ways we do [19:05] that is now in our referral queue. So you get  a referral to nefrology for example for kidney [19:11] instead of just a first in first out queue where  you went in you had the referral on date X and [19:15] some you know access technician is allocating you  and they're not the doctor they don't really know [19:20] the the AI will review that referral your chart  and be like is it urgent is it not how urgent [19:24] should it go now and kind of move stuff to the top  of the queue. So even with the existing workforce, [19:29] the things aren't languishing that could become  consequential by virtue of sitting in a queue [19:34] for four to six weeks. And so and to me that's a  material benefit to patients because they are more [19:38] seen as they should be in that more less balls  dropped to the cracks and healthcare is at risk [19:43] of more of that happening the more things there  are to be done. Isn't that funny what you said? [19:48] The the more efficient you get, the more problems  you solve, the the more challenges that typically [19:51] come up and the more opportunities that come up. I  I think we've seen that left and right as well. Um [19:57] I I really liked your scribe point. I think that  was spot on because there's been such a wave of [20:01] back office efficiency, automation, um analytics  is another one over the past few years, right? [20:07] Where it is everything from process to scheduling  to analysis and insight all the way to simply um [20:13] augmenting day-to-day workflows. Like if if anyone  out there uses granola for example as a notetaker, [20:17] right? I think that's a great consumer example of  of the scribe example. Um, one of the things that [20:22] we're starting to see as well is this inflection  on AI deployments starting to be more humanf [20:27] facing versus purely back office as well. We've  seen really big implementations around learning [20:32] and development where you can do roleplay and  simulation and have a proper back and forth with [20:36] someone and get feedback in real time on posture  on, you know, what you're doing and how you're [20:40] saying it even all the way down to tone. Um, all  the way to medical intake, patient onboarding [20:45] or pre or posttop care. Um, and then another one  being potentially even interviewing, right? making [20:50] sure you can have that first touch and really  build brand image. And so one of the trends I [20:53] think we'll start to see and it's still early, but  one of the trends we'll start to see over the next [20:56] year is a lot of the AI deployments are going  to start to shift from purely being efficiency [21:01] optimizations to really being experience  optimizations as well and replacing a lot of these [21:05] like previously regressed IVR systems and things  like that. Thank you. Um we've seen many AI pilots [21:12] uh that look promising but never scale. So what  made the difference between a successful pilot and [21:19] uh something that people actually used at  scale? Um yeah, that's my question. So uh David, [21:26] you want to go first? Yeah, sure. I think  for the technology organizations I work with, [21:31] a lot of has to do with um the ergonomics and uh  paying attention to the experience. Um it's funny [21:38] because you know we were uh we deploy enterprise  technology and there's a time when enterprise [21:43] technology it's just assumed the ergonomics is  bad but I think the one thing like the AI has [21:48] changed is that AI is a consumer product like if  we remember what drove the revolution like chat [21:54] GPT was the thing that brought AI on the map  and the operative word of chat GPT is chat and [22:00] it's an amazingly good ergonomic uh interface for  people to see the power of AI. So that has I think [22:09] um set a new standard for what people expect the  eye experience to be and therefore just from a [22:16] pure aside all the usual stuff like leadership  alignment. Yes. Yes. Yes. But when it comes to [22:21] actually rolling out a piece of AI technology  the ergonomics of that is really important. [22:27] That's something new that I have not seen before.  So I'll take a little bit different approach but [22:33] definitely agree with with what you said. Um, for  us it's I think first really identifying what the [22:38] use case is, the problem statement. Um, and this  stuff I think is the same on any project, program, [22:44] product, whatever, you know, analogy you want to  use for it. Um, but it's it's getting the right [22:51] use case. It's understanding what your success  criteria looks like before you start. Meaning, [22:56] what is going to be good? In a lot of cases, it's  a financial calculation. It's having the right [23:01] executive sponsorship. So if I'm doing something  in sales, I need sales to support it. If I need [23:06] I'm doing something in finance, I need them to  support it. If it's something consumerf facing, [23:10] I need that team to support it as well, too. So  it's executive sponsorship. It's the right use [23:16] case. It's knowing what good looks like. And also,  it's stagegating through the process to know that [23:20] if I can't make it work on paper, I probably  shouldn't be attempting it, right? So knowing [23:25] when you've you're going to fail and pulling back  early, right? that builds credibility with the [23:32] organization to let's well let's try something  else let's try something different let's look [23:35] at this scenario right but uh you know I think  a lot of tech teams push too hard on something [23:40] that will never work and then they