Bland CEO Warns Enterprise AI Wrappers Face Ruin as Subsidies End

Bland CEO警告：随着补贴终结，企业级AI套壳产品面临灭顶之灾 联络中心自动化唯有在供应商构建专有基础设施并清理混乱遗留架构的前提下才能生存 在Andy Tang的主持下，该圆桌讨论对比了AI创始人（Bland、Maven AGI）与企业客户体验（CX）采购方（The Cigna Group、FCP Euro）关于联络中心自动化的现实经济学与架构层面的不同视角。与会嘉宾探讨了为何轻量级AI套壳产品面临不可持续的单位经济效益，并指出真正的企业价值在于自动化关键任务及高风险对话，而非构建僵化的分流决策树。企业管理者透露，混乱的遗留基础设施与缺失的元数据仍是规模化落地的主要瓶颈，并得出结论：成功需要建立注重纪律性与结果导向的合作关系，而非单纯为了软件而做软件。 依赖第三方前沿LLM API的企业级AI套壳产品在风投资助结束后面临严峻的单位经济风险，这将使拥有自研专有模型的提供商更具优势。 (4:39) 客户体验（CX）与开发者代码Copilot是目前能为企业带来立竿见影且可衡量生产力ROI的两大主要领域。 (4:39) 真正的业务转型需要将AI与整体企业架构保持一致，并修复底层存在缺陷的流程，而非推行孤立的“为AI而AI”项目。 (8:54) 企业应严格执行成本纪律，首先审计重复性任务，并在轻量级内部自研方案或人工审核能提供更优ROI时拒绝供应商的高昂报价。 (8:54) 消费者对更快、更具共情力服务的普遍需求以及严峻的运营瓶颈，正在推动各垂直领域加速采用CX AI，其速度超越了传统的技术采纳曲线。 (16:16)

Enterprise customer experience adoption is moving faster than prior technology cycles, but industry leaders argue that long-term viability requires moving beyond venture-subsidized wrappers and superficial deflection toward sound unit economics, unified data infrastructure, and autonomous execution of high-stakes workflows. Unit Economics and the Vendor Subsidy Trap (4:39) Bland CEO Isaiah Granet cautions that the current generative AI cycle exhibits the hallmarks of past tech bubbles, warning that startups commanding valuations larger than Walmart while operating on fragile margins face eventual collapse. Granet compares third-party frontier API wrappers to early venture-funded ride-sharing subsidies: once introductory API pricing ends, companies lacking proprietary models and disciplined unit economics will fail. Bland opted to build its models in-house to protect gross margins, despite early skepticism from 198 venture capital investors who initially rejected the company. Maven AGI CEO Jonathan Corbin presents a complementary outlook, conceding that while private tech valuations may be inflated, enterprise productivity gains in customer experience and software engineering are already delivering immediate, measurable return on investment. Corbin notes that generative AI adoption is maturing significantly faster than previous five-to-ten-year enterprise technology transitions. Enterprise Restraint Against AI for AI's Sake (8:54) Enterprise CX executives emphasize that deploying software tools without addressing fundamentally flawed operational processes creates zero strategic advantage. Kirby Forlin of The Cigna Group points out that while individual productivity tools like Claude Opus can draft slide decks in seconds, slide generation does not constitute fundamental enterprise transformation. Real business impact requires strict enterprise architectural discipline and strategic alignment between business leaders and engineering teams. Demonstrating this cost discipline, Alejandro Villacis of FCP Euro outlines a pragmatic adoption framework that audits routine support tasks and assesses internal data before looking at external software. Villacis notes that FCP Euro rejected vendor call-analysis platforms with $100,000 baseline fees because paying more for experimental call analysis than for their entire AWS web hosting infrastructure or warehouse management system made no financial sense when human review remained cheaper. Automating High-Stakes Interactions Over Basic Deflection (18:52) Rather than relying on legacy interactive voice response (IVR) phone trees—which Granet actively bans on Bland's platform—the primary enterprise opportunity lies in executing complex, mission-critical calls end-to-end. Granet reports that over 70% of Bland's callers do not know or care that they are interacting with AI, as the platform guides 90-year-olds through 45-minute remote medical checkups and conducts volatile overseas polling. Granet envisions voice AI eventually powering 911 dispatch calls, arguing that the deepest business value comes from resolving regulated, high-risk conversations where failures risk lawsuits or death. Corbin reinforces that consumer demand for immediate, frictionless service pairs with severe internal enterprise bottlenecks. Citing his experience at HubSpot—which handled hundreds of millions of interactions across 220,000 organizations—Corbin argues that scaling support through human labor alone is impossible. As platforms secure HIPAA compliance and FedRAMP authorizations, automated execution becomes essential across both consumer touchpoints and heavily regulated sectors. Resolving Data Bottlenecks and Vendor Scope Creep (36:09) Deploying agentic systems remains severely constrained by fragmented data architecture and missing metadata. Forlin explains that healthcare's slower adoption is primarily caused by non-standardized electronic medical record systems and fragmented insurance databases rather than regulation. While Maven AGI applies direct connectors and synthetic metadata via a live 'graph of record,' and Cigna uses AI to map data warehouses, Forlin stresses that human-in-the-loop validation remains mandatory to verify data context before models can consume clean inputs. Because enterprise clients frequently grapple with data deficits, Granet contends that vendors must emulate Palantir by delivering consultative outcomes rather than hands-off software licenses. He cautions that vendors must actively combat enterprise 'AI psychosis,' a tendency where corporate leaders prematurely expand into mobile and web channels before proving stability. Vendors must enforce operational focus, refusing secondary expansions until core contact center workflows achieve validated outcomes. While the panel demonstrates that contact centers offer immediate, measurable ROI for generative AI, long-term industry success remains contingent on unresolved structural issues: bridging broken legacy metadata architectures, surviving the expiration of subsidized compute pricing, and maintaining strict execution discipline rather than chasing disjointed multi-channel experiments.

企业联络中心正从外围的呼叫分流工具演进为生成式人工智能落地的核心阵地，但产业内部在技术架构与商业可行性上出现深度分歧。与会专家指出，真正创造价值的AI部署必须摆脱脆弱的API包装模式与工具堆砌，直面遗留数据碎片化与高风险交互场景，通过自研核心能力与交付确定性业务结果实现突围。 估值泡沫、单位经济学与自研模型防线 (4:39) 早期风险投资人Andy Tang指出，Bland创始人Isaiah Granet在2023年从Y Combinator毕业后曾遭到198家风险投资机构拒绝，当时资方普遍误判语音交互将被纯文字取代。针对当前生成式AI的投资狂潮，Granet明确表示当前周期并未脱离历史科技泡沫规律，部分估值超越沃尔玛但单位经济模型脆弱的企业无法长期维系。他将依赖第三方前沿大模型API的应用比作早期的网约车模式，警告一旦资本补贴期结束、API调用成本上升，缺乏底层毛利支撑与架构壁垒的中间层企业将面临崩溃，这也是Bland选择自研全套语音模型与电话技术栈的原因。 Maven AGI首席执行官Jonathan Corbin提出了互补视角，他承认非公开市场的局部估值存在虚高，但强调企业级宏观生产力提升确凿存在。以往的技术范式转移通常需要5至10年才能在大型企业中完全成熟，而生成式AI在客户体验（CX）与开发者编程辅助两大场景中，展现出了远超以往周期的渗透速度与投资回报兑现能力。 高风险交互与复杂场景的自动化替代 (18:52) 传统观念认为受强监管制约的行业会严重拖慢AI采纳速度，但Corbin指出各行业普遍存在的恶劣客户服务痛点正在打破这一惯性。以HubSpot为例，22万家客户企业产生了数亿次支持请求，庞大的业务规模已无法单纯依赖人工坐席覆盖，这种现实瓶颈倒逼企业迅速推进HIPAA合规与FedRAMP授权，将AI部署至关键通路。Granet进一步驳斥了AI仅应处理低风险任务的常规假设，指出传统联络中心耗费三十年研发的交互式语音应答（IVR）电话树彻底失败，真正具备深度商业价值的恰恰是高风险、高复杂度的端到端全流程接管。 Bland披露其实际落地的AI代理已能够为90岁以上的老人执行长达45分钟的远程患者监测通话，自主指导佩戴血压计、诊断读数异常并排查突发紧急状况，甚至在海外动荡区域执行极高风险的民意调查。针对用户心理，Granet表示Bland对僵化的电话树实行零容忍策略，通过中西部友好口音等自适应声线化解用户抵触情绪，其实测数据显示超过70%的呼叫者并未意识到或不在意对方是AI；此外，针对年轻群体因电话社交焦虑而放弃预约看病的痛点，无摩擦的AI交互正显著改善医疗服务的可及性。 遗留系统泥潭与元数据治理的现实瓶颈 (26:07) 大型企业推行AI代理的主要阻碍并非来自算法瓶颈，而是非数字化原生实体内部严重碎片化的遗留数据基础设施。The Cigna Group数字化高级副总裁Kirby Forlin分析称，虽然30秒内即可解决的简单诉求（如核磁共振预约）理应完全数字化，但实际执行需要实时打通复杂的上下文环境，包括调和非标准化的Epic等不同电子病历系统（EMR）、核对保险细则、查询网络内定点机构及动态计算自费费用。FCP Euro副总裁Alejandro Villacis亦表示汽车电商领域面临同样的配件匹配难题，由于新车型与零部件层出不穷，坐席人员不得不持续在割裂的数据库与外部目录之间手动核对车辆识别码（VIN）。 元数据的缺失构成了企业部署自主Agent的核心断层。Forlin指出没有元数据就无法理解数据资产本身，Cigna团队内部虽利用AI自动化开展数据仓库映射与文档记录，但仍须保留主观领域专家（HITL）进行跨系统数据校验。Corbin介绍Maven AGI通过动态构建“记录图谱”（Graph of Record）并生成合成元数据来规避传统ETL管道的繁重复制，从而在通话中实现毫秒级的跨系统权限检索与个性化执行。 抵御AI精神病：以结果交付规避盲目扩张 (37:51) 在企业采购与商业交付模式上，Granet提出大型企业购买的是确定的业务结果而非单纯的软件系统。面对数据底座缺失的客户，标准化SaaS已无法奏效，厂商必须转向类似Palantir的高接触度咨询工程交付，派驻团队直接协助客户补齐元数据短板。他特别指出大企业内部经常受人事更迭和政治转向干扰，许多企业在OpenAI发布新版本后盲目中止外部合作转入内部自研，最终耗费巨额成本却未能达到业务预期。 针对技术管理层在各类新技术之间疲于奔命的“AI精神病”（AI psychosis）现象，Granet强调厂商必须发挥战略锚定作用，严格遏制范围蔓延（Scope Creep），在联络中心单点业务指标彻底跑通前，坚决拒绝向网页或移动端二次触点盲目扩散。Villacis从汽车后市场零售视角验证了这一成本纪律，FCP Euro直接拒绝了基础年费高达10万美元且附带额外交互抽成的未经验证分析工具，强调当AI实验性成本超越企业底层的仓储管理系统（WMS）或AWS云基建时，盲目采购在商业逻辑上完全不成立。 该讨论表明联络中心正从辅助人类坐席向全自主工作流演进，但其技术变现严重受制于底层数据治理成熟度与严苛的采购ROI考核。尽管四位嘉宾一致看好开源模型生态，但在激进的全自动化远景（如零人工兜底的911紧急调度）与当下监管滞后、数据断层的现实冲突之间，如何明确法律责任归属并平衡高昂的定制交付成本，仍是未决的关键课题。

