Tokio Marine, Hartford Reject AI Staff Cuts Over Catastrophic Risk

忧心巨灾风险，东京海上与Hartford拒绝通过AI裁员 保险业高管警告称，以AI取代核保人员来削减处理成本，恐将面临数百万美元的灾难性赔付风险，令任何运营开支的节省相形见绌。 专家小组避开生成式AI的炒作噱头，明确区分了基础生产力工具与依靠专有企业语境及专业空间智能驱动的真正核保“超额收益”。发言者指出，在核保队伍老龄化与全球风险波动加剧的双重背景下，企业在AI领域的成功取决于对工作流程进行彻底的重新设计，而非流于表面的工具替换。行业领袖强调，巨灾赔付损失的影响远大于微小的运营成本节省，主张打造“超级核保员”，并呼吁保险科技创业者集中精力解决具体且影响重大的风险难题，而不是推销臃肿庞杂的通用平台。 核保领域应用 AI 的核心目标是实现数据搜集与文件整理的自动化，让核保人员摆脱繁重的手工汇总工作，专注于高价值的风险评估、定制承保方案以及经纪人关系维护。 (2:34) 企业级 AI 落地涵盖三大阶段：桌面端工具仅是基础门槛，流程周转速度优化只能带来 1 到 2 年的短暂先发优势，而将顶尖核保人员的经验与智慧转化为专有的风险智能，才是获取真正市场超额收益（Alpha）的关键。 (4:40) 随着基础大模型日趋同质化，保险机构的核心竞争壁垒已转向专有上下文数据，例如引入车道级动态地理空间信息即可实现更精准的风险定价与理赔欺诈识别。 (7:36) 自主型 AI 智能体在保险业的规模化应用仍受严格监管与决策可解释性的制约；同时，由于资深商业核保人员正面临退休潮且后备人才不足，引入 AI 已成为应对行业人才断层的结构性刚需。 (9:11) 将 AI 单纯用于削减核保人员编制注定是得不偿失的策略，因为节省的微小运营处理费用，一旦因自动化定价失误承保了灾难性风险，就会被巨额赔付瞬间吞噬。 (27:03)

Executives from The Hartford, Tokio Marine, and TomTom argue that artificial intelligence delivers meaningful enterprise value only when it moves beyond generic productivity tools to generate true underwriting 'alpha.' They assert that carriers must prioritize proprietary risk context and loss-ratio performance over headcount reductions and superficial operational savings. The Three Horizons of Underwriting Alpha (4:40) Patrick Gallic of Tokio Marine outlines a three-tier model of enterprise AI adoption. Tier 1 encompasses desktop productivity tools like ChatGPT and Microsoft Copilot; Gallic views these as organizational hygiene that repel talent if withheld, but generate zero market outperformance. Tier 2 focuses on workflow optimization, such as compressing submission-to-quote cycles from one week to 24 hours. While necessary to avoid broker attrition and adverse selection, Gallic notes this speed advantage is temporary, yielding only a one- to two-year edge before industry standards normalize. True market alpha, Gallic argues, resides in Tier 3: developing organizational capabilities by codifying institutional underwriting expertise and leveraging proprietary data to refine risk appetite, attachment points, and policy terms. TomTom's Leo Sei emphasizes that because foundational large language models are rapidly commoditizing, outperformance depends entirely on proprietary context. For example, Sei cites an auto insurance pilot where replacing broad census-block groupings with dynamic, road-level geospatial data—incorporating live traffic flow and vehicle speeds—delivered a 2% underwriting performance lift. Protecting Loss Ratios Over Squeezing Headcount (27:03) The panel firmly rejects using AI to cut staff, characterizing workforce reduction as a losing strategy. In property and casualty insurance, operational processing expenses are negligible compared to potential claim payouts. For instance, saving $10 of operational processing on a $3,000 homeowner policy is wiped out if an automated pipeline binds a bad risk that triggers a $1,000,000 catastrophic loss. Underwriting profitability is governed primarily by loss ratios within the combined ratio, making risk selection accuracy far more consequential than minor back-office savings. Consequently, carriers aim to develop 'super underwriters' by eliminating manual data entry, spreadsheet cross-referencing, and summary drafting. Andrew Zarkowsky of The Hartford points out that commercial underwriting faces an impending demographic cliff: the majority of senior underwriters are in their 50s and approaching retirement, with few junior professionals entering the field. Automating standard, low-complexity properties—which often constitute 80% of a portfolio submission—frees experienced underwriters to evaluate volatile, non-standard exposures like California wildfire risk. Scaling Bottlenecks and Process Redesign (9:11) Despite widespread proofs of concept, Gallic observes that the industry remains far from enterprise scale: out of roughly 120 lines of business across an insurer, automated workflows are typically restricted to only one or two. The primary scaling bottleneck is liability: deploying non-deterministic, agentic models in heavily regulated environments shifts decision-making accountability away from humans without established control frameworks. To mitigate this risk, carriers currently confine autonomous agents to operational sandboxes, such as internal support, where supervisory AI systems monitor peer models. Achieving deeper integration requires restructuring entire workflows rather than swapping tools. Sei compares the current transition to industrial electrification, where early factories merely replaced basement steam engines with electric motors while keeping complex networks of shafts and belts. Significant productivity emerged only after engineers redesigned plants around standalone motorized tools. Furthermore, the panel emphasizes that underwriting requires auditability and procedural justice; policyholders and regulators will not accept black-box algorithms that fail to demonstrate transparent, rule-bound evaluation. In-House Data Harnesses versus Vendor Sprawl (36:23) The panel distinguishes an insurer's core business identity from commoditized enterprise overhead. A carrier's true differentiation lies strictly in risk engineering, underwriting evaluation, and claims adjusting. General administrative software, HR platforms, and user interface maintenance are best outsourced to enterprise SaaS providers with dedicated roadmaps and security maintenance. Rather than building software utilities internally, carriers are directing resources into proprietary 'data harnesses' that capture decades of loss history, distribution insights, and broker relationship nuances. The panelists caution founders against pitching broad platforms that attempt to duplicate internal carrier capabilities, recommending instead that vendors excel in narrow, data-intensive bottlenecks. Gallic also warns against relying on third-party startups that pivot from predictable per-seat licensing to consumption-based token pricing, which can create costly vendor lock-in for critical workflows. While demographic pressure and expanding systemic risks make AI adoption inevitable, carriers remain constrained by the challenge of governing non-deterministic models in strictly regulated lines, leaving enterprise deployment concentrated in tightly bounded pilot workflows.

在保险行业中，人工智能的战略价值并非压缩行政开支或裁减核保编制，而在于通过沉淀组织专有智慧与高精度上下文数据获取“承保超额收益”（Alpha）；来自The Hartford、Tokio Marine与TomTom的高管指出，行业正处于从表层效率工具迈向深层业务流程重构的关键节点，必须审慎平衡非确定性代理与严格监管合规之间的张力。 采纳三阶段与承保Alpha (4:40) Patrick Gallic（Tokio Marine）将企业AI应用划分为三个阶段：第一层是个体生产力工具（如ChatGPT与Microsoft Copilot），这如同现代办公中的基础书写技能，仅属于吸引并留存人才的“基本门槛”，无法提供超额市场收益；第二层是工作流时效压缩（如将出单审核周期从一周缩短至24小时），但这仅能提供1至2年的暂时性先发优势，随后会被全行业标准拉平；第三层则是将资深核保师的专有智慧、风险偏好与复杂条款制定能力沉淀为组织级智能，实现真正的承保Alpha。 研讨嘉宾强调，保险业盈利的核心在于由赔付率主导的综合成本率（Combined Ratio）。以一份年保费3000美元的家财险保单为例，若发生全损赔付，险企将承担高达100万美元的负债。单纯利用AI将单笔保单处理成本削减10美元毫无战略意义；若因自动化管道的筛选失误而承保了数笔劣质风险，产生的百万级赔付将彻底吞噬微小的运营成本结余。因此，AI的根本目标在于校准风险选择与定价精度，而非单纯追求廉价的运营降本。 规模化阻力与工业级流程重构 (9:11) Patrick Gallic坦言，保险业目前虽已跨越概念验证阶段，但远未实现真正的企业级规模化落地。多数机构在旗下一百余条业务线中，通常仅有一至两项实现了端到端自动化。阻碍规模化的主要瓶颈在于严格监管环境下将决策权移交给非确定性AI代理所引发的法律与运营责任转移风险。为此，机构普遍在内部客服或“AI监督AI”的受控沙箱中探索风控机制。 Leo Sei（TomTom）以工业电气化历史作比：早期工厂仅用中央电动机替换地下蒸汽机，却保留了低效的齿轮与传动皮带，直到全厂围绕独立电动机器彻底重组工序，才真正迎来生产力跃升。险企不能仅将AI视为替换旧工具的零星修补，而需在隔离的“净室环境”中从底层重新设计核保全流程。同时，企业应摒弃向董事会展示Token消耗量的虚荣指标，通过预先编译与压缩非结构化文档来消除原始数据的冗余开销。 空间数据向量化与客观理赔验证 (7:36) Leo Sei指出，随着基础大语言模型迅速同质化，承保的竞争优势几乎完全依赖于专有的高维度上下文数据。过去核保人员需逐一将地址复制到地图中肉眼观察卫星影像，如今通过将地理空间数据向量化为AI代理原生可调用的嵌入向量，微观区位差异得以精准量化。例如，道路两侧可能分属完全不同的洪水或山火风险带，商业地块内部建筑与密林的间距也会显著改变暴露程度。试点数据显示，以实时车流、实际车速与拥堵动态路级评分取代粗糙的邮编分区，直接为高频险种带来了2%的承保表现提升。 高精度数据流同时重塑了理赔环节的信任机制。在传统农作物冰雹险中，农户与查勘员常因各自选取受灾严重或完好的田块抽样而陷入零和博弈。引入无人机与卫星遥感结合的计算机视觉模型后，系统可客观测算出全场精确损失（如全田综合减产37%），以无可争议的透明审计留痕建立“程序正义”，从根本上消除欺诈摩擦与理赔诉讼成本。 人口断崖危机与超级专业人员 (15:43) 针对部分上市险企宣称以AI裁减人力的动向，Andrew Zarkowsky（The Hartford）与研讨专家坚决反对将专业承保团队边缘化。当下保险业面临严峻的人口老龄化断崖：大批资深商业核保师年过五旬并逼近退休，而青年后备人才供给严重不足。与此同时，极端天气频发使得加州山火或系统性网络安全等新型巨灾风险极度难以承保，亟需扩充精细化承保能力。 AI的核心定位是打造“超级核保人”与“超级查勘员”。以一份包含十处标的物业的商业财产险组合为例，AI可自动清分并预先核准八处低复杂度的常规资产，使专业核保师从繁重的手动检索与表格汇总中解脱出来，集中精力对剩余两处高波动复杂资产进行定制化建模，并深耕防灾工程建议与经纪人关系维护。 核心自研边界与垂直技术聚焦 (34:23) 与会专家清晰界定了保险公司的自研与外包边界：险企真正的核心护城河仅涵盖风险工程、承保定价与理赔裁决三大支柱，其技术自研重心应始终聚焦于积淀数十年的损失数据、分销网络与核保智慧治理层；至于界面交互设计、安全补丁与通用IT基础设施，则应交由成熟的SaaS厂商维护，避免企业内部工程团队陷入原型系统难以维保的技术泥潭。 对于面向险企的InsurTech初创企业，专家建议切忌盲目拓展平庸的平台化功能，或在确立业务后激进推行按Token计费的收费陷阱。面对精算严谨、注重百年历史模型的险企，初创公司只有专注攻克单一垂直环节的深层瓶颈、提供独家数据洞察并切实贡献承保Alpha，才能赢得大型机构的长期战略采购预算。 面对气候巨灾频发与日益复杂的商业环境，保险机构推进AI应用已非单纯的运营效率选择，而是填补老龄化人才缺口与释放承保容量的制度性必然。然而，保险的本质建立在契约信任与程序正义之上；如何在非确定性自主算法与确定性法律财务责任之间建立稳固的监督控制框架，仍是全行业规模化落地前必须攻破的制度难题。

## Background 本视频由人工智能行业峰会主办方 Ai4 于 2026 年 8 月 23 日发布在 YouTube 平台。页面简介显示，来自 The Hartford、东京海上集团（Tokio Marine）、TomTom 以及风险投资机构 Scale Venture Partners 的高管与投资人齐聚圆桌，深入探讨保险机构如何在严苛监管与合规审视下平衡速度与精度，利用 AI 变革核保、理赔与风险评估流程，并在坚守信任与可解释性的前提下实现业务现代化。 ## TL;DR 圆桌嘉宾一致指出，保险业应用 AI 必须跨越单纯的工具替用与“刷 Token 指标”的浅层阶段，转向通过专有上下文数据（如道路级地理空间信息）与承保经验代码化来创造真正的超额收益（Alpha）。专家们明确驳斥了将 AI 用于裁员降本的短视思维，指出保险的本质是风险定价，节省微薄运营费远不及防范百万级理赔失误重要，AI 的使命是对冲顶尖核保师退休潮并赋能“超级核保员”。险企应严格以核保、理赔与风险工程三大核心能力划定自研与外采边界，通过“清洁室”机制重构原生工作流并维系客户的“程序正义”感；同时告诫 InsurTech 初创企业必须“恪守自身航道”，打磨具备独家数据壁垒的单点极致方案。 ## Atmosphere & style 整场对谈风格极其务实、克制且充满商业洞见，主持人的提问直切投资回报率与商业博弈的痛点，而传统险企老将的拆解则透着久经市场周期的冷静与老辣。对话中不断涌现“为省10美元运营成本却招致百万美元巨亏”或“花千元买二手车蓝皮书”等辛辣生动的商业隐喻，彻底剥离了围绕大模型的浮躁炒作。整场交流既充斥着精算级的严密逻辑，又兼顾了对法律公平与同理心等社会契约本质的深层反思，带给观众一种置身于行业顶层战略闭门会般的清醒与厚重感。 ## Key takeaways - [02:34](https://www.youtube.com/watch?v=1qUzAfOphGI&t=154s) 核保的核心价值在于深度理解客户风险、定制方案、防灾防损与客户信任，AI 的首要任务是剥离繁琐的数据检索和整理工作，将专家精力释放到高价值业务上。 - [04:40](https://www.youtube.com/watch?v=1qUzAfOphGI&t=280s) 保险企业应用 AI 分为个人生产力工具、工作流流程优化以及专有组织能力/智慧代码化三个层次，只有将资深核保智慧转化为系统化能力才能带来超越市场的超额收益（Alpha）。 - [07:36](https://www.youtube.com/watch?v=1qUzAfOphGI&t=456s) 底层基础大模型正迅速走向大宗商品化，模型之间的性能差距不断收窄，输入模型的专属上下文数据与空间地理情报才是决定胜负的关键壁垒。 - [09:11](https://www.youtube.com/watch?v=1qUzAfOphGI&t=551s) 保险 AI 目前处于“跨过试点期但尚未规模化”的瓶颈，智能体带来的非确定性流程与责任转移机制在强监管环境下缺乏审计和控制手段。 - [10:38](https://www.youtube.com/watch?v=1qUzAfOphGI&t=638s) AI 落地不能停留在单点工具替换，而必须像工厂动力从集中式蒸汽机向独立电机演进一样，对业务工作流进行彻底的范式重构。 - [15:31](https://www.youtube.com/watch?v=1qUzAfOphGI&t=931s) 面对流程僵化与核保师老龄化退休潮的双重冲击，引入 AI 能拓宽风险接纳视野、开拓以往不可保的新市场，并将效率红利转化为更优保费。 - [19:40](https://www.youtube.com/watch?v=1qUzAfOphGI&t=1180s) 地理空间数据通过向量化和嵌入后极大降低了调用门槛，AI 能轻松实现道路级与地块级的精细化风险建模以及秒级理赔反欺诈核验。 - [22:46](https://www.youtube.com/watch?v=1qUzAfOphGI&t=1366s) 结合卫星遥感影像与 AI 图像分析开展灾后定损，消除了传统人工抽样勘查带来的主观博弈与猜忌，重塑了保司与客户间的信任基石。 - [27:03](https://www.youtube.com/watch?v=1qUzAfOphGI&t=1623s) 保险的商业本质是承担巨额负债风险而非节省微小运营开销，以裁员为导向的 AI 策略是走向归零的逐底竞争，AI 的真正目标是打造数据驱动的“超级核保员”。 - [31:12](https://www.youtube.com/watch?v=1qUzAfOphGI&t=1872s) 险企通过“双模型架构”应对组织噪音与算力成本：用轻量低阶模型作为全员思维沙盒过滤天马行空的无效想法，仅将验证成熟的高价值方案接入昂贵的高阶定制模型。 - [34:33](https://www.youtube.com/watch?v=1qUzAfOphGI&t=2073s) 在自研与外采决策上，险企应聚焦于风险工程、理赔定损和核保三大核心业务，构建固化数十年承保经验的专有评估框架（harness），而将通用软件维护留给垂直 SaaS 厂商。 - [37:23](https://www.youtube.com/watch?v=1qUzAfOphGI&t=2243s) 保险具备深厚的社会契约属性，程序正义与客户被倾听的权利不可或缺，单纯依靠冰冷的算法二元裁决会破坏维系保险商业模式的根本信任。 - [39:23](https://www.youtube.com/watch?v=1qUzAfOphGI&t=2363s) 对初创企业而言，单纯提供单点降本工具的市场空间有限，只有恪守自身细分赛道、利用专有数据和深度洞察协助保司优化综合成本率（创造 Alpha），才能赢得保司决策层青睐。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=1qUzAfOphGI&t=4s) 嘉宾开场与背景介绍 主持人 Alex Niehenke 首先用轻松幽默的话题开场，感谢大家在周二下午到场，并调侃台下观众可以随时往前坐，笑称“外面至少还有5个人对保险话题感兴趣，等会场子里肯定能坐满”。随后他介绍了自己的背景：他来自旧金山的风投机构 Scale Venture Partners（管理着数十亿美元资产），专注于 AI 技术栈中垂直行业的投资。尽管“热爱保险”在全美范围内可能不算大众爱好，但在本场会议中却极具共鸣；他主导投资了包括 Root Insurance、Archipelago Analytics、Sixfold 和 Spruce 等多家保险科技企业。 紧接着，三位来自不同领域的行业专家依次进行了自我介绍： - **Andrew Zarkowsky（The Hartford）**：供职于这家拥有200余年历史的老牌保险巨头（业务涵盖财产险、意外险及各类商业险等，不含健康险）。Andrew 在公司已有约20年工龄，一直深耕在商业险中端市场及大型商业险业务的核保（Underwriting）一线，如今专注于将 AI 深度引入核保体系。 - **Leo Sei（TomTom）**：负责 TomTom 的产品与市场。他解释了 TomTom 为什么会出现在保险峰会上：虽然大家熟知 TomTom 是因为其25年前的卫星导航业务，但如今它已转型为一家 B2B 地理空间数据（Geospatial Data）公司。TomTom 拥有路网结构、实时交通、地址、历史事故、车道封闭以及车速等多维度数据，而在过去两年中，AI 极大地降低了调用和分析这类空间数据的技术门槛，让保险机构能够在日常工作流中即插即用地提取空间情报。 - **Patrick Gallic（Tokio Marine Group）**：来自日本历史最悠久、规模最大的保险集团东京海上（Tokio Marine）。在过去20年里，东京海上在全球范围内收购了约100家保险承保机构，并整合成数十家集团成员公司；Patrick 的职责正是主导日本本土以外所有海外集团成员公司的 AI 战略落地。 --- ### [02:34](https://www.youtube.com/watch?v=1qUzAfOphGI&t=154s) 摆脱繁琐杂务：AI 赋能下的核心核保价值重构 Alex Niehenke 抛出了第一个破冰问题，希望大家跳出空洞抽象的口号，从实际投资回报率（ROI）以及能否为企业创造超越市场平均水平的“超额收益（Alpha）”出发，分享近期在 AI 落地方面真正令人兴奋的实质性成果。 ```mermaid flowchart TD accTitle: 核保工作流与AI介入对比 subgraph Legacy ["传统核保痛点（低价值/耗时耗力）"] A1["多方搜集原始信息"] A2["跨系统检索与比对"] A3["手动总结与整理材料"] end subgraph AI_Engine ["AI 自动化赋能层"] B["自动化数据提取与智能汇总"] end subgraph Core_Value ["真正的高价值核保（核心竞争力）"] C1["定制化保险方案"] C2["风险工程与防灾防损"] C3["理赔衔接与客户关系"] end Legacy --> AI_Engine AI_Engine --> Core_Value ``` Andrew Zarkowsky 接过话题坦言，他的头衔“AI 核保主管（Head of AI Underwriting）”在公司完全是“新造出来的职位”。作为在核保领域摸爬滚打了20年的老将，他指出真正的核保价值在于：深入理解客户风险、定制个性化保单方案、解决疑难问题、提供风险工程（Risk Engineering）解决方案以防患于未然、在发生损失时提供精准的理赔对接，以及建立深厚的客户信任关系。 然而在现实中，核保员绝大部分时间和精力都被迫消耗在“收集信息、检索数据、编写摘要”等繁重枯燥的基础杂务上，这些流程既不创造溢价，也毫无乐趣。The Hartford 目前正全力推进 AI 自动化来处理这些低效节点，目标是当核保员坐下办公时，所有关键信息已经提炼就绪，让他们能够将100%的精力投入到真正解决客户问题的“高价值核保活动”中。 --- ### [04:40](https://www.youtube.com/watch?v=1qUzAfOphGI&t=280s) AI 赋能三层次与寻找超越市场的“Alpha” Patrick Gallic 进一步从资本回报与超额收益（Alpha）的视角拆解了保险企业应用 AI 的三个核心层次。他指出，仅仅谈论 ROI 是不够的，关键在于企业如何依靠 AI 跑赢市场基准： ```mermaid flowchart LR accTitle: AI 落地的三个层次与 Alpha 创造能力 subgraph L1 ["第一层：个人生产力"] direction TB N1["ChatGPT / Copilots"] N2["基础工具标配<br/>（无 Alpha 溢价）"] end subgraph L2 ["第二层：工作流优化"] direction TB N3["核保周期提速<br/>（从一周压缩至一天）"] N4["短期优势（1-2年）<br/>后期成行业基线"] end subgraph L3 ["第三层：组织能力与数据智能"] direction TB N5["核保智慧代码化<br/>（Codify Wisdom）"] N6["精准风险偏好与定价<br/>（创造真正的 Alpha）"] end L1 --> L2 --> L3 ``` - **第一层：个人生产力工具（Individual Productivity）**：例如部署 ChatGPT 或各类 Copilot 工具。这类工具可能存在一定 ROI，但“绝对不会带来 Alpha”。Patrick 用书写历史做类比：古代中世纪僧侣探索逗号、句号和书写规范时那是重大创新，但放到今天，如果强迫一个人全靠手写，那就是一种惩罚；企业若不向员工提供 AI 工具，同样会沦为惩罚，优秀人才也会流失。 - **第二层：工作流流程优化（Workflow Optimization）**：例如将投保提交到出具报价（Submission to Quote）的时间从一周压缩至一天。这能在 1 到 2 年内建立先发竞争壁垒；但随着整个行业普及提速，服务等级协议（SLA）的时间窗口被普遍拉平，不跟进的企业只会面临“逆向选择（Adverse Selection）”，因此流程提速最终会变为行业标配而非长期 Alpha。 - **第三层：构建专有组织能力与智慧代码化（Codifying Wisdom）**：真正的超额收益源于如何利用自有数据与智能工具，精准把控风险胃口（Risk Appetite）、承保责任起赔层级（Attachment Levels）以及保单条款表述。他提到一个经典现象：“把同一份保单分发给 10 位核保员，你可能会得到 10 种不同的核保结论。”核保员的价值在于其脑海中的隐性经验与智慧；AI 的关键突破在于将这些顶尖智慧转化为系统化的组织能力，并反哺核保员。 Patrick 用“天气预报的发展史”生动比喻了这种演进：正如初中科学课上靠立一根木棍观察升降、或是单看气压计来预测阴晴，到如今气象学家全面利用卫星遥感与高阶 AI 模型，公众消费气象数据的方式发生了质的飞跃；AI 同样正在让保险机构从传统经验判断迈向高精度决策时代。 --- ### [07:36](https://www.youtube.com/watch?v=1qUzAfOphGI&t=456s) 基础模型同质化下，专有“上下文数据”即核心壁垒 曾任职于 Google 十余年、深耕 AI 领域的 Leo Sei 对 Patrick 的“Alpha 理论”表示强烈赞同。Leo 指出，目前各大底层大语言模型（LLM）正快速走向通用商品化（Commoditized），大模型之间的性能差距正不断收窄，决定胜负的终极分水岭在于**输入模型的上下文数据质量与深度（Context）**。 Leo 结合 TomTom 的落地案例说明了丰富地理空间上下文为保险承保与理赔带来的巨大 Alpha 突破： - **道路级精细化定价**：某合作保险公司将其风险评估维度从传统粗放的“人口普查区/邮政编码（Census / Zip Code）”颗粒度，升级为基于 TomTom 真实车速和实时路况的“道路级（Road-level）”评分模型，直接让承保回报率提升了 2%，这在成熟精算体系中是极其惊人的提升。 - **理赔欺诈甄别与秒级核验**：TomTom 目前正与一家车险公司开展理赔核验（Claim Validation）试点。当投保人声称“在某特定时间于某特定路段发生碰撞事故”时，系统利用精确的历史交通流和事故事件数据进行交叉验证，能够迅速判定该理赔请求是真实合规还是涉嫌欺诈，为车险反欺诈与自动核赔开辟了全新路径。 ### [09:11](https://www.youtube.com/watch?v=1qUzAfOphGI&t=551s) 保险 AI 现状：跨过试点阶段，但在规模化落地与责任转移中受阻 当前保险核保领域的 AI 处于“已过试点期，但尚未真正实现规模化（past pilot, not yet at scale）”的夹缝阶段。行业内的普遍现状是：几乎每家大型险企都能举手表态自己拥有一套端到端的全自动化智能决策流程，但一旦深入追问便会发现，在旗下多达 120 条业务线中，真正落地投产的往往只有区区一两条。 ```mermaid flowchart TD accTitle: AI in Underwriting Scalability Bottlenecks subgraph Current["现状矛盾"] A["自动化试点成功<br/>(1-2 条业务线)"] -->|难以横向复制| B["全公司 120+ 条业务线<br/>普遍仍依赖传统模式"] end subgraph Bottlenecks["规模化阻碍"] B --> C["责任转移风险<br/>(Shift of Liability)"] B --> D["非确定性流程<br/>(Non-deterministic vs. 严苛监管)"] B --> E["缺乏有效管控机制<br/>(AI 监管 AI 仍在探索中)"] end ``` 这种规模化推进的困难，核心在于智能体架构（Agentic AI）所引发的责任转移机制。当企业将决策权交给 AI 智能体时，实质上是将履职责任转移到了非确定性系统（non-deterministic process）中，而强监管的保险行业目前根本缺乏针对此类系统的有效控制与审计手段。因此，当前大部分创新仍局限在内部客服、聊天支持或让“AI 充当其他 AI 的监督者”等外围尝试上。 ### [10:38](https://www.youtube.com/watch?v=1qUzAfOphGI&t=638s) 蒸汽机到电机的隐喻：从工具单点替换到工作流范式重构 AI 落地不应仅仅停留在“把现有工具换成 AI”，而必须经历全面的业务流程再造。这就像历史上工厂动力从蒸汽机向电力转型的过程： * **第一次工业改造的误区**：工厂最初引入电力时，只是简单地把底层产生动力的大型蒸汽机拆掉、换上一台大型电动机，但头顶上依旧保留着错综复杂的齿轮、皮带与传动轴，系统复杂度丝毫未减。 * **真正的流程重构**：直到行业意识到电动机可以小型化，并直接为每一台缝纫机或独立机床单独配备专属电机，才真正释放了电力的生产力。 在保险核保中，很多初创企业的方案只是在传统管道的某一个切片做局部打补丁（例如单点的数据增强）。从两年前宣称“在核保管道中插入第三方数据”，到一年前主打“自动发邮件找主管审批或找客户要材料的异常处理智能体”，再到近期热炒的“核保工具架（Underwriting Harness）”，概念层出不穷，但底层实质的业务重构远未完成。 ```mermaid flowchart LR accTitle: Workflow Paradigm Shift Analogy subgraph SteamAge["单点替换模式 (类似蒸汽机直接换电机)"] S1["集中式动力源/单点AI"] --> S2["保留原有繁冗复杂的<br/>传动轴/官僚报批流程"] end subgraph ElectricAge["流程重构模式 (每台机器独立电机)"] E1["端到端解耦工作流"] --> E2["分布式轻量智能体<br/>直接赋能具体业务节点"] end ``` ### [12:34](https://www.youtube.com/watch?v=1qUzAfOphGI&t=754s) 险企的确定性诉求与初创企业定价暴涨的商业博弈 保险业本质上是一个极度崇尚“可预测性、一致性与例行常规”的行业。在巨灾（Catastrophe）和财产险核保中，从业者依赖的是 100 年甚至 250 年一遇的洪水图谱与长期历史数据，绝不会因为科技圈 3 个月前流行的新概念就盲目买单。 核保业务规模极其庞大（例如 Hartford 这类险企拥有 19,000 名员工、数百名专业核保师），很多仅有 4 名研发人员、试图从理赔跨界复制到核保的初创团队，根本无法应对这种体量与行业厚度。 与此同时，保险机构与 AI 初创公司合作时还面临着剧烈的商业模式变动风险： - **收费模式剧变**：初创团队在验证了产品市场契合度（PMF）后，收费模式往往迅速从传统的“按席位付费（Seat License）”跳变至“基于 Token”或“额度积分制（Credit System）”。 - **成本面临 10x 激增**：当险企将核心业务或知识流程交由外部 AI 代劳后，该环节就从企业的资产演变为了对外的依赖与潜在风险。一旦一年合同期满，供应商因自身未实现盈利而将价格上涨 10 倍，险企将被迫陷入巨大的被动局面。 - **行业特殊性**：普通企业软件算错可能只是“糟糕，出了点小偏差”，但在保险领域，由于严格的可审计性（Auditability）和信任机制，任何错误都会演变成真金白银的赔付责任。 ### [15:31](https://www.youtube.com/watch?v=1qUzAfOphGI&t=931s) 迫在眉睫的银发退休潮与 AI 拓展可保边界的潜力 面对是否应该因为 AI 太难落地而放弃的质疑，答案是绝对不能放弃。保险行业目前正遭受两大严峻现实的夹击，AI 的引入是必由之路： 1. **流程停滞 25 年**：现行的核保流程与 25 年前相比几乎毫无变化，内部充斥着大量被制度化固化的官僚审批流程。 2. **人才断层与老龄化危机**：顶尖核保师的年龄普遍在 50 岁以上，随着退休潮席卷而来，作为传统行业的保险业对大学毕业生的吸引力有限，行业正面临不可逆转的专业人才缺口。 从商业本质来看，保险的核心在于承担实体经济的风险，以支撑其正常运转。然而当前世界面临的风险日益复杂（如加利福尼亚州等局部区域的财产险承保难题，以及网络安全险的风险累积效应/Aggregation），大量商业投保需求（Submissions）因核保能力不足而根本无法完成出单（Placed）。 通过引入 AI 实现核保自动化与流程提效，不仅能对冲人才流失危机，更重要的是能够“拓宽风险接纳的视野范围（Open up the aperture）”，建立全新的可预测模型，开拓原先无法承保的新市场。由此释放的效率红利最终将平摊并反馈给市场，转化为更具竞争力的综合成本率与更实惠的保费。 ### [18:18](https://www.youtube.com/watch?v=1qUzAfOphGI&t=1098s) 释放核保精力：从简单业务走向高难度风险定价 利用 AI 优化资源配置，能让保险公司以更合理的价格为客户提供更优质的保障，实现双赢。以财产险为例，假设一家客户企业拥有 10 处物业网点，通常其中 8 处情况非常简单明了，仅有 1 到 2 处属于高难度风险标的。在传统模式下，核保员必须逐一审查全部 10 处地点，导致真正能花在攻克复杂风险上的时间和精力极为有限。而在未来，AI 完全能够自动处理这 8 个简单网点，甚至协助分析那 2 个难点。 这样一来，核保人员便能腾出宝贵的时间去攻关以往难以承保的复杂领域，例如加利福尼亚州的森林火灾险。如果核保员每天的大部分时间都被俄亥俄州那些简单明了的标准业务所占据，行业就永远无法在野火这类极端风险的定价上取得突破。 - 传统核保痛点：核保员在收到资产价值清单（Statement of Values）后，甚至需要手动在 Google Maps 上逐一比对建筑外观与表格地址，流程低效且机械。 - AI 的实际价值：接管基础性、重复性的核验工作，释放核保专业人员的时间与精力去攻克高难度风险定价。 --- ### [19:40](https://www.youtube.com/watch?v=1qUzAfOphGI&t=1180s) 地理空间数据与 AI 结合带来的全新维度 TomTom 在保险领域并非新玩家，但以往的合作模式主要是向大型保险公司单向输出地理数据，由对方的核保或理赔团队自行运算处理。由于过去地理位置数据（Location Data）非常复杂，需要专用的分析工具与专业技能，导致业务管理层往往难以预估投资回报率（ROI），测试和接入的门槛极高。 随着 AI 技术的进步以及地理数据向量化（Vectorization）与嵌入（Embedding）技术的应用，智能体（AI Agent）即便没有深厚的 GIS 专业背景也能轻松理解和调用地理信息。无论是核验道路两侧的洪水淹没线与防火隔离带，还是评估大型地块中靠近森林深处与靠近马路边缘的建筑在火灾风险上的差异，AI 都能进行极高精度的空间风险建模。 ```mermaid flowchart LR accTitle: 地理空间数据在保险中的演进 subgraph 传统模式 A["传统复杂位置数据"] --> B["专业 GIS 工具与团队"] B --> C["高成本测试与低 ROI"] end subgraph AI 赋能模式 D["位置数据向量化/Embedding"] --> E["AI 智能体直接调用"] E --> F["精细化风险建模<br/>(洪水/火灾/多网点比对)"] end ``` - 个人与企业端全景视图（Insurance 360）：利用 AI 自动通读所有保单，分析重复投保与保障盲区，未来更可能催生出 AI 责任险等全新的 B2B 险种。 - 克服数据过载：核保员日常要处理大量销售、沟通中介和维护客户关系的杂务，并非数据科学家。单纯堆砌海量数据只会导致核保员无所适从，AI 的关键作用在于精准提炼并推荐对当前客户真正有价值的核心数据。 --- ### [22:46](https://www.youtube.com/watch?v=1qUzAfOphGI&t=1366s) 卫星遥感与 AI 查勘：消除零和博弈，重塑理赔信任 保险公司与被保险人之间的核心基石在于“信任”，理赔环节绝非双方互相博弈或“算计”的战场。以农作物冰雹灾害保险为例：在一片上万英亩的玉米农场遭遇冰雹袭击后，传统查勘员开着皮卡来到现场，农户往往倾向于引导查勘员走向受灾最严重的田块，宣称“损失高达 50%”；而心存疑虑的查勘员若走向另一侧完好的田块，则可能得出几乎无损的结论。这种人工抽样模式必然导致双方互不信任。 引入地理卫星遥感影像（Geo Satellite Imagery）并结合 AI 图像识别后，系统能直接对全田块的受损面积与折断比例进行客观定量评估。查勘员只需出示客观的分析图并核实：“根据卫星图像显示受损率为 37%，您是否认可？”双方快速签字确认，不仅显著提升了定损效率，更消除了人为主观臆断带来的猜忌与对立。 - 农作物定损信任机制：通过卫星影像与 AI 分析代替人工粗略估测，让定损数据透明公开，消除农户与保司之间的博弈心理。 - 自动化与遥感技术：利用卫星或无人机（如 DroneDeploy 等技术）进行农田大面积灾损评估，正在成为重塑理赔体验的标准工具。 --- ### [24:38](https://www.youtube.com/watch?v=1qUzAfOphGI&t=1478s) 争鸣与共识：AI 是替代人力还是打造“超级核保员”？ 关于 AI 是否会引发行业裁员的争议，各家保险公司的战略定位存在明显分歧。部分保司在财报和监管披露中明确提及未来因 AI 替代而缩减的人工比例，但现场嘉宾对此持坚决反对态度。 真正的核心价值在于业务领域专家（SME）——包括核保员、理赔查勘员和风险工程师。AI 的使命不是淘汰核保员，而是为他们赋能，将其打造成拥有全维度数据支撑的“超级核保员”（Super Underwriter）与“超级理赔员”。 - 赋能专家而非裁员：优秀的核保员拥有深厚的业务认知，在与分销渠道和客户交流时能提供极具创造力的定制化保障方案；AI 负责剥离繁琐的数据搜集与录入工作，让专家专注于发挥专业洞察与服务价值。 - 规模效应：保司的目标不是借 AI 削减人员编制，而是希望借此拥有更多高素质的核保专家，从而为客户创造无可替代的卓越服务体验。 ### [27:03](https://www.youtube.com/watch?v=1qUzAfOphGI&t=1623s) 效率提升 vs. 裁员：保险的核心是承保风险而非节省几块钱运营成本 讨论开篇聚焦于一个业界普遍焦虑的话题：AI 是否会导致保险从业人员的大规模缩减。发言嘉宾明确反驳了这一观点，指出保险行业不仅不会裁员，反而需要更多的人才。在他们的团队内部，从来没有任何“用 AI 取代员工”的讨论。这种把 AI 纯粹当作减员工具的思路被直斥为“注定失败的主张，是一场走向归零的逐底竞争”（"a losing proposition and a drive to zero"）。 究其根本，保险业务的商业逻辑在于收取保费以承担潜在的巨额负债。承保决策做对还是做错，其财务影响远远压倒了内部运营成本的微小节省。以一张普通的房屋保险保单为例： - 客户可能每年仅支付 $3,000 美元的保费，但一旦发生重大事故，赔偿金额可能高达 $1,000,000 美元； - 如果依靠自动化流程在处理这张保单时节省了 $10 美元的运营开销，却因为模型失误错误地吸纳了几个劣质风险，保险公司面临的将是数百万美元的净亏损。 因此，AI 在保险公司的核心价值绝不是取代人力，而在于协助更精准地评估风险与提升整体业务韧性。 --- ### [28:00](https://www.youtube.com/watch?v=1qUzAfOphGI&t=1680s) 基础大模型的使用与成本焦虑：Token 计费模式下的分层治理 Scale Venture Partners 的合伙人 Alexander Niehenke 随后抛出一个关键问题：各位是否是大语言模型（LLM）基础模型供应商的大客户？这些算力支出是仅发生在工程研发端，还是已经深入到了日常的核保团队？董事会是否已经开始对着暴涨的账单惊呼“我们必须把这笔开支控制住”？ 嘉宾们确认，大模型的使用范围和账单金额确实在激增，目前内部的推进主要分为三个层次： 1. **全员企业访问层（Enterprise Access）**：几乎所有员工都能使用某种基础聊天界面，实现基本的生产力赋能。 2. **工作流自动化层（Workflow Automation）**：正处于加速扩展期。在单兵作战时投资回报率（ROI）非常明显，但一旦扩展到规模化业务流中，伴随而来的责任承担（Liability）与合规风险使得“到底该扩多大、推多快”成为高层深入博弈的核心议题。 3. **组织变革层（Organizational Change）**：目前最大的预算并非直接花在算力上，而是投入到了推动全员组织适应和转型的软性成本中。 在探讨“Token Maxing”（无节制消耗 Token）和成本失控问题时，嘉宾提出了一套以研发性质为边界的分阶段把关机制（Stage-gate）： ```mermaid flowchart TD accTitle: LLM 资源分配与成本控制策略 A["AI 应用场景分流"] --> B["探索性研发 / 创新场景 (R&D)"] A --> C["确定性业务流程 (Deterministic Process)"] subgraph RND ["高容忍度算力供给"] B --> D["采用非确定性 AI 探索新方案"] D --> E["实行 Stage-gate 阶段把关，允许高 Token 消耗"] end subgraph Process ["严格成本效益管控"] C --> F["固定流程优化 / 既定规则操作"] F --> G["严禁无节制调用昂贵 Token，强调单产 ROI"] end ``` 如果 AI 是用来作为非确定性工具去探索全新流程、进行代码重构或研发发现，给予充沛的 Token 额度是非常值得的投资；但如果员工仅仅是在一个既定的确定性业务流程中机械工作，却消耗无上限的 Token 来产生相同的业务结果，这种开销在管理层面前将变得极难自圆其说。 --- ### [29:45](https://www.youtube.com/watch?v=1qUzAfOphGI&t=1785s) Token 成本是短期阵痛：知识压缩、开源模型与“灯泡演进”隐喻 TomTom 的首席产品官 Leo Sei 对大模型成本表现出相当乐观的态度。尽管当前很多企业还享受着按坐席（Per-seat）付费的红利，而供应商正逐步转向按 Token 计费，这可能导致明年的预算谈判变得艰难，但他认为两点核心逻辑决定了成本不会成为长期瓶颈： 首先，输入端（Token In）的极大浪费即将被技术消除。目前整个行业仍处于“粗暴地将几百页的 PowerPoint 和 PDF 直接扔给 LLM”的蛮荒阶段。通过知识压缩（Compressing and Compiling Knowledge）和上下文检索优化，可以轻松斩获大量成本优化的红利。Leo 形象地打了一个“电灯泡演进”的隐喻： > 早期的白炽灯是通过直接燃烧线圈来发光，发光效率极低且大量能量转化为无用热量；而经历技术迭代后的现代 LED 灯，以极低的电耗就能提供更优质的光源。当前的 LLM 调用就处在第一代白炽灯阶段，未来对能耗与 Token 的极致优化是必然趋势。 其次，开源权重模型（如各类高性价比开源架构）的爆发正在大幅拉低市场价格。LLM 本质上正在演化为标准的“推理机器”（Reasoning Machines）并迅速大宗商品化（Commoditization）。未来的主战场绝不在于单次调用的 Token 价格，而在于数据上下文（Context）以及模型周边的工具生态系统建设。 --- ### [31:12](https://www.youtube.com/watch?v=1qUzAfOphGI&t=1872s) 双模型架构：低阶沙盒过滤噪音，高阶定制模型落地业务 针对模型落地的治理，有险企展示了极具实操性的“双模型机制”（Two-Tier Model Architecture），巧妙解决了组织内部“创新噪音过大”与“算力开销昂贵”的矛盾： ```mermaid flowchart LR accTitle: 企业双模型治理架构 User["全员员工"] --> Sandbox["轻量低阶模型 (全员沙盒)"] Sandbox --> Filter{"想法与落地验证"} Filter -- "未玩懂技术 / 天马行空" --> Discard["过滤消除无效方案"] Filter -- "具备可行性 / 高价值业务创意" --> CoreModel["高阶定制基础模型 (核心工作流)"] CoreModel --> Production["正式生产与核保系统集成"] ``` - **第一层（轻量低阶模型）**：全员开放，成本低廉。这个层级的核心价值在于充当员工的“思维演练场”。管理层透露，业务部门每隔一天就会有人兴冲冲跑来推销“绝妙点子”。那些从没真正用过 AI 的人，提出的想法往往脱离实际、荒谬得令人啼笑皆非；而真正借助低阶模型日常摸索过的员工，哪怕用的只是低精度版本，也能敏锐意识到 AI 的能力边界，提出切中痛点的可行建议。 - **第二层（高阶定制基础模型）**：仅用于承接经过第一层实际筛选、验证过逻辑的高价值创意，并将其整合进公司定制的工作流程中。这套筛选漏斗既保护了昂贵模型的算力开支，又最大化了落地效率。 --- ### [32:19](https://www.youtube.com/watch?v=1qUzAfOphGI&t=1939s) 监管与开源模型之争：险企为何难以转向前沿开源生态 主持人将话题引向硅谷当时的热点：以 Kimi 以及开源大模型为代表的新势力迅速缩小了与顶级闭源模型的差距。但在强监管、重合规的保险业，数据驻留（Data Residency）与供应商注册地限制是无法逾越的红线。这些高性价比模型是否真的能成为主流保险载体（Carriers）的可行方案？ 嘉宾们的反馈呈现出高度的现实主义与克制： - **迁移壁垒与生态绑定**：大型保险机构目前深度绑定于美国本土的头部云服务商。模型一旦深度嵌入生产环境，在交易层面实现整体底层切换是不可想象的工程灾难。 - **综合成本并无优势**：开源模型虽然自身权重免费或价格极低，但当大型保险公司通过受合规认可的主流企业级云服务商去采购托管算力时，算上合规审查与治理成本，最终的折扣幅度根本不足以抵消其引入的潜在运营与数据风险。 - **真正的受益场景**：险企更期待展厅里的第三方 SaaS 供应商能够积极集成这些高性价比模型。一旦供应商因采用开源技术降低了研发与运营成本，险企便能以更优惠的坐席单价采购外部工具，从而间接享受开源生态的技术红利。 正如嘉宾所总结，企业从来不是在单纯购买一个大模型本身，而是购买模型周围的上下文建模能力、安全基础设施和配套工具链。大模型最终会像底层的 CPU、TPU 硬件一样被深度抽象化，具体是哪家模型在底层运行将不再是最核心的关注点。 --- ### [34:33](https://www.youtube.com/watch?v=1qUzAfOphGI&t=2073s) 自研还是外采？用“核心业务（Core）”重新划定组织边界 讨论尾声转向了组织推进 AI 的顶层战略：保险公司究竟应当全面自研自建（甚至激进到用“氛围编码/Vibe Coding”把昂贵的行业核心系统如 Guidewire 替换掉），还是应当聚焦主业、将非核心模块交给外部供应商？ Hartford 与东京海上等代表给出了高度一致的战略定力原则：**严格以“是否属于保险公司的核心竞争力”来决定技术边界**。 保险公司的三大绝对核心业务支柱包括： 1. **风险工程（Risk Engineering）**：如何通过前置的工程与技术手段协助客户防范损失，从源头上减少理赔出险； 2. **理赔查勘定损（Claims Adjusting）**：在灾难或事故发生后，以最高效、最公平的方式协助客户完成定损与赔付修复； 3. **核保承保（Underwriting）**：准确评估并定价各类复杂风险，筑牢保司资产与负债的生命线。 对于这三项核心业务，保险公司必须倾注全力自主掌控；而对于人力资源系统（如 Workday）等通用企业应用或边缘支撑工具，在外部已有专业供应商持续收集全行业需求、不断迭代功能的前提下，保险公司绝不应分散宝贵的工程精力去闭门造车。界定清楚什么是核心、什么是非核心，是避免在 AI 浪潮中盲目消耗预算的关键准绳。 ### [36:23](https://www.youtube.com/watch?v=1qUzAfOphGI&t=2183s) 自研测试基准与 SaaS 分工：将数十年承保经验固化为核心护城河 在探讨保险科技与大模型工程的落地时，嘉宾们深入剖析了保险公司与软件服务商（SaaS）的权责边界。即便团队具备使用 AI 快速生成原型或辅助编码（vibe code）的能力，维护一个完整的生产级软件体系却涉及截然不同的繁重工程：长期的代码维护、严格的合规安全审查、人员培训、市场宣发材料的制作、系统排错、用户入职引导（onboarding）、新功能发布规划、Beta 测试，乃至界面上“按钮该用什么颜色才能提升用户采纳率”等细微体验设计。这些工作天然属于外部专业 SaaS 厂商的范畴，保险机构不应将有限的技术资源消耗在基础软件运维上。 保险机构真正必须自主掌握并重度投入的，是其底层数据的使用层与逻辑评估框架（harness）。这套框架代表了保险公司经过二十年乃至五十年沉淀下来的智慧与专业认知（codification of wisdom and knowledge），包括历史赔付记录（loss history）、各级核保经验、对经纪人网络的认知以及对分销渠道的深刻理解。如果无法将这些无形资产转化为专有的调度与评估基准，单纯利用通用 AI 优化工作流并不会带来独特的竞争壁垒。 ```mermaid flowchart TD accTitle: 保险公司与 SaaS 服务商的分工架构 subgraph SaaS["SaaS 服务商职责"] S1["基础软件开发与维护"] S2["系统安全合规与排错"] S3["UI/UX 与用户采纳体验"] end subgraph Carrier["保险公司核心壁垒 (Harness)"] C1["20-50 年承保智慧与经验固化"] C2["历史损失数据与赔付洞察"] C3["经纪人与渠道专属评估逻辑"] end SaaS -->|赋能工作流| Carrier ``` - 保险公司自研的精力应聚焦于构建反映业务认知与历史风险数据的专有评估架构（harness），而非造通用软件轮子。 - 软件的功能迭代、漏洞修复、安全合规与人机交互设计，应交给垂直 SaaS 提供商负责。 ### [37:23](https://www.youtube.com/watch?v=1qUzAfOphGI&t=2243s) 程序正义与“被倾听的权利”：保险本质上不是冰冷的算法裁决 当 AI 能够将核保与理赔流程压缩至毫秒级时，保险行业深层次的社会契约属性随之显现。嘉宾从法律与社会学概念切入，指出保险不仅仅是一个输出最终结果的黑盒，其核心机制在于“程序正义”（procedural justice）和“发言权”（voice）。在纠纷解决中，过程本身往往就是目的；如同两方对簿公堂，即便判决结果对其中一方不利，只要双方都拥有陈述自身诉求的权利、遵守相同的规则、感受到程序公开公正，败诉方通常也会信服并遵守裁决。 如果由 AI 直接替代这一过程，仅仅给出一个绝对精确但冷冰冰的二元裁决——按下按钮便立刻告知“你对了”或“你错了”——绝大多数客户和经纪人都无法心悦诚服地接受。保险机构在日常业务中投入巨资所维护的，恰恰是一套能让各方声音被充分听取、所有情理与非标细节被综合权衡的公正流程。无论底层算法如何演进，这种蕴含同理心与沟通机制的程序价值在保险业务中永远不可或缺。 - 保险交易兼具契约执行与人际互信，程序的公正与被倾听感比单纯的“算法正确率”更能获得客户认可。 - 若用黑盒 AI 彻底剥离人工审核与沟通环节，可能会破坏维系保险商业模式的根本信任。 ### [38:30](https://www.youtube.com/watch?v=1qUzAfOphGI&t=2310s) 告别“刷 Token 指标”：通过清洁室重构核心业务流程 回顾过往的 AI 探索历程，与会代表坦言走过类似的弯路。在生成式 AI 爆发初期，许多团队为了向董事会快速展示进展，往往会追踪易于量化的高频指标，例如“Token 消耗量最大化”（token maxing）或表层工具的使用频率。然而，大家很快意识到这种唯使用量论的方法完全走偏了方向——企业不能只盯着那些看似容易切入的边缘功能，而必须直击核心业务价值。 为此，团队全面转向了一种被称为“清洁室”（clean rooms）的重构模式：抽调专注于核心难题的关键人员，完全脱离原有业务流程的路径依赖，从零起步思考“如果在 AI 原生环境下，这个核心问题应该如何被彻底重构”。这种机制旨在催生根本性的新型解决方案，待验证成熟后再重新嵌入主营业务体系，从而彻底摒弃了仅在老旧流程表面修修补补或机械堆叠 Token 的无效内耗。 - 摒弃以 Token 消耗量或使用频率向董事会汇报成果的浅层衡量方式。 - 建立跨职能的“清洁室”机制，摆脱原有流程包袱，从底层探索 AI 原生的业务解法。 ### [39:23](https://www.youtube.com/watch?v=1qUzAfOphGI&t=2363s) 降本还是产生 Alpha：保险巨头如何审视初创企业的商业提案 Scale Venture Partners 的合伙人 Alexander Niehenke 抛出了一个直击早期创业者痛点的问题：保险公司的损益表（P&L）中通常充斥着大量的第三方外包成本——例如由于蒙大拿州的核保极其复杂，机构往往选择将业务外包至菲律宾的人工团队，或者长期依赖收费高昂的传统供应商。那么，初创公司究竟应该专注于切入这些高额支出场景，帮助保司把流程做得更便宜、更高效，还是应该彻底重塑业务流程？ The Hartford 的 Andrew Zarkowsky 从保司最关键的财务指标——综合成本率（Combined Ratio）进行了剖析。综合成本率由运营费用（operational expenses）和赔付损失（losses）两部分构成。如果承保 1 美元保费，综合成本率为 85%，保险公司便赚取 15 美分的承保利润。在整体成本结构中，单纯的运营费用占比相对较小，单点环节的降本虽然有意义，但其对应的市场规模（TAM）往往非常有限。Andrew 打了一个生动的比方： > “这就好比我要去买一辆 40,000 美元的福特野马，你跑过来跟我说：‘我有这辆车的凯利蓝皮书（Kelley Blue Book）估值报告，你得付我 1,000 美元。’但我顶多只愿意为此付 20 美元，它根本不值 1,000 美元。” 能够真正打动保司决策层的，是能够带来超额承保收益（Alpha）的产品。例如像 TomTom 那样提供独家且无法通过公开渠道获取的地理与行车数据，或是提供前所未有的洞察维度与获客渠道。另一位嘉宾补充道，当 InsurTech 团队前来推介时，最关键的是“恪守自身航道”（stay in their lane），将某一个单点场景做到极致。保司最反感的是初创公司声称自己把一个痛点做好了，紧接着就要兜售另外 20 个功能——事实上大型保司内部早已有庞大人手在处理那些周边事务。只要外部厂商能在某个高价值单点做到无人能及，帮保司剥离繁复工作并跑赢竞争对手，就足以汇聚成巨大的 Alpha 竞争优势。 | 评估维度 | 单纯降本型提案（Cost Reduction） | 创造 Alpha 型提案（Alpha Generation） | |:--- |:--- |:--- | | **关注指标** | 运营费用（P&L 中的外包/系统成本） | 综合成本率（Combined Ratio）与损失赔付控制 | | **典型缺陷** | 针对单点省钱，价值天花板低（如“花1000美元买车况报告”类比） | 若战线拉得过长，无法在核心单点形成绝对壁垒 | | **保司期待** | 价格极度透明、便宜，直接替代昂贵的遗留供应商 | 独家专有数据（如 TomTom 级数据）、深度行业洞察与核心痛点极致解决 | | **最佳策略** | 仅作为通用工具嵌入 | **恪守自身航道**，在一个关键卡点做到顶尖，协助保司跑赢同行 | - 初创公司若想在保险行业立足，不应过度夸大单点降本工具的市场空间，而应专注于通过专有数据与深入洞察提升核保与赔付表现（创造 Alpha）。 - “恪守自身航道”并把单点业务打磨到极致，远比提供大而无当、横跨数十个周边功能的通用套件更具吸引力。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=1qUzAfOphGI&t=4s) 外面至少还有5个人对保险话题感兴趣，等会场子里肯定能坐满 — Alexander Niehenke > [04:40](https://www.youtube.com/watch?v=1qUzAfOphGI&t=280s) 把同一份保单分发给 10 位核保员，你可能会得到 10 种不同的核保结论。 — Patrick Gallic > [09:11](https://www.youtube.com/watch?v=1qUzAfOphGI&t=551s) 已过试点期，但尚未真正实现规模化（past pilot, not yet at scale） > [22:46](https://www.youtube.com/watch?v=1qUzAfOphGI&t=1366s) 根据卫星图像显示受损率为 37%，您是否认可？ > [27:03](https://www.youtube.com/watch?v=1qUzAfOphGI&t=1623s) 注定失败的主张，是一场走向归零的逐底竞争（a losing proposition and a drive to zero） > [28:00](https://www.youtube.com/watch?v=1qUzAfOphGI&t=1680s) 我们必须把这笔开支控制住 — Alexander Niehenke > [29:45](https://www.youtube.com/watch?v=1qUzAfOphGI&t=1785s) 早期的白炽灯是通过直接燃烧线圈来发光，发光效率极低且大量能量转化为无用热量；而经历技术迭代后的现代 LED 灯，以极低的电耗就能提供更优质的光源。当前的 LLM 调用就处在第一代白炽灯阶段，未来对能耗与 Token 的极致优化是必然趋势。 — Leo Sei > [39:23](https://www.youtube.com/watch?v=1qUzAfOphGI&t=2363s) 这就好比我要去买一辆 40,000 美元的福特野马，你跑过来跟我说：‘我有这辆车的凯利蓝皮书（Kelley Blue Book）估值报告，你得付我 1,000 美元。’但我顶多只愿意为此付 20 美元，它根本不值 1,000 美元。 — Andrew Zarkowsky > [39:23](https://www.youtube.com/watch?v=1qUzAfOphGI&t=2363s) 恪守自身航道（stay in their lane） ## Mentioned Alexander Niehenke, Andrew Zarkowsky, Leo Sei, Patrick Gallic, Scale Venture Partners, The Hartford, TomTom, Tokio Marine Group（东京海上）, Root Insurance, Archipelago Analytics, Sixfold, Spruce, ChatGPT, Copilot, Google, Google Maps, DroneDeploy, Kimi, Guidewire, Workday, Kelley Blue Book（凯利蓝皮书）, Ford Mustang（福特野马） ## Video description Andrew Zarkowsky - Head of AI Underwriting, The Hartford Patrick Gallic - Head of International AI Hub, Tokio Marine Group Alexander Niehenke - Partner, Scale Venture Partners Leo Sei - Chief Product Officer, TomTom Insurers are balancing speed and accuracy with regulatory scrutiny. In this panel, carriers, insurtechs and AI leaders share how AI is transforming underwriting, claims, and risk evaluation—while building safeguards around fairness and explainability. Expect insights into real deployments and how insurers are modernizing without losing trust. [Source](https://www.youtube.com/watch?v=1qUzAfOphGI)

[0:04] Alexander Niehenke: 外面至少还有5个人对保险话题感兴趣，等会场子里肯定能坐满 [4:40] Patrick Gallic: 把同一份保单分发给 10 位核保员，你可能会得到 10 种不同的核保结论。 [9:11] 已过试点期，但尚未真正实现规模化（past pilot, not yet at scale） [22:46] 根据卫星图像显示受损率为 37%，您是否认可？ [27:03] 注定失败的主张，是一场走向归零的逐底竞争（a losing proposition and a drive to zero） [28:00] Alexander Niehenke: 我们必须把这笔开支控制住 [29:45] Leo Sei: 早期的白炽灯是通过直接燃烧线圈来发光，发光效率极低且大量能量转化为无用热量；而经历技术迭代后的现代 LED 灯，以极低的电耗就能提供更优质的光源。当前的 LLM 调用就处在第一代白炽灯阶段，未来对能耗与 Token 的极致优化是必然趋势。 [39:23] Andrew Zarkowsky: 这就好比我要去买一辆 40,000 美元的福特野马，你跑过来跟我说：‘我有这辆车的凯利蓝皮书（Kelley Blue Book）估值报告，你得付我 1,000 美元。’但我顶多只愿意为此付 20 美元，它根本不值 1,000 美元。 [39:23] 恪守自身航道（stay in their lane）

[0:04] Alexander Niehenke: there are at least five more people out there that want to hear about insurance [2:34] Andrew Zarkowsky: actual underwriting [9:11] Patrick Gallic: AI is becoming a supervisor of other AI [10:38] Leo Sei: spot-and-spot replacement [11:45] Alexander Niehenke: underwriting harnesses [14:45] Leo Sei: oops [15:43] Andrew Zarkowsky: just AI that [24:35] Patrick Gallic: super underwriters [27:03] Patrick Gallic: losing proposition and a drive to zero. [32:23] Alexander Niehenke: Dario appreciates you. [34:23] Alexander Niehenke: vibe code [39:23] Patrick Gallic: It would be like if I'm trying to buy a car and you come in and say, 'Hey, I have the Kelley Blue Book on this car. You're going to buy a car for $40,000. You should pay me $1,000 for a review of this Ford Mustang.'... Maybe I'll pay you 20 bucks for it, but that's all it's sort of worth.

