Hartford, IMA Deploy AI Underwriting to Slash Millions in Labor Costs

Hartford与IMA应用AI核保削减数百万美元人力成本 “克隆”专业核保能力与保单审核自动化带来了数百万美元的运营节省，但传统大型机的技术负债与广泛的LLM除外条款或将阻碍全行业的普及。 来自主要经纪公司和保险机构的行业领袖探讨了AI如何深刻变革商业保险的整个投保分流与核保生命周期。讨论重点介绍了切实的运营成果，例如通过保单审核自动化大幅削减第三方业务流程外包（BPO）成本，以及对复杂的财产投资组合进行分类分流，以便核保人员能够专注于高危风险。尽管大型机和中介数据库中的传统技术负债仍是老牌机构面临的一大阻碍，但演讲者强调，AI的真正威力在于捕获并普及整个核保团队的专业领域知识。讨论小组还探讨了建立核保人员信任的必要性，以及正在兴起的针对AI的责任风险敞口与保险公司除外条款浪潮。 行业预测显示，AI 将改变或自动化 100% 的保险工作流程；商业经纪人正借助 AI 提炼并构建结构化风险档案，在向最终客户推介方案前先向承保机构完成关键的“初次销售”。 (5:27) 传统承保机构的主机定价引擎与经纪公司老旧的 SQL 系统构成严重技术债，使得基于现代 API 架构起步的新兴经纪商，预计将比管理百亿美元保费的成熟机构提前数年实现端到端自主业务流。 (11:40) 数据驱动的智能分派引擎正在取代传统基于人情偏好的选单方式，通过调取承保历史数据，在出单时效、条款灵活性与利润分成上算法化匹配最佳承保机构。 (18:10) AI 自动化保单审计与合同合规比对将单次耗时 6 至 8 小时的人工审核压缩至即时完成，帮助大型经纪机构收回昂贵的外部 BPO 业务，节省了数百万美元的运营成本。 (21:57)

Commercial insurance leaders view artificial intelligence primarily as an augmentation tool to synthesize unstructured data, triage complex risks, and capture specialized domain expertise rather than an autonomous replacement for human practitioners. However, enterprise adoption faces friction from legacy mainframe infrastructure, stringent requirements for underwriter trust, and the emergence of broad policy exclusions around algorithmic liabilities. Augmenting Underwriter Judgment and Accelerating Risk Triage (23:38) Andrew Zarkowsky of The Hartford explains that commercial underwriting has evolved from manual data collection to managing overwhelming volumes of unstructured digital reports, third-party datasets, and complex guidelines. AI models provide immediate value by synthesizing fragmented files into structured briefs and accelerating account triage. In property underwriting, for example, AI can automatically evaluate straightforward exposures—such as nine low-risk inland locations—enabling human underwriters to concentrate their time on catastrophic hazard zones like hurricane-exposed coastal properties. Carriers and brokerages are also using AI to capture and scale institutional expertise. Zarkowsky notes efforts to encode the methodology of elite real estate specialists into digital assistants, democratizing top-tier underwriting judgment across the company. Similarly, Garrett Droege of IMA Financial Group is digitizing his specialized knowledge in high-volatility sectors such as Web3 and AI compute infrastructure, removing individual technical specialists as operational bottlenecks while freeing professionals to design tailored risk structures. Legacy Mainframes and Technical Debt Constrain Automation (11:40) Despite aggressive projections—including consulting forecasts cited by Zarkowsky that AI will impact 100% of insurance workflows—adoption is constrained by deep architectural debt. Core actuarial rating engines at major carriers still operate on legacy mainframes that resist modern API and agentic integrations. Similarly, established brokerages rely on agency management systems with unstructured database backends and rely heavily on scanned PDF files. While Droege forecasts that autonomous AI agents will manage end-to-end insurance workflows from intake to policy binding within three to five years, he argues this will primarily be achieved by greenfield brokerages built from scratch on modern stacks. For established firms managing $10 billion to $11 billion in market premium, migrating decades of unstructured documents into standardized data environments remains a substantial operational hurdle. Underwriter Accountability and Measurable Operational Returns (16:58) External AI deployment requires clearing the strict threshold of underwriter trust. Zarkowsky emphasizes the "name under" principle: because underwriters sign their names to policies and bear direct professional liability for portfolio losses, they will reject any tool suspected of hallucinations. Given that insurance represents an intangible, trust-based contract to indemnify future losses, internal compliance and model explainability are mandatory prerequisites. Where deployed, AI delivers clear financial and operational efficiencies. Droege highlights that IMA replaced subjective, relationship-based carrier selection with placement algorithms that analyze historical turnaround speeds, coverage breadth, and profit-sharing terms. Furthermore, IMA automated commercial policy reviews—reducing six-to-eight-hour manual audits to instantaneous checks—and brought third-party contract compliance reviews back in-house from external business process outsourcing firms, generating millions of dollars in operational savings. Emerging AI Liabilities and Sweeping Carrier Exclusions (28:26) As commercial enterprises adopt artificial intelligence, uncertainty regarding liability is creating new underwriting challenges. Droege observes that many traditional carriers are introducing blanket AI exclusions that deny claims arising from third-party large language models, even when resulting losses involve traditional bodily injury or property damage. Because corporate buyers often lack the internal capacity to verify the safety of vendor-supplied AI software, they are shifting liability back onto developers through contract warranties. This has catalyzed a growing market for specialized AI liability insurance, where software vendors purchase dedicated coverage policies to assure enterprise buyers that algorithmic failures will not result in unbacked catastrophic exposures. While artificial intelligence provides immediate operational savings in document extraction, contract auditing, and administrative triage, the industry remains split between modern greenfield entrants and incumbents managing mainframe technical debt. Furthermore, as carriers introduce sweeping exclusions for third-party algorithmic failures, defining and underwriting AI risks themselves remains an evolving commercial challenge.