## Background 本内容由 Ai4 于 2026 年 8 月 22 日发布在 YouTube 平台上，题为《The Augmented Contact Center》。视频围绕 AI 如何重塑现代联络中心展开，邀请了来自 Maven AGI、Bland、FCP Euro、Cigna 及 Draper Associates 的创始人与高管，共同探讨 AI Agent 与 Copilot 在提升坐席效率、打通跨渠道上下文以及在复杂受监管场景中实现端到端自动化落地的最新策略。 ## TL;DR 圆桌嘉宾围绕企业联络中心的 AI 变革展开了多维度的深度拆解。创业者与大企业高管首先审视了当前的 AI 经济周期与估值泡沫，强调必须依靠自研基础设施和健康的单位经济效益来穿透市场周期，同时警惕非数字化企业“为了做 AI 而做 AI”的表面自动化陷阱。随后，讨论深入探讨了 AI 在医疗健康、高难度电话（如 911 报警、远程患者监测）及汽车配件等高壁垒、强监管场景中的自主化落地实践，直面数据孤岛与元数据缺失等底层瓶颈。最后，嘉宾们分析了企业在构建 AI 智能体时自研与外采的博弈，主张以“交付业务结果”为导向单点打透，并对开源模型、监管时机等行业关键议题进行了鲜明表态。 ## Atmosphere & style 整场座谈会节奏明快、观点犀利，充满了从业者一线实战与风投敏锐洞察的碰撞火花。主持人的幽默串场与嘉宾们毫不避讳的直言不讳形成了绝妙平衡——从拆穿大模型 API 补贴背后的虚假繁荣，到调侃 CTO 们的“AI 精神病”，现场互动既充满商业务实感又频频引发现场笑声。在严肃讨论高危医疗呼叫与复杂数据孤岛时，对话展现出极高的专业严谨度；而在终局快问快答阶段，气氛则迅速切换至轻快诙谐，给观众带来兼具干货深度与沉浸观赏性的精彩体验。 ## Key takeaways - [04:39](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=279s) 当前 AI 行业存在依赖补贴的估值泡沫，企业必须注重底层单位经济效益（Unit Economics）；目前真正产生明确规模化落地 ROI 的核心赛道仅有客户体验（CX）与代码生成（Coding）。 - [08:54](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=534s) 企业应警惕“为 AI 而 AI”的表面自动化，单纯使用 Copilot 进行个人微小提效不等于业务模式重塑，切忌用 AI 工具去固化本就支离破碎的旧业务流程。 - [12:15](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=735s) 评估 AI 采买方案需严格核算总体拥有成本（TCO），如果 AI 分析工具的费用逼近核心仓储系统（WMS）或全站云基建（AWS），则在经济模型上缺乏合理性，应采取小步快跑的务实策略。 - [16:10](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=970s) AI 采纳并未遵循“受监管行业最后落地”的传统互联网规律，普遍恶劣的客户体验痛点正在倒逼医疗等强监管领域加速合规化（如通过 HIPAA/FedRAMP 认证）并引入 AI。 - [20:02](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=1202s) AI 最具商业价值与技术护城河的场景恰恰在高度受监管与高危领域（如高龄患者监护、敏感民调与 911 报警），其终极目标是实现完全无人工干预的端到端自主接管。 - [22:07](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=1327s) 2025 年将是企业工作流的分水岭：未来原生型企业将不再局限于用 AI 辅助人类的既有流程，而是直接让 AI 全权主导并端到端执行业务流程本身。 - [23:35](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=1415s) 联络中心应彻底摒弃低效的传统电话按键菜单（IVR），通过拟真自然的语音语调与用户建立信任，并将“单次发言字数（Utterances）”作为衡量用户是否建立信任的核心指标。 - [26:07](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=1567s) 医疗等传统行业 AI 落地的最大瓶颈并非监管本身，而是非数字化原生的数据基础设施，以及各大电子病历（EMR）系统之间严重的孤岛割裂与缺乏互操作性。 - [30:00](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=1800s) 联络中心存在的本质是企业当前技术交付能力与用户理想体验之间的“代偿机制”，用于人工弥补如高精度汽车配件适配等无法完全产品化自助的复杂业务缺口。 - [32:47](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=1967s) 企业自研 AI Agent 往往低估了多轮上下文调度与动态执行的工程复杂度，浅层知识库问答与真正代表客户跨系统执行后台操作之间存在巨大鸿沟。 - [35:01](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=2101s) 通过“记录图谱”（Graph of Record）直接关联异构数据源可免去漫长繁琐的 ETL 清洗，并可通过 AI 辅助标注与人类在环核验解决企业元数据缺失的问题。 - [37:51](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=2271s) 大型企业级 AI 采购的本质是“购买业务结果而非软件代码”（Palantir 模式），盲目跟风自研底层大模型接口往往会付出数倍于采购的成本却难以交付可用系统。 - [39:57](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=2397s) 必须抵御技术决策者的“AI 精神病”（在各类新技术概念间盲目切换），保持战略纪律，在呼叫中心等核心单点场景彻底打透并跑通正向 ROI 后，再向 Web/App 等全渠道拓展。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=4s) 开场与嘉宾背景：从客服基层经验到 AI 创业前沿 主持人 Andy Tang 幽默地指出本次座谈会精心安排的座次格局：两位企业客户代表与两位 AI 创业专家并排而坐，台上的创业者们显然都在向客户展示能力。谈到为何主动选择主持这场关于客户体验（CX）的圆桌时，Andy 分享了自己的投资与人生感悟——他曾建议暑期实习的六名分析师全去从事客户服务或餐厅服务员工作。他认为服务基层是极为关键的技能锻炼，而非单纯的兼职糊口；而如今，这项古老的人际沟通技能正与当下最火热的 AI 深度结合。 随后四位嘉宾分别进行了自我介绍： - **Isaiah Granet**（Bland 联合创始人兼 CEO）：Bland 专注于将“极度复杂、极度棘手”的电话场景进行端到端自动化。Isaiah 形象地将其描述为“如果出错就会导致诉讼或闹出人命”的高难度通话。Bland 坚持全栈自研模型，刚刚发布了评测表现逼真的文本转语音（TTS）模型。公司迄今融资超 1 亿美元，服务 250 家企业客户及 35 万自助开发者用户。 - **Jonathan Corbin**（Maven AGI 联合创始人兼 CEO）：Maven AGI 致力于打造用于客户体验的 AI Agent，目前已服务包括 Dell、Hinge、Hard Rock 等在内的约 100 家知名客户。团队规模约 100 人，累计融资约 7800 万美元。 - **Kirby Forlin**（The Cigna Group 数字与分析高级副总裁）：从传统大型企业的数字化视角出发，Cigna 今年的核心工作是实现跨应用、跨触点客户上下文（Context）的无缝打通，确保客户在 App 操作后转接人工客服时，所有数据与信息完全对齐一致。 - **Alejandro Villacis**（FCP Euro 客户体验副总裁）：FCP Euro 是一家拥有 40 年历史（线上运营 25 年）的欧洲汽车零部件家族企业，秉持“自己不装车就不卖给客户”的高品质承诺。Alejandro 过去一年重点在联络中心落地 AI，旨在替座席分担机械重复的琐事，使员工能专注于创造营收与提升客户留存的高价值对话。 ```mermaid flowchart LR accTitle: Contact Center Context and AI Collaboration accDescription: Flow showing how AI helps customer agents by integrating context from multiple digital touchpoints into high-value interactions. subgraph UserJourney ["用户旅程与触点"] A["数字化应用操作 (App/Web)"] --> B["遇到复杂问题转接呼叫"] end subgraph AIPlatform ["AI 联络中心平台"] B --> C["上下文数据打通与意图识别"] C --> D["端到端语音/智能体自动化"] C --> E["座席辅助 (Copilot) 消除机械劳动"] end subgraph BusinessValue ["业务价值输出"] D --> F["高难度通话即时解决"] E --> G["高价值沟通：销售转化与客户留存"] end ``` --- ### [04:39](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=279s) 创业者审视 AI 经济周期：泡沫预警与底层商业模式 作为拥有 20 年经验的早期风投，Andy 回顾了多次技术周期。他提到 Isaiah 创业初期在硅谷遭遇 198 次投资人拒绝的轶事——在 2023 年，许多投资人曾笃定地断言“一年后没人会再打电话，语音交互已死”。Andy 同样曾在 2015 年种子轮投资过 Otter.ai，当时也有人觉得录制全球音频对话“既诡异又不切实际”。针对当前的 AI 周期，Andy 向两位创始人抛出核心问题：这次真的有所不同吗？还是处于另一场泡沫中？ ```mermaid flowchart TD accTitle: AI Economic Cycle and Risk Comparison accDescription: Comparing the risk of relying on subsidized frontier models versus building fundamental self-hosted capabilities. subgraph SubsidizedModel ["依赖补贴的 API 包装模式"] A1["高昂的前沿模型 API 成本"] --> B1["补贴期结束 / 算力提价"] B1 --> C1["毛利急剧恶化，陷入经营危机"] end subgraph FundamentalsModel ["注重商业本质的自研模式"] A2["自研模型与基础设施 (如 Bland)"] --> B2["可控的单位经济效益 (Unit Economics)"] B2 --> C2["穿透泡沫周期，具备长期健康度"] end ``` 两位创始人对当前的经济周期给出了深刻而冷静的剖析： - **Isaiah Granet** 明确指出当前大概率存在巨大泡沫：“当某些公司的估值超过沃尔玛，但单位经济效益（Unit Economics）却与沃尔玛相差无几时，你就必须质疑其底层商业逻辑是否健全。”他将现状类比于早期 Uber 等公司的“补贴期”——一旦前沿大模型背后的资金补贴结束，缺乏核心壁垒和健康毛利的公司将难以为继。这正是 Bland 选择不依赖前沿大模型 API、坚持自研模型的原因。他提醒企业客户，必须警惕那些挥金如土、不顾财务健康的软件供应商的长期生存能力。 - **Jonathan Corbin** 则从宏观与微观两个维度进行拆解：从宏观来看，AI 必然会重塑各行各业；但从微观实际落地来看，目前真正产生巨大实效（ROI）的领域只有两个——**客户体验（CX）**与**代码编写（Coding）**。尽管部分初创公司估值虚高，但 AI 技术的渗透和演化速度远超以往任何一次 5 到 10 年的技术成熟周期，其对企业运营模式的深远影响是毋庸置疑的。 --- ### [08:54](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=534s) 传统企业落地的现实痛点：警惕“为 AI 而 AI”的表面自动化 面对 AI 热潮，来自大型医疗健康巨头 Cigna 的 Kirby Forlin 分享了传统非数字化原生企业的真实采买与部署现状。企业内部虽然充斥着对 AI 的投资热情，但普遍面临“为了做 AI 而做 AI”的误区，往往缺乏结合业务本质的重塑思考。 Kirby 坦言，目前很多企业在落地 AI 时陷入了碎片化与重复采购的怪圈： - **效率工具的局限性**：虽然引入 Copilot 或大模型（如使用 Claude Opus 生成演示文稿）确实极大节省了员工制作 PPT 的时间，从而释放了部分战略思考精力，但这仅仅是个人层面的微小效率提升，并不等于“业务模式的重塑”（Business Reinvention）。 - **固化破损流程的伪自动化**：许多所谓的流程自动化，只是在用 AI 去自动化一个原本就存在缺陷或支离破碎的旧流程，不仅没有解决核心业务痛点，反而采购了一堆功能重叠的工具。 - **战略思考的错位**：真正的 AI 落地绝非“给全员分发一套 AI 工具”即可交差，企业必须从全局视角出发，重新构想在生成式 AI 时代下业务生态与用户体验该如何协同运作。 ### [11:16](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=676s) 战略先行与避免盲目投入：技术与业务视角的深度融合 企业在迈向 AI 驱动的转型过程中，最忌讳的是本末倒置。当前许多企业往往陷入了技术驱动的误区——盲目跟风引入各种 AI 工具，却缺乏清晰的顶层业务战略。真正的正确路径应当是“先明确业务战略，再决定需要匹配何种 AI 解决方案”。由于许多传统业务人员对前沿技术的实际边界并不熟悉，很难具象化地构想出转型的可能性，因此技术团队与业务团队必须从战略高度紧密携手。 宏观层面的技术巨变固然已经发生，但真正的全面落地仍需时日。在此过程中，必然会有许多企业在错误的方向上过度投资，甚至遭遇大量试错与挫折。对于非数字化原生（Non-digitally native）的企业而言，要想在这一轮浪潮中胜出，核心关键在于引入那些既懂技术又能彻底重新构想业务底层运作模式的高端人才。 --- ### [12:15](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=735s) FCP Euro 的务实落地：从代码提效到电话转录的成本账 在汽车配件电商平台 FCP Euro，AI 投资采取了极其审慎且务实的策略。技术与软件工程团队率先尝试使用大语言模型进行辅助编程，提升代码交付效率并减少重复劳动。客户体验（CX）团队则在大约三到六个月后跟进，但他们没有盲目接触市面上那些千篇一律、同质化严重的 AI 软件供应商，而是先从自身的实际业务痛点出发制定落地战略。 ```mermaid flowchart TD accTitle: FCP Euro 务实 AI 落地与成本评估流程 subgraph 痛点识别与优先级排序 A["梳理客服重复性工作"] --> B["分析 18 个月积累的通话转录数据"] end subgraph 方案选型与成本核算 B --> C{"优先复用现有工具<br/>还是外部采购？"} C -->|自研/轻量方案| D["小步快跑 (Scrappy)<br/>成本可控且灵活"] C -->|高价供应商| E["评估总体拥有成本 (TCO)"] end subgraph 决策底线 E --> F{"费用是否超过<br/>WMS 或 AWS 基础成本？"} F -->|是| G["果断否决：缺乏差异化价值"] F -->|否| H["谨慎试点推进"] end ``` 在具体的客服场景中，FCP Euro 积累了长达 18 个月的电话录音与转录文本，但早期缺乏足够的人手去分析这些数据并提炼洞察。在评估外部供应商时，部分厂商报出了“每年 10 万美元基础底费 + 额外按调用量计费”的天价方案。对于电商业务而言，这种未经充分验证的解决方案其总体拥有成本（TCO）甚至高过了核心仓储管理系统（WMS），几乎快赶上托管整个网站基础设施的 AWS 云服务账单。面对这种“卖梦想”的浮夸销售，FCP Euro 果断选择拒绝，转而依托清晰的优先级，从小规模、低成本（Scrappy）的务实切入点起步。 - **自研与采买平衡**：优先盘点企业现存软件与工具资产，严禁购买功能重叠的冗余系统。 - **价值与成本对齐**：若 AI 分析工具的成本逼近核心业务系统（如 WMS）或全站云基建（AWS），且无法证明超额 ROI，则在经济模型上完全不合理。 --- ### [16:10](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=970s) 监管周期 vs 痛点驱动：AI 采纳路径是否遵循互联网规律？ 投资人提出了一种行业采纳周期的假设：回顾互联网的普及历史，往往是从通信、零售、企业级软件起步，再逐步渗透到工业制造、金融服务，最后才是受严格监管的教育与政府部门——监管越严的行业，采纳周期越慢。那么，AI 的落地是否也将遵循这一规律？ 对此，现场嘉宾给出了强烈的反驳。当今无论在哪个行业，极其糟糕的客户体验已经成为普遍现象——从入住酒店时的繁琐排队，到打篮球扭伤膝盖后联络保险公司与医生却被告知“挂号排到一年后”。庞大的用户痛点正在倒逼所有行业做出改变，消费者甚至会用脚投票：“如果不改进体验，我就把钱花到别处去。” 以 HubSpot 为例，其当年服务 22 万家企业客户，面对数以亿计的用户互动需求，仅靠堆砌人工客服根本无法维系。AI 带来的“更少人力支撑更大规模互动”的经济效益，构成了颠覆性的商业闭环。更重要的是，现代企业早已掌握了合规推进的标准打法，例如 Maven AGI 从创立之初便迅速通过了 HIPAA 医疗隐私认证，并积极推进美国联邦政府的 FedRAMP 认证，合规资质已不再是阻碍 AI 进入受监管行业的绊脚石。 --- ### [20:02](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=1202s) 攻坚高难度与高危场景：Bland 在极端复杂与受监管领域的突破 针对受监管行业采纳滞后的传统偏见，语音 AI 平台 Bland 提出了完全相反的实践观察：**AI 最具商业价值、护城河最深的领域，恰恰存在于那些高度受监管、任务极其繁重的复杂通话场景中。** | 典型应用场景 | 传统人工/旧系统痛点 | Bland 智能语音 AI 的实现能力 | |:--- |:--- |:--- | | **高龄患者远程监测 (RPM)** | 90岁以上老人操作困难，极易引发医患沟通脱节与误诊。 | 进行长达 45 分钟的深度通话，指导老人佩戴血压仪、确认读数，全程拉家常（如聊孙辈）保持参与感，并在读数异常时精准识别医疗紧急事件并触发送医。 | | **地缘政治敏感地区民意调研** | 数据极易被当地敌对政权截获，调研员与受访者面临人身搜查风险。 | 自动化执行高安全性民调，安全归集高敏感数据，避免人工接触带来的泄密与人身危险。 | | **复杂应急响应 (如 911 报警)** | 传统 IVR 试图分流挂机，人工坐席受限于情绪与承载峰值。 | 目标实现完全无人工干预的自动化接警，在极端高压下精准提取生命关键信息并调度资源。 | 在过去三十年中，企业为了缩减呼叫中心成本，引入了层层递进的 IVR（交互式语音应答）菜单、转接流转和网页重定向，竭尽所能降低呼叫时长与呼叫量。如今仍流转到人工坐席的电话，全都是自动化规则无法处理的“硬骨头”，依赖的是人类的判断力。Bland 当前在获得 FedRAMP 资质后，直接切入这类极高难度的核心业务，其终极愿景是未来全面接管 911 紧急报警电话——构建一个完全由 AI 实时承接、无需反复呼叫“转人工”、能够独立完成高危决策的自主系统。 ### [22:07](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=1327s) 2025年分水岭：从辅助人类工作流到完全接管复杂场景 在极其紧急或复杂的呼叫场景中——比如用户在电话中情绪崩溃、大声呼救并需要被准确转接到特定消防或警察部门时，联络中心背后很可能根本没有人工坐席，因此 AI 必须具备极高的可靠性与精准度。Bland 之所以选择自研底层模型甚至内部构建全套电话通信协议栈（Telephony Stack），正是为了在不依赖外部第三方模型的前提下，能够稳定承载这些最棘手、最复杂的电话交互。 谈及企业的智能化进程，近期一位 YC 合伙人的观点引发了广泛共鸣：“企业可以划分为2025年12月之前创立的公司，和2025年12月之后创立的公司。”这两类企业有着本质区别：2025年底之后的新型企业不再试图用 AI 去“辅助”人类已有的工作流程，而是直接让 AI 全权执行工作流本身。如今许多团队虽然声称在使用 AI，但其实只是用来辅助生成 PPT；然而真正的变革在于——为什么人类还需要留在流程中去制作 PPT？同样，在财务领域，与其让人类使用 AI 辅助开具发票，不如思考如果今天创办一家新公司，是否还需要设立传统财务部门。绝大多数人目前尚无法想象这种范式转移，而技术的使命正是将这种愿景落地到最严苛、最可怕的真实通话中，让用户在电话另一端建立起绝对的信任。 ```mermaid flowchart LR accTitle: 传统辅助模式与未来原生AI工作流对比 subgraph Traditional["传统模式 (2025年12月前)"] direction TB A1["人工主导"] --> A2["AI 辅助提效"] --> A3["人工审核交付"] end subgraph Autonomous["原生模式 (2025年12月后)"] direction TB B1["业务目标输入"] --> B2["AI 端到端全自动执行"] --> B3["直接交付结果"] end ``` --- ### [23:35](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=1415s) 告别传统按键菜单：“Bland Tone”与用发言字数衡量用户信任 许多用户在拨打客服电话时有一种肌肉记忆式的本能反应：一听到机器声音就会立刻连续喊上15遍“人工客服（Operator）”，这完全源于过去被低劣交互系统反复踢皮球的糟糕体验。针对这一痛点，Bland 对传统电话按键菜单（Phone Tree / IVR）深恶痛绝，甚至立下规矩：如果客户试图用其产品搭建传统按键分流菜单，Bland 会直接与其解约，因为语音交互体验必须既优秀又安全。 为了消除机械感并建立信任，团队研发了名为“Bland Tone”的交互语调——其声音特质听起来就像一位五十多岁、极具亲和力与耐心、语气温柔的美国中西部女性。在新语音模型的支持下，超过70%的通话用户完全没有意识到或根本不在乎电话另一端是 AI。评估这种体验指标的方法非常独特——“单次发言字数”（Utterances）： - **负面/受挫状态**：如果用户仅回复两三个单字，如“不，人工”、“转人工”、“帮助”，说明用户正处于防御或不耐烦状态，体验较差。 - **积极/深度互动**：如果用户开始进行长句交流，回复字数超过两到三个词，说明用户已经完全沉浸在自然的对话中，对 AI 产生了高度信任。 --- ### [25:08](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=1508s) Z世代的电话社交焦虑与医疗交互变革 代际差异对交互渠道的偏好产生了深远影响。