[0:04] All right, good afternoon everybody. Thank  you for coming out on this Tuesday. Um, [0:09] get up at any point in time and come closer. We're  going to have about 500 more people coming in and [0:14] this is all going to fill out and so you're  going to want those front seats. All right, [0:16] I promise promise promise you that there are at  least five more people out there that want to [0:21] hear about insurance. So, we'll get those people  in here. Uh, we're going to start with a brief [0:24] round of introductions. My name is Alex Nihank.  I'll be moderating this panel today. I'm with a [0:29] San Francisco based venture firm, scale venture  partners. We manage a few billion dollars. Uh [0:33] we invest across different uh areas within the  AI stack. I focus particularly on the vertical [0:38] industries. I love insurance. I know that that is  a a common statement for everybody in this room, [0:43] maybe not broader across the US, but that has led  me to investing in companies like Root Insurance, [0:48] Archipelago Analytics, Sixfold, Spruce, a  few others. And with that, let's just kind [0:53] of go across the line, let everybody introduce  themselves. Yeah, thank you very much. Uh Andrew [0:57] Zarowski, it's great to be here. I work with the  Hartford Insurance Company. Uh we are a 200 plus [1:02] year old company that does a lot of different  types of insurance, not health insurance, [1:05] but everything from property casualty, corporate  insurance. Uh this is my fourth year here. I'm [1:11] very excited to be on stage this year. Uh I  have been with the company for about 20 years [1:16] and that entire time I'm on the underwriting side  within our commercial, middle market, and large [1:21] commercial practice. All right. Thanks. Uh so  my name is Leo Se. I lead product and marketing [1:28] at TomTom. Um you might wonder why is TomTom on  the stage um because we're very well known for [1:34] having satellite navigation 25 years ago. But  fast forward to today we're a B2B geospatial [1:40] uh data company. And what's very interesting at  this moment is we have representations of traffic, [1:45] road addresses, uh accidents, lane closure, you  name it, speeds and all of this is now extremely [1:52] useful in a lot of context including insurance.  Um because with AI you can use this data in your [1:57] workflow a lot more easily than before where  you required special skills and special tools. [2:02] So now with the event of AI in the last two years  we've seen that location data is extremely useful [2:07] especially in insurance where we work a lot. Good  afternoon everyone. My name is Patrick Alec. I'm [2:12] part of Tokyo Marine. If you haven't heard  of Tokyo Marine, it's the largest and oldest [2:16] Japanese insurance company. And over the last 20  years, they've purchased about a hundred other [2:21] carriers globally um and rolled them up into um a  couple dozen group companies. So my role is I lead [2:28] AI across all of the group companies um outside  of Japan. Awesome. Thank you. Let's start with a [2:34] layup. Um just broadly, you guys all work at large  companies. uh we can be insurance related, it can [2:40] not be insurance related, but what is something  that that you've done lately on the AI side that's [2:44] that's excited you that's blown your mind. Maybe  we contextualize it in the concept of of ROI or [2:49] even you know what kind of alpha it's generating  for your organizations and I know we're going to [2:53] bounce around some of these questions and some but  like I think this is one that almost all of you [2:56] can kind of tackle and and I think it's just kind  of fun because I don't know I've sat on a bunch [3:00] of these panels and sometimes it feels a little  abstract what folks are doing and the more we can [3:03] get tactical I think it's just a great way to open  things up. I I'll start. I appreciate that. So, [3:10] at the Hartford, one of the things that we really  focus on, so as I mentioned, I'm an underwriter. I [3:14] and I transitioned from being an underwriter for  about 20 years into the head of AI underwriting. [3:19] And if you're wondering what that title means, we  literally made it up because we didn't have that [3:23] role before. I I now run a group that what we're  trying to do is trying to figure out how do we [3:28] take AI and and help our underwriters. And when I  think about this and some of the stuff that we're [3:33] doing today is that in the underwriting process  for anybody in the room who's been an underwriter, [3:37] you think about the really high value stuff  like you get information, you try to customize [3:42] a program, you try to solve problems, you  try to give risk engineering solutions, [3:46] how do you prevent a loss from happening if a  loss does happen, how do you get to the right [3:51] types of claim service? Uh how do you build  relationships? All those high value things. [3:56] But for any of the underwriters in the room, you  know that to get to that, you have to do a whole [4:00] bunch of other work. And that tends to be a lot  of gathering information, finding information, [4:06] summarizing information, which a doesn't add  value, and just frankly, it's not fun. And [4:11] so what we're trying to build, and this is this  is what's really interesting and exciting about [4:16] what we're doing right now in the organization is  trying to get those tasks automated so that when [4:21] you're an underwriter, you sit down and you would  do what in my opinion is actually underwriting. [4:26] You don't do any of the gathering. You don't  do the summarizing. It's just sitting in front [4:29] of you and you're able to pursue the highv value  activities like working with customers to solve [4:34] their problems. That's what our underwriters want  to do and that's what AI is allowing us to do. [4:40] Yeah, I'll just jump in um a little bit because  I I love that question around like what are you [4:44] actually doing? Um but more importantly like so we  talk about ROI, but it's really like what's alpha? [4:49] How is your return better than sort of the market  average? So when we think about it, there's three [4:53] different ways that we're implementing AI. Um  the first one is like individual productivity, [4:58] but I will say that there's um maybe an ROI,  but there's definitely no alpha on that. So [5:02] your chat GPTs, your co-pilots, all those  tools. The way that we think about it is [5:06] um you know sort of like handwriting. At one  point handwriting was very innovative, right? [5:12] People were trying to figure out like commas and  periods and spaces and and like the monks that [5:15] was like a a big deal. And now it's a form of  punishment. If you have to handwrite something, [5:20] maybe in school you guys did that where you had to  like handwrite stuff. Um I think if you don't give [5:24] access to the tools to your employees, basically  it's a form of punishment and people aren't going [5:28] to want to work there. And the second area is  on workflow optimization. And I think you you [5:31] touched on some of that. Um, and I think that's  really important. It's good. I think eventually [5:35] though in terms of alpha for the company, you're  not going to really create alpha on just straight [5:39] up workflow optimization. Um, because the the SLAs  or whatever your time frames are just going to [5:44] compress. If you think about your like submission  to quote process, if it typically takes like a [5:49] week um, and the rest of the market now can do  it in a day, then you're going to be adversely [5:53] selected because they're going to say, "How come  you can't compete with everyone?" So you might get [5:57] like a one to two years competitive advantage um  to create alpha but then after that if you kind [6:02] of just don't keep a pace with it you're going  to be adversely selected. So where I think it's [6:06] where I'm most excited about it is where we're  building out actually organizational capability [6:12] where you have a differentiated product offering  or approach to the market primarily because it's [6:18] how you use your data. Um and I think it's a use  of your data to sort of look at your risk appetite [6:23] to look at where your attachment levels are where  your policy wording sort of where that I think you [6:27] were talking about this bit as well where you  differentiate yourself from the market because [6:31] you have access to that intelligence. Um and the  part is sort of like codifying the wisdom of your [6:36] individual underwriters. Individual underwriters  as you guys know if there's any underwriters um [6:41] you give the same policy to 10 underwriters you're  probably going to get 10 different answers. Um but [6:46] what do you do? You hire an underwriter because  they're really smart. you trust their wisdom. Um, [6:50] and that wisdom sits with the underwriter. And the  idea here is how do you actually codify that and [6:54] bring it in and then make them upskilled. Um,  my my last analogy on that one is sort of like [6:59] the the weatherman. Um, I don't know if any of you  guys uh when I was like in seventh grade, we put a [7:04] stick outside and if the weather if the stick went  up or down, you could see if it was going to rain [7:08] or not rain. Um, and then it was like barometric  pressure. You looked at like barometric pressure. [7:13] But now the weathermen have access to satellite  imagery. They have all these advanced AI models. [7:18] that can do lots of stuff. And what happens as a  result, we consume a lot more weather data. Um, [7:22] we get alerts on our phones, we see what's coming,  how we're going to access it. I think that's [7:27] where I'm most excited. And we're seeing that um  definitely emerge now across many of our different [7:32] um lines of business where where they're starting  to be to create alpha. I would say it's the [7:36] beginning of the journey. And I would um jump on  this because I really like your explanation of [7:41] alpha and really the differentiated context um as  opposed to just ARI. For me, it's very exciting. [7:47] And I've been the AI world for a long time. Spent  a decade at Google before that. Um and and as I [7:53] discover the uh insurance but also the overall uh  world uh outside of the geospatial navigation. If [8:00] you look at AI, the LLM models are getting  commoditized. We we hear all of them getting [8:05] better and better, but they're getting very close.  And so it boils down to context. And I think to [8:08] your point, the differentiation is going to be  can you get a better context for your model? And [8:12] that's where I've been very surprised and very  excited to see that what I thought in terms of [8:17] underwriting being already extremely optimized.  You come in with richer context and you can still [8:23] uh yield a better return. So we have a good  example of a insurance moving from census and zip [8:28] code to road level uh scoring with speeds and and  traffic that we have and yielding 2% better which [8:35] for me at the insurance world seems huge. And  we're also working um with an autoinsur to pilot [8:41] some uh claim validation where someone sends okay  I got to you know crash this particular road this [8:45] particular time and thanks to the very accurate  traffic data that we have and incident data we're [8:50] able to say well this uh claim is likely valid  or likely fraudulent and suddenly it's opening [8:55] a huge set of uh uh uh claim validation and fraud  uh verification for them. So I agree I think going [9:01] for the alpha is where do you differentiate  context? Can you go further? And that's where [9:06] it's going to be uh your differentiating factor in  that space. Where are we in the hype cycle as it [9:11] comes to insurance underwriting? Um well, I'll go  first. Give me give me your quick hot take. Yeah. [9:17] Yeah. Yeah. Uh I would say that we are past pilot.  