Cigna, Optum Rx Pivot AI Strategy From Back-Office Cuts to Top Staff

Cigna 与 Optum Rx 转变 AI 战略：从削减后勤转向赋能核心员工 医疗支付方正在放弃低回报的后勤自动化，转而采用多智能体预先授权和知识工作者工具，因为微弱的边际单位成本削减已无法带来企业级回报。 医疗支付方与风投领袖探讨了医疗 AI 经济效益的转变，即从基础行政成本削减，走向高影响力的人才效能倍增及多智能体系统构建。专家们重点介绍了实际落地案例，如 Optum Rx 的领域专属智能体架构与实时预先授权副驾驶系统，同时强调须设立严格的“人机协同”安全防线以杜绝 AI 自主拒赔。此外，他们还深入分析了智能体评估与推理预算管理中的隐性成本，倡导利用分层开源模型路由和持续学习框架，以实现可持续的运营与临床投资回报。 全球医疗 AI 市场预计到 2030 年将增至 1100 亿美元，同时 CMS 推进的 WISeR 试点确立了“AI 提供建议、医生拥有最终决断权”的合规原则，推动生成式 AI 在支付方运营与预先授权流程中加速落地。 (3:20) 保险支付机构的企业级 ROI 正发生范式转移：由于成熟的后台单据流程单位成本已被高度压缩，AI 的核心价值从挤压微小边际成本转向赋能顶尖 10% 的核心知识工作者，成为组织内部的高效能放大器与基础底座。 (6:45) Optum Rx 正在从单体模型转向由 50 至 70 个专业领域微智能体构成的协作生态系统，结合内部“AI Builder”赋能计划与 MCP 网关，建立中心化的安全治理与智能体交易市场。 (14:42) 在预先授权审核中，实战部署的 AI 临床副驾驶使人员处理吞吐量提高 4 倍并将裁决成本缩减 75%，但治理体系设立了严格红线：AI 绝不能在缺乏执业医生或护士人工复核的情况下自主发出拒付决定。 (16:52)

Healthcare payor executives and investors argue that artificial intelligence delivers the greatest return when deployed as an operational multiplier for skilled knowledge workers and clinical decision-makers rather than as an incremental tool for micro-optimizing low-cost back-office tasks. Reframing Payor ROI Beyond Incremental Automation (6:45) Rather than seeking fractional-cent savings in claims queues—which Dan Mackey of The Cigna Group observes are already heavily optimized down to roughly four cents per human document check—payors find significant economic value in accelerating high-performing knowledge workers. Mackey equates generative AI to foundational infrastructure like email, where direct monetary return is difficult to isolate, but organizational throughput surges when senior personnel can prototype functional tools in 45 minutes instead of two weeks. Dr. Aisha Rahim of Johns Hopkins Health Plans adds that return on investment must also be measured non-monetarily through reduced clinician burnout and talent retention, pointing to ambient clinical documentation tools that relieve providers from off-hours charting burdens. Prior Authorization Automation and Non-Autonomous Guardrails (16:52) Prior authorization remains one of the healthcare system's most expensive administrative bottlenecks, routinely stalling care through weeks of document exchange. Automated systems, such as Optum Rx's guideline assistant and Obvious Ventures' portfolio company Gen Health AI, ingest patient records and surface relevant clinical evidence in real time, with Gen Health demonstrating a fourfold productivity increase and a 75% reduction in adjudication costs across daily production workloads. Despite these operational gains, healthcare payors emphasize strict governance boundaries. Dr. Rahim stresses that Johns Hopkins Health Plans prohibits AI from autonomously issuing adverse determinations or claim denials; any non-approval must be routed to nurses and medical directors for formal evaluation. This human-in-the-loop requirement aligns with the Centers for Medicare & Medicaid Services' upcoming 2026 WISeR pilot, operating under the mandate that artificial intelligence may recommend while physicians decide. Modular Multi-Agent Frameworks and Tiered Model Economics (14:42) To prevent monolithic failures and avoid runaway compute expenses, payors are shifting toward distributed multi-agent systems and tiered model architectures. Bosky Mathew outlines Optum Rx's strategy to deploy 50 to 70 domain-specific micro-agents covering discrete pharmacy benefit workflows, including claims explanations, formulary inquiries, and prior authorizations, coordinated through an internal agent marketplace and standardized Model Context Protocol (MCP) gateways. From an infrastructure perspective, Kahini Shah of Obvious Ventures notes that production efficiency requires routing narrow tasks—such as call-center summaries and member FAQs—to smaller, open-source models fine-tuned and hosted on-premises. Reserving expensive, third-party frontier language models exclusively for complex reasoning helps contain inference expenses while ensuring sensitive health records remain securely inside enterprise data perimeters. Clinical Diagnostic Efficacy and Latent Operational Expenses (33:01) Beyond administrative efficiency, venture investment is increasingly targeting clinical diagnostic platforms that directly reduce downstream acute care expenses. Shah highlights Bayesian Health's FDA-cleared sepsis prediction chassis, which operates within hospital workflows under physician supervision and has demonstrated a 27% decrease in patient mortality alongside a 10% reduction in overall hospital utilization. However, expanding these agentic and diagnostic systems exposes significant hidden costs. Mathew points out that building rigorous evaluation frameworks for autonomous agents remains difficult due to healthcare's zero-tolerance threshold for clinical and operational errors. Furthermore, forecasting ongoing GPU and token operational costs remains unresolved across the sector, with an audience poll conducted by Dr. Rahim revealing that virtually no participating healthcare organizations had completed a formal token cost analysis. While panel members foresee long-term disruption from continuous learning architectures and unforeseen second-order industry shifts, near-term payor adoption remains constrained by evaluation difficulties, token cost unpredictability, and necessary compliance checkpoints that balance rapid prototyping against strict clinical safeguards.

在医疗支付方（Payer）领域，人工智能的战略重心正从简单的外包成本削减转向放大核心知识工作者效能以及构建多智能体（Multi-agent）工作流。行业专家与投资者指出，通过在预先授权、会员服务和临床辅助中实施严格的人在回路（Human-in-the-loop）治理，AI正重构运营模式并带来可量化的效率与临床收益。 重估投资回报率：放大核心产能而非微调边际成本 (6:45) 医疗支付方对于AI投资回报率（ROI）的衡量方式正在发生转变。The Cigna Group的Dan Mackey指出，成熟的大型支付方后台流程历经数十年优化，人工处理单份文档的成本可能低至0.04美元，AI很难在这些微小成本上实现显著削减。相反，真正的价值在于为前10%的高绩效知识工作者提供杠杆，例如帮助资深管理者在45分钟内完成以往需要跨部门协作两周才能成型的原型验证，将AI视为如同电子邮件一般的底层基础设施。 预先授权再造与人在回路治理红线 (16:52) 传统预先授权流程严重依赖纸质信函往来，常导致数周或数月的治疗延误。Cigna与Johns Hopkins Health Plans（JHHP）正在通过AI实现文档提交时的实时解析与缺失数据（如MRI流水号）即时校验。来自Obvious Ventures的Kahini Shah引用的Gen Health AI案例显示，其实时辅助系统可将日常审核成本降低75%，处理量提升4倍，并通过可视化溯源逻辑树解决信任难题。但在治理层面，JHHP的Aisha Rahim博士强调了一条不可逾越的红线：AI绝不允许自主签发拒绝赔付或负面决定，所有潜在拒赔必须交由执业护士和医疗总监人工裁决，以确保临床合规与算法安全性。 多智能体架构与企业级分层模型部署 (14:42) 为避免单一庞大模型的局限性与高昂推理成本，行业正在转向专用微型智能体矩阵与分层调度架构。Optum Rx的Bosky Mathew介绍，他们正在规划部署50至70个覆盖药品目录、理赔拆解和预审的领域微智能体，并通过内部“AI Builder”项目提升员工技能，同时借助MCP（模型上下文协议）网关与智能体市场（Agent Marketplace）进行治理与互联。Kahini Shah与Aisha Rahim补充指出，鉴于企业中仅约10%的数据属于高价值决策数据，企业正采用分级路由机制：常规客服总结与知识问答交由内部微调的开源小模型处理，复杂的综合推理才调用前沿大模型，以控制Token与算力支出并保护患者数据隐私。 临床诊断前沿与自适应学习演进 (33:01) 除了管理流程自动化，风险投资与医疗机构正将视线拓展至临床AI诊断领域。Kahini Shah以约翰霍普金斯大学孵化的Bayesian Health为例，该平台在克利夫兰诊所等机构上线败血症预警系统，在医生参与验证的闭环下，实现了患者死亡率降低27%、医疗资源利用率下降10%的成效。展望未来架构，与会专家一致认为AI系统必须突破当前静态微调与检索增强生成（RAG）的局限，向能够根据生产环境中的遥测信号与临床医生反馈实时自我改进的连续自适应学习（Continuous Learning）系统演进。 尽管行业对AI的前景保持乐观，但在零容忍的医疗差错红线下，多智能体系统的评测框架匮乏、长期Token与算力运维成本不可预测，依然是企业规模化落地的核心阻碍；同时，行业需平衡前期的快速原型构建与后续严格的合规接入。

## Background 该视频由行业会议组织机构 Ai4 于 2025 年 8 月 28 日发布在 YouTube 平台。根据官方简介，本场研讨会汇聚了医疗保险与投资领域的领军人物，重点探讨医疗支付方（Payor）如何借助 AI 提升运营效率、改善会员护理并降低成本，涵盖理赔自动化、反欺诈、个性化健康管理以及兼顾监管与数据隐私的系统集成最佳实践。 ## TL;DR 来自约翰斯·霍普金斯健康计划、Cigna、Optum Rx 与 Obvious Ventures 的专家们深入拆解了 AI 在医疗支付方落地的实际 ROI 与工程战略。讨论指出当前最大回报并不在于微观流程降本，而在于赋能核心高绩效人才以及重塑预先授权（Prior Authorization）等高痛点流程，同时必须坚守拒赔环节“人机协同”（Human-in-the-loop）的安全红线。面对算力成本与数据隐私挑战，行业正快速转向多模型分层编排与私有化部署，并加速构建涵盖 50 至 70 个垂直智能体的多智能体平台（Agentic Platform）；未来 AI 将从静态检索向具备自我演进能力的持续学习系统跨越，全面重构医疗价值链。 ## Atmosphere & style 整场圆桌研讨气氛务实、敏锐且兼具幽默感，主持人在临床医生与 AI 治理者的双重身份下游刃有余地引导节奏。嘉宾们抛弃了空洞的概念宣传，在分享“边开飞机边造飞机”的真实阵痛时频频引发全场共鸣与笑声（例如将合规博弈比作敏捷突围、将技术范式转移比作“外卖专车”）。不同背景的碰撞火花四溅——既有老牌保险巨头严苛的技术经济账，也有架构师对 PB 级数据智能体生态的硬核推演与投资人对自主演进系统的前瞻洞察，带给观众极具实操价值与思想深度的行业纵深感。 ## Key takeaways - [03:20](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=200s) 医疗 AI 市场正在快速扩张，预计到 2030 年将达到 1100 亿美元，超七成医疗机构已在积极部署生成式 AI 解决方案。 - [03:20](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=200s) 2025 年政策监管呈现转向：联邦行政令 4179 号倾向于促进创新并扶持开源，CMS 的 WISER 计划则探索在预先授权中践行“AI 推荐、医生决策”机制以削减医疗浪费。 - [06:45](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=405s) 大型保险支付方对成熟流程的微观降本空间有限，AI 当前最大的 ROI 并非在已精简的流程中抠出几美分，而是赋能顶尖的 10% 核心人才产生十倍的爆发力。 - [11:13](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=673s) 生成式 AI 在软件开发与呼叫中心两大场景变革显著：既大幅提升了代码生成效率与系统迁移速度，又支撑了低成本、大规模的主动随访与慢病关怀。 - [13:34](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=814s) Optum Rx 正在推行多智能体（Agentic AI）平台化战略，规划构建数十个垂直专用智能体，并通过智能体规范（Agent Spec）与 MCP 网关落实安全治理。 - [16:48](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=1008s) 预先授权（Prior Auth）的自动化大幅压缩了医院与保险机构间的流转时间，但行业必须恪守“人机协同”红线，AI 严禁自主执行拒赔决定。 - [22:40](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=1360s) AI 在医疗落地的隐性成本极高，尤其是零容忍业务下的测试验证（Eval）与长周期微调成本，多智能体系统的科学评估方法行业仍在摸索。 - [27:06](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=1626s) 医疗架构正从单纯调用通用大模型快速转向“多模型分层编排与私有化本地部署”，以此控制推理成本并保障敏感数据合规与隐私。 - [31:51](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=1911s) 在传统大企业落地 AI 项目中，敏捷推进哲学主张“先快速做成原型跑通业务，再引入法务与 HR 进行合规与降本”，避免创新在早期被流程卡死。 - [33:05](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=1985s) 临床 AI 诊断与监测（如 Bayesian Health 的早期脓毒症检测）已实现显著的死亡率与医疗成本双降，但建立全流程行业评估标准是公私合作的关键。 - [36:31](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=2191s) 除了提升运营效率，智能体（AI Companions）还能向会员清晰解析药品定价底层逻辑与明细，推动药房福利管理的费用透明度。 - [38:42](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=2322s) AI 系统正在从“静态预训练 + RAG”向能够根据环境指标与人类反馈实现“持续在线学习与自我进化”的智能体范式演进，并将催生深远的二阶商业颠覆效应。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=4s) 研讨会开场与嘉宾介绍 主持人 Aisha Rahim 博士（约翰斯·霍普金斯健康计划 Johns Hopkins Health Plans 医学总监兼 AI 治理委员会联合负责人）向全场观众致意并拉开研讨会帷幕。作为一名接受过内科专业训练的临床医生，Rahim 博士在五年前转型加入健康保险计划，主导该机构的 AI 战略规划并一手建立了人工智能治理委员会（AI Governance Council）。除此职务外，她还担任哈佛梅西学院（Harvard Macy Institute，由哈佛医学院与哈佛商学院联合设立、面向医疗管理者及医护人员提供专业培训的机构）的客座教员，经常在全球各地就 AI 与医疗变革进行演讲与交流。 随后，Rahim 博士邀请三位来自医疗投资与保险科技领域的领军嘉宾依次进行自我介绍： - **Kahini Shah**（Obvious Ventures 合伙人）：来自硅谷早期风险投资机构 Obvious Ventures，专注 Seed 轮与 A 轮投资。其基金投资组合涵盖 Medicare Advantage 领域的 Devoted Health、基础医疗平台 Galileo、逆转 2 型糖尿病的数字健康企业 Virta Health，以及多家专注于预先授权（Prior Authorization）与 AI 诊断行政流程的医疗 AI 公司。Shah 拥有十余年 AI 从业经验，早年曾任软件工程师、产品经理及 Google 早期 AI 基金投资人。 - **Dan Mackey**（The Cigna Group 全球 AI 创新总监）：负责 Cigna 全球产品线的人工智能创新布局。在加入 Cigna 之前，他曾供职于科罗拉多大学健康科学中心（CU Health Sciences Center），从事核能工程（Nuclear Engineer）；更早之前他还做过物理安全渗透测试员（幽默地自称为“合法破门而入的银行劫匪”并向企业指出安防漏洞）。这段经历引来了全场的热烈掌声与笑声。 - **Bosky Mathew**（Optum Rx 架构副总裁兼首席架构师）：来自联合健康集团（UnitedHealth Group）旗下的药房福利管理平台 Optum Rx。Mathew 目前正在全面主导 Optum Rx 的生成式 AI（GenAI）与智能体系统（Agentic AI）落地与技术战略。 ```mermaid flowchart TD accTitle: 嘉宾阵容与所属机构背景 subgraph Panelists["圆桌嘉宾与所属机构"] direction TB AR["Aisha Rahim 博士<br/>Johns Hopkins Health Plans<br/>医学总监 / AI 治理委员会联合负责人"] KS["Kahini Shah<br/>Obvious Ventures 合伙人<br/>早期医疗与 AI 投资"] DM["Dan Mackey<br/>The Cigna Group<br/>全球 AI 创新总监"] BM["Bosky Mathew<br/>Optum Rx<br/>架构副总裁兼首席架构师"] end ``` --- ### [03:20](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=200s) 行业现状、市场规模与 2025 年政策宏图 Rahim 博士首先向现场观众进行了两轮举手调查。现场观众中不仅 100% 来自医疗健康行业，而且几乎全体（甚至夸张地说有“101%”）都已在个人生活或日常工作中使用 AI 工具。这与 ChatGPT 诞生初期仅有一两人举手的场景形成了鲜明对比，展现出生成式 AI 在医疗领域极高的渗透率。 根据麦肯锡（McKinsey）等机构的行业数据，2024 年全球医疗 AI 市场规模预计在 140 亿至 290 亿美元之间，到 2030 年有望进一步攀升至 1100 亿美元。其增长引擎主要来自： 1. 保险支付方（Payor）运营自动化； 2. 医疗服务提供方（Provider）工作流优化； 3. 患者与会员参与度（Patient Engagement）的技术突破。 目前已有超过 70% 的医疗机构（涵盖支付方、提供方与整合医疗系统）正在积极部署生成式 AI 解决方案。此外，NVIDIA 在 2025 年 7 月发布的医疗行业调查也证实，AI 的采用已经显著降低了行政负担、加速了产品创新并直接拉动了营收增长。 在政策监管方面，2025 年迎来了重大转向： - **联邦行政令 4179 号与《赢得竞赛：美国 AI 行动计划》**：2025 年 1 月，特朗普政府签署第 4179 号行政命令，回调了拜登时期的部分强监管框架，并于 7 月发布包含了 90 多项政策建议的联邦 AI 行动计划。核心重点在于强调创新竞争、扶持开源生态、简化数据中心审批流程，并推动联邦法规在特定州级 AI 立法上的优先适用权（Preemption）。 - **CMS 的 WISER 计划**：美国联邦医疗保险和医疗补助服务中心（CMS）计划于 2026 年在 6 个州启动 WISER（Waste, Inefficiency, and Systemic Errors Reduction）试点项目，探索利用 AI 辅助预先授权（Prior Authorization）。该机制遵循“AI 推荐，医生决策（AI will recommend, physician will decide）”的原则，旨在减少不必要的医疗浪费，但同时也引发了行业内部关于潜在看病与审批延迟的讨论。 --- ### [06:45](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=405s) 保险支付方（Payor）的真实 ROI 落在何处？ 围绕“支付方机构在何处能获得可衡量的实际投资回报率（Measurable ROI）”这一核心议题，Bosky Mathew 与 Dan Mackey 分别结合各自庞大机构的实操经验给出了深刻见解。 **Optum Rx 的智能体战略与人机协同：** Bosky Mathew 指出，Optum Rx 的核心策略并非介入临床决策，而是聚焦于运营效率、质量改善以及提升会员体验： - **会员端伙伴（Member Companions/Assistants）**：帮助会员轻松理解福利条款、用药计划并改善健康产出。 - **呼叫中心座席助手（Call Center Agent Assistants）**：在座席接听会员电话时，实时提供辅助信息支持，快速化解复杂疑问。 - **客户数据导航**：企业与大型客户在 Optum 极其复杂的底层应用与系统网络中检索数据十分困难，通过 AI 智能体可大幅降低数据提取门槛。 **Cigna 的“十倍加速器”与 ROI 现实考量：** Dan Mackey 提出了发人深省的成本与生产力视角。他将 AI 形象地比作马里奥游戏里的“无敌星加速道具（Power-up）”，是一枚十倍放大器（10x accelerator）。然而对于具有数百年历史的老牌保险巨头（Cigna 的历史甚至早于白宫落成）而言，传统流程早已被每季度抠出 3% 成本的管理层优化到了极致： - **微观流程的经济账**：例如在呼叫中心或文档校验（Document Validation）环节，人工确认一份 JSON 格式文档并点击“下一步”可能只需耗费 4 美分（0.04 美元），目前调用 AI API 与工程支持完成同等精度的操作很难把成本压至 4 美分以下。 - **赋能顶尖 10% 核心人才的爆发力**：因此，当前阶段最大的 ROI 并不在于将已经极其精简的微观流程降本几美分，而在于将企业内部最顶尖、最敏锐的 10% 高效员工赋能 10 倍。以往一位高管提出新业务设想，需要指派跨部门团队耗时两周沟通与拉锯；现在借助 AI，高管在 45 分钟内就能独立探索构想，甚至能带着已经接近成型（Almost Done）的应用概念原型来找技术团队。“这才是真正的 ROI 核心：你是想让呼叫中心效率提升 10 倍，还是想让你最核心的 10 个人爆发 10 倍的战力？” 接续这一观点，Rahim 博士补充表示，约翰斯·霍普金斯健康计划（JHHP）同样将核心精力投向了会员互动与赋能，通过聊天机器人（Chatbots/Bots）向会员清晰解读福利方案以提高粘性，并正在反欺诈检测（Fraud Detection）等高价值领域积极布局。 ### [11:13](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=673s) 投资人视角：软件工程提效与呼叫中心变革 在医保机构（Payor）如何运用 AI 防范欺诈、浪费与滥用（FWA）以及优化重复用药和索赔审核之后，投资人 Kahini Shah（来自 Obvious Ventures）分享了她最兴奋的两个投资方向。第一个是软件工程领域的开发者效率提升。虽然现场举手的软件工程师并不多，但像 Cursor、GitHub Copilot 这类代码补全工具已迅速普及；同时，Devin、Factory、All Hands 等智能体解决方案（Agentic Solutions）也在快速兴起，让 AI 能够端到端承接工程师的复杂任务。Kahini 引用了 Sundar Pichai 的数据指出，Google 内部目前已有 25% 的代码由 AI 生成。她好奇医疗机构内部 AI 生成代码的比例究竟有多少，但同时也敏锐地指出，当 AI 承担大部分工作时，真正的瓶颈转向了“如何审查和监管 AI 的产出”。尽管如此，AI 在处理工程师普遍厌倦的枯燥任务（如日常维护、系统迁移）上带来了巨大的生产力跃升。 第二个备受关注的方向是呼叫中心（Call Center）。随着语音交互成本持续下降且能力日益精准，AI 能够接管低复杂度、低风险的事务性请求（例如更新个人信息、查询保单承保范围等），从而为人工座席腾出时间处理更复杂的个性化诉求。更重要的是，这让支付方能以“前所未有的规模”开展个性化会员触达与关怀服务——例如自动推送疫苗接种提醒、癌症筛查提醒以及慢性病全流程管理。由于几乎人手一部手机，这种普惠型 AI 方案不仅大幅拓展了医疗可及性，降低了服务成本，更有望在本质上改善患者健康预后。 - **Google 代码现状**：Google 内部已有 25% 的代码由 AI 自动生成。 - **代表性工具/平台**：代码辅助工具 Cursor、GitHub Copilot；自主智能体 Devin、Factory、All Hands。 - **呼叫中心的核心价值**：低成本语音处理低风险咨询（修改信息、查保单），释放人力；支持大规模主动随访（疫苗、筛查、慢病关怀）。 --- ### [13:34](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=814s) “边开飞机边造飞机”：高风险与高回报项目的幕后探索 针对非技术人员通过 AI 快速搭建工具的现象，主持人 Aisha Rahim 分享了一个生动的轶事：她的一位朋友完全不懂编程，却全凭 AI 平台搭建出了一个完整的网站；然而系统上线后后端突然崩溃，AI 却无能为力，无法自行修复。这恰恰印证了当下的技术现状——大家都在“边开飞机边造飞机”（building the plane as we are flying it）。但随着工具日益成熟，未来哪怕是医护人员或业务人员，也能在不具备深厚编码知识的情况下自主构建解决方案。紧接着，讨论转向核心议题：各大机构在幕后正在推进哪些高风险、高回报的代表性 AI 项目？ Optum Rx 首席架构师 Bosky Mathew 透露，他们的核心战略是在全域业务中构建专业助手或伴侣（Assistants / Companions），覆盖运营流转、呼叫中心、区域业务到面向会员的全场景。Optum Rx 计划构建 50、60 甚至 70 个高度垂直的专用智能体，分别精通福利构成核算、处方目录（Formulary）查询、报销理赔解释以及预先授权（Prior Authorization）审核等。为此，他们正全力培养内部的“AI 构建者”（AI Builder），让工程师乃至业务专家都具备构建和微调智能体的技术能力。在系统架构上，为了规范化管理海量智能体，团队正在推行平台化策略，从制定“智能体规范”（Agent Spec）起步，规划引入 MCP（Model Context Protocol）网关，并严格落实安全隔离、治理与负责任 AI（Responsible AI）规范，同时紧密协同 Optum 集团企业级的 Centralized AI 团队。 ```mermaid flowchart TD accTitle: Optum Rx 多智能体生态与治理平台架构 subgraph Users ["业务构建端与用户端"] B["AI Builder 开发者<br/>(工程师与业务人员)"] M["终端使用者<br/>(运营、坐席、会员)"] end subgraph Platform ["平台治理与通信层 (Platform Approach)"] AS["Agent Spec 规范"] MCP["MCP Gateway (网关)"] GOV["Responsible AI / 安全治理"] end subgraph Agents ["垂直业务智能体集群 (50-70 个专用 Agent)"] A1["福利组件智能体<br/>(Benefit Components)"] A2["处方目录智能体<br/>(Formulary Agent)"] A3["理赔解释智能体<br/>(Claim Explainer)"] A4["预先授权智能体<br/>(Prior Auth Agent)"] end B --> AS AS --> MCP MCP --> GOV GOV --> A1 GOV --> A2 GOV --> A3 GOV --> A4 A1 & A2 & A3 & A4 --> M ``` --- ### [16:48](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=1008s) 预先授权（Prior Auth）的自动化破局与“人机协同”红线 The Cigna Group 的 Dan Mackey 直言，比起纯粹技术酷炫的项目，对社会产生深远影响的落地应用更令他心潮澎湃。其中最典型、最亟待改造的场景便是健康险中的“预先授权”（Prior Authorization, 简称 Pre-Auth / Prior Auth）。在传统流程中，患者需要某项手术时，医院必须提前向保险公司申请批准；而这一流转过程效率极低，大量机构甚至还在用纸质信封、贴邮票来回邮寄单据。患者往往在等待手术的煎熬中耗费数月之久。Cigna 目前正致力于实现即时化预审：医院端上传理赔申请资料后，AI 系统立刻审核并实时反馈结论；若未获通过，会当场精确定位缺失信息（例如缺少某张核磁共振 MRI 的编号），医院补齐后重新上传，即可把原本耗时数周甚至数月的折磨缩短至几分钟。未来随着医院端也引入 AI，系统间将演变为“两台计算机直接对话”，从根源上消灭低效文书。 Aisha Rahim 完全赞同这一愿景，并从支付方医疗总监的角度补充了风控考量。在理想工作流中，医院提交材料后，AI 直接调取该险种的具体临床标准进行比对。然而，涉及拒赔或负面医疗决定时，必须坚守“人机协同”（Human-in-the-loop）的负责任底线： 1. **负面权益红线**：AI 绝不能自动执行拒赔；任何对会员不利、不予批准的决定，必须交由专业护士和医疗总监复核。 2. **早期双向质检**：在系统投产初期，不仅拒赔需要人工审核，AI 给出的“批准”结论同样需要严格的人工抽检，以防模型误判而漏批不合理的赔付。 Kahini Shah 紧接着分享了 Obvious Ventures 投资的明星初创公司 Gen Health AI（由医疗互操作 FHIR 服务器 OneUp Health 的前创始人创办）。该公司专门针对利用率管理（Utilization Management）打造端到端 Co-Pilot： - **实时辅助与规则匹配**：在资料录入阶段常驻在审核员身旁，自动提取关键病历信息，自动对齐保险公司规则库，由人类最终拍板。 - **业务落地成效**：目前已在生产环境每日处理数千例案件，将人工效率提升了 4 倍，审核时间压缩到数分钟，裁决成本直接骤降 75%。 - **白盒化透明审计**：通过极其友好的 UI 界面逐步展现 AI 推理逻辑链条，并附带精确的病历原文引文（Citations），使得人工复核极为轻松直观，且审核员可即时向模型提交纠错反馈。 ### [21:32](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=1292s) 先前授权（Prior Authorization）助手与医疗浪费的削减 在讨论通过强化学习（Reinforcement Learning）将人工反馈闭环输入模型以持续提升效率后，Bosky Mathew 分享了 Optum Rx 在药房端构建的第一个 AI 智能体——先前授权助手（Prior Authorization Assistant）。在药房业务中，每种药物都有严格的临床指南，先前授权的审核流程正是基于这些指南进行。该智能体能够针对临床指南中的关键问题，自动在病历和图表记录（Chart notes）中检索支持性证据，从而极大地赋能了药剂师的工作效率。 ```mermaid flowchart LR accTitle: Prior Authorization Assistant Workflow subgraph Input["数据输入与规则"] A["临床指南问题"] B["患者病历与图表"] end subgraph Agent["AI 智能体处理"] C["先前授权助手 (Agent)"] D["证据检索与匹配"] end subgraph Output["赋能与优化"] E["赋能药剂师决策"] F["强化学习反馈闭环"] end A --> C B --> C C --> D D --> E E --> F F -.-> C ``` 主持人和嘉宾们对这类举措表示高度认可。传统的事先授权流程极为繁琐，直接造成了数以百万美元计的医疗资金浪费。通过引入自动化辅助工具，不仅加速了审核周期，也切实减少了由于流程冗余而产生的巨额医疗管理浪费。 ### [22:40](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=1360s) AI 落地的隐性成本与严苛的评估挑战 尽管 AI 应用的前景看似一片光明，但从治理和落地的幕后视角来看，实际推行过程中充满了真实的痛苦与阻碍。Aisha Rahim 引导大家探讨 AI 采纳背后的隐性成本。Bosky Mathew 指出，最大的痛点之一在于医疗业务对错误零容忍，因此必须在 AI 应用的测试与验证（Testing & Evaluation）上投入重金。尤其在智能体（Agent）和多智能体系统（Multi-agent systems）的架构下，行业目前仍在摸索怎样才是真正科学的评估模型。 此外，大语言模型（LLM）的运营成本极难预估：到底是直接调用大模型，还是投资去微调更小、成本更低的模型来压降长期的运维成本（Operationalization cost），目前整个行业都处于边探索边建立认知的阶段。现场主持人甚至询问台下观众是否有人精准测算过企业内部 GPU 或 Token 的消耗占比，结果全场仅有一人举手，生动体现了现阶段 AI 成本核算普遍处于“边飞边造飞机”（Build this plane as we are flying it）的探索初期。 ### [25:00](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=1500s) 重新审视 ROI：赋能顶级人才 vs. 替代客服中心 Dan Mackey 针对投资回报率（ROI）提出了独到见解。许多企业高管在立项时往往抱着非常传统且机械的预期，希望把 AI 丢进呼叫中心或文档处理部门，指望“投入十亿美元就能直接砍掉五亿成本”。然而，Dan 认为 AI 更像电子邮件（Email）这样的基础使能技术——几乎没有企业能精确量化出“花一百万在电子邮件上究竟帮公司赚了多少钱”。 真正的顶级 ROI 往往来自赋能核心高绩效人才（Top performers）。例如在制药公司，AI 如果能让最优秀的 10 名核心研发人员拥有 10 倍的生产力，就相当于瞬间拥有了一支百人顶级团队，这种价值是颠覆性的；反之，哪怕通过 AI 让药瓶背面根本没人拨打的 800 客服电话彻底免费，对公司的底层净利润也没有实质改变。Aisha Rahim 对此表示强烈共鸣，并以临床文档为例指出，AI 带来的 ROI 不能单纯用金钱衡量，它更关乎医生满意度与人才留存。医生们往往被繁重的病历记录压得喘不过气，AI 文档工具能够减轻这一负担，留住核心医疗专家，这才是医疗机构持续创新的根本。 ### [27:06](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=1626s) 多模型分层编排与私有化本地部署的兴起 Kahini Shah 从投资和行业趋势角度分析指出，随着 AI 从概念创新走向实际生产，推理成本（Inference cost）正成为核心考量。