如今许多 Z 世代年轻人（甚至是更年轻的 Alpha 世代）几乎从不主动去医院看病，其核心原因在于他们对“打电话”有着普遍且被广泛证实的社交焦虑（Social Anxiety）。在他们的社交认知里，直接拨打电话是一种侵入式的压力，连与朋友联络前都必须先发短信确认。 如果医疗系统（例如加州大学医疗系统 UC Health）能将问诊与挂号入口做成手机通讯录里的一个 AI 联系人，允许年轻人随时通过短信或轻量语音直接沟通，挂号和就医的意愿将成倍提升。如果传统医疗行业不主动拥抱这种未来交互模式，年轻一代甚至会选择直接放弃就医。 --- ### [26:07](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=1567s) 医疗AI落地的真正瓶颈：并非监管，而是混乱的数据与EMR割裂 拥有消费品行业（CPG）与数据科学背景的行业从业者在进入医疗健康领域时，起初常抱有利用海量数据攻克绝症、解决全球问题的宏大愿景，但深入后才会发现医疗行业 AI 落地缓慢的真正症结：**并非单纯因为行业监管严格，而是因为底层数据基础设施极度混乱且非数字化原生（Non-digitally Native）**。 医疗健康服务的理想体验应当是消除一切不必要的摩擦——任何能在30秒内完成的数字化操作（如在线预约 MRI 检查）都应完全自助化，只有数字化渠道无法解决的复杂诉求才应流转至人工联络中心。然而，实现这一愿景背后需要巨大的数据基建投入。系统必须实时厘清并打通极其复杂的上下文： - 用户的具体医疗保险计划（Plan & Coverage）与自付费用计算 - 绑定的电子病历系统（EMR） - 签约的医疗提供者网络（Provider Network）及其合规转诊路径 虽然法规要求医疗机构必须配备 EMR，但并未强制统一标准，导致市场上除 Epic 等巨头外还存在大量互不兼容的病历软件，数据互操作性（Interoperability）极差。这种错综复杂的数据孤岛，才是制约医疗 AI 发展的核心障碍。 | 核心维度 | 理想数字化体验 | 医疗行业当前现实困境 | |:--- |:--- |:--- | | **就医/检查预约** | 30秒内线上自主完成（如 MRI 预约） | 必须依赖人工坐席或繁琐电话反复核对 | | **数据与系统集成** | 保险计划、自付金额、EMR 与服务网络实时打通 | EMR 品牌繁多且数据割裂，缺乏互操作性标准 | | **联络中心定位** | 仅承接高度复杂、无法在线解决的疑难个案 | 充当数据系统不完善时的“人工代偿补丁” | --- ### [30:00](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=1800s) 电商与汽配场景的启示：联络中心存在的本质是数字体验的代偿 联络中心之所以持续存在，本质上是因为企业**当前能够交付的技术能力**与**用户理想的交付体验**之间存在差距。在电商与高精度汽配（如 FCP Euro 所处的欧洲汽车零部件）领域，这一矛盾尤为典型。 汽车零配件的适配性（Fitment）要求极高，用户往往花费数百美元购买零件来修理通勤或家庭用车；一旦买错零件，车辆就面临多瘫痪数天的风险。尽管平台希望推动用户在线自助选配，但欧洲车型结构高度复杂，很多时候必须进行车辆识别代码（VIN）深度校验，而这一功能尚未能在前端网站实现完全的产品化自助。因此，目前只能依靠专家团队在后台调取并交叉比对多个数据库来人工解答。联络中心此时扮演的是一种“高附加值的人工代偿服务”。随着每年各大车厂不断推出全新车型，零部件的品质等级与适配数据呈爆炸式增长，这种数据复杂性与医疗行业面临的挑战如出一辙。 --- ### [32:08](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=1928s) 自研还是外购：早期企业构建AI Agent的维护困境 随着企业对智能化服务需求的提升，行业内关于“自研还是外购（Build vs. Buy）”的讨论日益激烈。很多企业客户在面对市场方案时，最初往往抱有“我们可以内部自己构建一套 AI Agent”的想法。然而，由于当前的 AI Agent 领域尚未完全成熟，很多自行组装了几个 Agent 的团队很快就会陷入现实困境：当构建的 Agent 数量逐渐增多时，其长期的系统运维、数据对齐与持续维护成本将迅速失控。 ### [32:47](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=1967s) 自研还是外采：企业级全流程客户体验的“数据鸿沟” 在讨论企业构建智能体时，关于“自研还是外采”（Build vs. Buy）的争论往往不可避免。对于像 Cigna 这样的大型医疗健康企业而言，数据高度碎片化地分散在整个集团的各个孤岛系统中。要想真正落地 AI，企业必须把这些数据集汇集起来，通过繁琐的 ETL 流程清洗，再喂给大模型。许多企业误以为搭建 AI 智能体非常简单，只要让模型坐在知识库上充当“复读机”，把现成的问答信息念给用户即可。 ```mermaid flowchart TD accTitle: Enterprise Context Integration Flow subgraph S1["浅层知识问答 (易于搭建)"] A["分散企业数据"] --> B["传统 ETL 流程"] B --> C["静态知识库"] C --> D["浅层 AI 智能体 (复读信息)"] end subgraph S2["全流程服务体验 (高复杂度)"] E["用户多维上下文<br/>(历史记录/订阅级别/账户余额)"] --> G["动态上下文编排引擎"] F["跨系统权限与操作接口"] --> G G --> H["全自动代客执行 (多轮交互)"] end ``` 然而，客户体验的核心在于解决实际业务痛点，而不仅仅是提供通用信息。真正的全流程服务体验要求 AI 在每一轮对话中引入极其丰富的上下文：准确辨别用户身份、追溯历史互动记录、核验客户的订阅级别与软件授权数量、查询名下账户余额，并在此基础上代表客户在后台系统直接执行操作。这就需要对庞杂的企业数据进行全量映射，在每次多轮对话交互中精准调用对应的上下文片段。这种架构极具复杂度，但每一位新入职的产品经理总会跃跃欲试地宣称“我可以自己把它造出来”，直到碰壁之后才不得不向专业厂商求助。 - **浅层问答与深度服务的本质区别**：前者仅依赖静态文档检索，后者需要动态组合用户历史、资费等级、账户余额等多维上下文并触发后台操作。 - **产品经理的“自研执念”**：企业内部团队往往低估了实时多轮对话上下文调度的工程难度，盲目自研导致资源与时间的双重浪费。 --- ### [35:01](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=2101s) Maven AGI 的“记录图谱”与元数据缺失的破解之道 现存的企业客户旅程体系根本不是为了客户而设计的，而是受限于传统 IT 架构与历史系统技术瓶颈的产物。理想的现代化客户交互应当是双向互动的（Bimodal），兼具主动预警与被动响应。为了实现这一构想，Maven AGI 开发了一套“记录图谱”（Graph of Record）。该架构无需企业将所有数据强行复制或重新经历漫长脆弱的 ETL 清洗，而是直接与企业生态系统中的各数据源建立连接。通过将搜索技术与大语言模型结合，系统能瞬间构建起描述“用户与所有企业数据资产关联”的即时知识图谱，无缝适配电话语音、在线客服以及生成式搜索等多种触点模态。 面对大型企业内部普遍存在的“缺少元数据（Metadata）”痛点——即数据散落但缺乏业务上下文描述——Maven AGI 会通过数据评估算法生成合成数据选项来进行填补。而来自 Cigna 的实战经验表明，他们正在利用 AI 工具反向推进元数据标注与数据映射工作。原本静置在数据仓库中未被充分利用的海量数据，首先由 AI 生成说明文档与元数据标签，随后引入熟悉底层系统架构的“人类在环”（Human-in-the-loop）业务专家进行准确性核验，验证通过后再接入下游模型生态。 - **记录图谱（Graph of Record）**：直连跨系统异构数据，不复制数据底座、免去全量 ETL，通过图关联实时捕捉业务上下文。 - **AI 辅助元数据治理**：在 Cigna 的实践中，先由 AI 解析数据仓库并自动补全元数据，再由业务专家进行二次校准。 --- ### [37:51](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=2271s) 买软件还是买结果？企业级交付的“Palantir 模式”与认知陷阱 对于超大型集团而言，采购技术的本质是“购买业务结果”，而不是单纯买一套软件代码。在企业规模达到一定量级时，软件交付在很大程度上呈现出咨询服务的形态。如果客户缺少元数据，AI 供应商甚至需要派遣驻场部署团队深入业务现场，与客户共同构建元数据底座。然而，企业在落地过程中面临的最大阻碍，往往是内部缺乏战略定力与政治意愿——高管频繁换届，每次人员变动就会推翻既有选型重新采购。 两年前 OpenAI 推出实时 API 等底层能力时，许多企业高管叫停外部采购，坚称“我们可以基于新模型自己造”。然而两年过去，这些企业消耗的自研研发预算远超采购成熟方案的成本，却依然无法交付可用系统。外部专业团队的核心价值在于以交付确定性的商业结果为导向，用极致的落地速度为客户创造价值。现场嘉宾打趣称，这正是经典的“Palantir 模式”——如同当年深入政府部门驻场开发软件协助抓捕本·拉登一样，靠深度下沉与极强的执行力确保结果达成。 - **购买结果而非代码（Buying Outcomes, Not Software）**：企业级交付需要深度结合业务流程定制与现场实施，纯工具交付难以产生闭环价值。 - **自研陷阱与高昂试错成本**：因底层大模型接口更新而盲目上马自研项目的企业，在周期延宕与人力消耗后往往付出数倍于采购的代价。 --- ### [39:57](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=2397s) 抵御 CTO 的“AI 精神病”：从呼叫中心单点打透到全渠道协同 当企业试图将呼叫中心的上下文数据体系扩展至 Web 端、移动 App 等全渠道时，往往会陷入注意力发散的困境。当前不少企业的 CTO 正处于一种“AI 精神病”（AI psychosis）状态——在各种眼花缭乱的新技术概念之间频繁横跳，缺乏聚焦。作为外部赋能方，AI 供应商扮演的重要角色就是为企业提供变革的合规借口，打破陈旧的内部操作守则（Playbook），迫使团队保持战略专注。 在落地策略上，必须坚持“单点彻底打透”的纪律性。必须先将呼叫中心的业务成效做到极致，在完成既定目标之前，坚决不盲目过早地将战线拉长至网页客服或聊天机器人。尽管这种克制有时会让供应商在商业竞争中暂时面临被全渠道概念产品截流的风险，但行业历史经验证明，那些缺乏自律、试图跨阶段全线铺开的企业最终往往满盘皆输。 - **AI 精神病（AI psychosis）**：形象比喻技术决策者在快速更迭的 AI 概念前频繁切换方向、缺乏战略定力的现象。 - **单点突破原则**：呼叫中心场景落地极其依赖深度的业务打磨，必须在客服场景跑通正向 ROI 后再向 Web 和 App 触点延伸。 --- ### [40:58](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=2458s) 终局快问快答：工时、开源模型与 AI 监管的即时抉择 在圆桌论坛的最后阶段，主持人 Andy Tang 组织全场嘉宾进行了高密度的快问快答，嘉宾们就行业热点、工作哲学与技术路线做出了即时表态： | 提问议题 | 选项 A | 选项 B | 嘉宾最终表态与趣味互动 | |:--- |:--- |:--- |:--- | | **工作时长** | 每周工作 5 小时 | 每周工作 35 小时 | 全员一致选择 **5 小时**；嘉宾表示“这样可以多陪家人”，现场更有人幽默反问“什么是 996？”引来全场会心大笑。 | | **职业/学术路径** | 电气工程学士 (B.S. EE) | 资深持证电工 (Master Electrician) | 形成 **2:2 对立**：Isaiah 等两位倾向理论工程学士，另两位嘉宾则坚定选择高薪且不可替代的资深电工。 | | **模型技术路线** | 开源模型 (Open Source) | 闭源专有模型 (Proprietary) | **全员一致支持开源模型**（4 票秒选）。 | | **AI 监管时机** | 立即监管 (Now) | 三年后监管 (3 Years Later) | 观点 **平分秋色 (2:2)**：半数嘉宾主张当下立刻出台规范，半数主张给予技术三年发展缓冲期。 | | **芯片出口管制** | 英伟达出口管制 | 自由市场竞争 | 前三位嘉宾果断支持**出口管制**，最后一位嘉宾幽默表示不甚了解该领域故跟随大众，论坛在轻松掌声中圆满结束。 | ## Notable quotes > [00:04](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=4s) 如果出错就会导致诉讼或闹出人命 — Isaiah Granet > [04:39](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=279s) 当某些公司的估值超过沃尔玛，但单位经济效益（Unit Economics）却与沃尔玛相差无几时，你就必须质疑其底层商业逻辑是否健全。 — Isaiah Granet > [04:39](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=279s) 目前真正产生巨大实效（ROI）的领域只有两个——客户体验（CX）与代码编写（Coding）。 — Jonathan Corbin > [20:02](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=1202s) AI 最具商业价值、护城河最深的领域，恰恰存在于那些高度受监管、任务极其繁重的复杂通话场景中。 — Isaiah Granet > [22:07](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=1327s) 企业可以划分为2025年12月之前创立的公司，和2025年12月之后创立的公司。 — Isaiah Granet > [26:07](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=1567s) 并非单纯因为行业监管严格，而是因为底层数据基础设施极度混乱且非数字化原生（Non-digitally Native）。 — Kirby Forlin > [30:00](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=1800s) 联络中心之所以持续存在，本质上是因为企业当前能够交付的技术能力与用户理想的交付体验之间存在差距。 — Alejandro Villacis > [37:51](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=2271s) 购买业务结果，而不是单纯买一套软件代码。 — Jonathan Corbin > [39:57](https://www.youtube.com/watch?v=tcijaaGmmBQ&t=2397s) 当前不少企业的 CTO 正处于一种‘AI 精神病’（AI psychosis）状态——在各种眼花缭乱的新技术概念之间频繁横跳，缺乏聚焦。 — Isaiah Granet ## Mentioned Jonathan Corbin, Alejandro Villacis, Isaiah Granet, Andy Tang, Kirby Forlin, Maven AGI, FCP Euro, Bland, Draper Associates, The Cigna Group, Otter.ai, Claude Opus, AWS, HubSpot, HIPAA, FedRAMP, YC (Y Combinator), UC Health, Epic, Palantir, OpenAI, Nvidia (英伟达) ## Video description Jonathan Corbin - CEO/Co-Founder, Maven AGI Alejandro Villacis - Vice President Customer Experience, FCP Euro Isaiah Granet - Co-Founder, CEO, Bland Andy Tang - Partner, Draper Associates Kirby Forlin - VP Digital & Analytics Performance, The Cigna Group AI isn’t just transforming customer-facing chatbots—it’s also empowering human agents behind the scenes. This panel will explore how AI-powered copilots are providing real-time recommendations, automatic call summarization, sentiment analysis, and instant knowledge retrieval to boost agent productivity. Panelists will discuss how companies are deploying these tools in contact centers, measuring improvements in resolution time and customer satisfaction, and balancing automation with the human touch. Expect insights into the platforms and strategies that are redefining the modern contact center and helping organizations deliver faster, more consistent, and more empathetic support at scale. [Source](https://www.youtube.com/watch?v=tcijaaGmmBQ)

[0:04] Isaiah Granet: 如果出错就会导致诉讼或闹出人命 [4:39] Isaiah Granet: 当某些公司的估值超过沃尔玛，但单位经济效益（Unit Economics）却与沃尔玛相差无几时，你就必须质疑其底层商业逻辑是否健全。 [4:39] Jonathan Corbin: 目前真正产生巨大实效（ROI）的领域只有两个——客户体验（CX）与代码编写（Coding）。 [20:02] Isaiah Granet: AI 最具商业价值、护城河最深的领域，恰恰存在于那些高度受监管、任务极其繁重的复杂通话场景中。 [22:07] Isaiah Granet: 企业可以划分为2025年12月之前创立的公司，和2025年12月之后创立的公司。 [26:07] Kirby Forlin: 并非单纯因为行业监管严格，而是因为底层数据基础设施极度混乱且非数字化原生（Non-digitally Native）。 [30:00] Alejandro Villacis: 联络中心之所以持续存在，本质上是因为企业当前能够交付的技术能力与用户理想的交付体验之间存在差距。 [37:51] Jonathan Corbin: 购买业务结果，而不是单纯买一套软件代码。 [39:57] Isaiah Granet: 当前不少企业的 CTO 正处于一种‘AI 精神病’（AI psychosis）状态——在各种眼花缭乱的新技术概念之间频繁横跳，缺乏聚焦。

[0:04] Andy Tang: We're all pitching to these guys. That's where the money, the demands are. [0:04] Alejandro Villacis: if we wouldn't put the part on our cars, we wouldn't sell it to our customers. [0:04] Isaiah Granet: if we get it wrong, someone's going to sue or die. [4:39] Andy Tang: nobody will be doing phone calls a year from now. [18:52] Isaiah Granet: There are companies before December 2025 and there are companies after. [23:35] Isaiah Granet: we will fire you as a customer. [36:09] Kirby Forlin: without metadata, there's no information about the data itself. [37:51] Isaiah Granet: buying an outcome, not a software. [39:34] Isaiah Granet: we're throwing out the playbook