Um we are not yet at scale. Um and and this is I I [9:23] think there's two parts to that. One is um I think  every single organization probably can raise their [9:28] hand if I say do you have an endto-end process  that's fully automated? AI is there. it's making, [9:32] you know, you know, intelligent decisions.  Probably everyone in this room will raise their [9:36] hand. And then I ask, how many lines of business  do you have? And they go, I have 120. And how many [9:40] have you scaled this out to? One or two. Um, so  I think that we're beyond sort of the pilot. And [9:45] I think you are scaling in areas. Um, but it's  not sort of widespread. This is my view. It's [9:50] not sort of um we're at the scale phase and and  I think there's a challenge because of that. um [9:55] mostly that with uh Agentic what you're doing is  you're shifting the liability to some type of an [10:00] agent to take that action. Um and in that process  right now we really don't have the right controls [10:06] in place to be able to scale that efficiently and  using a nondeterministic process for a regulated [10:11] market. Um and so I think that there's a lot of  sort of internal work that's done and what you're [10:16] seeing is that there's experimentation on more of  like the chat customer support internal internal [10:20] features uh where AI is becoming a supervisor of  other AI and trying to figure out is that a way [10:26] to scale um but in terms of like broad across the  market you know that my view is that we are beyond [10:32] pilot we have not yet fully scaled I sorry I  was just going to say well or you go Leo and [10:38] then I got a question for Andrew. Okay. No. So, so  I agree. I think uh what's interesting is the hype [10:43] usually you still don't see the sign. And here for  us, we have all the proof point that it's working. [10:47] But I agree with you. I think the the difference  now is how do you scale this and how do you make [10:51] this a reality? And there's an analogy that I that  I heard that I really like and applies to this is [10:56] um it's it's the difference between now you saw  that it work but you need to resist the urge of [11:00] just replacing an existing tool of AI and instead  transform your whole process and that takes time [11:05] and it's complex. Um someone uh gave me an example  that might be interesting when when electricity [11:10] came around and before that it was steam engine.  And so you had this big steam engine at the bottom [11:14] tower and then huge geared and shafts and belts  and gears all the way to the top with all the [11:19] machineries. And so when electricity come around  the first thing that the people that is like well [11:22] they replace the bottom engine with an electric  uh engine instead. Uh but you still had all the [11:27] belts and all the gears and all the complexity  instead of saying well actually now I can have [11:30] a swing machine with its own engine. So you need  to move from okay we know that this new technology [11:36] work but between that and having it replace  the work the the workflows the processes it's [11:41] not just a spot-and-spot replacement and that is  where it's hard and so I think that's the phase [11:45] we're in I was going to ask you like if I look  at the um the startups that I've seen over the [11:50] last couple years maybe two years ago the startups  would come in and they'd say we're going to insert [11:54] ourselves somewhere in the underwriting pipeline  um you know the submission comes in and then we're [11:59] we're going to augment that we're going to add  a bunch of data maybe based on internal maybe a [12:02] third party resources. Then maybe a year ago,  that conversation changed to actually we have [12:06] these agents because there's exception handling  and we're going to send an email to your manager [12:10] and we're going to go grab this from the customer.  We're going to bring it in. And then in the last [12:13] 3 months, people started talking to me about and  actually we're building an underwriting harness [12:17] because obviously that's the invogue word as of  3 months ago. And I don't even know if we've if [12:22] we're even got the thing that we were talking  about 2 years ago productized yet in most and [12:26] like like like you you walk around some of these  conferences and you see the booths and you see the [12:30] words and I'm like are we are we where we were two  years ago? Are we where we are? What what's your [12:34] read on that? Yeah, there there's a lot of back  and forth right now going on and and I've been [12:39] doing this now for about two and a half years and  you do see even in the last two and a half years [12:43] you see companies that are really trying to push  one way and then push another. I I actually had [12:47] a a vendor that we work with and they do great  things for us in our claims department. And they [12:51] came up to us and said, you know, we're going to  do what we do in claims and we're going to do it [12:54] in underwriting. And I said, that's awesome. I  said, how many people do you have in that part [12:58] of your group? And they said, we just hired  our fourth person. And and I said, you know, [13:02] we we're we have 19,000 people. We have hundreds  of underwriters. You you [snorts] you can't solve [13:08] with four people. You're just you're just not  there yet. But that's what you try to get. You [13:12] get into that mentality of we could just take this  and move it over there. Uh we we are an industry [13:18] that likes predictability. We like consistency.  We like routine. Um if you know anything about [13:24] uh catastrophe or property underwriting, you  know that we use things called flood planes. [13:28] We look at a 100redyear data, 250-year data.  So we want to have something that is we know [13:35] is reliable. We know is there. So if it's the new  buzzword as of 3 months ago, that doesn't really [13:40] resonate with our industry. Yeah. And and and  there's one other thing I want to add on that. [13:44] 100% agree. Um is that when you're talking to the  startups, the problem is that once they get really [13:49] good product market fit and you go, "Hey, this is  great." Their pricing changes very radically. Um [13:55] so I don't know. You mean it gets more expensive?  Yeah. Yeah. And so the whole the whole structure, [13:59] right, it moves from, you know, a seat license  to a token base to a a you know, a credit system. [14:05] Um and where the challenge here is that if you  really want to create alpha, what you're doing [14:09] is you're transferring some sort of knowledge or  process over to AI. And so when you do that, it [14:14] really instead of becoming an asset that you have  internally, it has to move over to the liability [14:18] and risk area. So now once I transfer that over  to some, you know, a startup here and say, "Hey, [14:23] you guys can do this process or you can take this  off." And I have a one-year contract with you, [14:28] it works really well. And then you come back  and say, "Hey, we realize we're not making money [14:32] on this. We're going to have to 10x the cost,"  which is happening across the board. Um, and you [14:36] sit there and go, "Well, hang on. If we really  optimize around this new process, we're going [14:40] to put ourselves really at a disadvantage. Leo,  Leo, do you think this conversation is distinct [14:45] for insurance or I mean you cover a broader  set of industries or do do you think this is [14:48] just the enterprise conversation across multiple  verticals? So I think there there are stuff that [14:53] is common to the enterprise. So what you mentioned  between the the pilot we passed the pilot phase [14:57] in the adoption. I think what you mention about  the um you know predictability uh auditability [15:03] trustability all of those things are very specific  to insurance and so I think that there are common [15:08] uh challenges and there are challenges that are  very unique um around this uh auditability of of [15:13] all the decision because to your point you cannot  just transfer your liability uh to an agent like [15:18] in in some industry it's fine you know they got  they got that wrong oops in your industry it's not [15:23] it's not uh it's not that easy how do we solve  that like I let me play devil's advocate like [15:31] should we just say f it we're not going to do AI  in the insurance industry just too hard right like [15:36] uh what are we all doing so I so that that is a  very it's a very interesting question we do come [15:43] across this all the time and I think that the  last presentation that I was at talking about [15:47] do you need a gentic AI for everything I think  that people have the tendency who are who are [15:53] not in the space who are more executive level  or leadership level that are not so deep in AI. [15:58] They have the mentality of, oh, just AI that, you  should just AI that. And and clearly that's not [16:02] the right answer. What we do need though is we  need efficiency and we need process change. Um, [16:08] I started underwriting almost 25 years ago and  basically we do the same thing we did 25 years [16:14] ago and and it might be even longer than that. I  just know the last 25 years and that just doesn't [16:18] make any sense whatsoever. The other issue that's  hitting our industry, and I this is probably a lot [16:23] of other industries as well, but the majority of  really good underwriters are probably 50 something [16:31] and up. And we do not have people as fun and  as exciting as insurance is, as you mentioned, [16:39] um it's not the thing that most college grads want  to do right away. So, we don't have the talent. [16:44] We're going to need that amount of resource. We're  going to need some more or we're going to need [16:48] to get better and more efficient. And I I don't  see I see the retirements coming and I don't see [16:52] the as many people coming in to fill those gaps.  Yeah. And I think so like does does insurance need [16:59] AI? Absolutely. Um I'm I'm like the most excited  person about AI. I think it's awesome. Um and and [17:05] why good good because you know we're spend a lot  of time here. [laughter] Um [clears throat] well [17:10] and and why is it because fundamentally like at  its basics what is insurance? Insurance is about [17:15] taking risk on from other businesses so that way  they can operate and they can move it over. And [17:20] I think the world is getting increasingly risky.  