尽管推理成本整体在下降，但调用通用大模型的绝对支出依然偏高。因此，行业正在快速转向以下架构模式： - **专用与开源小模型**：对于特定单一任务（如呼叫中心对话摘要、FAQ 问答、内部检索），针对性训练或微调更小的专用模型，不仅成本极低，且在该任务上的准确率反而更高。 - **多模型编排层（Orchestration Layer）**：由轻量级模型处理常规任务，仅在需要复杂泛化或强逻辑推理时才将请求路由至通用大语言模型。例如由前 Google Gemini AI 工程师创立的开源机器学习平台 Umei，正帮助医疗客户落地这种多模型协同架构。 - **数据隐私与本地化驻留**：医疗数据高度敏感，机构通常不愿让数据离开自身网络或被外部拿去训练模型。结合中小型模型进行本地私有化部署，不仅降低了算力和存储开销，还满足了严格的数据合规要求。 Aisha Rahim 进一步补充，在企业数据中，通常只有约 10% 是真正可用于决策的高质量有效数据，其余 80% 至 90% 并不具备决策价值。未来医疗企业无需无休止地采购云存储来堆积无用数据，而是将算力与存储聚焦在高效的小型/中型模型上，建立分层架构。 ### [29:43](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=1783s) Optum 的规模数据优势与智能预测型呼叫中心 Bosky Mathew 阐述了 Optum Rx 在“药房福利管理（PBM）+ 药房服务（包括送药上门与特药药房）+ Optum Care 医疗服务交付 + Optum Insight 数据洞察”全链条一体化布局下的独特优势。该体系覆盖了超过 6000 万会员，沉淀了 6 到 9 PB（Petabytes）的海量真实世界数据。 庞大的真实数据储备成为了 Optum 构建和测试 Agentic 系统的核心护城河。基于海量样本的验证，模型能达到极高的准确度。Bosky 透露，他们最新研发并即将投入生产的一款联络中心（Contact Center）AI 产品，能够在会员刚拨通电话的瞬间，精准预测该会员来电的具体诉求。这不仅大幅提升了运营流转效率，更极大地优化了会员的咨询体验。 ### [31:51](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=1911s) 跨国落地策略与敏捷开发哲学：先做成，再合规 在谈及美国本土与国际市场在 AI 落地策略上的差异时，Dan Mackey 直言从他的视角来看“两者并没有本质区别”。他的核心原则始终是寻找适合产品的最佳工具，只要能把事情办成，过程中的路径并不重要。Dan 指出，大企业中有太多项目往往深陷泥潭——大家反复纠结“我们是不是该评估一下这个”或者“先跟法务法规团队聊聊”，导致项目最终根本无法上线。Dan 的做事哲学是“先让它跑通，然后再去搞清楚如何让它变便宜、合规并满足所有要求”（"make it work and then figure out how to make it cheap and compliant and all these things"）。在他的流程中，团队会先集中精力把原型与项目做出来，随后再交由法务和人力资源部门进行层层筛选与合规打磨，最终以某种可行的方式推向落地。 主持人 Aisha Rahim 随即回应道，这种将法务与 HR 放在合适时间节点介入的做法非常巧妙，避免了项目在一开始就被直接扼杀。在很多大型传统组织中，HR 和法务往往是第一批前去评估外部供应商的部门，结果在业务团队甚至还没看到供应商真正能做出什么成果之前，项目就因为一句“不符合合规要求”而被直接否决。 ```mermaid flowchart LR accTitle: 传统与敏捷 AI 项目推进模式对比 subgraph Traditional["传统大企业模式"] direction TB T1["评估供应商"] --> T2["法务/HR 前置卡点审查"] T2 -->|"合规受阻 / 流程拖沓"| T3["项目胎死腹中"] end subgraph Agile["Dan 的敏捷推进哲学"] direction TB A1["业务与工程验证"] --> A2["快速构建原型跑通功能"] A2 --> A3["引入法务/HR 进行合规化与降本"] A3 --> A4["成功部署落地"] end ``` --- ### [33:05](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=1985s) 风投资金流向与临床 AI 诊断突破：Bayesian Health 与行业标准 Aisha 进而向投资人 Kahini Shah 抛出问题，询问当前风险投资在医疗 AI 领域的具体流向、投资偏好以及预示颠覆的行业信号，甚至开玩笑询问是否有“股票内幕推荐”。Kahini 结合麦肯锡的行业数据回应称，VC 对整个 AI 技术栈的不同层面都充满兴趣，包括行政工作流优化、高级数据分析以及提升内部生产力的工具。然而，风投界目前最感到兴奋的下一个前沿阵地是 **AI 诊断（AI Diagnostics）**。 虽然诊断并不完全属于保险支付方的核心范畴，但 Kahini 以其投资的 Bayesian Health 为例展示了 AI 技术的落地力量。该团队脱胎于约翰斯·霍普金斯大学（Johns Hopkins），他们构建了一整套从 AI 模型构建、验证、获取 FDA 批准，到最终上线部署与实时监控的底层平台架构（Chassis）。整个系统严格保留了“医生在环”（Doctor-in-the-loop）机制，由专业医生进行监控并提供反馈，驱动模型持续进化。其率先落地的早期脓毒症（Sepsis）检测模型已在包括克利夫兰医学中心（Cleveland Clinic）在内的多家知名医院上线运行。 - **显著的临床与财务收益**：临床数据显示，该脓毒症检测模型实现了 27% 的患者死亡率下降（挽救生命），同时带来了 10% 的医疗资源利用率下降（降低成本）。 - **智能体减负**：在此基础上，平台结合了当下炙手可热的智能体 AI（Agentic AI），利用智能体协助医生完成临床文档记录与医嘱录入，极大地减轻了繁琐的工作流负担，缓解医护人员的职业倦怠（Clinician burnout）。 - **行业标准化建设**：Aisha 补充指出，公私合作在该领域至关重要，例如健康 AI 联盟（Coalition for Health AI, CHAI）正在积极制定针对脓毒症等医疗模型的评估标准，涵盖从研发前、研发中到部署后的数据选用规范与性能基准衡量。 --- ### [36:31](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=2191s) 智能体（Agentic AI）演进：多智能体系统、市场机制与药品价格透明化 探讨到支付方领域的未来重大技术趋势时，Bosky Mathew 强调智能体 AI 毫无疑问是当前最核心的投资方向。Optum Rx 正在采取平台化策略来应对未来大量涌现的 AI 智能体，确保其开发过程具备高度的一致性、标准化、安全性以及负责任 AI（Responsible AI）的治理要求。 Bosky 透露，未来的技术架构将深度依赖多智能体系统（Multi-agent systems），智能体之间需要相互协同合作。为了解决智能体的发现与调用问题，他们正在构建一个内部的“智能体应用市场”（Agent Marketplace）。在这个市场中，各类具备特定技能的专业智能体会对外发布，供各个业务线查找并调用所需的能力。在接下来的 12 个月内，团队将重点夯实这些基础设施组件，同时推进低垂果实（Low-hanging fruits）业务场景的试点。 除了大幅提升运营效率和为会员提供更无缝的互联体验外，Bosky 特别强调了“透明度”（Transparency）这一关键命题。通过构建 AI 助手与陪伴式智能体（AI Companions），系统能够向会员及企业客户清晰解释药品定价的底层逻辑与费用明细，解决“为什么某种药品在某地是这个价格”、“为什么用户需要支付这一金额”等疑虑，从而彻底改善药品价格的透明度。 --- ### [38:42](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=2322s) 二阶效应与持续学习系统：从“外卖专车”比喻到自主演进架构 Dan Mackey 在展望 AI 智能体的远期未来时指出，虽然近期的路线图相对明晰，但其更深远的长期冲击却极难预测，人们必须关注技术的“二阶和三阶效应”（Second and third order effects）。Dan 引用了一段 20 世纪 80 年代比尔·盖茨谈论互联网的经典影像作为类比：盖茨当年兴奋地向大众描绘互联网能让大家“在线和朋友下国际象棋”，但现实中几乎没有人仅仅为了下棋而用互联网。互联网真正引发的颠覆性变革，是催生了 Uber、DoorDash 这类按需配送和外卖服务（"burrito taxis"），它摧毁了传统出租车行业，重构了低成本的物流供应链。同理，AI 也将重塑商业价值链，吞噬现有业务的既有模式，并催生出当下人类无法想象的全新业态。 ```mermaid flowchart TD accTitle: 静态模型到持续演进智能体系统的转变 subgraph StaticPhase["静态与增强阶段"] direction LR S1["大语言模型预训练"] --> S2["静态微调 (Fine-tuning)"] S2 --> S3["RAG (检索增强生成 / 静态数据注入)"] end subgraph AgenticPhase["自主进化阶段 (Agentic Paradigm)"] direction LR A1["多智能体协同系统"] --> A2["环境与交互反馈信号 (人类点赞/踩/系统指标)"] A2 --> A3["持续在线学习与自我进化 (Continual Learning)"] end StaticPhase -->|"架构演进"| AgenticPhase ``` Kahini Shah 在最后总结了纯技术层面的演进趋势。她指出，当前的系统（无论是 OpenAI 模型还是企业自建系统）在训练或微调后本质上仍是静态的，即便是借助 RAG（检索增强生成）技术引入实时数据，底层能力依然缺乏主动迭代。随着行业迈入智能体范式（Agentic Paradigm），前沿研究的焦点正转向如何让系统实现“持续在线学习与自我进化”（Continuously learning and self-improving）。系统需要能够根据真实环境中的反馈信号（如人类的点赞/点踩反馈、系统的运行指标等）进行不间断的自我完善。随着现场工作人员给出时间耗尽的信号，讨论在对持续学习架构的期待中圆满收尾。 ## Notable quotes > [03:20](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=200s) AI will recommend, physician will decide（AI 推荐，医生决策）。 — Aisha Rahim > [06:45](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=405s) 这才是真正的 ROI 核心：你是想让呼叫中心效率提升 10 倍，还是想让你最核心的 10 个人爆发 10 倍的战力？ — Dan Mackey > [11:13](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=673s) Google 内部目前已有 25% 的代码由 AI 生成。 — Kahini Shah > [13:34](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=814s) 大家都在“边开飞机边造飞机”（building the plane as we are flying it）。 — Aisha Rahim > [25:00](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=1500s) 几乎没有企业能精确量化出‘花一百万在电子邮件上究竟帮公司赚了多少钱’。 — Dan Mackey > [31:51](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=1911s) 先让它跑通，然后再去搞清楚如何让它变便宜、合规并满足所有要求（make it work and then figure out how to make it cheap and compliant and all these things）。 — Dan Mackey > [38:42](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=2322s) 盖茨当年兴奋地向大众描绘互联网能让大家‘在线和朋友下国际象棋’，但互联网真正引发的颠覆性变革是催生了 Uber、DoorDash 这类按需配送和外卖服务（burrito taxis）。 — Dan Mackey > [38:42](https://www.youtube.com/watch?v=R2oxtNeNJ38&t=2322s) 前沿研究的焦点正转向如何让系统实现“持续在线学习与自我进化”（Continuously learning and self-improving）。 — Kahini Shah ## Mentioned Aisha Rahim, Johns Hopkins Health Plans (JHHP), Harvard Macy Institute, Kahini Shah, Obvious Ventures, Devoted Health, Galileo, Virta Health, Dan Mackey, The Cigna Group, CU Health Sciences Center, Bosky Mathew, Optum Rx, UnitedHealth Group, McKinsey, NVIDIA, CMS (Centers for Medicare & Medicaid Services), WISER, Cursor, GitHub Copilot, Devin, Factory, All Hands, Google, Sundar Pichai, MCP (Model Context Protocol), Gen Health AI, OneUp Health, Umei, Bayesian Health, Cleveland Clinic, CHAI (Coalition for Health AI), Bill Gates, DoorDash, Uber, OpenAI ## Video description Aisha Rahim, Medical Director and Co-Lead at AI Governance Council, Johns Hopkins Health Plans (JHHP) Dan Mackey, Director of AI Innovation, Worldwide, The Cigna Group Bosky Mathew, VP Architecture & Chief Architect, Optum RX Kahini Shah, Partner, Obvious Ventures Abstracts This session will explore how healthcare payors are using AI to enhance operational efficiency, improve member care, and drive cost savings. Panelists will discuss AI-driven solutions for automating claims processing, detecting fraud, and predicting member needs to offer more personalized care plans. Learn how AI is transforming payor operations, from streamlining back-office functions to improving customer service through advanced analytics. The discussion will also focus on best practices for integrating AI into payor systems while navigating regulatory requirements and ensuring data privacy. [Source](https://www.youtube.com/watch?v=R2oxtNeNJ38)

[3:20] Aisha Rahim: AI will recommend, physician will decide（AI 推荐，医生决策）。 [6:45] Dan Mackey: 这才是真正的 ROI 核心：你是想让呼叫中心效率提升 10 倍，还是想让你最核心的 10 个人爆发 10 倍的战力？ [11:13] Kahini Shah: Google 内部目前已有 25% 的代码由 AI 生成。 [13:34] Aisha Rahim: 大家都在“边开飞机边造飞机”（building the plane as we are flying it）。 [25:00] Dan Mackey: 几乎没有企业能精确量化出‘花一百万在电子邮件上究竟帮公司赚了多少钱’。 [31:51] Dan Mackey: 先让它跑通，然后再去搞清楚如何让它变便宜、合规并满足所有要求（make it work and then figure out how to make it cheap and compliant and all these things）。 [38:42] Dan Mackey: 盖茨当年兴奋地向大众描绘互联网能让大家‘在线和朋友下国际象棋’，但互联网真正引发的颠覆性变革是催生了 Uber、DoorDash 这类按需配送和外卖服务（burrito taxis）。 [38:42] Kahini Shah: 前沿研究的焦点正转向如何让系统实现“持续在线学习与自我进化”（Continuously learning and self-improving）。

[0:04] Dan Mackey: Prior to that, I was a bank robber. I would break into companies and tell them how I did it. [3:20] CMS WISeR Initiative: AI will recommend, physician will decide. [6:45] Dan Mackey: Cigna is the seventh-oldest company in the United States. It was incorporated and running before the White House was built. So ever since then, every quarter a manager has tried to save 3% somewhere... It may only cost four pennies to have a guy look at a document and confirm the JSON file and hit next. And getting AI to do the same job for less than 4 cents is not on the table right now. [11:13] Dr. Aisha Rahim: building the plane as we are flying it [16:52] Dr. Aisha Rahim: no adverse benefit determinations or denials can be issued autonomously by AI. [25:00] Dr. Aisha Rahim: Our physicians are drowning. I'm a physician myself, and I spend so much time taking home the documentation and still not able to finish it at the time of point of care. Artificial intelligence ROI is not just monetary; it's also satisfaction, making sure you are keeping your top talent and not losing that. [31:51] Dan Mackey: My philosophy is make it work and then figure out how to make it cheap and compliant and all these things. [22:40] Dr. Aisha Rahim: to build this plane as we are flying it