商业保险经纪人与承保人正利用人工智能技术加速工作流、降低常规人为差错并提炼专家经验，但其规模化落地受制于老旧大型机与非结构化数据的技术债务，同时在算法可信度与新型AI责任认定上面临严峻挑战。 投保申请重塑与风险智能分流 (5:27) Andrew Zarkowsky指出，传统商业承保人长期被淹没在第三方报告、非结构化PDF及复杂准则中；AI当前最直接的高价值应用在于抓取并提炼分散的数据堆，输出精简高效的承保摘要。Garrett Droege从经纪端补充道，经纪人本质上面临“把保单卖两次”的流程——必须先向承保人销售风险以争取最优价格与条款，随后才能向投保企业交付方案。高质量的AI材料能结构化展示安全控制与转嫁机制，替代低效的手工复制与纸质扫描。 在承保评估中，AI还赋能了风险分流：以The Hartford处理十处物业的账户为例，系统能快速核准并放行位于伊利诺伊州的九处常规低风险物业，使承保人集中资源深入评估极易遭受飓风等巨灾侵袭的迈阿密资产，提升整体审查效率。 遗留架构制约与数据就绪困境 (11:40) 在数据策略层面，Andrew强调在构建AI前必须先明确机构拥有何种数据；缺乏干净、结构化管道的用例将被暂缓推进。Garrett直言，传统承保人与经纪机构面临着沉重的技术债务：主流大型承保人的核心精算与定价模型仍托管在难以对接现代API与智能体的大型机上；管理着超百亿美元保费的大型经纪商，其代理管理系统（AMS）也普遍基于充满非结构化数据的旧版SQL Server。 尽管Garrett预测自主AI智能体将在三至五年内管理从录入到绑定的全流程，但他指出该变革更可能先由全新构建的独立机构实现；对于老牌机构而言，迁移数十年积累的非结构化数据与扫描文档仍是难以逾越的工程瓶颈。 承保签字原则与算法可信边界 (15:50) Andrew指出，保险业的基石是兑现灾后赔付的信任，算法的合规透明不可妥协。此外，承保人（Underwriter）在字面意义上需以个人名义签署合同并承担连带信誉责任，一旦AI发生幻觉或失误破坏了与经纪人的合作互信，工具将被彻底废弃，因此内部信任是外部部署的前提。 Garrett则主张在SOC 2 Type II等既有隐私标准下积极探索AI，避免过度等待明确监管而错失良机。在日常经纪实务中，针对数千万美元规模的技术企业申报，智能匹配系统能基于历史承保速度与利润分配机制推荐最优承保人，取代传统依赖人际关系或“赠送甜甜圈”的主观方式。他强调，相比容易因疲劳或喝咖啡而分心出错的人类员工，AI能消除常规事务性失误并实现对全部工作流的协同赋能。 降本增效与专家经验数字化 (21:57) 在具体投资回报方面，Garrett表示自动化审计显著改善了利润率。传统上经纪人为挖掘保障漏洞所提供的全套保单审查需耗费六至八小时人工，现已实现AI即时审计；合同合规审查平台还能直接对保单与商业合同的比对结果给出红黄绿三级合规判定。通过将此类以往依赖第三方业务流程外包（BPO）的核对工作收回内部执行，IMA实现了数百万美元的运营节省。 Andrew与Garrett进一步强调了专长数字化的深远价值：The Hartford将顶尖房地产承保专家Tracy的方法论输入模型，使全司人员均可调用其决策逻辑；Garrett亦在其负责的Web3与高性能算力等前沿数字风险领域将个人专业经验数字化，消除了自身作为专业审批瓶颈的局限，让团队能够自主配置高度复杂的特种保障方案。 新型技术责任与除外条款博弈 (28:26) 随着企业客户自身广泛引入AI，传统保单正面临前所未有的责任界定挑战。Garrett指出，由于对技术演进缺乏把握，主流承保人正普遍出台全面的AI除外责任批单，明确规定若损失涉及第三方大语言模型（LLM），即便导致了人身伤害或财产损失也将被拒赔。由于终端企业缺乏评估外部AI软件安全性的专业能力，买方强烈要求将风险逆向转嫁给软件供应商。 如同电影《Tommy Boy》中通过保证书消除买方疑虑的类比，AI开发者正被迫购买专属的AI责任保险以作为销售背书；在医疗诊断等高风险场景中，漏诊引发的过失致死诉讼往往直接指向模型开发者，这种将技术风险刚性外包的诉求正催生出一个崭新的特种商业保险细分市场。 讨论表明，AI在保险产业链中的应用正从文档摘要与外包降本，深入至资深专家经验的复制与前沿特种风险的重构。然而，遗留大型机系统的架构债务、承保人对模型幻觉的零容忍态度，以及承保端对生成式AI风险采取的一刀切除外策略，反映出技术集成与系统信任之间仍存在尚未解决的深层博弈。

## Background 本期内容由行业会议组织 Ai4 于 2025 年 8 月 29 日在 YouTube 平台发布。视频汇聚了来自知名承保公司与经纪机构的行业领袖，围绕人工智能在保险风险管理、核保流程、理赔自动化中的前沿应用，以及传统老旧系统改造与监管合规挑战展开了深度圆桌探讨。 ## TL;DR 本场讨论汇集了来自 The Hartford 的承保专家与来自 IMA Financial Group 的资深经纪人，共同剖析了 AI 对保险行业全链条的颠覆性重塑。对谈详细拆解了经纪端如何利用 AI 赋能复杂的“双重销售”流程与保单审计，以及承保端如何通过数据智能聚合消除人工瓶颈、复刻顶级专家的核保经验并实现动态风险分流。与此同时，嘉宾们深入探讨了老旧大型机与 SQL 数据库的技术债务痛点、核保人员对工具信任的底线要求，以及险企引入 AI 免责条款所催生出的全新“AI 专属责任险”蓝海市场。 ## Atmosphere & style 整场对谈节奏轻快且充满机智的幽默互动，嘉宾们频频借用“希格斯玻色子发现经历”、“头发秃了”以及“买没人想买的产品”等自嘲段子调侃彼此，现场气氛轻松融洽。然而在专业议题上，对话展现出极高的行业洞察力与现实审慎感，既有对 AI 自动化降本增效的兴奋展望，也毫不避讳地直面行业历史技术包袱与职业信誉底线，呈现出一种在务实克制与前沿创新之间取得微妙平衡的专业探讨张力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=4s) 保险业的本质是“预测不可预测之事”，AI 虽能带来高效率的精准预测，但可解释性、业务合规与建立可信赖的决策机制是首要挑战。 - [02:43](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=163s) 咨询机构评估显示保险业受 AI 影响与转型的比例高达 100%，从经纪人到承保端的全链条岗位都将被 AI 深度重塑与赋能。 - [05:08](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=308s) 在承保端，AI 当前最核心的落地价值是扮演智能信息中枢，将分散在各处的复杂指引与非结构化卷宗快速提炼摘要，大幅降低核保前期的阅读负担。 - [07:29](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=449s) 保险经纪人需要“把保险卖两次”（先向承保人争取最优条款，再向客户交付方案），AI 能协助反推核保疑点，显著提升投保申请方案（Submission）的质量。 - [11:41](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=701s) 数据就绪度决定 AI 成败，但行业面临严重的历史技术债——旧式 SQL 数据库及甚至仍在运行的大型机（Mainframe），使得老牌机构端到端部署 Agent 极为困难。 - [16:58](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=1018s) 核保人以个人声誉与签字为保单背书，因此 AI 工具的准确度与合规把控必须达到极高标准、赢得内部核保员的深度信任后才能推向业务一线。 - [19:20](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=1160s) AI 能取代传统基于人情关系（“谁送了甜甜圈”）的随意派单模式，依托历史数据实现承保机构的智能匹配，并避免人工录入中因疲劳和分心产生的低级错误。 - [21:57](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=1317s) 在保单漏洞审计与商业合同合规校验场景中，AI 将 6 到 8 小时的人工作业转变为全天候自动化处理，直接替代昂贵的 BPO 外包并为机构节省数百万美元。 - [23:38](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=1418s) AI 的核心价值在于克隆并沉淀顶尖专家的行业智慧，打破垂直细分领域因少数专家档期限制所导致的业务瓶颈，实现专家知识全员“民主化”。 - [26:20](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=1580s) 通过 AI 实现动态风险分流，系统能够自动化极速核验低风险常规资产，让资深核保专家将精力聚焦在高灾害风险等复杂标的上。 - [28:26](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=1706s) 传统险企正在将 AI 造成的损失列入免责条款，这倒逼 AI 软件开发商通过采购专属“AI 责任险”为自身产品背书，为保险业孕育出全新的细分险种市场。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=4s) 圓桌開場：當物理學遇上保險與 AI 的預測本質 會議在熱烈的氛圍中開場。主持人 Lovedeep Saini 幽默地調侃道，大家剛聽完「AI 教父」以及 AI 在好萊塢應用的演講，此刻應該都充滿了幹勁。他指出，保險業的核心本質就是「預測不可預測之事」（predicting the unpredictable），而 AI 當前展現出的巨大承諾正是快速且精準地進行預測。然而，核心挑戰隨之而來：我們是否能夠真正信任 AI 給出的決策？這些決策是否具備可解釋性？以及業務流程是否完全合規？ Lovedeep 介紹了自己的背景。作為 Conner Strong & Buckelew 的首席分析官（Chief Analytics Officer），他的職責是為運營與銷售團隊配備由 AI 驅動的工具，將繁瑣的簡單任務委派給 AI，進而更好地服務客戶。談及個人經歷時，Lovedeep 透露自己擁有物理學博士學位，在進入保險經紀行業前曾有近十年的物理學科研經歷，並作為實驗物理學家參與了 2013 年榮獲諾貝爾物理學獎的「希格斯玻色子」（Higgs Boson）發現實驗。正是這段處理海量前沿實驗數據的科研背景，使他深諳如何利用數據、AI 與先進技術從複雜模式中提煉洞察，並將其遷移到保險風險評估與定價之中。 ```mermaid flowchart TD accTitle: 保險經紀與承保流程中的 AI 協同 subgraph S1["客戶與經紀人端 (Broker)"] A["企業客戶 (Insured)"] --> B["保險經紀人 (Broker/IMA, CS&B)"] B --> C["收集非結構化數據 & 風險畫像"] C --> D["AI 賦能生成高質量投保方案 (Submission)"] end subgraph S2["承保端 (Carrier)"] D --> E["保險承保人 (Underwriter/The Hartford)"] E --> F["AI 匯總多源第三方數據與規則"] F --> G["精準核保、定價與出具條款"] end ``` --- ### [02:43](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=163s) 嘉賓陣容：經紀人與承保公司的雙重視角 嘉賓介紹環節在一陣輕鬆的互相調侃中展開。IMA Financial Group 的創新負責人兼數字風險實踐負責人 Garrett Droege 開玩笑地說，自己的背景遠沒有物理學家發現諾貝爾獎粒子那麼驚艷，大家可以「降低期待」。Garrett 擁有 20 年的保險從業經驗，專注於技術風險，歷任經紀公司的運營、銷售、創新及領導崗位。他介紹 IMA 與 Conner Strong 類似，是全美第二或第三大獨立、員工持股的大型中端市場保險經紀機構。他強調，本次對談涵蓋了「保險公司（Carrier）」與「經紀人（Broker）」兩種互補的視角，能為行業帶來更為立體的全景思考。 隨後發言的是 The Hartford（哈特福德保險）中大型商業業務 AI 核保負責人 Andrew Zarkowsky。Andrew 同樣在保險核保一線深耕了近 20 年。他感謝主辦方設立保險專題，並直言保險常被外人誤解為枯燥的行業，但他堅決反對這種偏見。Andrew 引用了一家知名大型諮詢公司的評估數據：在各行業受 AI 衝擊與轉型的預測中，有的行業是 30%，有的是 50%，而保險行業受影響的比例被評定為 **100%**。這意味著保險全鏈條都將被 AI 深刻重塑（無論是增強還是自動化）。現場兩位嘉賓還就資歷開起了年齡玩笑——Andrew 指著自己的灰白頭髮自嘲資深，Garrett 則幽默回擊道：「你說你有白頭髮，而我連頭髮都沒了，這才叫年紀大。」 - **諮詢公司行業影響度評估**：保險業被列為受 AI 影響程度最高（100%）的行業榜首。 - **嘉賓視角互補**：Andrew 代表擁有 200 多年歷史的老牌承保機構（The Hartford），Garrett 則代表大型獨立經紀機構（IMA Financial Group）。 --- ### [05:08](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=308s) 承保端痛點：從紙質卷宗到 AI 信息智能聚合 討論切入核心議題：AI 究竟在哪裡釋放最大價值？Andrew 回顧了核保工作模式的演進史。在互聯網普及前，經紀人必須親手把所有信息遞交給核保人，核保人依賴紙質申請表和線下面對面訪談；互聯網時代到來後，數據獲取模式轉向「你自己去網上找」，核保人需要花費大量時間檢索第三方數據、互聯網信息。 隨著現代保險業的分工日益精細，不同行業、不同風險規模、不同險種的核保指引（Underwriting Guidelines）極其複雜。如果把審核一家客戶所需的各類資料打印出來，會是一整疊厚重的紙質卷宗。核保人員往往需要耗費大量精力閱讀並理解這家企業，才能進入實質性的風險評估階段。Andrew 認為，AI 在當下及短期內最核心的價值，就是充當智能的信息中樞——將分散在各處的海量數據匯集、提煉、智能摘要，在核保一開始就將核心要點精準呈現在核保人眼前。 --- ### [07:29](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=449s) 經紀端實踐：用 AI 打贏「第二次銷售」 Garrett 從保險經紀人的角度提出了深刻的洞察。他指出，經紀人的工作本質上需要**「把保險賣兩次」（Sell it twice）**： 1. **第一次銷售（後台）**：經紀人把客戶的風險項目呈報給承保公司（如 Andrew 所在的 The Hartford 及其競爭對手）。 2. **第二次銷售（前端）**：在承保公司出具條款後，經紀人協助客戶挑選最優方案。 Garrett 強調，第一次銷售直接決定了保單的定價高低與保障範圍的寬窄。