[0:04] All right, great. Welcome. Um, I'm excited. This  is going to be great. Uh, you know, this, uh, [0:10] I took a lot of care in the seating arrangement.  I hope you appreciate that. Two customers and two [0:16] AI experts. So, it's going to be fun when we,  you know, ask these guys questions. We can see [0:21] your list of priorities. Yes. [laughter] Yeah.  Exactly. Yeah. Yeah. Exactly. We're all We're [0:25] all pitching to these guys. That's where the money  the mans are. Um yeah, so but before before this [0:32] I want to tell you when they asked me they gave  me a list of panels to moderate um I picked this [0:39] one and I think it surprised a lot of people um  because it's not very often I get to talk about [0:46] customer related things and it occurred to me um  it is an area of my passion because our latest [0:58] uh analyst class from the summer. There were  six of them. They asked me what they should [1:03] do. I told every single one of them to go get  a job in customer service or um restaurant way [1:10] staff. And in fact, I'm curious how many of you  have worked, you know, when you were younger [1:16] worked in some sort of customer service related  industry or way staff, right? Yeah. Great. So, [1:22] I think passed on that message to the our Gen  Alpha friends. um it is an important skill set. [1:29] It's not just a way to pay for, you know, college  and whatnot, right? So, I thought it's kind of [1:34] interesting. Uh, yet now we're combining that  with this um, you know, flavor of the day AI. So, [1:41] it's a it's a great topic that I want us to to  think about. Uh, but without further ado, I would [1:47] like all of us to introduce ourselves. Why don't  we start from Isaiah? Perfect. Uh, hi everyone. [1:53] I'm Isaiah. I'm the co-founder and CEO at Bland.  Um what we do is we automate the really difficult, [1:59] really ugly phone calls. So the way I like to say  it is we want to be doing the phone calls where [2:04] um if we get it wrong, someone's going to sue  or die. So we're doing the really hard stuff, [2:08] fully automating end to end. Um we build all  of our models inhouse. So some of you might [2:12] have seen actually we just released our new  texttospech model yesterday, which is now um [2:17] EVEL does the most realistic in the world. We've  raised a little bit over a 100 million to date. [2:22] We work with 250 enterprises and 350,000 self-s  served users today. Hey everyone, I'm Jonathan [2:28] Corbin, CEO and co-founder of Maven AGI. Uh we  create AI agents for customer experience working [2:35] with about a 100 customers including some that  you've probably heard of, folks like Dell, Hinge, [2:40] Hard Rock, uh and many others. Um and we've have  about 100 employees. Um raised about $78 million. [2:50] um and excited to share more about our journey. Hi  all, I'm Kirby Forland. I um in the SVP of digital [3:00] data and analytics. I'm going to have a little bit  of a different perspective here because I'm really [3:04] come at this from a digital perspective. But the  thing that we've been working on this year is how [3:08] do we make sure all the context about customers  is connected. So within all the applications [3:14] that they might be going to and data might be  going over there and they might then eventually [3:19] call. How do we make sure that everybody's on the  same page and has the same information from which [3:23] that we can give a better customer experience? Hi  everyone, I'm Alejandro Vijas. I am with FCP Euro [3:31] family-owned company 40 years old online for  about 25 years. Um, we sell automotive parts [3:38] for European vehicles and our brand promise is  that if we wouldn't put the part on our cars, [3:43] we wouldn't sell it to our customers. So, we  stand behind every product we sell and and we [3:49] try to sell only highquality products that that  we can stand behind. Um, I've been in e-commerce [3:55] for about 15 years. um have been at FCP Euro  for about two two and a half years and have [4:02] really spent the last year implementing AI in the  contact center to try to augment the capabilities [4:08] of our agents and focus more on the value added  um types of conversations that they can have where [4:14] we're generating revenue and where we're driving  retention. So our focus for the next year at least [4:23] is continuing to find what are those opportunities  where we can save time from agents doing road [4:29] activities so that we can dedicate them to more  sales and retention activities. Great. Um so you [4:39] know I'm an early stage venture capitalist  been doing this for 20 years. Uh I've seen [4:43] multiple cycles. So, one of the things I want to  ask you guys, especially the two of you, Isaiah is [4:48] smiling because you know this guy is an incredible  entrepreneur. Uh, I'm glad I wasn't one of the 198 [4:57] pitches you made coming out of IC. I think he got  turned down by 198 investors. I'm glad I wasn't [5:02] one of them. They were complaining to him about  not wanting to back audiobased um companies. Phone [5:09] calls will be dead. The exact quote was, "Nobody  will be doing phone calls a year from now." And [5:15] that was in 2023. [laughter] Wow. It was uh Yeah.  Yeah. By the way, uh fun fact, I backed a company [5:21] called uh seeded a company called Otter AI out  of a um a cafe in downtown Powto in 2015. They [5:30] wanted to record all the audio conversations in  the world. I thought they were both creepy and [5:37] uh and what else I say? uh uh unrealistic, but  I'm glad both you guys proved me wrong. Um but [5:45] I'm very disappointed I didn't I wasn't the 199th  investor you talked to. But um uh that aside, [5:54] what do the two of you think about the um  AI economic cycle? You guys are founders, [6:00] you are CEOs. Um you have you're you're looking  at the big picture. You're guiding your company. [6:08] So tell tell us what you think. Is this is this  time different? No. Um I [snorts] mean I think [6:16] ultimately when you have companies that are valued  at more than Walmart with unit economics that look [6:21] about the same you start to ask questions and say  hey is this fundamentally sound? And then when [6:27] you start looking at down market and see all the  companies that are built on top of them. I think [6:31] what's particularly dangerous and part of why we  built bland without frontier models. This was a [6:36] really difficult decision was that we looked at  historically what happens with Uber with other [6:40] companies. The subsidy period eventually expires  and at that point if your business doesn't have [6:46] real lags and real fundamentals it gets really  ugly. I think that there's a probably a very [6:50] high chance that we're in a big bubble right  now. Um and I think that uh it doesn't mean AI [6:55] isn't valuable just like the internet wasn't  valuable in the early 2000s. Uh but it does [7:00] mean that companies need to be really careful and  mindful and that the people that are spending like [7:04] there's no tomorrow and are not paying attention  to the fundamentals, gross margin um are going to [7:09] be in hot water and the companies that are buying  these software from that companies um they need to [7:14] be concerned about the long-term health of those  businesses because otherwise it could be a really [7:19] ugly situation. Yeah. Yeah, I agree with you.  I think this time you know if I compare this to [7:23] internet or um you know um share economy and you  know a few other cycles after uh sort of before [7:32] this people actually paying now right it's not  just the eyeballs the question is are they paying [7:37] enough to keep you guys afloat but I'm pleased to  hear you guys all have a lot of customers Jonathan [7:43] your your view on this I think it depends on  what you're looking at right if you look at [7:46] it from a macro perspective I think that AI is  going to fundamentally change the way that we do [7:51] everything. If you look at it from like what is  the impact of AI from a micro perspective, where [7:58] are we seeing the most results of it? There's like  two areas. One of them is the reason why we're all [8:04] here today, which is customer experience. The  other is coding. Now, will we see that trickle [8:09] over into other areas of the economy and other  places within like white collar and blue collar [8:15] jobs? I think so. Like how long does that take?  