And if you look at um like the percentage of of [17:25] commercial submissions that are actually getting  placed um there there's [clears throat] a large [17:30] amount of risk that exists out there that  is not being placed. And I think that if [17:35] insurance if insurance using AI can actually  open up our aperture and create new markets [17:40] or better underwriting or more predictability, I  think that's amazing. I think a lot of like you [17:44] were saying in our workflows how they've stayed  the same um a lot of that is like institutional [17:49] just codification of bure bureaucracy um and as  we can automate that that'll be great and the [17:54] reason why I think everyone in this room even if  you're not in insurance you probably buy insurance [17:59] um and all of that will be passed on to you um  because ultimately that the market will have a [18:03] combined rate that you know they won't let us  go above and in a soft market the rates are [18:08] going to drop and I think that we're seeing as  as there's markets that we can't ensure right [18:13] um like in California and other property markets,  there's risks like cyber with aggregation that are [18:18] more difficult to ensure. Um using AI, if we can  actually allocate our resources better to provide [18:23] better coverage to people at a more reasonable  price point, I think it's a win-win for everyone. [18:27] Yeah. And I I could just tell you how how it's  going to happen, at least from my view. So today, [18:32] and I and I just use uh property insurance as  a very easy example. So today, you could have a [18:37] company that let's say has 10 locations. Normally,  if you have 10 locations, there's usually one or [18:41] two really hard ones and probably like eight of  them that are just super easy. Right? Today, an [18:46] underwriter needs to go through all 10 locations.  So, they can spend only a certain amount of time [18:50] on those hard ones. In the future, AI is for sure  going to be able to handle the eight easy ones, [18:56] maybe even help you out on the 10 hard ones,  which means that you can pursue harder things. So, [19:01] how do we understand how to ensure wildfire in  California? That's a really hard thing to do, [19:06] but you're never going to get there if  you have to spend all of your time doing, [19:10] you know, Ohio easy stuff. You you have to be  able to have the time and energy to do that. [19:15] And AI is going to open up that time. That's how  these things are practically going to get done. [19:19] I remember I was talking to an underwriter a  few years ago and he was telling me that his [19:23] superpower when the statement of values come in  that have all the addresses of the properties, [19:26] his superpower was to go into Google Maps, look at  the building and crossverify it to the information [19:32] in the spreadsheet. And I thought to myself,  there might be an opportunity there. And this [19:36] was somebody who had been in their 50s, had been  in the industry in a long time. Leo, you've kind [19:40] of alluded to it, but like TomTom seems pretty  excited about insurance. Um um are you guys a new [19:46] vendor in this industry and and and how do this  all come to buy and how does that narrative fit [19:49] into the AI narrative? So So we're not new. We've  been working with insurance for a while, but kind [19:54] of the old way. So we supply data and and some of  those big insurance, they take that data and they [19:58] uh crunch it to their underwriter or to the claim.  Um I think similar to what you said I think for us [20:04] it's also a new opportunity um because as I said  in the past location data is very complex so you [20:09] need special scale special tools and so as a as  a business executive or as a business function [20:15] sometime you you don't really know whether  the ROI will be there and the cost to test [20:19] is too high and so now with AI and and the work  that we're doing to make the location data very [20:24] um um vectorzed and and with embedding so that any  agent with no location data can can understand it [20:30] any business especially in the insurance can  test it out. Um and so to your point you just [20:34] mentioned someone was just receiving addresses  and validating on Google maps that's something [20:38] we can do quite easily and and sometime it will  matter right if you're on the right side or left [20:42] side of the road for a flood line or a fire line  that's going to be a big difference if you're on [20:46] a big complex lot and you're at the far out on the  forest or at the forefront on the streets in your [20:50] fire risk it's going to be different so all those  things cannot be uh uh very very interesting and [20:56] I really uh on the previous question I wanted to  add something that you said that I think was very [20:59] interesting to your point I think there's many  ways to look at this one is like okay looking for [21:05] uh efficiency and that's what a lot of people  do try to improve the process and automate [21:08] them but I agree there's definitely a lot of new  opportunities that that arise whether it's more [21:13] complex risk scoring that you couldn't do in the  past like flood modeling through the terrain and [21:18] the the things that you have to try to predict it  uh or even just like I'm going to take a personal [21:23] example here for a second but like as individuals  and maybe even more as businesses figuring out [21:28] what's your current coverage where are your risk  spots where are you double insured like I build [21:32] my own insurance 360 to just read through all  my policies and figure out what am I covered [21:36] and paying twice for, what do I miss, what's my  score. You apply this to insurance to the business [21:41] world and B2B insurance to your point. Maybe you  eventually will have AI liability insurance like [21:46] all those things will open up new businesses. So  I think there's there's many ways to look at this [21:49] and to be excited. By the way, plug uh he wrote  a really cool blog post on on just context for [21:55] models. Spends I'm googling his name. It's worth  a read. You were going to say something. Yeah. [21:59] So one of the things that's exciting about some  of the data and I think TomT time there's a lot [22:02] of vendors out there that have really good data  is that as an underwriter you have to understand [22:08] that they're doing 99 things and a lot of it is  not what I would call underwriting. They a lot [22:13] of folks in the what we call underwriters are out  talking to insurance brokers. They're out selling. [22:18] They're building relationships. They're doing a  lot of things. They are not data scientists. So [22:21] if you throw a hundred pieces of data at them they  don't know what to do with it. So it's not just [22:26] about give them the data. It's give them the data  and then help them figure out what is the data [22:30] that's actually important that is in meaningful to  this specific customer. Uh that is something that [22:35] we really struggle with right now in the industry  because we just give we overload our underwriters [22:41] with data and then they don't know what to do with  it or they do the wrong thing with it. Yeah. Can I [22:46] add one other story on on insurance? Absolutely.  Um [clears throat] so I think the other thing [22:50] that AI can do is help build trust. I think for  in for a carrier um our relationship with the [22:56] insured is really based on trust and it's really  not about like a one-time hey we're going to get [23:00] you when there's a claim. Um and I'll just give  you like a little example. So in crop insurance [23:05] uh the way that I don't know if you guys know but  the the process of if there's hail damage in a [23:10] field of corn what it looks like. So what would  happen is you have a you know a farm with maybe [23:14] 10,000 acres on it. um you know, it's this massive  farm and this storm came through and and there's [23:20] hail damage all over it. And so the adjuster  shows up, they drive out in their pickup truck [23:24] and they walk out and they say, "Okay, which way  should we go?" And they walk through the field and [23:28] they measure where the breakage is on the corn,  what area has been hit and how much has been hit. [23:33] What always happens is that the farmer knows, hey,  I'm going to put the guy to the left in the field [23:38] that looks completely decimated, right? And he's  going to go, "Hey, it's a 50% of my crop is lost, [23:43] right?" the the adjuster might say, "I'm going to  go to the right, right, not knowing." And in that [23:47] area, there's actually no loss. And so what what  is there? There's like a mistrust between the the [23:52] farmer and the carrier. And they both think that  each other is trying to take advantage of them. Um [23:56] just using like geo satellite imagery with AI on  top of that and detecting what percentage of the [24:02] crop has actually been damaged and how it's been  damaged means that your adjusting process goes, [24:06] hey, it looks like it's 37%. Here's a picture.  