如果經紀人無法在承保人面前把客戶的優勢與風險控制措施充分展示出來，就拿不到有競爭力的條件。AI 在此處猶如「加了興奮劑」（on steroids），極大地增強了投保申請方案（Submission Building）的質量。經紀人可以利用 AI 站在核保人的視角反推疑點，主動利用風險管理手段、合同風險轉移等措施來打消承保人的顧慮，從而遞交一份全方位、高標準的申請文件。 ```mermaid flowchart LR accTitle: 經紀人雙重銷售模型 subgraph FirstSale["第一次銷售 (經紀人對承保人)"] direction TB B1["經紀人提煉企業全貌"] --> B2["利用 AI 進行風險分析與疑慮化解"] B2 --> B3["向核保人提交優質方案 (Submission)"] end subgraph SecondSale["第二次銷售 (經紀人對客戶)"] direction TB C1["爭取到更優費率與條款"] --> C2["為企業客戶量身定製最優保單"] end FirstSale --> SecondSale ``` 與此同時，Garrett 犀利地吐槽了保險經紀行業極其落後的內部工作流。他直言經紀行業擁有他見過「所有行業中最糟糕的手工流程」：時至今日，仍有人手寫打印申請單、重新掃描並在不同的 Excel 表格和 PDF 之間手動複製貼上。Andrew 也對此表示贊同，並直言市場上的經紀人水平參差不齊，優質經紀人與普通經紀人的分水嶺就在於遞交的 Submission 質量，AI 將推動整體質量的提升。Garrett 特別補充，大型商業保險（如科技或生命科學企業）絕非標準化的車險或家財險，每一份保單都是高度定製的，AI 正在幫助經紀人更精準地捕捉這類非標企業的真實保障需求。 - **核心概念「雙重銷售」**：先向承保人「銷售」客戶的風險管理能力以爭取最優條款，再向客戶「銷售」保單方案。 - **非標業務定製化**：高科技、生命科學等複雜商業客戶需要極度精細的非標準保單，依賴高質量數據輸入。 --- ### [11:41](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=701s) 數據戰略與現實碰撞：大型機、老舊 SQL 與歷史技術債 針對「數據是 AI 的石油」這一觀點，兩位嘉賓深入探討了企業的數據治理與歷史技術包袱（Legacy Tech Debt）。 | 維度 | Andrew（承保端 / The Hartford） | Garrett（經紀端 / IMA Financial Group） | |:--- |:--- |:--- | | **數據策略優先級** | 評估所有 AI 用例的首要標準是「手頭現有數據的完備度」，優先落地數據就緒的項目。 | 主張徹底重構，但受制於龐大的現有業務規模與歷史系統阻力。 | | **核心技術障礙** | 多源數據的對齊、定義與初期有效輸入。 | 代理管理系統（AMS）依賴過時 SQL 架構，充斥大量非結構化數據；部分承保端甚至仍運行在大型機（Mainframe）上。 | | **體量與轉型難度** | 200 年老牌機構，流程極端重視合規審慎。 | 管理 100-110 億美元市場保費、數萬名客戶，跨系統數據遷移代價高昂。 | Andrew 表示，儘管人人都高喊「數據就是一切」，但在落地 AI 核保工具時，最致命的循環是：你無法指導 AI 該做什麼，除非你清楚 AI 擁有什麼數據。面對內部湧現的各種 AI 用例，他篩選的首要原則就是「看數據」——與其在數據匱乏的用例上死磕，不如優先把資源傾斜給數據儲備紮實、架構清晰的業務場景。 Garrett 則從技術架構底層剖析了行業頑疾。他自嘲自己從小就是個喜歡拆解玩具馬達並徹底重建的「破壞型」思維者，但在保險業，現實卻非常骨感。行業內至今仍有保險公司的核保與精算模型運行在**大型機計算機（Mainframe）**上，根本無法直接架設現代 API 或 AI Agent 層。對於大型經紀機構而言，核心客戶數據存放在被稱為「代理管理系統（Agency Management System, AMS）」的舊式 SQL 數據庫中，其中充斥著大量無法直接調用的非結構化數據與掃描件 PDF。儘管 IMA 建立了自身的數據湖，但也僅能掌控部分切片。Garrett 預測，雖然行業期望在 3 到 5 年內實現 AI Agent 端到端的全自動工作流，但這種變革大概率會由完全沒有歷史包袱、從零起步（Greenfield）的新興經紀公司率先實現；對於手握 100 至 110 億美元保費規模、服務數萬家客戶的老牌機構來說，數據遷移與架構改造無疑是一場硬仗。 --- ### [15:51](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=951s) 信任與合規：保險業不可逾越的底線 Lovedeep 提出了一個尖銳的問題：客戶既要求更快的回覆與更高的效率，而保險業又受制於老舊系統與嚴苛監管，如何在加速推進 AI 的同時確保「信任（Trust）」與「合規（Compliance）」不被稀釋？ Andrew 堅定地回應，對於 The Hartford 這樣擁有 200 多年歷史的品牌而言，商業道德、信任與品牌聲譽是**絕對不可妥協的底線（Non-negotiables）**。他指出，保險業的基石本就是信任契約——客戶支付保費購買一份承諾，在事故發生時獲得賠付，一旦信任瓦解，整個行業的根基將不復存在。因此，公司建立了極其嚴苛的 AI 合規、治理與監管審查流程，業務團隊在推進任何 AI 創新時都會與合規部門進行深度且高頻的協同，絕不允許以犧牲公信力為代價換取盲目的技術冒進。 ### [16:58](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=1018s) 核保员的信任底线与以客户为中心的技术落地 Andrew Zarkowsky 从核保员（Underwriter）的职业心理切入，剖析了为什么 AI 工具在保险核心流程中的信任度是“绝对不可妥协的底线”。核保员的英文直译是“在文件下方签署名字的人”，这意味着每一笔保单都背负着个人的声誉与专业诚信。核保员在公司内外的立足之本就是其评估风险的信誉；如果与经纪人合作时因为工具错误而“坑”了对方一次，职业关系就可能彻底破裂。因此，如果 AI 工具存在不可控的错误率，核保员从一开始就会拒绝使用，这类工具根本没有机会流向外部客户。The Hartford 在研发 AI 应用时，将建立核保员的内部信任与合规把控列为最高优先级。 ```mermaid flowchart TD accTitle: 核保信任与风控闭环 subgraph Trust ["信任与核保逻辑"] A["核保员个人声誉与签字背书"] --> B["对 AI 工具的严格检验与信任构建"] B --> C["内部合规与风险控制"] end subgraph Execution ["业务落地与价值实现"] C --> D["赋能经纪人与客户体验"] D --> E["业务转化与精准承保"] end ``` 随后 Garrett Droege 幽默地援引了 IMA Financial Group 首席执行官的一句名言：“我们卖的是一种没人想买的产品，而且用起来还特别麻烦（a pain in the ass to use）。” 现实中几乎没人喜欢保险，直到灾难发生时大家才会感叹“幸好买了”。 Garrett 强调，经纪公司的核心在于改善客户体验，满足客户的即时需求，而不是消极等待监管法规的尘埃落定。企业只要恪守数据隐私规范并达到 SOC 2 Type II 等安全认证标准，就应该果断推进 AI 实验与应用落地。 ### [19:20](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=1160s) 从“谁送了甜甜圈”到数据驱动的智能分发与差错控制 Garrett Droege 举例说明了代理管理系统（Agency Management System）在引入 AI 后的巨大变革。假设销售端接洽了一家年营收 7500 万美元的软件开发公司，传统模式下，服务团队面对海量承保机构往往无所适从。过去的经纪业务撮合往往极具随意性——“上周谁给办公室带了甜甜圈，我可能就欠他人情；哎呀 Sandra 没分到最后一个甜甜圈，我觉得过意不去，这次就把单子派给她吧”。这种依赖人情和零食的决策显然不符合客户的最佳利益。AI 能够基于真实历史数据，迅速分析出针对同类客户哪家核保机构最有竞争力、响应最快、愿意协商保障条款并提供灵活定价，同时兼顾经纪人与保险公司之间的后端利益分成协议。 针对业内对大模型“幻觉”（Hallucination）的担忧，Garrett 分享了一段长达 18 个月的数字化转型实地观察经历。在那段追踪员工日常案头操作的审计中，他目睹了大量不可避免的人为低级失误： - 复制粘贴错误 - 员工去接杯咖啡回来就忘记刚才处理到哪一步 - 突然被一通电话打断导致流程遗漏 Garrett 指出，人类无时无刻不在犯错，虽然 AI 也会犯错并需要人类持续审计，但 AI 不会因为分心或疲劳而出错。AI 不会完全取代人类，但 100% 会深度重塑（Augment）每一个岗位。 --- ### [21:57](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=1317s) 降本增效的真实 ROI：保单审计自动化与数百万美元外包回流 在探讨 AI 的投资回报率（ROI）指标时，Garrett Droege 直言“资金与利润就是一切”，并分享了两个已实现显著降本增效的成熟应用场景： ```mermaid flowchart LR accTitle: 核心业务审计自动化流程 subgraph Prospecting ["业务竞标阶段"] A["输入客户现有保单"] --> B["AI 自动比对条款与背书"] B --> C["即时生成漏洞诊断报告"] end subgraph Compliance ["合同合规阶段"] D["上传商业合同与保单"] --> E["AI 红黄绿灯合规校验"] E --> F["替代高成本 BPO 外包"] end ``` - **竞标阶段的保障方案审计（Coverage Audits）**：在争夺新客户时，经纪团队通常需要找出客户现有保单中的漏洞、缺失的附加背书（Endorsements）或不利条款，证明自己能提供更优方案。以往人工完成一次端到端审计需要 6 到 8 个小时，如今 AI 每天都在全天候自动执行这一繁琐流程。 - **合同合规性审查（Contractual Compliance）**：核对保单条款是否满足商业合同的具体要求。系统在上传合同和保单后，能以“红灯、黄灯、绿灯”直观反馈合规状态。该流程以往大多外包给昂贵的业务流程外包（BPO）服务商，如今将这项职能完全收回内部自主处理，直接为公司节省了数百万美元的成本。 --- ### [23:38](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=1418s) “核保员是英雄”：克隆行业专家的经验与消除业务瓶颈 Andrew Zarkowsky 强调，相比于单纯的效率提升，他更看重质量的飞跃。当高水平的 AI 工具与资深核保员相结合时，错误率会大幅趋近于零。以 The Hartford 团队内部一位名叫 Tracy 的顶级房地产行业核保专家为例：她是全行业最顶尖的专家之一，但作为个人，她不可能亲自承保每一个涌入的房地产项目。如果能通过 AI 将“Tracy 级别”的专业知识与判定逻辑沉淀并分发给全员，就能让所有承保人员摆脱繁重的文书工作，把精力聚焦在客户服务、风险管理建议及理赔方案设计上。正如 Andrew 所总结的理念： > “核保员是英雄，AI 是让英雄拯救世界的工具，我们希望他们去拯救客户。” Garrett Droege 对此深表赞同。他自身负责 IMA 的数字风险业务（涵盖区块链、Web3、加密货币交易所、比特币矿场、AI 算力基础设施与高性能计算 HPC 等新兴复杂领域），由于全公司只有他深谙此类前沿风险的设计逻辑，他长期成为业务流转中的单一瓶颈。如今，他正积极将自己的专业大脑“下载”并训练到系统知识库中，打造虚拟助手。他不担心自己会被取代，因为新型风险模型始终需要专家进行微调与迭代；通过将专业知识民主化（Democratize），团队不再会因为专家个人的时间档期（Availability）而错失宝贵的商业机会。 --- ### [26:20](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=1580s) 动态风险分流：将专家精力聚焦于高风险场景 针对 AI 是否会改变现有风险评估与定价方式的问题，Andrew Zarkowsky 说明了大型险企如何利用 AI 实现智能化风险分流与精准资源配置。 以一份包含 10 处经营地点的商业投保申请为例： - **低风险常规资产**：其中 9 处地点位于伊利诺伊州，处于非灾害高发区，且保险限额较低。 - **高风险复杂资产**：第 10 处地点位于佛罗里达州迈阿密，面临严重的飓风、龙卷风和冰雹灾害风险。 在传统作业模式下，核保员必须在所有 10 处地点上平均分配时间和精力。借助 AI 工具后，系统能够自动化、极速核验前 9 处低风险地点，让专业核保员将核心精力全部集中在迈阿密这一处复杂的风险点上，从而更快、更具针对性地为客户设计解决方案。 --- ### [28:26](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=1706s) AI 自身带来的全新风险：免责条款与“AI 专属保险”新蓝海 讨论的尾声转向了 AI 技术自身引发的新型商业风险。Garrett Droege 透露，当前许多企业客户都在关切引入 AI 是否会改变其风险敞口，而承保机构对此已迅速出台应对策略——在保单中加入 **AI 免责条款（AI Exclusions）**。这意味着，只要某起索赔是由第三方大语言模型（LLM）引起的，哪怕属于传统的身体伤害（Bodily Injury）或财产损失（Property Damage），保险公司也一概不予赔付。 ```mermaid flowchart TD accTitle: AI 风险转嫁与专属保险链条 A["企业采购 AI 平台与工具"] --> B["险企在常规保单中加入 AI 免责条款"] B --> C["企业面临 AI 误诊/故障风险敞口"] C --> D["企业要求 AI 厂商提供兜底担保"] D --> E["AI 软件公司采购专属 AI 责任险"] E --> F["保险机构开拓全新细分险种蓝海"] ``` 由于下游企业无法自行验证供应商 AI 模型的安全性与可靠性，它们正在设法将这一责任倒逼转嫁给 AI 开发商。这与网络安全行业的发展轨迹如出一辙——当年网络安全公司推销防御系统时客户缺乏评估能力，如今 AI 软件公司同样需要向客户出具履约担保。 Garrett 引用了经典喜剧电影《乌龙兄弟》（*Tommy Boy*）中“在包装盒上贴保证标签”的桥段，指出 AI 专属保险如今已成为 AI 科技公司在市场竞争中的重要“营销与背书工具”。例如在医疗领域，若 AI 影像扫描系统遗漏了病灶导致患者病情恶化甚至死亡，AI 开发商必然会首当其冲遭遇巨额诉讼。通过购买专门的 AI 责任保险，AI 厂商能够向企业客户证明自身具备足够的抗风险能力。