just lose their  entire credibility with the organization as well [23:47] yeah definitely and I think uh you know I would  definitely agree everything that's been said you [23:52] know sometimes you get lucky and things are just  the right ergonomic fit like ambient was this way [23:57] yes we we drive you know CMI is great at driving  education, but it's a product people like and are [24:03] personally motivated by virtue of their burnout to  pursue. And so you don't have to convince them to [24:07] do it. And that's really great. And often times  in organizations, the things that are valuable [24:12] to do aren't the things that the workforce might  just spontaneously prefer to do anyway. And there [24:19] there's some friction to create value. You know,  it's like, oh, if you're going to catch this extra [24:24] billing opportunity in the back office, there's  a net new work ask that has to go somewhere. [24:29] And I think where this fails to scale is when  people do not understand how what you're doing [24:35] impacts or ties to value. And it sounds kind of  ambiguous, but I think it lends itself to a very [24:40] specific exercise. And we've tried to do a lot  more recently as say if you're a VP level person, [24:45] you own a business area, you have a thing that  creates value. You have to be able to articulate [24:52] what the value is. Is it yield? Is it, you  know, cost savings and what are your P&L levers? [24:57] Pick the top three. Don't stop there though.  So you have the lever. What action produces the [25:03] output and you go backwards. What produces that  action? What produce that action until you stop at [25:07] a thing where you know it it's not your workflow  or AI can you know come in and take that over. It [25:12] has to be very much line of sight to value because  things just get totally derailed otherwise like oh [25:17] I don't like this or I would rather do it this way  and it's like valid but this will just never end [25:24] unless everyone can say yes but I'm doing this  because it gets me to value. So if you can't [25:28] make that rationalization uh the ship will just go  off course inevitably and get stuck somewhere. We [25:34] sit on the other side of the fence where we're  not the ones actually necessarily in the weeds [25:37] with implementation but instead are building the  solutions for the enterprise to implement. And so [25:41] I'll start by saying I love what you all said. Um  what we see across enterprises and successful PC's [25:46] are the people who come in and are able to deploy  quickest, understand what the success metric is [25:52] they're trying to drive, iterate fast and and  innovate along it are often the ones who deploy [25:56] AI months if not years before their counterparts.  Right? And so I completely agree with everything [26:00] you all said on actually outlining everything  really clearly, making sure you have the success [26:05] metrics, but I actually think a lot of this goes  back to the the Schroinger's paradox earlier on, [26:09] right? where AI is neither um you know deployed  or not deployed at any given time and while PC's [26:15] are a really helpful tool for leadership to you  know instill a sense of structure and a experiment [26:21] within a uh within an organization ultimately AI  is evolving fast enough where PC's are becoming [26:28] fluid right AI models AI tools are being replaced  week by week month by month year by year and so [26:34] being able to consistently iterate push across  multiple axes at the same time and drive that [26:39] frontier forward um wherever you fall on that  frontier line is also one of the things that's [26:43] making enterprises either successful or candidly  not successful today as well and so um I don't [26:48] think PC's are going away for what it's worth  because that gives the structure to roll these [26:51] out within organizations but the way PC's are seen  or discussed internally are also fundamentally [26:56] changing you're working from different industry  but have a similar same challenges so Peter I [27:03] have a question for you uh healthcare is one of  the most regulated and highstake industry ries [27:09] environment. So where do you where do you draw the  line between letting AI take on certain tasks and [27:15] um using AI to support human decision making?  That's a great question. So I have many thoughts [27:21] on this topic. I won't I won't uh impose them  all upon the group here but uh definitely we're [27:27] a high-risisk high regulated industry. We should  respect the risk. One thing I think often happens [27:32] in and why this conference is very nice to talk to  non-healthcare people that happens kind of inside [27:37] of healthcare is a lot of admiration of the risk  without an acknowledgement that not doing anything [27:43] is also risk because the state of healthcare today  is woefully incomplete to actually affectuate [27:50] positive care in a lot of patients lives and  so there's it could be more urgency I think for [27:55] healthcare to say yes it's a challenging road to  figure out the compliant path But we have to do [28:01] it with the same imperative that we have to find a  chemotherapy that can cure cancer X. This can just [28:07] sit there and say well it can be hard you know and  on the on the back of that I think that these are [28:12] solvable challenges. So we think of things like  compliance and can I use this model here and there [28:16] yeah you don't want to get it wrong but you can  get it right and you can't let the org culture [28:22] stop at enumerating the ways you get it wrong cuz  it's like yes so let's do it correctly. So there's [28:28] a rigor of like we'll do it the right way. We will  be specific about what that is. We will articulate [28:33] why. We will pursue the path forward. And the last  part about decision-m. So for those who are not in [28:38] healthcare, you know, 5-second history lesson.  