[0:04] Good afternoon everyone. It's so [0:07] exciting to see all of you here today. [0:09] Uh I'm Dr. Aisha Raheem. I'm a medical [0:12] director at Johns Hopkins Health Plan. [0:15] I'm internal medicine trained physician [0:17] and decided to take a little bit of [0:18] different pathway and join the health [0:20] plan more than five years ago. I lead [0:23] their artificial intelligence and [0:25] strategy uh and I help them set up the [0:28] artificial intelligence governance [0:29] council. In my spare time, I do travel a [0:33] lot and speak about AI and healthc care [0:37] transformation globally, nationally, [0:40] internationally and I'm also a faculty [0:43] at Harvard Macy Institute. Uh if anyone [0:46] knows about Harvard Mason Institute, [0:47] it's a it's a combination of Harvard [0:50] Medicine and Harvard uh business school [0:54] which has come together and offer [0:56] educational courses to our providers and [1:00] folks that are interested in healthcare. [1:03] With that [1:04] um I have a privilege of having our very [1:10] three key leaders from different [1:12] organizations. [1:14] uh these are very bright leaders and [1:16] they are leading artificial intelligence [1:18] strategy in their respective [1:19] organization. So I'm going to pass this [1:22] mic around and do a round robin of [1:25] starting with Kahin Sha. [1:28] Please describe your role and what you [1:30] do regarding AI and pass it on to Dan [1:33] and Bossi. [1:35] Hi everyone, I'm Kahinisha, partner at [1:38] Obvious Ventures. We're an early stage [1:40] firm based in Silicon Valley. We do seed [1:42] series A investments and we've invested [1:45] in healthcare companies like Devoted [1:47] Health in the Medicare Advantage space, [1:49] Galileo in the primary health space, [1:52] Verta Health, reversing type 2 diabetes [1:55] and a number of companies in healthcare [1:57] AI helping with the prior oath process, [2:01] AI diagnostic administration, um, etc. [2:04] And I've worked in AI for over a decade. [2:07] First as a software engineer, product [2:10] manager, and investor. I was previously [2:12] at Google's early stage AI fund before [2:14] obvious ventures. Excited to be here. [2:20] >> No one clapped for you. [2:23] >> There we go. [2:23] >> Thank you. [2:25] [Applause] [2:28] >> My name is Dan. I work for Sigma. Um I [2:31] do AI for Sigma's entire product [2:33] portfolio worldwide. Before that, I [2:35] worked at U Health Sciences Center. I [2:38] was a nuclear engineer. And then prior [2:40] to that, I was a bank robber. I would [2:42] break into companies and tell them how I [2:43] did it. So hopefully that diverse skill [2:47] set will be entertaining for you all. [2:51] >> That's the clapping. [2:53] >> Okay. [2:55] Hello everyone. Uh my name is Loski [2:56] Matthew. I'm the chief architect and uh [2:59] VP of architecture for Optimarox. Um uh [3:02] Optimarox is part of United Health [3:03] Group, the largest healthcare company. [3:05] Um I currently right now um I'm driving [3:08] the AI strategy for you know optimum [3:11] specifically optim building and [3:13] implementing um genai and agentic [3:15] strategy for optimics. [3:17] Okay. Awesome. [3:20] All right. We're getting the room warmed [3:22] up a little bit here. Okay. So before we [3:25] dive into questions here, I would like [3:28] to know who's here in the audience. I [3:31] would like to know by show of hands how [3:34] many folks are in healthcare industry [3:36] here. [3:37] Oo 100% awesome. Okay, next question. By [3:42] show of hands, how many folks use AI in [3:46] their personal or in their work life? [3:51] Uh that's 101%. [3:54] So this is great. You know what? Since [3:57] the start of like chat GPT era, I've [4:00] been traveling and doing conferences and [4:02] sessions. It used to be like one or two [4:05] folks in the room using AI and now it's [4:08] amazing to see pretty much 10 and1% [4:11] adoption or use of artificial [4:12] intelligence. So that's great. So before [4:15] we jump into the question wanted to set [4:18] a little bit of stage of what's [4:19] happening in the industry [4:21] and talk a little bit about what's [4:23] happening in the policy front. Um the [4:25] global healthcare AI market is valued [4:28] between 14 to 29 billion in 2024. [4:32] 14 to 29 billion 2024. That's according [4:36] to McKenzie. And then some estimate says [4:39] it could go up to uh 110 billion by [4:42] 2030. That's quite a bit of numbers [4:45] fair. We're talking about billions in in [4:47] terms of investment and the growth is [4:50] driven by peer automation, provider [4:53] workflow optimization and breakthroughs [4:55] in patients engagement. [4:58] 70% of organization including peers and [5:00] providers and integrated systems are [5:02] actively implementing Gen AI solution to [5:04] improve efficiency, accuracy and member [5:07] satisfaction. And according to Nvidia's [5:10] July 2025 healthcare survey, AI adoption [5:13] is already boosting revenue, reducing [5:16] administrative burden and uh [5:18] accelerating innovation. [5:21] In terms of regulatory uh up uh you know [5:24] forefront in January 2025, President [5:27] Trump issued executive order for 4179 [5:31] rolling back Biden's era guard rails and [5:34] launching our federal AI action plan. [5:37] And this was to pri prioritize [5:39] innovation, open-source development and [5:42] reduce regulation. Just last month in [5:45] July, administration uh Trump [5:47] administration released winning the race [5:50] uh and America's AI action plan with [5:52] more than 90 policy recommendations [5:54] including preeemption of certain states [5:57] AI laws, streamed data center [6:00] permitting, and incentives for US-based [6:02] AI and research and development. On the [6:05] CMS side, the wiser initiative with [6:07] pilot AI system prior authorization in [6:10] six states in 2026. AI will recommend [6:14] physician will decide this is meant to [6:16] reduce waste but also raising debates [6:18] over potential care delays. Does anybody [6:21] know about the wiser program by CMS? [6:24] Okay, so yeah, we should go and read [6:27] about it and what does that mean and how [6:30] that innovation will affect peers and [6:32] providers. So with with that sort of [6:35] every person involved in this room in [6:37] healthcare with that kind of policy and [6:40] innovation [6:42] push I'm going to start with my first [6:45] panel panelist Bosski where is AI [6:48] delivering measurable ROI [6:51] especially for peers today like where [6:54] where do you see that the most impact? [6:57] Yeah. So um in optimaroxics I can start [6:59] with um what's our strategy where we are [7:02] where we are investing and then can talk [7:04] about the ROI and all that. So in in uh [7:07] in optimarics what we are how we are [7:09] thinking uh the the geni or agentic [7:11] strategy is uh going to be improving our [7:14] operational efficiencies uh mainly like [7:16] creating assistance in every areas of [7:18] our operations not just only for [7:20] operations also for member outcome right [7:22] so we will be uh creating the assistance [7:25] or companions for member uh which will [7:27] help member to like you know um get [7:29] better better outcome better experience [7:32] and also like call center agents will be [7:34] creating the same kind of comb companion [7:36] and assistance for calls center agents [7:38] to help members when you know when they [7:40] get calls from members um and also our [7:42] clients we have a large number of [7:44] clients that we service so for um [7:46] clients always look for data right and [7:48] it's very hard to navigate the complex [7:50] ecosystem that we have um with really [7:53] like you know large um set of [7:55] application systems that they need to [7:57] navigate to find data so that is where [7:59] we are we see the value and you know [8:01] like that's going to definitely improve [8:03] our operational efficiencies is uh the [8:06] quality of operations and again uh the [8:08] member outcome the member outcome is the [8:10] most important thing for the better [8:11] member experience. So that is how we we [8:13] we believe uh we will become we we will [8:15] be like becoming a market differentiator [8:18] in the world in this in this area being [8:20] the largest player in you know in the [8:22] healthcare ecosystem. [8:24] >> Sounds good. So really it seems like a [8:28] lot of AI adoption are places where AI [8:31] is solving problems and once it's [8:34] solving problems it's adding value and [8:36] that's where the ROI is coming to. [8:38] >> Yeah. So not really specifically you [8:40] know uh we don't really uh go into do [8:42] any AI based solutions for clinical [8:44] decisioning anything like that. So that [8:46] is not going to be there mostly focusing [8:48] in like improving member experience you [8:50] know improving our operational quality [8:52] efficiencies and that sorts of areas. [8:54] Okay, makes sense. Dan, do you have a [8:56] take on that question? The question [8:58] really is where is AI delivering [9:00] measurable ROI for peers. [9:02] >> Yeah. So, we're working on the along the [9:04] same lines. I think everybody is, but [9:06] the big thing about AI is that it's a [9:08] 10x accelerator. It's a multiplier. It's [9:11] a power up. Like, if you're playing [9:13] Mario and you get the little star thing, [9:14] that's what AI is doing for you. So, [9:16] there's a lot of conversations about can [9:18] you put AI in the contact center? Can [9:21] you use it for document processing? [9:23] where can you, you know, best use it? [9:26] And the issue that we're running into is [9:28] that a lot of these processes are [9:30] already optimized. Sigma, for example, [9:32] is the seventh orders company in the [9:34] United States. It was incorporated and [9:36] running before the White House was [9:38] built. So, ever since then, you know, [9:40] every quarter, a manager has tried to [9:42] save 3% somewhere and the entire [9:44] workflow is as optimized as possible. So [9:47] we have people who are doing the call [9:50] centers, people who are doing document [9:52] validation, and it may only cost four [9:55] pennies to have a guy look at a document [9:58] and confirm the JSON file and hit next. [10:00] And getting AI to do the same job for [10:02] less than 4 cents is not on the table [10:06] right now. It's coming, but not today. [10:08] The the greater RORI that we're getting [10:11] is in empowering the top 10% most [10:14] productive people. So there will be I [10:17] see executives who have ideas and they [10:19] would have given them to a team to work [10:21] on and the team spends two weeks on it [10:24] brings it back questions goes back and [10:25] forth but now they do the entire thing [10:27] in 45 minutes they explore their own [10:30] ideas. So what you end up with is these [10:33] guys who I saw one executive who brought [10:36] me an app that was almost done with his [10:38] concept on it. That's where the true [10:40] value is going to come from. Like, do [10:41] you want to have your call center be 10 [10:44] times more efficient or do you want your [10:46] best 10 people to be 10 times more [10:48] powerful? That's the the true RARI [10:51] breakdown. [10:53] >> Sounds great. And at Johns Hopkins [10:55] Health Plan, I'm just want to add that [10:57] we are actively putting our energy into [11:00] helping our members be more informed in [11:03] terms of member engagement as well as [11:05] having a chatbot or a bot to explain [11:09] member the benefits to drive more [11:11] engagement there. Also on the fraud, [11:13] waste and abuse side, we are act heavily [11:16] investing to make sure that uh we are [11:19] not losing money on duplicate uh [11:22] medications and claims and that's where [11:25] another investment is going and then [11:26] obviously I think there is a very heavy [11:29] potential on um optimization side as [11:31] well and we can talk a bit later as [11:33] well. Kahini do you have any insights on [11:35] this one? [11:36] >> Absolutely. From an investment [11:39] perspective, the two areas I'm really [11:41] excited about is one software [11:43] engineering and making developers more [11:46] efficient. Do we have any developers in [11:48] the audience here? [11:50] Um, less um but we're seeing code [11:54] complete tools like cursor or github [11:56] copilot gain popularity. We're seeing [11:59] agentic solutions um like Devon factory [12:03] all hands where um AI can do tasks for [12:07] engineers taking off and Sundar said [12:10] that 25% of their code at Google now is [12:13] AI generated. I'm kind of curious what [12:15] percentage of code within healthcare [12:17] organizations is now AI generated. But [12:20] now we're seeing that the bottleneck is [12:22] if AI is doing the work, how do we [12:24] review and police AI, which I think will [12:26] be a topic for later. But this is a [12:29] great boost in productivity. It also [12:32] kind of helps do boring tasks that [12:34] engineers don't often want to do like [12:36] maintenances or migrations. Um the other [12:40] area that we're jazzed about is call [12:42] center. Um so it is a great tool right [12:46] as voice gets less expensive which it [12:48] will over time and more apt I think AI [12:52] can handle low complexity lowrisk task [12:55] right update information tell me about [12:56] my coverage leaving room and space um [13:00] for people to do more complex use cases [13:04] and I also think it can help payers do [13:07] personalized [13:09] member outreach and engagement at an [13:12] unprecedented unpredented scale, right? [13:14] Imagine being able to send vaccination [13:17] reminders or screening reminders or help [13:19] manage chronic care and chronic [13:22] conditions. Um, so I think all of this [13:24] will become possible at scale. And what [13:26] I love about this is it improves access. [13:28] Everybody has a phone. It reduces costs [13:31] and hopefully hopefully improves [13:33] outcomes. [13:34] >> That sounds amazing. And you know [13:37] building upon it is that I recently had [13:40] a friend who built up his whole website [13:43] on um artificial intelligence platform. [13:48] He's not a coder but he did that. Having [13:51] said that after doing that the whole [13:53] building the the back end collapsed and [13:56] the artificial intelligence was not able [13:58] to fix it. So I think this is happening [14:01] as we are speaking and we're sort of [14:04] building the plane as we are flying it [14:06] at the moment. More the technology is [14:10] going to become advanced. The more we [14:11] are able to build these solutions out [14:13] without having much knowledge of coding [14:15] and we're going to be able to code even [14:18] as healthcare providers and other folks [14:20] as well. So that brings to our next [14:24] question. What are the high-risk and [14:26] high reward projects that you guys are [14:28] working on? And I know innovation sort [14:30] of sometimes starts quietly. Uh could [14:33] each one of you share behind the scene [14:35] the AI initiatives that you're working [14:38] and are your most excited about? [14:42] >> Yeah, I can start. Um so in like in [14:44] Optimarox uh we are actually as I [14:47] mentioned before so we are our strategy [14:50] is to build um assistance or companions [14:52] for all different like all all different [14:54] areas right including um you know [14:56] different areas of operations calls and [14:59] regions members member facing all all [15:01] sorts of things. So we really now in the [15:04] process of designing it uh in like you [15:06] know uh expanding on it what does what [15:08] does that really mean to us? Um so we [15:10] will be building like kind of you know [15:12] 50 60 70 agents um and that specialized [15:15] in certain areas you know some of the [15:17] agents will be specialized in answering [15:19] questions about benefit components or [15:21] formulary or claim explaining the claim [15:23] or prior oath and all sorts of areas. So [15:26] we'll be like building so many like [15:28] agents and also we are investing heavily [15:30] in our engineers and also our business [15:32] users to like you know be equipped with [15:34] the right technology and also like [15:36] trained in how to build them. we call [15:38] them call them as AI builder. Um in [15:41] internally we call them as AI builder. [15:43] So for once we have uh you know once [15:46] once we have these many like you know [15:48] engineers ready to like build AI [15:50] projects uh also you know as I said we [15:53] are going to build we need to build so [15:54] many of this you know specialized agents [15:56] and also like you know functional [15:57] agents. We need to be like building this [16:00] is um what do you call in a in a [16:03] specific way right? we need to have some [16:05] kind of policies and right practices [16:07] that we need to implement. So we are in [16:09] the process of investing like kind of in [16:10] a platform approach uh building a kind [16:13] of an agent spec starting with an agent [16:16] spec and also like you know other uh [16:18] maybe like now MCB is uh in there like [16:21] we probably will be building an MCP [16:22] gateway and figuring out the security [16:25] and all other you know governance and [16:28] and responsible AI practices all of that [16:30] are there. Now the good thing is like we [16:33] have a centralized or enterprise AI team [16:35] like optimum AI team which is helping us [16:37] but partnering with us even in optimum [16:39] RX we need to like heavily invest in it [16:42] that is what we are we are actually now [16:43] currently working on which I'm super [16:45] excited about [16:48] >> and do you want to share any insights on [16:49] that any high-risisk higher reward [16:51] projects [16:52] >> yeah we got a bunch there's some that [16:54] are very technically cool and attract me [16:56] but the ones that are more impactful to [16:58] society will be things Like in [17:02] insurance, there's a concept called a [17:03] pre-o, a pre-authorization. If you go to [17:05] your doctor, your hospital, and you want [17:07] a surgery, they ask the