这种新型风险需求正在为保险业孕育出一个充满潜力的新兴险种市场（Cottage Industry）。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=4s) predicting the unpredictable — Lovedeep Saini > [02:43](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=163s) 100% — Andrew Zarkowsky > [07:29](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=449s) Sell it twice — Garrett Droege > [07:29](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=449s) on steroids — Garrett Droege > [15:51](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=951s) Non-negotiables — Andrew Zarkowsky > [16:58](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=1018s) a pain in the ass to use — Garrett Droege > [23:38](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=1418s) 核保员是英雄，AI 是让英雄拯救世界的工具，我们希望他们去拯救客户。 — Andrew Zarkowsky > [28:26](https://www.youtube.com/watch?v=5VvG9Rc3fE4&t=1706s) AI Exclusions — Garrett Droege ## Mentioned Lovedeep Saini, Conner Strong & Buckelew, Higgs Boson, Garrett Droege, IMA Financial Group, Andrew Zarkowsky, The Hartford, Agency Management System (AMS), SQL, Mainframe, SOC 2 Type II, Tommy Boy ## Video description Andrew Zarkowsky, Head of AI Underwriting, The Hartford Garrett Droege, SVP, Director of Innovation and Digital Risk Practice Leader, IMA Financial Group, Inc. Lovedeep Saini, Chief Analytics Officer, Conner Strong & Buckelew Abstract This session will explore how AI is transforming the insurance industry by streamlining operations and improving risk management, underwriting, and claims automation. Panelists will discuss how AI-powered tools are helping insurers assess risk more accurately, automate the underwriting process, and detect fraud in real time. Learn how leading insurance companies are leveraging AI to enhance customer experiences, expedite claims processing, and drive operational efficiency. The panel will also examine the challenges of integrating AI into legacy systems and navigating the regulatory landscape in the evolving world of AI-driven insurance. [Source](https://www.youtube.com/watch?v=5VvG9Rc3fE4)

[0:04] Lovedeep Saini: predicting the unpredictable [2:43] Andrew Zarkowsky: 100% [7:29] Garrett Droege: Sell it twice [7:29] Garrett Droege: on steroids [15:51] Andrew Zarkowsky: Non-negotiables [16:58] Garrett Droege: a pain in the ass to use [23:38] Andrew Zarkowsky: 核保员是英雄，AI 是让英雄拯救世界的工具，我们希望他们去拯救客户。 [28:26] Garrett Droege: AI Exclusions

[0:04] Lovedeep Saini: predicting the unpredictable [2:43] Andrew Zarkowsky: 100% impact [5:27] Garrett Droege: We think of our jobs as we have to sell insurance, but we have to sell it twice. [11:40] Andrew Zarkowsky: data is everything [15:50] Andrew Zarkowsky: The entire insurance industry is built on trust... There's a policy, there's a payment and then if something goes wrong, we pay out. It's all trust. So if we break that trust, it will devastate our industry. [18:10] Garrett Droege: we sell a product that nobody wants to buy and it's a pain in the ass to use. [21:57] Garrett Droege: money is everything [23:38] Andrew Zarkowsky: the underwriter is the hero [23:38] Garrett Droege: downloading my brain

[0:04] Okay, good morning again. Um now we are [0:06] sitting in a session which is about AI [0:10] in insurance and I'm sure we are already [0:13] quite motivated after meeting uh [0:16] godfather of AI and AI in Hollywood and [0:19] we are energetic enough. So let's talk [0:21] about AI in insurance and we all know [0:23] that insurance is all about predicting [0:25] the unpredictable and AI is already [0:28] promising enough that it says that it [0:31] can do uh predictions quite accurately [0:35] fastly but the problem is uh do we trust [0:39] that decision that AI gives us or uh is [0:44] it explainable are we compliant all [0:46] these things we'll discuss here so my [0:49] name is loved Sani and I'm working as [0:51] chief analytics officer at Connor Strong [0:53] and Bucklau and that's where uh we serve [0:57] our clients. Connor Strong and Buckloo [0:59] is all about insurance brokerage and we [1:02] help our clients to cover their risk [1:04] while working with insurers. uh my role [1:08] there is to [1:11] provide my you know operations and sales [1:14] teams with the tools that are powered by [1:17] AI and uh so that they could serve their [1:20] clients better by delegating simple task [1:23] to these AI tools. So that's me. I have [1:26] a close to 10 years of experience in [1:29] insurance brokerage and I was trying to [1:31] explain this to Garrett earlier that I [1:34] came from physics background uh where I [1:36] worked almost for 10 years uh with the [1:39] physics collaboration where we [1:41] participated in discovering Higs Bzon if [1:44] does not ring any bell here. So that was [1:47] 2013 Nobel Prize in physics. uh but that [1:52] discovery was led by some [1:53] experimentalist and I'm one of them. I'm [1:56] proud participant of that experiment. Uh [1:58] and that's where I earned my PhD in [2:00] physics too. So I'm coming from a [2:03] background where I know how to use data [2:06] AI technology and all this kind of fancy [2:10] stuff to help uncover the patterns from [2:15] our data and bring insights there. So [2:17] now that's me but that's what we're [2:19] going to discuss here in this panel too [2:21] that how to use data and AI to help with [2:26] insurance risk and uh [2:29] um pricing the risk there. So we have [2:32] experts here. Let's start with your [2:35] introductions and your name role and how [2:38] AI intersect with your work today. Let's [2:42] start from here. Get it? [2:43] >> Yeah. Good morning. My background is not [2:45] anywhere near as impressive. Um so lower [2:48] your expectations. My name is Garrett [2:50] Droy. Um I lead innovation. I also lead [2:52] digital risk for IMA Financial Group. We [2:55] are um like Connor Strong um a large [2:58] mid-market focused in uh independent [3:00] insurance brokerage. I think we're the [3:02] second or third largest privately held [3:04] employeeowned brokerage. Um my [3:07] background is I've been in insurance for [3:09] 20 years focused mostly on technology [3:12] risk. I've worked in multiple parts of [3:14] brokerages, operations, sales, uh, [3:17] innovation, leadership, and so I I come [3:20] with a lot of different perspectives. [3:22] Um, yeah, look forward to the chat this [3:24] morning and and I think it's it's really [3:26] good. We've got we've got kind of a [3:27] carrier and a broker perspective here, [3:30] which is hopefully a little bit more [3:32] holistic than some panels. So, uh, [3:34] thanks for having us. [3:35] >> Yes, I'm