Do you agree with it? Yes, you sign off. [24:11] Now there's actually trust between the farmer that  hey we don't think you're trying to take advantage [24:15] of us. You're not trying to take advantage of  us. We're really trying to create trust in the [24:18] system. Um and that's another example I think AI  can be helpful. Yeah. Uh I invested in a company [24:23] called Drone Deploy that was doing this with  with with drone data. We sold them last week [24:27] so I don't need to plug them anymore but you know  if this was two weeks ago I'd say buy that that [24:32] technology. Um hopefully this will be entertaining  for the uh for the audience. Uh, Andrew Patrick, [24:38] uh, uh, forgive me, but I'm going to see if  I can get this pulled off. Where do you think [24:42] your two organizations disagree about AI and  underwriting or where are you taking different [24:46] approaches where, if you had to guess, I'm curious  where this goes, let's see, you know, so I I don't [24:53] know much about your organization, so I I can't  really speak about where we where we disagree. Uh, [24:57] but where are you controversial in your approach  maybe? So, so one of the things that we are [25:02] uh and I again I'm not going to talk about any  other of my competitors, but one of the things [25:06] that we are very very focused on at our company  is that we believe I personally believe and we [25:11] believe in the value of the subject matter expert  which which I would define as the underwriter or [25:16] the claims adjuster or the risk engineer whoever  is a person providing value. We do not look at [25:21] this as a situation of how do we get rid of our  underwriters. We look at as how do we empower our [25:25] underwriters and make them super underwriters. Uh  that's where I think that there's value. Uh when [25:30] I look at our claims adjusters, we want to make  them super claims adjusters so they could have [25:35] all the information respond back immediately. We  I saw this in my own personal career when I was an [25:41] underwriter and then eventually I became a leader  of other underwriters. When you are an underwriter [25:46] and you have that knowledge, you go out into  the marketplace, you're talking to distribution [25:50] partners, you're talking to customers, you will  wow them with knowledge. If you're deep enough [25:54] and you're you're dedicated to your craft, you're  going to give them solutions that they can't get [25:59] from anywhere else because you're spending all  that time and effort and energy figuring out how [26:03] to wow them and how to create custom solutions for  them. That cannot happen if you spend all of your [26:08] time sitting going through data, gathering data,  and all that mundane stuff that doesn't add value. [26:13] So what what I see as the value proposition which  I I know that there are companies out there that [26:18] are are you know we're going to say we're going  to use AI and then get rid of our underwriters and [26:22] claims justice or not get rid of but have a lot  less of them. We are not looking at like that at [26:26] all. We are trying to empower the underwriters. We  want to have more underwriters because we want to [26:30] have more of those people that are going to create  the better experience for our customers. Um first [26:35] I would say we actually partner on on a couple  things. So try guys. I'm trying real hard here. I [26:40] will say though um in [clears throat] terms of in  terms of the controversial take I think there's a [26:45] number of insurance carriers who have come out and  said that there is um job displacement because of [26:49] AI. Um and you can see that that um I'm not going  to name names but you can look in sort of their [26:54] annual reports and and their filings of like what  percentage of their staff they think are going to [26:58] get replaced because of AI. Um I don't believe  that at all. Um I think that it will probably [27:03] be the opposite. I really don't think that we're  going to have a reduction of staff or replacement [27:07] of staff. I think if anything, like you were  saying, we need more people in insurance. Um, [27:11] and I think that there's a lot more that we can  do to be more efficient. Uh, across our groups, [27:16] we that's not part of any of the conversation  where we think that we're using AI to replace [27:20] people. Um, and I think honestly that's like a  losing proposition and a drive to to zero because [27:26] um ultimately insurance the the product that we're  so the revenue that we take in for the amount of [27:32] liability that we're taking on um getting that  right or getting that wrong is much more important [27:37] than sort of the operational expenses that go  into it. Uh if you think about just like your [27:40] home policy, if you pay $3,000, what if you have  a loss and it's a million dollars? Um that $3,000, [27:47] how much operational expense is there in us sort  of processing that? Maybe we can save like $10 by [27:52] automating a process, but if it means that we  take on a couple bad risks, um that's that's a [27:56] really terrible thing. So I think it will be  it will not really be on the the replacement [28:00] of people. Do you guys um are are any three of you  major consumers of the um the large foundational [28:07] model providers? Um and and is it is it purely on  your engineering teams? Is it in your underwriting [28:11] teams? Obviously for Tom Tom that question  is going to be slightly different, but I'd [28:14] be kind of curious where you might get yourself on  adoption usage spend of those companies. You don't [28:19] have to disclose numbers. Yeah. Uh we use we use  everyone um like large big dollars. Is it a big [28:24] expense for you where your boardroom sitting and  saying, "Man, we got to get these under control." [28:27] Uh it's definitely growing. Um so I think that if  you think about it adoption sort of in my three [28:32] layers um for enterprise access um nearly  everyone has access to some type of tool um some [28:39] sort of chat interface. I think on the workflow  automation side that is growing um and I think [28:43] as you hit scale sort of the return on investment  is there as an individual but when you scale it's [28:48] taking on the liability and all the the factors  related to that. Um and that's becoming a very [28:53] big conversation around like how much do we scale  and how quickly. Um, and then on on sort of the [28:58] the change, that's where the biggest budget goes  right now into sort of like the organizational [29:02] change. I will I just just say one thing because  I know that this is going to come up what you're [29:06] getting at which is like token maxing um and  and the cost. I think this is where it's really [29:11] interesting. Uh AI as a non-deterministic solution  to create like a workflow for a deterministic [29:17] process or for R&D or for new discovery or for  programming something that's not just like a [29:22] codified process that you're optimizing. Um, it's  great to spend money and I think the way that you [29:27] spend money is like in stagegate and so you can do  your token maxing sort of giving it to people that [29:32] are doing pure R&D. If someone's role is just  on um the process part uh having them consume [29:39] like an unlimited amount of tokens to do the same  amount of outcome um becomes a much more difficult [29:45] conversation. Yeah, I think we also we use them  quite heavily. Um and it's definitely a discussion [29:51] in the boardroom. I I I'm not too worried for  two reasons. Um but but I think uh first right [29:57] now we're still lucky because we're paying per  seats. Um and to your point earlier like we know [30:01] a lot of the vendors are moving per token and we  know that next year is going to be a difficult [30:05] discussion. Um but there's a couple things that  make me not too worried. One is I think it's it's [30:10] um we're still at the age where a lot of people  will you know dump PowerPoints and PDF in in [30:15] essence token in is not optimized today. Um  and and there's a lot of easy wins there by [30:20] uh compressing and compiling your knowledge. Um  and and to the point that you made earlier, I I'm [30:26] deeply passionate about this topic. So if you want  to know more, you can Google my name. I'm writing [30:29] a lot about this. But I'm I'm not too worried. I  think it's going to get better. Um um the way the [30:34] analogy that I look at is okay, when electricity  came around and you had the first, you know, [30:37] light, it was a light bulb and it was burning  to make light, right? And extremely inefficient. [30:41] And you fast forward to today, you have the  same light to LED and it consumes, you know, [30:45] uh a lot less. So we're going to optimize our  usage for sure. I'm not worried about this. The [30:49] second is the market itself is going to uh uh you  know uh be healthy. We see all the all the uh open [30:55] source weight model coming slashing the prices. So  I think LLM are are just being reasoning machines [31:00] and they're going to commoditize and the price are  going to go down. The the fight's going to be on [31:04] the context and the ecosystem around and not so  much on the token and the LLM because we're going [31:07] to learn how to optimize and I think the market's  going to regulate. So yes, it's a topic but I'm [31:12] not too sure about it. It's definitely something  that we're we're very conscious of. We're thinking [31:17] about how to optimize, but we actually have two  different approaches. We have one foundational [31:21] model that we use for all of our custom processes  and and and a workflow processes and things that [31:25] we're really trying to build. And then we have a a  different model that we give really to everybody. [31:30] It's a relatively low weight model that um allows  people just to essentially play around with [31:35] it. And I I kind of like those two approaches  because what we found was is that when I could [31:40] tell immediately when somebody comes up to me I I  get every other day somebody has a great idea. You [31:45] could tell if they've really used AI or not. You  know the people who use it once or twice they come [31:50] up with me with these ideas that make absolutely  no sense. I mean obviously I'm not going to laugh [31:55] at them but they just they just really don't make  any sense at all. the people who've played around [31:59] with it. Even the lowfidelity not great versions  of AI, if they've played around with it, they get [32:05] the what is a good idea and what is a meaningful  idea. And those ideas, what we've actually already [32:09] seen is that those are the ones that we then  take and then we put into our custombuilt model [32:14] that it's a more expensive model, but they've  kind of been vetted to some degree already. Uh, [32:19] which is making it a lot more efficient for us.  