If we look at other cycles, we'll usually start [8:21] off with like you know the people who are early  innovators kicking the tires, people who want to [8:26] like test out new technology and you see a trickle  down effect. Typically see it takes about 5 to 10 [8:31] years for to reach full maturity. Yeah. Now we're  seeing the evolution of this moving much faster [8:37] than anything we've ever seen before. Like where  it's actually impacting what customers are doing, [8:43] companies are doing, and how they're thinking  about the future. So are companies overvalued? [8:48] probably like but is there a massive macro impact  for sure. Yeah. Yeah. Yeah. Now from the customer [8:54] side what do you what are you guys seeing from  just from a corporate perspective? You are you [8:59] know buyers of AI products right? Um you guys  have very healthy balance sheet. Um are you guys [9:09] making investments in the AI areas? Do you want  to start? I mean definitely making investments [9:17] in the AI in the AI areas. Um I mean I think  it's an interesting time in business because I [9:22] think especially in a non-digitally native company  like Sigma is I think um they're they're trying to [9:29] figure out right now um let's do AI and I think it  needs at some point to flip in that AI isn't like [9:37] you have to change your business strategy that  it that AI helps you with a different business [9:41] strategy. I think the problem right now is it's  sort of AI for AI sake instead of like the whole [9:47] business reimagined about what it could be. So I  think that's probably the thing I'm seeing most in [9:52] business. A lot of investment, a lot of things  happening but sort of a disconnected ecosystem [9:58] and I think that stems from not reimagining how  it could be in the in the world where you know [10:04] Gen AI and of course all the other aspects of  AI exist. What I' what I've seen from a lot of [10:10] companies is they're like we're going to transform  our business with AI and they're like let's give [10:15] some AI to everyone and it has absolutely no  impact. What are you guys doing like how are [10:20] you thinking about implementing that? I'm really  curious to get it from your perspective. I mean [10:23] I think it's similar right like and I I mean I  think there's things that are doing impact. So [10:27] we have like I mean even just having you know  uh co-pilot or something that's helping people [10:32] become more efficient and you know in corporate  America you spend so much time writing PowerPoint [10:37] decks. I'm sure everybody here can appreciate how  much time you have spent like one of the greatest [10:42] efficiencies for me is the opus model and being  able to quickly generate slides which seems like [10:48] uh like seems kind of crazy but like it does  allow me to do a lot more strategic thinking [10:52] because I'm not spending so much time generating  slides but that isn't business reinvention. I [10:56] think the other thing I see is that there's a lot  of process automation trying to be able to use AI [11:01] but we're not re like we're just automating the  process that already exists that was probably [11:05] kind of broken. So I think there's like something  to be there. So I I and I I see a lot of tools [11:10] and duplicative tools being bought um instead  of you know I think this is where enterprise u [11:16] architecture comes in and we need a much more  thoughtful view on how we're going to actually [11:20] go forward on this. But again I think it goes back  to a lot of times it's like landing in the digital [11:26] world the tech world and what I keep saying is  like hey guys like we actually need the strategy [11:31] and then we can tell you what AI we need. And  I think it's ver I think it's upside down a [11:36] lot of the times right now. And I think we need  to get like tech and the business needs to come [11:40] together from a strategic viewpoint of what's  possible. And I think it's very hard because [11:46] if you're not familiar with it, it's very hard to  envision what's possible. So I think there's a I [11:50] I think the idea that there's a big macro change  happening is true. I think it's going to take a [11:56] little while longer and I do think companies  will overinvest invest in the wrong things. [12:01] we're going to make a lot of missteps. And I think  the non-digitally native companies that will win [12:05] are the people that bring in the talent that can  really think about reinvisioning a business and [12:10] and how it might be totally different. And and  I don't want to um go into a tangent too much. [12:15] Maybe you should go next and then I'll follow up.  Yeah. Yeah. So I'll say at FCP Euro, we're we're [12:20] very cautious about what investments we're making.  So we're being very thoughtful with AI investments [12:25] and I think we started in the technology teams or  software engineering particularly to look at how [12:33] can we leverage some of these models for coding  to make ourselves more efficient to be able to [12:37] deliver more more um updates and and and stuff.  We've seen some success there. The technology [12:46] team and the data team they've taken their  own approach to this is these are the things [12:50] that we're going to invest in. These are the  things that we want to um eliminate repetitive [12:55] work and we want to be able to deliver faster.  Other teams such as my own customer experience, [13:03] we started probably with a 3 to six month lag  from the technology teams thinking about okay [13:09] to your point what is our strategy for how we're  going to implement this. We don't want to go out [13:15] there and just start talking to every vendor that  is showing us tools that seem very similar and [13:24] to to be honest in some cases wasting our time on  conversations that all sound the same where it is [13:30] difficult to differentiate between what different  vendors might be able to provide. And so we focus [13:36] on the strategy like what do we want to try to  do if we implement AI? There were a couple things [13:43] that were clear to us. We spent a lot of time on  things that are repetitive that we can save time [13:49] on. Okay, great. Let's identify what all of those  are. Let's prioritize them. Let's figure out what [13:55] AI solutions we can use or buy or what do we  already have access to so that we're not buying [14:01] duplicative software and start to leverage that.  And then we said, okay, well, what is another pain [14:07] point? We don't have enough resources to devote to  gathering insights and analyzing transcripts and [14:14] conversations. We started recording transcripts  about 18 months ago and we hadn't implemented [14:21] any tools to try to analyze them or extract  value from them until about 6 months ago. And [14:27] so first we focused I hope you were using Otter  AI. [laughter] Just kidding. We're not. But but [14:37] um we we um we started transcribing all  our calls and we started thinking about [14:42] okay there's a lot of value that we can derive  here. What tools do we already have in place [14:47] that we can use? What are the new tools and  and models that we might be able to use? And [14:53] then we started little by little thinking about  okay like what is the investment that we have to [14:57] make and let's make sure that it aligns with  the strategy and that the costs make sense. [15:04] I've had some vendors come to me and say, "Hey,  like we can analyze all of your phone calls for, [15:09] I don't know, a base fee of $100,000 a year  plus however many transactions we're going to [15:16] evaluate." And and I'm like, when I start thinking  about how many conversations we have and I start [15:22] building out a total cost of ownership on this,  you're asking me to pay more for a new solution [15:26] that is unvetted, unproved than we pay for our  WMS software that we're that we're adding to the [15:34] to to the to cheaper to do it by human. Yeah. and  and it's like where what what is the differential [15:39] value that you're going to to bring to the table  when you're asking me to pay almost as much as [15:45] we pay for AWS to host our entire like website  infrastructure for call analysis like that doesn't [15:52] make sense right so coming back to the strategy  it's like okay what did we prioritize what do [15:58] we want to focus on and what are the tools that we  can use and can we start small and scrappy yes use [16:04] new technology and new tools but don't go to the  to the flashy Yeah. vendor that is trying to sell [16:10] you the dream. Yeah. Well, let me ask you this.  So, we we have um Kirby coming from, you know, [16:16] healthcare, Alejandro coming from e-commerce.  Um I'm I'm sort of looking and I'm going to say [16:23] something. I have a feeling you guys are going  to disagree. So, this is a this is a a green [16:28] light for disagreement. If I look at internet, the  adoption of internet started in communication. Um [16:36] and then and then in retail and in uh enterprise  and then in um let's see maybe industrial and then [16:46] financial services uh you know education they  probably would never adopt and then government [16:53] um right so if you kind of look at the order  of adoption this is essentially the level [17:00] of regulation they have essentially the more  regulated industry is it took them a little bit [17:05] longer [snorts] Right. Are you guys seeing that um  in AI adoption? Uh let's start from um um Jonathan [17:15] and Isaiah and I'm also so curious to see what the  healthcare expert has to say specifically about [17:20] this because you're very famous for HIPPA, right?  