Andrew Zarowski. I work for the [3:37] Hartford. Um and I actually wanted to [3:39] first just thank uh AI4 team for having [3:41] an insurance specific discussion. We we [3:44] we don't have a lot of these uh within [3:46] the industry and insurance to a lot of [3:49] people is not the most exciting [3:50] industry. I disagree with that. I'm [3:52] probably biased but I do disagree with [3:54] that. It is a very interesting industry. [3:56] But one of the things that I think is [3:58] definitely going to be interesting is [4:00] how AI is going to impact the insurance [4:02] industry. It depends on who you listen [4:04] to, of course, but AI will have a [4:06] profound impact on insurance, maybe as [4:09] much or maybe more than any other [4:11] industry that there is. There's a large [4:14] consulting firm that that we've met with [4:16] before that gave percentages on [4:18] industries. You know, this industry is [4:19] going to be 30% impacted, this 50%. The [4:22] insurance industry, they gave 100%. [4:25] Right? Now, that is not necessarily [4:27] automated. That might that's augmented. [4:29] How is AI going to help? but they [4:31] believe that insurance will be 100% [4:33] impacted and that it can't get any any [4:36] higher than that. So it puts us on the [4:37] top of the list. Um so like I said, I'm [4:39] Andrew Zarowski. I'm the head of AI for [4:41] underwriting within our middle and large [4:43] areas, middle and large business areas. [4:46] Um I've worked for the Hartford for [4:47] about 20 years. Uh almost that entire [4:49] time is within the underwriting ranks. [4:52] And it's a it's an interesting [4:54] opportunity to to come at this not from [4:56] a technical perspective but really from [4:58] an underwriting and business perspective [4:59] which is what I'm going to share. [5:01] >> Okay. We have a lot of experience here. [5:04] Okay. So let's start. [5:05] >> Nice way of saying we're old. [5:07] >> Okay. I didn't say that. Uh and everyone [5:10] knows Garrett I guess because uh he goes [5:12] everywhere to all the conferences in [5:15] about AI data. I have noticed a lot you [5:18] going everywhere. So let's touch on [5:21] this. [5:21] >> So have the people that monitor my [5:22] expense account. [5:25] >> Okay. So Andrew touched on this point [5:27] already that AI is impacting us, right? [5:30] So our next section is that how is uh AI [5:34] creating most value and where is it [5:36] creating? [5:37] >> Yeah. [5:38] >> In your personal experiences as well as [5:40] in your u [5:41] >> Yeah. And this is uh I think Garrett had [5:43] mentioned this is so Garrett and I know [5:44] each other by the way in case you didn't [5:45] figure that out from before. But yeah, [5:46] we we've worked together a fair amount [5:48] and it's nice. you're going to hear some [5:49] different perspective of how the broker [5:50] or the distribution partners think about [5:52] it, how the insurance carrier, so I'll [5:53] take the role of the insurance carrier, [5:55] but I want to bring everybody back and [5:57] and again, you know, may maybe not you, [5:59] but Garrett mentioned that I'm old, so [6:00] that's, you know, I've got gray hair. [6:01] Um, I've been in the insurance industry [6:03] for a while and actually does predate me [6:04] a little bit, but um before I got into [6:07] the insurance industry, it was a world [6:09] where the broker had to give you give [6:12] the underwriter all of the information [6:14] because there was no internet. You [6:15] couldn't go find stuff. You had to get [6:17] all the information. So that meant [6:18] applications, that meant interviews, [6:20] different things like that to be able to [6:21] get that information. So then comes the [6:23] internet and then comes the world where [6:26] hey, I'm not going to spend time giving [6:27] you the information. You go find it. You [6:30] use third party data. You use the [6:31] internet and go find that information. [6:34] So now as an underwriter, what you're [6:36] doing a lot of is you're going to find [6:38] stuff. You're gathering things, right? [6:40] You're reading a whole lot of [6:41] information. Also, the world of [6:43] insurance has gotten a lot more [6:45] complicated. were very specified in [6:47] industry, size of risk, different lines [6:49] of business with different underwriting [6:50] guidelines. And what you would have, now [6:53] I know nobody prints anything out [6:54] anymore, but if you did print out all [6:56] this information, it would be a huge [6:58] stack of papers that on every one of our [7:00] clients we need to review in order to [7:02] get to the point of being able to [7:04] understand the company because that's [7:05] step one and then assess their risk. And [7:08] this is where I think AI is really [7:10] coming into play now and in the very [7:12] short term is how do you get all that [7:15] information from all over the place. [7:17] Bring it in, summarize it to some degree [7:20] and give the underwriter what they need [7:21] to get started. And I think that that's [7:23] really where we're going and what we [7:24] have now. And we're starting to see that [7:26] being deployed even today. [7:28] >> Um yeah, I'd also like to point out you [7:30] said you have gray hair. I don't have [7:31] hair. So, um that's how old I am. Um, [7:36] yeah. So, as a broker, I I I tell our [7:38] clients this because I don't obviously [7:41] know everyone's perspective in the room. [7:42] Um, we think of our jobs as as we have [7:47] to sell insurance, but we have to sell [7:49] it twice. And so, I agree with Andrew. [7:54] um kind of some of the the magic behind [7:57] insurance broking is, [8:00] you know, people are used when they're [8:02] buying insurance, they're used to [8:03] getting the options and having someone [8:04] help them pick out which one is the best [8:06] option. [8:07] The first sale that happens kind of [8:09] behind the scenes is when I bring Andrew [8:12] and his his competitors the the [8:16] opportunity and h how well I present [8:20] that opportunity directly correlates [8:22] into the pricing, the scope of coverage. [8:25] And so if I don't do a good job selling [8:27] Andrew, then who cares what I'm bringing [8:30] back to the client because it's not as [8:31] good as it should be. And so we're [8:33] seeing the same thing with submission [8:36] building. You know, my my job is to [8:38] highlight a company, their risk profile, [8:42] everything they're doing well to get [8:44] Andrew and his competitors very excited [8:46] to release terms to write that