So in the early 2000s, we had this thing called [28:42] meaningful use where everyone believed naively in  retrospect that if you gave everyone an electronic [28:47] medical record would like fix healthcare. Oh,  it's going to be digital. Great. Every problem [28:50] solved. So these financial imperatives convinced  people to deploy it all. So we got digitized on an [28:57] arcane culture and system. So you have a digital  interface like a sea of PDF files basically that [29:03] a human agent has to like deal with to do patient  care. This is not a system at all really. And so [29:10] uh what happened to the physician in the '9s where  I'm a physician I think about your care and your [29:15] decision-m became I may get to that after 95% of  my mind is burned out doing all the logistical [29:22] information gathering and click streams to get to  think about your care. So to me, if we could just [29:28] get back to a point where what physicians did  was make decisions about your care, even half [29:34] the time, let alone most the time, healthcare  would be in a much better place than today, [29:37] where it's like a sliver of the time. And I think  really digging out of that hole is the mission [29:42] now. You bring up such a good point there, too,  because what happens is a lot of people compare [29:47] what AI is doing today versus what the experience  was a few years ago, but those experiences a few [29:52] years ago aren't actually good source of truth.  they're largely regressed, right? And you you take [29:57] the uh the healthcare example, when you used to go  to the doctor 20, 30 years ago, you'd go in, you'd [30:02] talk to the receptionist right off the bat, they  would know you, they would recognize you, they [30:06] would ask if you had any insurance change, etc.  You they would then pass the relevant information [30:09] on to maybe the nurse, right? The nurse comes  in, they take your your vitals, they understand, [30:14] they then give it to the doctor, and the doctor's  able to come in and in a very short amount of [30:18] time be incredibly effective, give the advice,  understand, etc. What you see now with tellahalth, [30:22] for example, is you come in, you fill in the same  form four different times. The doctor doesn't read [30:26] it because they know you just said no, no, no  to every one of them. They're asking the same [30:30] question again and suddenly your appointment  takes 20 minutes of uh an incredibly like high [30:35] value person's time and you're still not getting  the great output from it. And so so many of these [30:40] experiences have just fundamentally regressed  where uh for lack of a better way of put it, [30:44] we almost need to replace bad machines in a lot of  these places. Well, you you guys said two things [30:50] that popped in into my mind um that I I summarize  in you can automate a manual process or you can [30:57] create an automated process. And then the second  thing that really stuck with what you said is for [31:03] me if you're standing still you're falling behind  because everyone else is moving forward. And to [31:08] me that's kind of how you have to approach these  activities right and serving healthc care. Um I [31:14] definitely I feel your pain brother. or failure  panel. Yeah, I think the the one thing Pier said [31:20] is true is uh instead of enumerating all the ways  things can go wrong. Um think about what has to be [31:26] done right and I I mean I think AI I mean as we  talk about this we probably could play this back [31:31] and replace AI with digital or internet and it  would all fit or public cloud if you're from a [31:35] more nerdier industry. But I mean ultimately  for all those different um sort of changes [31:41] people figured out the right way to measure it and  govern it. So we're right now like we're kind of [31:46] like figuring it out like but I think the focus  should be on what's the right way to measure it, [31:50] what's the right way to monitor it versus it  simply just like cannot be done and therefore we [31:55] shouldn't do it. It's it's again going to happen.  It's going to infiltrate no matter what. So better [32:01] to focus on how to work it out safely and govern  it rather than the other way around. All right. [32:07] Uh Brian, I also have a question for you. So  Daxup operates in a very different environment. [32:13] uh like they're working in like kitchens and  facilities and frontline operations. How do you [32:19] bring AI tools to people who may not traditionally  interact with you know enterprise software? So we [32:27] still do have thousands of people at a like a site  level that do operate whe they're menu planning [32:33] whether they're tracking waste um things like  that. So there are quite a few site level things [32:39] that get done by a frontline of employ. Um I also  do think and and some of the multi capabilities [32:46] that are kind of coming to fruition now will be  very interesting in a health and safety uh point [32:52] of view in um in shrinkage prevention. Uh and I  think there's a series of use cases there that are [32:59] starting to grow and develop that we'll we'll be  taking advantage of in the next couple of years. [33:05] Great. Um, and for to jump to jump in there as  well, I think actually Anthropic is a really [33:11] good example of someone who's done a great job  at this. Now, Anthropic might not be the company [33:16] you imagine going to frontline workers in in a  kitchen or a factory, but the lesson to take from [33:20] them and and why they've become so successful is  they've successfully met people where they are, [33:25] right? They have help deploy the technology or  the AI to the consumer enterprise business in [33:32] the environment that they already have, right?  