So is AI following sort of the same footsteps? [17:29] I think it's interesting and I'll disagree with  the first part of it because depending on who you [17:33] talk to some people say the internet was invented  by the government so [laughter] at the beginning [17:38] but no I think like seriously I think that um what  we're seeing is like the business cases are there [17:44] for broadscale adoption like if so if we think  about customer experience most of us here have [17:50] had a terrible customer experience like I had one  today checking in my hotel I'm sure you all did as [17:58] Like everyone has has that bad terrible  customer experience. So companies hear [18:03] from their customers, hey improve, I'm taking my  money elsewhere. Like do something different or [18:09] like face the consequences. And so you have  people who have like broadscale consumer [18:14] interactions saying I need to do something  different and if I can save money doing it, [18:18] amazing. They're investing hundreds of millions  of dollars for large companies and customer [18:23] experience and not getting much further value. The  other thing is is like when you start to look at [18:28] like like healthare like not to pick on healthare  like [laughter] we were talking about it earlier [18:34] like I tweaked my knee playing basketball the  other day and I'm like trying to get a hold [18:38] of my insurance company and my doctor's like come  see me in a year and you're like okay this doesn't [18:42] feel like it's working really well. And I think  that everyone has these things where it's like, [18:48] wow, like if we just had a cheaper way to  be able to communicate with each other, [18:52] a better way to interact, like this would actually  make life a lot better for all of us. Yeah. And so [18:57] there's a lot of consumer pressure for people  to improve. The other is like how do I improve [19:02] my operational efficiency? So prior to starting  Maven, I worked at a bunch of companies that were [19:07] market leaders in the customer experience space.  Companies like Adobe and Marquetto and HubSpot. [19:13] And the challenges that we had is like we had  so many customers and we wanted to talk to all [19:17] of them but we couldn't afford to. Like we had at  HubSpot we had 220,000 companies we were working [19:23] with hundreds of millions of people who wanted to  interact and talk to our customer support team. We [19:27] just didn't have enough people. And so we said,  how do we like change this and how do we improve [19:31] this at scale? And so like the efficiencies  that you realize of actually less people, [19:37] more customer interactions makes the business  case really meaningful for broadspread adoption [19:42] even in heavily regulated industries because  now we know what the playbook is in terms of [19:47] adoption. Like like we went out when we started  the company a few years ago, we started doing [19:52] getting certifications. We're HIPPA certified.  We're in the process of getting Fed RAM. So you [19:56] can move much faster if you have a clear path to  adoption. So this time could be different. Isaiah, [20:02] what I'm I'm dying to find out what these um  phone calls you've been fielding that that avoid [20:09] lawsuits and what is that su or dies. Death,  man. This is sounds like healthcare industry, [20:14] but go ahead. Well, [laughter] maybe a couple of  examples, right? One is um we do remote patient [20:19] monitoring calls where we're doing 45 minute long  calls with patients that are sometimes over 90 [20:24] years old, helping them find their blood pressure  device, helping them put it on, take the reading, [20:29] and understand if they're having a medical  emergency or if they just didn't tighten it [20:32] enough. Um and and the entire time talking  about their grandkids, keeping them engaged, [20:37] making sure that all those things are happening.  And there are medical emergencies on those calls. [20:41] We do realize that people need to go to the  hospital. We also have um worked with companies [20:45] to do surveys and polls of countries that we've  all been hearing about in the news recently [20:50] um and collecting data that if that was to be in  the hands of the governments that could see that [20:55] they might actually go doortodoor and find those  people. We deal with really serious calls. And the [21:00] interesting thing is that is the exact inverse of  what you were saying in my mind which is the real [21:08] value we drive is the regulated industries. It  is the difficult calls because it turns out that [21:14] um people in the business have been trying for  the last three decades to do everything they can [21:19] to shrink the contact center or get rid of it  in every way. They put IVRs, they try and get [21:24] you off, they call deflection time, they send  you the website, they give you the runaround, [21:27] they transfer you a thousand places. Um the humans  that are there today exist because the business [21:33] has a gap between what the business wants to do  what the business can do today and that gap can [21:38] only be filled by human judgment right now. And  so our AI um where you know also now Fed ramp [21:45] uh in the marketplace we're working with these  really complex customers the value that we see [21:50] is actually when we're doing those hard calls. Um  and what I always talk about is one day we want [21:55] to be doing 911 calls. And so for everyone in the  audience, just to imagine if you called 911 and an [22:01] AI picked up, but there is no human. There is no  you can't say human 50 times and get one. There [22:07] is no there's nobody in that call center. That  AI has to be really damn good. Right. Right. But [22:14] there's not a whole lot of value in between that  AI actually being able to handle you, you know, [22:19] screaming on the phone in an emergency and and  like routing you to the right fire department or [22:25] police department. And so what we think about is  our job is to handle those most complex difficult [22:30] calls. That's part of why we don't use external  models um and even are building things like the [22:35] telefan stack in house because we have to be able  to deliver that level of reliability. And what we [22:40] see from a lot of enterprises right now is um I  was at a dinner with uh YC partner recently. He [22:47] said something interesting. There are companies  before December 2025 and there are companies [22:51] after. The companies after December 2025 look  really different. because they're not trying to [22:56] use AI to assist human workflows. They're just  using AI to do the workflow. And it started me [23:02] thinking like a lot of us think that we're using  AI, but we're not actually. Like using it for [23:06] your slides is awesome, but what if you weren't in  the loop to even have to do the slides? Uh or for [23:13] finance, AI assisted invoicing, but if you were  starting a company today, would you even have a [23:18] finance department? How would that work? And most  people can't actually fathom what that looks like. [23:23] And our job is help deliver that reality and  deliver it for the ugliest, most difficult, scary [23:28] phone calls where uh you really have to trust  what's happening on the other end. Um I can't help [23:35] but jump in and say that I am that guy who gets on  the phone and my immediate reaction is operator. I [23:42] would say 15 times. Um it has to be that it's the  quality you're talking about, right? Um because I [23:50] think we all had some bad experience because  it's that runound is is there quality there [23:56] already and yes so bland we there's probably some  blandigans in the audience and they know that I [24:02] uh hate phone trees with a passion and it is like  the one thing that you can do that we will fire [24:07] you as a customer if you try to build a phone tree  with our product because the experience has to be [24:12] superior and it has to be safe and so we have this  thing called bland tone which is like how does our [24:17] AI talk to you and usually it turns out. It's like  some Midwestern sort of late50s like very sweet [24:24] like engagement tone and how does it redirect  you? And with our new voices, what we're seeing is [24:29] um for a lot of our customers over 70% of the  calls are doing people either don't know or [24:34] don't care it's AI. The way we measure that is  by utterances. So what we track is how many words [24:40] are you saying in reply to when our AI talks. If  it's more than two or three words, you're actually [24:45] probably engaging with it really positively  and it's highly linked. If it's two words of [24:50] no operator, yes operator, help, transfer, like  those small utterances, then you probably are [24:56] not having a great experience. And so what we're  seeing is starting to cross that threshold. Um, [25:02] but it takes time and it takes earning trust. Is  is this a a kind of a generational experience? [25:08] Like my kids are 12 and 15. Do you think they have  they're more comfortable with that? My favorite [25:14] example is that Gen Z doesn't go to the doctor.  