insurance [17:09] company in advance, hey, will you pay [17:11] for this? And this pre-o process, it can [17:14] take a very long time. The they'll put [17:17] letters in envelopes with a stamp and [17:19] mail it back and forth with these [17:20] people. Um, months will go by and it's [17:23] months while the patient is waiting for [17:24] the surgery, too. the a thing that we're [17:28] working on on accelerating this and I [17:29] think you guys working along the same [17:30] lines is you just if you're the hospital [17:34] you upload your claim submission and it [17:36] can immediately tell you if it's [17:38] approved or if it's not what the problem [17:40] is like it will be this photo doesn't [17:42] have an MRI number in it or here's the [17:44] issue and you can fix it right there [17:46] upload it again you can reduce a [17:48] multi-week multi-month process down to a [17:52] couple of minutes and that's good for [17:53] the patient it's for the people who [17:56] don't like paperwork. It's really and on [17:58] the hospital side, they can automate [17:59] their side too with AI. So, it's just be [18:01] two computers talking to each other. [18:02] It's going to be great. I'm very [18:03] excited. [18:04] >> Yeah, that's that's actually the future [18:06] and I'm pretty excited about it. In an [18:08] ideal world, once the provider uploads [18:12] that document on the systems, the exact [18:14] criteria from that specific insurance is [18:17] already there from the insurance side [18:20] and it pulls up that criteria. It [18:22] compares it against that criteria and [18:24] says approved or not approved. And if [18:27] it's not approved, who where does it go [18:29] to? Does it go to provider or nurse? And [18:32] then that's where the the further [18:35] decision making making happens. And [18:36] that's where us as medical directors [18:38] that health plan come in and see where [18:40] where whether we need to approve it [18:42] because there isn't enough information [18:43] or not. [18:44] >> Get it fixed so quick. [18:45] >> Yeah. Exactly. And it's automated and [18:47] it's quick. Having said that, one of the [18:50] responsible principles of AI is that [18:52] human has to be in the loop. Anything [18:54] that is going to going to affect our [18:56] members negatively like negative [18:58] benefit, we're not approving that has to [19:01] be reviewed by our nurses and our [19:03] medical directors and it cannot be [19:05] automatically denied. And that's where [19:07] we draw a red line where we say, okay, [19:09] we're going to use artificial [19:10] intelligence, but we're going to use it [19:12] responsibly in a way that is going to [19:14] help us make decision faster. But in any [19:17] decision where is the decisions against [19:19] our member in a sense that we're not [19:20] approving service we have to review it [19:23] uh by our physicians and medical [19:25] directors and a lot of time we end up [19:27] approving it anyway. So [19:28] >> do you do human in the loop for [19:30] approvals or just denials? [19:32] >> Mostly uh at the earlier stage when [19:34] you're developing the system, it's for [19:36] approvals and a denial [19:38] >> to make sure the approvals are [19:39] appropriate. Oh yeah because you know if [19:42] you start off with the process from [19:44] scratch the approvals that are done by [19:46] AI are these approvals are even the [19:48] approvals that should have happened and [19:50] then we do uh robust quality review and [19:53] so we we this is not implemented yet [19:55] we're working towards it and then for [19:57] the denials obviously that's a sort of [20:00] very important so just adding that [20:03] >> okay do you have any insight on that [20:05] >> yeah I maybe I'll just double down on [20:07] the prior a theme here um but [20:09] >> big problem [20:11] But I'll give an example. So we invested [20:14] in a company in the space. It's called [20:16] Gen Health AI. The founder previously [20:19] started a company that maybe some of you [20:21] are familiar with called OneUp Health. [20:23] It's a fire server. But what Gen Health [20:25] is doing is they're building a co-pilot [20:29] for the UN process. So they're building [20:31] the tool that sits besides the decision [20:34] maker for the intake process, gathers [20:38] and pulls out the correct medical [20:40] information, applies rule, the correct [20:42] rules to it. And again, there's a human [20:45] actually making the decision. They're a [20:48] co-pilot to the human through this [20:50] process. And they're live. They're [20:53] processing thousands of cases a day. [20:55] They're making the people 4x more [20:58] productive, dropping the time like you [21:00] said down to minutes. Um, and they've [21:02] reduced the cost of adjudication by 75% [21:06] uh with the organizations they're [21:08] working with. And similarly, right, like [21:11] the question is like how do you make it [21:12] safe? How do you know? And their [21:15] solution is really elegant. They've [21:17] built a beautiful UI that walks through [21:20] the logic of each step, shows citations, [21:25] so it makes it easier for the human to [21:27] audit what AI is doing. And the human [21:30] can also give the model feedback um [21:32] through a process called reinforcement [21:34] learning. So we take that human input [21:37] and feed it back into the model and [21:38] continue to make it more efficient over [21:40] time. [21:41] >> And where can I buy this? [21:44] >> Will you? I mean [21:47] >> okay perfect [21:48] >> that's great [21:48] >> just one point here right on stage [21:51] >> just one point to add there we also like [21:52] built a prior authorization assistant [21:55] that was our first agent that we built [21:57] so from that we actually learned a lot [21:59] it's uh the kind of the same way just [22:02] basically um for we are in the you know [22:04] pharmacy side just for every drug [22:06] there's a clinical guidelines and based [22:08] on the clinical guidelines is what we do [22:09] the prior review so basically for each [22:12] of the clinical guidelines based [22:13] questions finding evidences from chart [22:15] notes and that is helping that is [22:17] actually empowering the pharmacist a [22:18] lot. [22:19] >> Yeah, Jenhel just started a pharma [22:21] module too. So, [22:22] >> okay. Amazing. [22:23] >> Yeah, that's great. I'm I'm very excited [22:25] about this initiative because it can [22:27] make a dent in multi-million dollar um [22:31] healthc care waste that we waste our [22:34] money on. Basically, this prior process [22:36] waste a lot of money in real time. Now I [22:40] want to sh pivot a little bit on the [22:43] sort of the hidden cost of AI adoption [22:46] and what are the barriers because I know [22:49] we are here to talk about all the rosy [22:51] stuff and it's great and it's all [22:52] working for us but behind the scenes [22:54] there are real tears and pain uh and I [22:58] have personally after establishing a [23:01] console I've [23:03] met with that continually in terms of [23:05] adopting different technologies and so [23:08] in your opinion uh starting with you [23:11] bossi what are those pain points and [23:14] struggles behind the scene? [23:16] >> Yeah. So uh when I talk about pain [23:18] points I think um one of the main thing [23:20] is just basically we need to invest [23:22] heavily in uh like you know testing uh [23:25] the um you know AI applications that we [23:28] are building. That's very critical very [23:30] crucial for especially like you know [23:32] businesses like ours right we cannot [23:34] make mistakes. Um so that's just um you [23:36] know we we still figuring out what is [23:38] the the right evaluation models for [23:41] specifically when it comes to agent and [23:42] multi-agent systems things like that. Um [23:45] so that that's one area and also when it [23:48] comes to cost the operational cost we [23:50] don't have any any idea right now. So [23:52] how much is LLM's going to cost us and [23:54] are we going to invest in you know like [23:57] um fine-tuning smaller models and that's [23:59] going to help us in reducing the the [24:01] overall like operationalization cost and [24:04] all of that area. So basically lot you [24:06] know lot more to figure out even now. Um [24:09] so that's just I see there is a lot of [24:11] hidden cost that way. [24:12] >> That's a great uh point that you [24:14] mentioned what are the cost of large [24:16] language model. I mean if if you build [24:18] it inhouse versus you bring bring it [24:20] from outside. Has anybody any cost [24:23] analysis done here in this room on what [24:26] is the percentage of GPUs or tokens that [24:29] you've guys are used in your companies [24:32] but raise of hand [24:35] any cost analys okay so okay great [24:38] awesome we have one guy here so yeah [24:40] that's sort of upcoming like we have [24:41] some discussions going on in our [24:43] organization is like how much are we [24:45] going to use it how much are we going to [24:47] you know spend on this so those are real [24:50] time questions that we are asking and [24:52] trying to figure out answers