They help you integrate your tools into it. They [33:34] help you integrate your workflow etc. And so  I think when we talk about AI transformation, [33:39] there's two aspects of it. One of them is the AI  and the efficiency. But the other one is what is [33:44] the actual like users workflow and where are they  today? And it's a lot easier to fight one battle [33:49] at a time versus fight both of them at once. And  so if you're able to take these deployments and [33:52] meet someone where they already are, integrate  into their existing workflows and help minimize [33:57] the amount of change that a actual individual user  has to take to see the results. the faster things [34:02] will also move. Yeah. I mean, I think uh the  way AI impact even in nontraditional industries [34:08] uh there's uh a lot of opportunity because  I think you guys both touched on this. A [34:13] lot of it is if you strip away a lot of  the ceremony of how people do their jobs [34:18] u from like if you're in the food services  industry, it's about understanding your [34:21] customers and spending more time serving  your customers. If you're a physician, [34:24] more about making decisions about health than to  like document what was done, right? If you're like [34:29] a programmer, it's like an engineer is actually  solving a problem that's value added versus like [34:34] getting approval for a plan. And a lot of the  ceremony can get stripped away by AI. So it's [34:40] not quite it may be counterintuitive as to why  somebody who's in like a you know food service [34:45] industry uses AI but like you look under the  job like we've we've like stuffed many jobs with [34:50] ceremony and that ceremony can be uh stripped  away. your to ceremony is not a word on part. [35:00] Yeah. Okay. Uh and a lot of the AI conversation  today focuses on text. Uh but the where do you see [35:09] the biggest opportunities today in other areas  like um images, voice, video or other formats? [35:18] I'm hungry to jump in here and say something.  So So I think just go ahead. Um and you made [35:22] this point earlier. I think like for systemwide  benefits to patients, I think voice is a huge [35:27] thing. Having that kind of concierge that can  attend to you and interact with you is will be [35:34] the defining subtle interface of healthcare in  the future. I think so that's probably the right [35:38] answer. I think another answer I really interested  because I I'm a pathologist. I'm an imaging guy in [35:42] a way. What I think is really really cool and  I I'll talk radiology for a second is we know [35:47] now that AI can not just do things humans could  do like you can see the nodules so can AI. is [35:51] automating you seeing the nodule to help you see  it, you know, more reliably, but it can see things [35:56] you literally couldn't see. So, an example  of this is like we have an abdominal CT. You [36:01] can't see that there's pancreatic cancer there at  early stage, but the AI can see there's pancreatic [36:06] cancer at early stage. Well, this is a whole new  like diagnostic assay basically. And the fact that [36:12] AI even at the chip level is making imaging as a  modality always more ubiquitous, more scalable, [36:18] more look like the butterfly IQ and ultrasound  that anyone can buy. As more of this data exists [36:23] in the frontier, it can be leveraged as point of  care diagnostics by AI that can interpret it in [36:29] ways we never could before. And so our role then  changes to be less about I have to feel the lump [36:35] which is fading because most things are not like  that now to I need to adjudicate what happens when [36:40] this flag means this and means that. But that  will be a a revolutionary service to patients. [36:44] I think that we can know more about you earlier  doing things humans never were capable of doing. [36:51] I have a little bias here because we're the lab  that builds the multimodal models. So if you've [36:55] ever interacted with conversational audio or  conversational video that likely stemmed from [36:59] us. Um so you know obviously quite bullish on on  the space. Um I what you said there was really [37:06] interesting right about the image because you get  a higher degree of data that you can work with and [37:10] we see the same in audio and video deployments in  in the conversational context because compared to [37:15] text you have a higher degree of richness right  I'm able to see your face I'm able to see if [37:19] you're bored and falling asleep I'm able to see if  you're you know interested and leaning forward and [37:24] vice versa you're able to see my face and you can  take a lot of information from that right and and [37:28] voice is you know really good as well but a step  down from that on on the video side and so there [37:32] are use cases that voice and video uniquely  unlocks, right? Whether it is the ability to [37:37] understand someone's, you know, emotional state  and factor that in the conversation. If you're [37:41] in a tutoring session, is someone confused and  frustrated about their calculus problem