And the reason that Jenz does not go to the doctor [25:19] is because they have social anxiety about picking  up the phone. This is like pretty well documented. [25:25] Yeah. Um [laughter] and I say this as uh for  Gen Z, right? Potentially. Why don't we pick on [25:33] why don't we pick on Gen Alphas? There aren't any  in the Those kids don't get it. Uh but uh Gen Z, [25:39] right? Like they have social anxiety because  there's this pressure about picking up the phone. [25:43] And I guarantee you if your doctor was an AI  contact in your phone that you could text or call [25:48] with the UC health system um right I think about  California people would make way more appointments [25:52] and they'd be way more likely to engage and if  we don't make that the future then people just [25:58] aren't going to go to the doctor um they will  opt out and we've seen this it happens uh and [26:02] so I think it is better yeah right you guys are  sitting here too complacent so what do you guys [26:07] think especially Kirby on the healthcare regulated  industry you AI adoption would you be do you see [26:16] this as a potential leading other industries or  more kind of follower? Um I I think the thing [26:24] that's been interesting because before I was in  healthcare which I've been in healthcare for about [26:27] the last five years I was in the consumer package  goods industry. Yeah. And so I actually and I'm [26:32] a data science by background. And so what I was  interested in was like coming in and like taking [26:38] all their data and solving the world's problems um  with you know with curing cancer. I had these like [26:43] you know lofty ideas. The reality of why I think  healthcare is slow to adopt is because their data [26:50] is in such bad shape. And so you could say all day  long you know it's that they're regulated or it's [26:56] this or it's that. It's a very very complicated  ecosystem that um is is difficult and they are [27:03] non-digitally native and so they haven't spent  a lot of time getting what you actually need to [27:08] have together to be actually able to do the things  that AI is going to enable. I love the idea of and [27:15] I believe human in the loop or a really great a  you know AI agent that can actually handle those [27:20] calls. That's the right thing that I believe our  agents should be handling. And I honestly think [27:25] like you need an MRI, you should be able to do  that online and you should be able to book that. [27:29] And our goal is like and the friction we're trying  to take out of the consumer experience is anything [27:34] that you that you should be able to handle in 30  seconds or less digitally, you should. And the [27:39] only things that should go to the call center are  the things that cannot be handled digitally. The [27:43] problem that we run into is to do all of that.  There's a massive data lift that needs to get [27:48] invested in so that we can in fact do this so that  I we have all the context so we know your plan, [27:54] your insurance, your EMR is connected, we  understand who your provider is, we understand [27:58] what's in your provider network so we can we can  we can route you there that we understand your [28:03] plan and how much that's going to cost. It's a  really super complicated uh data problem that is [28:10] what's holding I think the healthcare industry  back and I think they're doing a lot of good [28:13] work to do it. But I think it's a much more messy  problem than I understood. Um and I I didn't have [28:19] a lot of empathy for this um industry. Um before I  came in and was like, "Woo, this is messy." Um you [28:26] know, and there's no regulation right now, right?  That like there's regulation that you need to have [28:30] an EMR, but they didn't regulate that everybody  had to have the same EMR. So sorry electronic [28:36] medical records like so that we can understand  what's happening everybody like there's tons of [28:40] them like yes epic is the biggest one but there's  so many more so how do you get interoperability [28:44] with that if we need to understand that to have  a really good you know conversation to understand [28:49] when it should escalate so I think that is the  thing that holds it's less it's less it's less [28:56] regulation maybe more regulation could help us get  there but uh maybe not I don't know but there's um [29:00] is a data problem yeah yeah I I I you're good  companies 10 years ago I invested in also YC [29:07] company it's a AI doctor AI companies were more  primitive back then but it's the same idea right [29:14] if you have this AI it was probably founded by  a Gen Z person uh because you don't have to go [29:21] to a doctor you can just call to your AI doctor  right no phone calls required of course it was [29:26] very difficult to get the payers to pay it's we  didn't I think we all underestimated the business [29:31] model well and like talk about Gen Z I think the  thing that they'd want is to actually be able to [29:35] book online, right? Like they don't even there's  like I have two Gen Z kids. So like they're like, [29:40] "Oh my gosh." Like I when I suggest they call  their friends, we're like what? Like that is [29:44] not You should text them first. Yeah. You you  don't call people. But um you know like the part [29:49] of the problem is like why aren't people like we  should be able they should be able to go online. [29:53] We should be able to find their doctor. We should  be able to book like that should be something [29:55] that's digital and you know and escalate to call  if there's something more important. Let's talk [30:00] about e-commerce. I I think something Isaiah said  earlier really resonated with me where he said the [30:06] contact center like only exists because there's a  difference between what you can deliver today and [30:11] what you would ideally want to be able to deliver  to these customers, right? And we see that all the [30:16] time. Um in our business, uh the fitment of a part  for a vehicle has to be right for the vehicle that [30:25] you own. You're spending hundreds of dollars on  fixing your vehicle because you need to use it to [30:30] get to work typically, right? Or you need to use  it to transport your family. So you we we have to [30:36] get you the right part. If we don't, your vehicle  stays broken down for a couple more days. our the [30:44] way that our website works today, we we try to  drive a lot of self-service and selecting the [30:49] right vehicle part, but because of the complexity  of the European vehicles and like what parts are [30:56] compatible with different models, sometimes  you need VIN verification. We haven't built [31:01] that out like that's not built out on our website.  There are tools that we use to connect different [31:07] data sets so that we can answer those questions  and that's where our um experts on the team add [31:14] value. Um but it exists because we haven't built  out the technology for self-service. Would we [31:20] like our our our customers to self-s serve more  often? Absolutely. They can't right now. And so [31:27] we're providing a value added service for now. Can  we structure the data in such a way that we could [31:35] to your example about healthcare like be able to  create productize it on the website and and get to [31:41] a self-service? I'd like to think yes in the next  5 years that should be possible. Um but there are [31:48] constantly new vehicle makes and models coming  out every single year. There's a proliferation [31:54] of parts with different qualities and different  fitments assigned to them. And so it's a it's like [32:00] an everexpanding data problem which I'm sure you  face a lot in the healthcare space as well. Um and [32:07] it just becomes challenging. Yeah. Yeah. Um along  those lines of um sort of building versus buying. [32:15] Uh I'm curious what you guys have seen both from  the customer side as well as a solution side. How [32:21] do people maybe start with the two of you? um  as you talk to customers is there's this idea [32:28] of we might be able to build this ourselves. In  fact, I actually have a startup. Um I'm convinced [32:34] it's because they're in that space where the agent  space is not quite mature yet. So they're forced [32:42] to build their own agents and then they realize if  I build a few agents and I have to maintain these [32:47] agents, right? this idea of coding you either have  that um um the um AI code it for you completely [32:56] this idea of human machine collaboration makes  maintenance very difficult so they have to build [33:01] their agent factory just for their own use  I'm very curious what you guys are seeing [33:07] um what customers are telling you build versus  buy I think I think one of the things that's [33:12] interesting is like what you just described for  healthcare and for your company in particular of [33:17] you have fragmented data that lives across the  entire enterprise. And in order for you to be [33:23] able to deploy and use AI within your business,  you need to bring together all of that data set, [33:28] run it through this like ETL process, and now  you're ready to like feed it into AI. And that is [33:34] the scenario that every major company is in. And  so it's really easy for you to set up an AI agent [33:40] that can sit on a knowledge base and can like like  repeat like pieces of information back to people. [33:46] Yeah. Yeah, it becomes much harder when you  actually say what is the problem that we want to [33:50] solve? Well, we want to solve like better customer  experience. How do we do that? Well, part of it [33:57] is giving people the information they need. But if  you think about what's the full service experience [34:02] that's required to do that, you need to be able  to bring context into the interactions. You need [34:07] to understand who is this person? What are the  historic interactions they had? Like what is the [34:12] subscription level that they have to my service?  