account. [8:49] And we're seeing we can do that on [8:51] steroids with AI. We can get a much more [8:53] holistic picture around what is their [8:56] risk profile. What are what are the [8:58] things if I think like an underwriter um [9:01] would concern me and how do how do how [9:03] does my client alleviate those concerns [9:05] with risk management techniques and [9:07] contractual risk transfer and the things [9:09] that they're doing. AI is just making us [9:12] so much smarter. So, we're giving Andrew [9:14] and his competitors a much better [9:18] submission to base pricing, base [9:20] coverage. And so, that's that's really [9:23] kind of uh top of mind for us. And then [9:25] everything It would shock no one in the [9:28] room. Brokers have the worst workflows [9:32] of any industry I've ever worked with. [9:34] It's the most manual process. You still [9:37] have people that do print off [9:39] applications and handwrite them and scan [9:42] them in. Like I I hadn't used a scanner [9:45] in a decade 20 years ago when I entered [9:47] insurance and that was part of a [9:49] workflow. And so, um, we've come a long [9:52] way, but AI is really changing how we're [9:55] we're able to take data and pump it [9:58] wherever we need to rather than having [10:00] people copy and paste from one [10:02] spreadsheet to a PDF and then email it [10:05] somewhere. Like that happens every day [10:08] in your office, in my office, probably [10:10] yours too. Um, but it's changing really [10:12] quickly. And and just uh interrupt for [10:15] one second. The I think this is also the [10:16] difference between a good broker or good [10:18] agent and a not so good agent. Good [10:20] agents and brokers we as the [10:22] underwriters get good submissions from. [10:24] And I think AI to your point is going to [10:26] really help with that. Uh because the [10:28] better submissions that we get, the [10:29] better that we're going to be able to do [10:30] a good job underwriting. Right now there [10:32] are not many great brokers and hopefully [10:36] the more that there are, the better it [10:38] will be for the customers. [10:39] >> Yeah. And I don't I don't mean to imply [10:40] that we're trying to trick Andrew into [10:43] releasing coverage and we're using AI to [10:45] like manufacture a good submission. It [10:48] really is to design the best program for [10:51] that insured. And you know, a large [10:55] commercial client is a very complex [10:58] entity to ensure, right? They're not [11:00] just doing one thing. They're doing a [11:02] ton of different things. They have a lot [11:04] of different risks. And not all [11:06] insurance policies are like home and [11:08] auto insurance which for the most part [11:10] are largely standardized. A lot of the [11:12] coverages that we work in with um sorry [11:15] that's my phone. Um with you know things [11:18] like technology companies or life [11:19] science companies these are all very [11:23] bespoke policies and you need to design [11:25] them with the insured in mind making [11:27] sure that their risk are being [11:28] appropriately covered by that policy. [11:31] And we feel like AI is helping us [11:33] deliver a better submission to you to [11:35] make sure you have the full picture on [11:37] what we're actually asking for and what [11:38] the client needs. [11:40] >> Yeah. So what I'm hearing Andrea is [11:41] saying that Garrett, give me good data. [11:44] That's all I need from you. And Garrett [11:46] is saying, I'm trying my best. There you [11:47] go. You have good data from me. So all [11:50] we all know that data is the oil here, [11:52] right? [11:53] >> AI is nothing without good data. And you [11:55] guys already touched on that. So what's [11:57] your data strategy, data quality, data [12:00] governance practice that makes your AI [12:03] applications more helpful, impactful? [12:06] >> Yeah. So, so everybody the the the kind [12:09] of the cool thing to say is data is [12:10] everything, right? That that's what [12:11] everybody says, but you really need to [12:13] dig down and what what does that [12:15] actually mean that data is everything? [12:17] So in in our world, we're trying to [12:19] build different AI tools around [12:21] underwriting. And to do that, you need [12:24] to tell the AI tool what to do. But you [12:27] can't tell the AI tool what to do unless [12:28] you know what the AI tool has in terms [12:30] of data, right? So it's sort of this [12:32] circular idea that you you need [12:35] everything and you need everything at [12:36] the beginning and you need to know what [12:38] it is at the beginning. Um, I get the [12:40] fortunate opportunity to hear a lot of [12:42] use cases from people within my company [12:44] and even people outside of my company [12:46] just bouncing ideas and usually the [12:48] first thing that I say to them is what [12:51] data do you have? Right? Because if the [12:53] answer is I don't have very much or I [12:55] don't have very good data. It's not that [12:57] you can't do it, but just think about [12:59] the prioritization. If I've got one use [13:01] case that has all the data lined up and [13:03] another use case that doesn't have the [13:04] lot data lined up, you could do both, [13:06] but but why? I mean, why wouldn't you [13:08] prioritize the one with data? So, that's [13:10] really what I think about when when it [13:12] comes to what is really data really [13:14] means to people. That's how I look at [13:16] it. [13:17] >> Yeah. And my I'm very much um if you got [13:20] to know me, like I I'm the kid that like [13:22] took apart his toys to figure out how [13:24] the motor worked. And um and then I [13:26] would like I don't like that. I'm going [13:28] to rebuild that. I like burning things [13:30] to the ground and building them back [13:32] bigger, better, stronger. That is my [13:35] belief. Unfortunately, I think in the [13:38] case of carriers, in the case of [13:39] brokers, we have we have a lot of legacy [13:43] tech debt that is based on very [13:46] antiquated technology. There are [13:49] insurance carriers, and I I don't I'm [13:51] not accusing you of this, but there are [13:54] carriers that have underwriting [13:56] actuarial models that are on mainframe [13:58] computers still, [14:00] >> right? So when you talk about things [14:02] like AI and these agentic layers um APIs [14:07] like okay but you're not going to build [14:09] an API into a mainframe computer like [14:12] it's just not it's not practical. Um [14:15] brokers have the same problem where we [14:17] have