as like [24:54] we're trying to do that to build this [24:56] plane as we are flying it. Dan, if you [24:58] have any insight on that. [25:00] >> Yeah, the biggest ROI situation, at [25:04] least from my perspective, is kind of a [25:06] a practical concern. A lot of times [25:09] executive leadership will want to apply [25:11] it to call centers or document [25:13] processing or areas like that. And what [25:15] they want to see is, okay, we'll put in [25:17] a billion dollars and then I want to see [25:19] a cost reduction of half that, something [25:20] like that, right? But AI is a technology [25:23] more like email where figuring out the [25:26] ROI of your email is very difficult. [25:28] Like if you spend a million dollars on [25:29] email, how much is that helping your [25:31] company? You can't really measure it. [25:33] Um, [25:35] most of the ROI is going to come from [25:38] enabling the top performers. Like if [25:40] you're a pharma company, would you [25:42] rather have your 10 best researchers be [25:45] 10 times more powerful? You have a team [25:47] of 100 now. Or do you want that 800 [25:49] number on the back of the bottle that no [25:51] one calls to be free? It doesn't make a [25:53] change in your your bottom line. So [25:55] finding the the right applications for [25:58] it that will actually return the [26:00] greatest investment and that may not be [26:03] quantifiable and getting the buy in for [26:05] it. That's really kind of where people [26:07] should be thinking. I feel like [26:09] >> that's a great point and I was as soon [26:11] as you talk about it reminded me of an [26:13] example of a bridge AI where really it's [26:16] not about cost reduction you know it's [26:18] about improving physician satisfaction [26:21] and the time that they're spending on [26:22] the documentation and that's not really [26:24] translating into a monetary ROI uh our [26:28] physicians are drowning I'm a physician [26:30] myself and I spent so much time taking [26:34] home the documentation and still not [26:36] able to finish it at the time a point of [26:38] care and so sometimes I think or I think [26:41] most of the time artificial intelligence [26:43] ROI is not just monitoring it's it's [26:46] also satisfaction making sure you are [26:49] keeping your top talent and not losing [26:51] that and especially for organizations [26:53] like Johns Hopkins it's very important [26:55] that we invest or we have we are [26:57] investing heavily in our uh people so [26:59] that they we can they can stay with us [27:01] and innovate and move the organization [27:03] forward uh any insight on that Nahini [27:06] Yeah. Um, we are seeing cost as a [27:10] growing concern now, especially as AI [27:13] moves from innovation to actual [27:16] production, right? Like Dan said, now [27:18] it's like, are you cheaper than my cost [27:21] call center? Even if you're making my [27:23] 10x engineer more productive or making [27:25] my 5x engineers 10x engineers, how how [27:27] do I measure that? Right? It's becoming [27:29] a question and the cost of inference is [27:33] high, dropping, but high. And what we're [27:36] seeing is the rise of specialized [27:40] models. So if a model only needs to do [27:42] one thing like call center [27:44] summarization, [27:46] it makes more sense to train a smaller [27:48] model. It's cheaper and then it's also [27:50] more accurate for that one use case. [27:54] We're seeing the rise of more opensource [27:57] models. It can be self-hosted. It can be [28:00] fine-tuned to the organization. And you [28:03] can do things like FAQs, chat bots, like [28:06] very effectively using that at a low [28:09] cost, better search, and then really [28:11] just using the large language models [28:13] where they're required for more [28:16] generalized tasks or more reasoning [28:18] tasks. And then kind of like building an [28:20] orchestration layer, right, that goes [28:22] across these. And um we have a company [28:25] um in the space serving healthcare [28:26] clients. It's called Umei. It's an [28:28] open-source machine learning platform [28:31] started by Google, Gemini, AI engineers, [28:34] and starting to do more of this like [28:37] multimodel thing within larger [28:39] organizations. [28:40] That's a great point. Multimodel [28:43] and tiered approach where you're [28:45] starting off with mediumsiz language [28:47] models which are trained on a smaller [28:50] data but use less GPU, less cost and [28:52] less data storage capability. And I know [28:56] cloud storing company will tell you [28:57] otherwise because they would like to buy [28:59] have you buy more storage for more data [29:02] storing. Uh having said that it's [29:06] certainly within a company only 10% of [29:08] the data is usable data meaning the rest [29:11] 80% or 90% is not something that you [29:14] take that data and make decisions off of [29:16] it. So I think that's where we're going [29:18] to be heading in future. We're going to [29:20] be using less data with the with the [29:23] with the vision of le storage capacity [29:25] that they're using and medium-sized [29:28] language models and then building up to [29:30] large language models. 100% right. And [29:33] if you're a healthcare organization, you [29:34] might not want data to leave your [29:36] premises. It's sensitive. You might not [29:38] want to give it away for training. So, I [29:40] think we'll see a lot a lot more of [29:42] that. [29:43] >> Okay, great. So I'm going to go a little [29:46] bit deeper into um the PBM and AI [29:49] integration and Bossi enlighten us in in [29:53] terms of Optum's unique position. They [29:55] have uh integrated pharmacy care [29:58] delivery pair functions. How is that [30:01] footprint accelerating AI's potential? [30:03] Uh can you add a little bit more color [30:05] to that? Yeah. So we are we actually um [30:08] you know have a unique opportunity here [30:11] um being like specifically in optimarx [30:13] itself we have um or home delivery [30:16] specialtyarmacies which is um which is a [30:18] pharmacy services and we also have [30:20] pharmacy benefit which is PBM. Um so we [30:22] have both and again like in if you look [30:24] at holistic optimum we have optimum care [30:27] and which is care delivery [30:28] organizations. We optimum insight which [30:30] we have lot of data and you know like um [30:32] other um kind of products. uh with all [30:35] of this like the data is our I think the [30:38] highest uh you know the the biggest [30:40] advantage that we have. Um so in our [30:42] products we serve 60 million membership [30:44] and we also have maybe data if you look [30:47] at it probably 6 to9 pabytes of data. So [30:50] when we really like build this agentic [30:52] systems we can really um test them on [30:55] you know on the real data with the large [30:57] amount of data that is going to help us [30:59] build you know um a better accuracy kind [31:02] of results and the products that that [31:04] provide u better accuracies and all [31:06] that. So recently like we built a [31:08] product which we are already like going [31:09] to um go in you know um going going in [31:12] production uh basically a contact center [31:14] uh product which actually can um you [31:17] know cor predict why the member is [31:19] calling us right that is a big advantage [31:21] uh for us and also it's a good [31:23] experience for the member who is calling [31:25] in that we know why the person is [31:27] calling so that is actually showing you [31:29] really high accuracy so because we were [31:31] able to test it with you know the large [31:33] amount of um like membership that we [31:35] support and amount of data that we have. [31:37] >> Great. That's very unique advantage that [31:40] you have your own data and you train [31:41] your own models and build on top of [31:43] that. Um I have a question for you Dan. [31:47] Um leading AI globally at Sigma. Um what [31:51] advant this is a unique vantage point [31:54] and I was just wondering how do adoption [31:57] strategies differ between US and [31:59] international markets? [32:03] >> From my perspective they don't. Um, I'm [32:05] always looking for the best tool for the [32:07] best product and I don't really care how [32:09] you get there as long as it gets done. [32:11] So many projects get caught up with, [32:13] well, we should look at this or talk to [32:15] legal about this and they never actually [32:16] get done. My philosophy is make it work [32:20] and then figure out how to make it cheap [32:21] and compliant and all these things. Um, [32:24] so from my perspective, we develop the [32:27] projects and then they filter down [32:29] through legal and HR and end up getting [32:31] implemented in some fashion. But um [32:34] that's my philosophy. [32:35] >> So that's great. So legal and HR are not [32:38] involved before the project. [32:40] >> Well, no, they don't kill it in the [32:42] >> you bring them in at the appropriate [32:44] time. [32:44] >> Okay. All right. That's amazing cuz you [32:47] know it's in bigger organization [32:49] >> HR and legal are the first to vet out a [32:53] vendor and before you even