I mean, one of the topics that we've touched on [32:23] a few times here is is trust, accountability,  auditability. Um, It seems like July the topic [32:29] in the valley was was was Kimmy and the open  source models and how they've just closed the [32:32] gap. Do you guys think that those models given  you know data residency where those vendors are [32:38] doiciled? Are those a plausible uh solution for  for a meaningful insurance carrier at this point [32:42] in time or is is that something that like you're  you're tracking it? It's curious but but yeah I I [32:48] um where's the conversation on those because it's  it's a major topic in the valley right now. Yeah, [32:55] I I would just say from our perspective, we are  uh heavily focused on large US-based companies. Um [33:02] and and I don't want to say that we're not looking  and paying attention, but you couldn't imagine [33:08] next month switching over. It it would be uh I  mean not not just from a a a cost perspective, [33:14] but just transactionally, how do you switch  when you're so embedded? Yeah. It it just I [33:19] can't see it making any sense. Dario appreciates  you. Yeah. Uh I think that there is optionality [33:26] um and and depending on like your model routing  where you're going to use it. The question is do [33:31] we need it right now uh at this price point? I'm  not sure that like Chem3 has a price point by the [33:36] time we we buy it from uh a major cloud provider  that we're getting that much of a discount for [33:41] sort of the risk that we're taking on. Um I  where I really am hopeful is [snorts] that [33:45] um everyone in the the ex in the uh exhibit hall  um who's selling us sort of seats and and that [33:51] kind of thing are going to use Kimmy and that's  going to reduce their costs. Um then that's going [33:55] to pass back on to us so we can have some more  priceability and and I would add I think I mean [34:01] reflect on what you were saying I don't think as  as corporation we're really buying the model we're [34:04] buying the context and infrastructure around and  the context modeling and and all the tool that [34:08] comes with it. So it might come a world where  you know the the ecosystem switch models around [34:13] but for you it's an abstraction just like at some  point you know TPU CPUs you just have you know an [34:19] infrastructure around doesn't really matter for  you because you're using a cloud. Um so I think [34:23] we'll get to that point. Um but yeah um how much  should we all in the audience and how much are you [34:33] all in your in your companies pushing AI generally  versus you know make I AI available like you kind [34:41] of touched on this too like how do we how do we  get our employees along do we need to push AI [34:46] products do we need to just empower employees  to use AI do we just need to use our engineers [34:51] um if I think about there's just so many different  ways that you can improve. Like you guys could [34:56] decide to quote unquote, you know, vi I don't  know what you use on your underwriting side, [35:00] but we we're going to vibe code guidewire  out, right? Because that's an expensive [35:03] relationship or you could just bring in really  really cool tools on on on top of that. Or you [35:08] could just allow your employees to to go out  and do whatever they want to do individually. [35:13] It's it's a complicated topic. I I mean, one of  the ways that we're looking at it is that we're [35:18] trying to say what is core to our organization and  what is not. Actually, that's a that's a great way [35:24] of phrasing that question. Maybe each of you  can kind of talk. What do you think is core [35:26] to your organization? I like that. Yeah. So, when  we think about core to our organization, we think [35:31] about risk engineering. How do we help our clients  not have claims? We think about claims adjusting, [35:36] which is how if they have the claim, how do we  help fix that as best as we can and we think about [35:40] underwriting and and there's some other obviously  there's some other things in the background too, [35:43] but those are our core operations. So, if you're  not that, uh, the opportunity for us to go find [35:49] a vendor to do that for us, uh, is is in our mind  more optimal because you just don't have the the [35:55] the need to go build all that knowledge inhouse  when it's not really part of our core operation. [36:02] I I completely concur with that. I I think here's  the the question is um maybe this is maybe this is [36:08] the controversial take here is that um if it's not  core to claims and underwriting and it's like your [36:13] workday application or if it's your other sort of  uh applications that you have there's a reason why [36:19] there's a salesperson who comes in who's looking  at all the feature requests who's rolling out new [36:23] features who's training people who has you know  the the marketing materials who's making sure when [36:29] there's bugs in their services and onboarding  people and the security all those aspects [36:34] um I don't think we should be taking on it.  And so yeah, you could vibe code a solution, [36:37] but actually to maintain it and and um secure it  and then think about the new features that are [36:42] going to be at release and then doing beta testing  and what kind of colors should you have on the [36:46] buttons so that way you get user adoption.  Um I think all that is going to sit within [36:51] uh the SAS providers. I think the question then  becomes to you is like well where is it? It has [36:56] to be at the at the use of your data layer. So  the the the harness I think is something that we [37:03] have to own. I think that is a a codification  of our wisdom and knowledge over 20 50 years. [37:09] Um all of our loss history, all of our our  knowledge of brokers, of distribution channels, [37:14] um the more that we can create a harness that  actually reflects that is going to be where we're [37:19] going to invest. And and the last part I'll just  add on that which is quite interesting is that [37:23] um if you can use AI to optimize your workflow,  the question is like what really sets you apart? [37:29] And I would argue that in um this gets a little  bit philosoph you know philosophical um but in [37:35] insurance there's like procedural justice and  there's like um voice um if you go into like uh [37:40] about process um with Merin I don't know if anyone  has studied this um but it's really like the the [37:46] process sometimes is the is the objective and not  necessarily the result at the end. So in other [37:51] words, if you go to court against someone else,  the fact that you believe this process is fair [37:55] and you go before the judge and the judge does a  ruling in the end, you're both going to abide by [37:59] it because you feel like your voice is heard, you  had the same rules. It was it was a fair process [38:04] that you were able to get to in the end. If AI  can come out there and say like, "Hey, what's your [38:08] court case? We're going to give complete accuracy  and you press a button and says you're right, [38:12] you're wrong." I don't know how many people  are going to abide by that. Um, and I think a [38:15] lot of times in the insurance process, that's what  we're investing in. We're investing in like a fair [38:19] process. where a voice is heard, um where where  there's procedural justice, where we're making [38:25] sure that um you know, everything's considered.  And so that's a that's a big part of insurance [38:30] that I think doesn't go away. And I think what  we the the path we went through I think is very [38:36] similar to you guys and and I was I thought I was  going with a controversial take, but it looks like [38:41] we're aligned. I was going to say initially went  with token maxing and usage because it's easy for [38:44] the board to measure and I think that's completely  wrong and and we moved to to exactly what you said [38:49] like look at what's core to the business don't  look at what's on the side and feels easy but [38:52] what's core and then what we did is okay we we  don't want to change that process and just plug in [38:58] AI we want to rethink that process entirely and so  we create those clean rooms where it's like okay [39:03] take a couple people's on that core problem but  try to rethink it from the from the complete uh uh [39:09] uh ground up and how would you change it with AI  and then come up with a solution that's novel and [39:14] then you plug that back into your your original  core problem. So token maxing was the wrong idea [39:19] and then uh looking at really what's core is what  we're doing now. You guys said something that that [39:23] sparked if I were a startup getting go now or  trying to build a big business in this market. [39:27] Maybe there's a few of you in the audience. You  kind of said hey I hope my vendors use Kimmy in [39:32] order to reduce costs. I've always been stunned  that when I spend time with with any insurance [39:36] carrier you discover that there's all these third  party costs in their P&L. maybe it's really hard [39:39] to underwrite XYZ in Montana and so we actually  ship that to the Philippines and there's humans [39:44] that do that or here we have this really legacy  vendor. Um it seems like a lot of companies that [39:49] come through my doors are trying to reinvent the  way that you guys are doing things. Is the real [39:54] opportunity to look at what are the big buckets  where you're already spending and just like do [39:59] that better, do that cheaper, reinvent that is  is is it like which one would you appreciate more [40:04] walking in your door and pitching you? I think the  one that's going to create alpha I think the like [40:10] I said there's um well if we're all insurance  people you have your combined ratio and your [40:15] combined ratio is how you make money right if you  um combined ratio is made up of your operational [40:19] expenses and your losses um and so if you take  in a dollar and your combined ratio is 85 um [40:25] then you make 15 cents on the dollar right that's  how it works um and so when I think about it the [40:30] operational expenses and sort of the efficiencies  are a pretty small bucket um and they're [40:35] meaningful and we want to sort of optimize. But  I think often times when people come and pitch, [40:39] they're like, "Hey, we can take one piece of  this process out." And they [snorts] think [40:43] there's going to get some great TAM from  it. Um, but like the analogy I've given many [40:47] u startups is to say it would be like if I'm  trying to buy a car and you come in and say, "Hey, [40:51] I have the the Kelly Blue Book on this car. You're  going to buy a car for $40,000. You should pay me [40:56] $1,000 for a review of, you know, this this uh  this this Ford Mustang, right?" And you're like, [41:02] "No, it's not really worth $1,000." It's like like  maybe I'll pay you 20 bucks for it, but that's [41:07] all it's sort of worth. So, I think if you had a  startup to come, I would say, can you like TomTom, [41:12] can you give me data that I can't get otherwise?  Can you give me insights? Can you give me access [41:17] um that I wouldn't otherwise get? I don't know.  What about you for you? Yeah. So, I I think about [41:21] it when I when we get insured techs come to us,  I like them to stay in their lane and I like them [41:27] to do unbelievably well in that one spot. Um, and  maybe maybe a little bit of a different point is [41:33] is that I I like the idea of kind of hitting a big  ticket item, but if you can just nail something, [41:38] that's awesome. What what what I tend to get,  what I don't like is we can nail this and now [41:44] we're going to sell you 20 other things and just  just leave those alone. We've got 30 people that [41:48] are doing those 20 other things. But if you can  nail something, uh, because if you could take [41:52] that and take that off of our plate and we can do  it better than everybody else, there's the alpha [41:56] that you're talking about. And it's all those  little things that are going to end up being a big