technology systems they call it an [14:20] agency management system that has all of [14:22] the client data that we need to sub [14:24] submit to the insurance carriers. The [14:26] problem is it's all unstructured data. [14:29] It's it's in a very old SQL server-based [14:34] uh architecture and it's it's [14:36] nonstructured. So you you can't really [14:40] give a complete picture of that data. [14:43] You can't manipulate it in the way that [14:44] we want to. And so we have built our own [14:47] data lake. We have made attempts at [14:51] gaining control of our own data, but [14:53] it's a slice. It's not the whole [14:55] picture. And so when when I think about [14:57] where insurance is going to be in what I [15:00] think three years, but certainly by [15:01] five, where we have AI agents that are [15:05] conducting entire end to-end workflows, [15:07] that's only going to happen with the [15:10] brokers that are probably starting fresh [15:12] today, building a brand new brokerage [15:15] from the ground up on today's [15:17] technology, then I think it's possible. [15:19] I think brokers like IMA and Connor [15:22] Strong are gonna have a real hard time [15:24] doing that because we have um we put [15:27] about 10 11 billion in premium into the [15:31] market. We have gosh tens of thousands [15:35] of clients. um that is a lot of data and [15:38] trying to get that into a new system uh [15:41] it's it's a challenge [15:44] >> and as you mentioned earlier scanned [15:45] PDFs they are the biggest challenge [15:47] right [15:48] >> it's real it's very real [15:50] >> yeah so we touched on this part uh that [15:53] customers they expect faster results [15:55] fast returns from us but we have legacy [15:58] systems we on top of that uh have [16:02] regulations that slow us down so [16:04] question is How are you using AI to [16:07] streamline your operations without [16:09] compromising trust and compliance? [16:13] >> Yeah. So, so the first thing that I'll [16:15] I'll start out with. So, the the [16:16] Hartford, if you don't know our brand, [16:18] we're a 200 plus year old company, [16:20] right? Our ethics trust brand [16:23] non-negotiables there. There's just no [16:24] way we're going to even attempt to to do [16:26] anything there. uh we we have a very [16:29] very focused AI compliance and [16:32] governance regulation process because we [16:35] don't want to even have an opportunity [16:37] to trust that the entire insurance [16:38] industry is built on trust, right? You [16:41] there's a there's a policy, there's a [16:42] payment and then if something goes [16:44] wrong, we pay out. It's all trust. So if [16:46] we break that trust, it it it will [16:48] devastate our industry. So we have a [16:50] mindset of that is absolutely not a [16:52] negotiable. But when I talk to a lot of [16:55] the folks in in the compliance team, [16:56] which I spend a lot of time with them, [16:58] we we often have this interesting [17:00] conversation about, I don't know that [17:02] it's even going to get there. And and [17:04] let me explain to you because from an [17:06] underwriter point of view, the [17:07] underwriter themselves has to absolutely [17:11] keep up their trust and they they have [17:14] to have their integrity because as an [17:16] underwriter, both internally and [17:17] externally, you're only as good as your [17:20] as your ability to underwrite and the [17:21] trust that you're a good underwriter. If [17:23] I'm working with one of my broker [17:25] friends, they're they're going to love [17:27] working with me until the day that I [17:28] burn them and then I'm done for forever. [17:31] So, you can't compromise on that. So, I [17:33] don't when when we develop these [17:35] programs, I explained to them that an [17:36] underwriter and literally you're writing [17:38] your name under, right? You're signing [17:40] off on something. You're not going to do [17:42] it if the AI tool may or may not have [17:45] made a mistake. You're just not going to [17:46] use the tool, right? So I don't even [17:48] know that there's a a scenario where [17:51] that tool gets out of our walls because [17:53] an underwriter won't even use it. So we [17:55] have built trust in from the very [17:57] beginning we have has the highest [17:59] priority. And when we think about [18:00] trusting and and compliance issues [18:02] that's where it it's a non-negotiable [18:05] because people won't even use the [18:06] product internally let alone thinking [18:08] about it outside of our walls. [18:10] >> Yeah. Our CEO has a has a saying like we [18:13] sell a product that nobody wants to buy [18:15] and it's a pain in the ass to use, which [18:18] is really true, right? Um, no one likes [18:21] insurance until something bad happens [18:22] and then it turns out, oh, that's pretty [18:24] good that we had that. Um, but like we [18:27] spend a lot of time I spend a lot of [18:29] time focused on customer experience. You [18:32] know, where where do our clients want us [18:35] to be assisting them in this process? [18:38] And you know, um, I do I I also care [18:41] about ethics and and regulations. Um, [18:44] I'm I'm sure ours is nowhere near as as [18:47] formal as as the Hartfords is. Um, but [18:50] at the same time, we we have to meet [18:53] clients where they want us to be. And I [18:56] don't think that waiting on regulatory [18:59] clarity is the is the right thing. We we [19:03] have data privacy standards uh that we [19:06] adhere to. We're SOCK 2 type two like we [19:09] we are very focused on privacy security [19:12] and we feel like that's probably good [19:15] enough. We need to start experimenting [19:17] with AI for our clients sake. And I'll [19:20] give you an example. Um you know a good [19:23] agency management system should start to [19:26] help our desk level associates in [19:29] building the right submission. So, um, a [19:32] an opportunity comes in from a producer [19:34] that's a salesperson. Um, and let's just [19:36] say it's a $75 million revenue software [19:40] developing company, right? Like that's [19:42] the client that we have to build a [19:43] program for. The system should [19:45] automatically start telling me based on [19:47] real data, where have we placed similar [19:51] clients, right? Who which underwriters [19:53] are most competitive? Which underwriters [19:55] give us the fastest response? which ones [19:57] are the ones that are willing to [19:58] negotiate coverage terms and give us [20:00] pricing flexibility, which which ones do [20:02] we have agreements