get to see [32:55] their real works, it's it's already [32:57] gone. They're not compliant. So I think [32:59] uh that that's an interesting approach. [33:01] I really like it. All right, Gi, coming [33:05] back to you. Um, in terms of VC funding, [33:09] um, where are the VC [33:12] putting their bets in AI, you know, in [33:14] terms of AI funding flows or is it life [33:18] sciences startups or in world? Where's [33:20] the venture money flowing and what does [33:23] the signal for disruption ahead? What is [33:25] the signal for disruption? [33:26] >> Any stock tips? [33:29] Um, [33:31] so I'd say from the VC side, right, just [33:34] going referring back to the McKenzie [33:37] numbers that you talked about, we're [33:39] excited about all pieces um, of the AI [33:43] stack, administrative um, workflows of [33:47] course, right? [33:48] um doing better analytics, internal [33:51] productivity, and I think the area that [33:55] is coming next um and that we're really [33:58] excited about is AI diagnostics. Um and [34:02] this is not exactly a payer example, but [34:05] we're investors in a company called [34:07] Basian Health. It actually came out of [34:09] John's Hopkins. It's a professor there. [34:12] And what they're doing is they're [34:15] building a platform that can help build [34:19] AI models, validate them, get them FDA [34:24] approved, and then deploy them and [34:26] monitor them in production. And once [34:28] again, note there is a doctor in the [34:30] loop also monitoring it, giving [34:32] feedback, making sure these models get [34:35] better and better over time. But they've [34:38] built what is the chassis to create AI [34:41] diagnostic models in um a really [34:44] responsible way. And their first model, [34:47] it's early sepsis detection. It's live [34:50] at a number of hospitals including the [34:53] Cleveland Clinic where they're seeing [34:55] 27% [34:57] decrease in mortality. So it's saving [35:00] lives and 10% decrease in utilization. [35:04] So also saving money. And I think we're [35:06] going to see this across use cases. And [35:09] then what they're doing, you know, all [35:10] of the rage, agentic AI, I'm sure you've [35:13] heard a lot about it already, and I [35:14] don't need to tell you, but they're [35:16] using their AI agents then to help [35:19] physicians documentation, [35:22] put in orders, and reduce the workflows [35:25] around it to really just help reduce [35:28] clinician burnout. So I think we're [35:31] going to see these like multiaceted [35:33] systems really improving care, improving [35:37] outcomes and helping reduce burnout and [35:39] cost. That's great. And I think that [35:42] brings about the question of private [35:45] public partnership and I think uh one of [35:48] the organization that's working in that [35:49] space is coalition of healthcare AI. You [35:52] probably must have heard about it. They [35:54] are working to develop some standards in [35:56] terms of these health care models and to [35:59] say okay if the healthcare model is [36:01] going to do sepsis detection what are [36:04] the ways that it should have been [36:05] developed what kind of data should be [36:07] develop developed on or trained on what [36:10] are the consideration from [36:12] pre-development to development to [36:14] post-development stage and what are the [36:16] benchmarks that we're going to use to [36:18] measure that this model is performing [36:20] well or not. So look it up. Coalition of [36:22] healthcare AI is doing quite a bit of [36:24] work in that area and I do believe there [36:26] might be some collaboration going on at [36:28] the backstage there as well. Looking [36:31] forward what are the big trends in [36:33] artificial intelligence and pair area. [36:37] Um are there any uh I know there's a lot [36:41] of talk of agentic AI. Bos give us a [36:43] little bit of insight into you talk [36:45] about agentic. Is that a newest trend [36:47] there in your company or do you think [36:49] there's more exciting stuff going on [36:51] other than that? [36:53] >> Definitely agentic AI is the new trend. [36:55] It is um that's where you know we are [36:57] investing heavily and I talked about our [37:00] platform approach in building as we are [37:02] going to build so many AI agents or [37:05] applying in agentic practices. So we [37:07] need to build them in a consistent way [37:09] or standard applying the standardization [37:11] governance um also the responsible AI [37:14] practices security all of that we are [37:16] you know the foundational areas that we [37:18] are um going to invest in um is going to [37:21] definitely going to be the future for us [37:23] in building you know so many of the [37:25] agents uh mainly like you know multi- [37:27] aent systems and then agents need to [37:29] collaborate each other we need to [37:30] discover you know what an agent can do [37:33] so we also have this um agent [37:36] marketplace where the agents will be [37:37] like published and you know others can [37:39] come go in and see like what agents are [37:41] available the specialist agents what are [37:43] the skill sets and how can I use it um [37:46] all of that like we are investing right [37:48] now the focus is building these [37:50] foundational components and also [37:52] piloting many of the lowhanging fruits [37:55] and that from that we going to learn and [37:57] I think next 12 months I think we will [37:59] be at a place that we should be able to [38:01] really you know like um improve the [38:04] operational efficiencies as we are [38:05] planning with and also for the member [38:07] experience and providing better [38:09] connected experience for the member and [38:11] also like transparency. Transparency is [38:13] another big area that you know member [38:15] needs to know or even like our clients [38:16] need to know when it when a drug get [38:18] priced somewhere you know what's what's [38:20] the how why is it why is it priced that [38:23] way why are they paying that much things [38:24] like that right the transparency is [38:26] another big thing uh that we are like [38:28] you know with this um assistants or [38:31] companions that we are building as a as [38:33] AI assistants or a agents is going to [38:35] help our members clients and all our [38:38] constituents [38:38] >> to improve the transparency in the [38:40] pricing okay great that's amazing [38:42] Then do you have any insight into the [38:45] future of AI agents or anything [38:47] different? [38:47] >> Yeah, a little bit. I mean that answer [38:49] is very comprehensive for the immediate [38:50] term future, but the the larger future [38:53] and the biggest impact is very hard to [38:56] predict. You you want the second and [38:58] third order effects. Think about like [39:00] the internet. There's this video of Bill [39:02] Gates from the 80s talking about the [39:03] internet and he's like look you can play [39:06] chess with your friends online with the [39:07] internet which no one does. What people [39:10] do is they use it to buy burrito taxis [39:12] and order food to their house. That was [39:13] a huge change. The the internet [39:16] destroyed the taxi industry and created [39:18] a a cheap logistical chain that you can [39:20] use for any purpose. That's what AI is [39:23] going to do. It's going to perform some [39:25] kind of function. It will consume the [39:28] value of other businesses and enable [39:30] things that we cannot do now. That's [39:32] what the future's going to be. So try to [39:34] try to keep your eye on that. Try to [39:36] imagine that. And um it's kind of what I [39:38] think about a lot. [39:40] >> Sounds good. I think you talk a little [39:42] bit about disruptive innovation and if [39:45] you look about disruptive innovation, [39:46] artificial intelligence itself has not [39:48] caused disruptive innovation. Having [39:50] said that, we're still looking for an [39:52] economic model which is going to be [39:54] wrapped around it and then we'll see if [39:57] it's going to has this disruptive [39:59] effect. So any last words, Kahini? I'd [40:03] say purely from a technology trend of [40:06] course we're looking at agents and then [40:08] we're also looking at how do these [40:11] models continuously update continuously [40:15] self-improve right so if we think about [40:18] systems right now open AI whoever you [40:21] train a really large model but then [40:23] that's static maybe you fine-tune it on [40:26] your data but again that's static [40:29] rag is is a solution here to you know [40:32] incorporate more real data. But as we [40:36] move to this agentic paradigm, how do we [40:39] make sure that the system like I said is [40:41] continuously learning from maybe human [40:43] feedback, pumps up, pumps down or [40:45] feedback signals it's getting from the [40:47] system? Um so I think this is kind of [40:50] the new area of like thinking and [40:53] research that folks are looking into [40:55] continual feedback and improvement in [40:57] the systems. I think I've been given [41:00] enough signals that we are up against [41:02] our time. Uh, do we have time for [41:04] questions or no?