How many licenses do they have? What is the [34:17] account balance within the accounts they have with  us? And then you have to be able to take actions [34:21] on their behalf. And what that requires is going  out mapping all the enterprise data, understanding [34:27] which piece of information I'm pulling into  each turn of the conversation. It's incredibly [34:32] complex. And that's why it's really hard for you  to go and build this. Even though it's tempting, [34:36] right? Because these are all your data. These are  every every like new product manager that joins [34:41] the company is like, I am going to go build this.  [laughter] But how do you and [snorts] that's when [34:46] they'll call you. But if if we're how how are you  coming in as a you know somebody who's already had [34:52] it built and then implementing it on the data  that actually needs to feed into your system, [34:57] I feel like the problem still exists. Yeah. Kind  of. So I I think when we went out there and we [35:02] started having conversations with people on how do  we change the customer experience and the journey [35:07] that you've constructed for companies today, it's  not built for customers. It is built for companies [35:13] using their existing ecosystem and the limitations  that they have. If we were to reimagine that, [35:17] what would it look like? It would be this like  biodal interaction. It's proactive and reactive. [35:23] And in order to be able to enable that, you  need to go out and connect and map out all the [35:28] different data that exist across the ecosystem.  And for us, we built, we call it a graph of record [35:33] where it connects to all the different pieces  of information across the ecosystem. We create [35:39] instant graphs that understand the relationship  between like me as a consumer all the pieces of [35:45] information across the enterprise that relate to  me and you can create this personalized experience [35:50] for each individual user and so we connect to  the data we don't copy it over we don't have to [35:55] run it through an ETL process and because like we  use a combination of search technology along with [35:59] large language models it allows us to be able  to have this like instant understanding of like [36:05] here's the information here's how it pertains to  this customer and you can do it on a phone call, [36:09] you can do it in chat, you can do it in generative  search, whatever modality you want to use. I can't [36:15] help myself. But what if the metadata like doesn't  exist, right? So there's no context to the data [36:20] which in these big companies I've been in and  someone who's pretty involved in data that would [36:25] be like the thing that I'm struggling with is like  on agents in general is I'm like you got to build [36:30] your metadata and then then you can buy but like  without metadata there's no information about the [36:34] data itself and I'm not sure how and I'd love to  hear but like so I'm curious like because that's [36:38] like a big problem. Yeah, I think it it definitely  is. And so there's there's actually like ways that [36:43] we've created for us to be able to evaluate data  and create synthetic options that replace that. [36:51] Yeah. So so I'll I'll jump in here. Actually,  one of my data colleagues is in the audience and [36:55] he's partnered with me a lot on this. To your  point about building out the meta metadata, [37:01] we've started to use AI to be able to do that.  Um because we had to build out a lot of data [37:07] connections and data mapping. to Jonathan's  point um that didn't exist previously. We had [37:13] to document all the data that we had. We had good  pipelines like everything sitting in our in our [37:19] data warehouse. We were just not leveraging a lot  of the the data and connecting it appropriately. [37:24] But AI tools can help us document can help us  add the metadata. Um and and we've actually [37:30] done that like our data team has taken the lead  on a lot of that. And then we do bring humans in [37:34] the loop who have more context about the data  structures and all the different systems that [37:39] they're coming from and people like myself can  help validate okay what what is here and is this [37:46] accurate and if it is accurate then we can start  feeding it into other models. Part of all of this [37:51] I think ultimately comes down to a company  the size of Sigma. you are buying an outcome, [37:57] not a software. And we talk to our deployed team  about this, but your job is to deliver outcomes. [38:03] And at some scale, software just starts to look  like consulting because that's kind of what it is [38:08] a little bit. And so if you said to us today, the  metadata isn't there. We'd say, okay, we're going [38:13] to go build that with you to get the metadata in  there. And we've done that with customers. Um, [38:18] and the biggest thing that we see is that does  Sigma have the political time, appetite, and [38:23] will internally to not go try and flip over and  buy something every year. And one of the biggest [38:30] struggles that we have is that enterprises are not  investing in a software and a tool right now and [38:35] everything around it. And every time someone gets  fired, they flip around and they try something [38:40] new. And we have people that we started working  with, you know, two years ago and they said, "Oh, [38:44] we're going to build this ourselves now that,  you know, OpenAI real timeout is whatever. Two [38:49] years later, they spent a lot more money than  they would have spent on us." Um, because a [38:53] lot of it is bringing in this outside party whose  job is to deliver outcomes. And our job is to use [38:58] that software to leverage and and deliver that  at a much higher pace and route. Um, and then [39:04] ultimately like make sure that we're driving  business value for you. And if that means going [39:09] and filling in the metadata, then that's what we  will actually go do to get it done. I mean, that's [39:15] actually the Palanteer model, right? Imagine going  to the government and build out software for them [39:20] and catch Osama bin Laden. [laughter] That's what  it took. Yeah. I I think that's I mean I think [39:26] that's actually a really interesting point in like  making some bets and investments in the companies [39:30] that you're going to stick with because I I agree  with that part. I think the other thing that like [39:34] for me because I work across quite a few different  departments um is like so if we built that for the [39:41] call center then I want to make sure that we're  thinking about how that then maps into like the [39:46] broader data ecosystem. So I'd be curious how I  can I ask a question um I'm curious no please but [39:53] u two-minute warning so make it fast and we are  going to wrap it up. Okay. Um how do you do that? [39:57] how when you guys are talking to these customers  specifically about the outcomes and call centers, [40:01] how do you have the broader conversation about  the web and app and you know everything else [40:05] that's trying to use that same context? Man, it  there's a lot of CTOs out there with AI psychosis [40:10] right now. Uh and they're basically jumping in  between different things very fast. And a big part [40:15] of our job is that we are there because to give  permission to the enterprise to go do something to [40:21] do it in a new way and to actually transform. um  we're giving them an excuse to say we're throwing [40:26] out the playbook uh because Xpland is helping to  do this and this is the way they say to do it. Um, [40:31] and what we have to do is drive focus. And so my  job is going to be to get the call center outcomes [40:38] right. And if someone wants to talk to me about  the website and chat, then our answer is going [40:42] to be not until we finish this call center piece  because it requires so much discipline and follow [40:47] through. And it burns us sometimes to be honest  because then someone comes in and does the web and [40:52] it trickles over. But those enterprises that don't  have the discipline to go piece by piece lose [40:56] a lot of times is what we see. Catch these guys  outside after the panel. So, I got rapid fire for [41:01] you guys. Okay. So, I'm going to start here. Five  hour work week versus 35 hours. Five hours. So, [41:08] I could spend more time with my family. Perfect.  Five. What is 996? Five. All right. Next one. Um, [41:16] Bachelor of Science, Electrical Engineering versus  Master Electrician. Isaiah. Bachelor of Science. [41:23] Bachelor of Science. Electrician. I think I'd take  electrician, too. All right. You do great. Uh, [41:30] open source versus proprietary models. Uh, open  source. Open source. Open source. Open source. [41:37] Right. Very good. Regulating AI now versus  three years from now. From here. Now. Now. [41:44] Three years from now. Three years from  now. All right. Um, last but not least, [41:50] Nvidia export control versus free market. Export  control. Export control. Export control. I am not [41:59] familiar with this one. So, I'll go with  the [laughter] Well, with that, thank you