or are [37:44] they, you know, excited and they want something  more challenging? All the way to, you know, [37:49] a support context where you're actually  on a video call, you're able to show them, [37:53] um, you know, them, right, being the AI, uh, hey,  here here's my printer. What's wrong with it? Um, [37:58] and it's able to guide you through because it has  that, you know, information uh, accessed. We have [38:02] deployments across the board where people are able  to reduce on call site visits by 50%. Right? Where [38:07] efficacy of treatment goes up 30 to 40% because  there's more information coming in. And obviously [38:11] there's many reasons for that. But the richness  of data that comes from these mediums is is vastly [38:15] different from what we previously had. This I yes  this excites me a lot also because you think of [38:21] the world of like so in Texas dementia prevention  is one of our big things that we have billions of [38:25] dollars of Texas for dementia prevention over the  next decade. This this is not a task that will be [38:31] done in the hospital. It's a task that will  be done by AI being woven into your everyday [38:36] interactions in your life in a way like voice that  is richer and the biomarkers are not going to come [38:41] from medicine proper. They're going to come from  Siri or whatever. They come from the tech inward. [38:46] And I think this is to the betterment of patients  without a doubt. And it's just a transformational [38:50] wave that we have to deal with because it's the  right thing to do as the profession. Yeah. I [38:56] think from my perspective multimoto is uh just  beginning to be understood of what's valuable [39:02] like I think in financial services and insurance  industries and anywhere data dominates the core [39:07] business competitive advantage a lot of there's  still a lot of um business data and therefore [39:13] value that is in uh formats that are not text uh  like cuz in financial service a lot of the charts [39:19] and historical sort of like numbers and figures  they're all images really. So if you think about [39:25] these companies that are trying to figure out  it's like phase two of digital like right like [39:30] people could maybe digitize some text but the rest  of the stuff that couldn't be digitized as text [39:34] is basically images and for a lot of financial  services industries uh folks in insurance it's [39:39] like it's either better data to help them offer  more accurate neural policies or it's their core [39:45] competitive advantage in terms of serving their  clients better because they understand their [39:48] history better. So lots of Yeah. Yeah. Yes.  Uh before we close, I'd like to ask each of [39:56] you for one quick takeaway. Uh for organizations  that are just start just getting started with AI, [40:04] what is one mistake you would tell them to avoid?  Uh and from the vendor perspective, what is the [40:10] biggest piece of the what is the biggest piece of  advice you give customers before they begin their [40:16] AI journey? Yeah, I'm happy to kick it off here.  Um I I mean there there's so much advice on either [40:24] end of the spectrum that it's it's hard to distill  down. Obviously ultimately I think what we're [40:29] seeing today though is that speed matters more  than anything else. It doesn't matter where you [40:33] at where you're at on the adoption curve. If you  are at the forefront, if you're falling behind, [40:37] the only thing that changes your effective ability  to benefit from AI is how fast you move to use it, [40:42] implement it. And often times what comes at  the contradiction of speed is perfection, [40:46] right? And so being able to balance those two  things effectively is what we've seen lead to [40:52] the best and and ultimately the best business  outcomes with deployments. I think that's spot [40:58] on. I think uh one of us said if you're going  to do something, focus on iterability above all [41:04] other virtues and pick a team that can participate  iterably in something. Especially if you're first [41:09] in don't pick a team that's like skeptical. Pick  an iterable partner and run with them. Yeah. Yeah, [41:14] I'm going to I'm going to say probably something  very similar. What to me it's focus, learn fast, [41:19] which is a positive way of saying fail fast,  right? Um, and be clear on what good looks [41:25] like. And I think, you know, you don't need to  throw spaghetti at the wall here. You just need [41:29] to be focused and try one thing at a time and  get to the end. Yeah. Yeah. I think uh I mean, [41:36] completely agree. Um, the one thing maybe I'll  add is there's so much being said about knowing [41:40] the outcome, knowing the goal. I would say  it's more important for from what I've seen [41:45] people who can be successful and iterate fast is  people who can all agree that they just have to [41:48] agree on something. It's very different to say  we have to agree on something versus we have [41:52] to agree on the best thing. Nobody knows what  best is but we can all agree on something and [41:57] um think about how you would quantify like  quantification is also there's a difference [42:01] between knowing what good looks like and being  able to quantify what good is. So it's like agree [42:05] on something not the best thing. Start think  about quantification rather than qualification. [42:10] And I think that's these are the uh things  I can avoid paying down the road. Awesome. [42:14] Thank you. Thank you so much you guys and thank  you to our panelists everyone. Yeah. Thanks.