with? Because brokers [20:04] and carriers structure these these kind [20:06] of back-end deals on like if we both [20:08] have a good year, we're going to both [20:10] participate in the success of that. My I [20:13] I can't tell you the the service teams [20:15] are overwhelmed with the choices. Um and [20:19] a lot of times like the old school way [20:21] of broking business is who brought [20:23] donuts into the office last week? um I [20:26] probably owe them and oh gosh Sandra [20:28] didn't get the last one and I felt bad [20:30] so I should probably send it to her. [20:32] Like I get that but that's not in the [20:35] best interest of the client and that's [20:36] our job is to provide the best product [20:39] for the client and I want that based on [20:41] real data and I'm going to do that with [20:43] AI more than I'm going to do that with [20:45] humans. And so and we do worry about AI [20:49] hallucinations and getting things wrong. [20:51] Um, but I sat in a 18-month digital [20:55] transformation audit where I literally [20:57] watched people conduct their day-to-day [20:59] lives, right? Like, how do you do your [21:01] desk work? And no judgment, we're just [21:03] taking notes and we're just kind of [21:04] trying to figure out how we're going to [21:06] grow. Um, and I watched people make just [21:09] careless mistake after careless mistake. [21:12] Copy paste something, forget uh where [21:14] they were because they got up and got [21:16] coffee, got distracted by a call. Like [21:19] AI doesn't have those challenges. Humans [21:21] make mistakes all the time. AI will [21:23] continue to make mistakes. We will do [21:25] our best to audit the AI, right? I don't [21:28] believe in AI just coming in and, you [21:31] know, fully replacing people. It will [21:34] certainly replace parts of everyone's [21:36] jobs. To your point, 100% will be will [21:39] be augmented. Um, so I think we have to [21:42] be using it today and um, yeah, we we'll [21:45] try and stay, you know, up with [21:47] regulations and making sure that we're [21:49] doing everything the right way. But I [21:50] think for the sake of our clients, we [21:53] have to be pulling AI into everything [21:55] that we do. [21:57] >> Do you guys have any ROI metrics? like [22:00] trust metric is one thing we just [22:03] touched on that but where is the most [22:05] impact as in uh accuracy efficiency [22:09] cost reduction [22:11] >> yeah both I mean every money is [22:13] everything um so if we can improve [22:16] margins somewhere uh so we have a couple [22:18] of platforms that we're using that that [22:20] do um coverage audits right so take um [22:25] take someone's insurance program audit [22:27] the program what's wrong with it? Is it [22:29] missing any endorsements? Are there, you [22:32] know, problematic language baked into [22:34] the the forms? That's typically part of [22:36] the the process when we are competing [22:39] for a piece of business. We are um, you [22:42] know, telling these clients, I think [22:43] we'd do a better job for you than these [22:45] guys would, and I can prove it. If I [22:47] audit their work, I'm going to find a [22:49] bunch of problems in your program. And [22:50] so, we do this big free audit. Probably [22:53] takes six to eight hours to to complete [22:55] end to end. Uh we have AI doing that [22:57] every day of the week. Now um we have AI [23:00] that is uh you know auditing let's say [23:04] for compliance purposes. We have clients [23:06] that have have to make sure that the [23:08] insurance program meets their [23:10] contractual requirements. So we upload [23:12] the contract, we upload the policy and [23:14] we get green light, yellow light, red [23:15] light and we are measuring it based on [23:19] efficiency and cost reduction because a [23:21] lot of that gets outsourced to BPOS and [23:24] people that um are not inexpensive and [23:27] we've estimated um pretty significant [23:30] savings I mean in the millions just by [23:33] pulling that one function back inhouse. [23:36] That's the important part, Andrew. [23:38] >> Yeah. And and so the efficiency and [23:40] quality conversation, I think, is a [23:42] really good one. Everybody likes talking [23:44] about efficiency and and there's real [23:46] gains there, I think, to be made. I'm [23:47] more excited about the quality. I think [23:49] it's a very exciting opportunity, like [23:51] Garrett said. There are people who make [23:53] mistakes. I mean, people make mistakes. [23:54] That's what it is. I think that AI also [23:57] makes mistakes. What we're finding is is [23:58] that when you match really good AI with [24:01] a really good underwriter, the amount of [24:04] mistakes go really, really small. And we [24:06] also have the ability, I think, to get [24:08] to more of the knowledge and more of the [24:11] expertise out there. So, we have this [24:12] one lady who works in our our group. Her [24:15] name is Tracy. She is a fantastic [24:17] underwriter of the real estate industry, [24:19] probably one of the best in the entire [24:20] business. She's one person, right? She [24:23] can't underwrite every single real [24:24] estate industry deal that we have. It's [24:26] not possible. But if we can get her [24:29] level of knowledge out to everybody, [24:31] that's going to make a better customer [24:32] experience. The customers are going to [24:34] spend more time. We're going to spend [24:36] more time thinking about customer [24:37] opportunities about how we can make [24:39] their program better, what kind of risk [24:41] management services we can offer to [24:43] them, what kind of claim services. So, [24:45] it's less on the paperwork and more on [24:47] solving customer problems. And that's [24:49] really what we're trying to bring here [24:50] because the the AI ultimately is is a [24:53] tool, right? It's a tool. We sort of [24:55] look at it and and I love the line um [24:58] the the underwriter is the hero, right? [25:00] That's how we really think about things. [25:02] The underwriter is the hero. AI is the [25:04] tool to allow the hero to save the day [25:06] and we wanted them to save the customer [25:07] and that's how we're thinking about at [25:08] the Hartford. [25:09] >> You want to build traces, bots, virtual [25:12] assistants, multiple [25:13] >> I I fully agree with that. I'm trying to [25:15] So, I also lead a practice called [25:16] digital risk. Um it's our focus on [25:19] blockchain web 3 digital assets. That's [25:22] a really complex type of insurance to [25:25] build programs around. Um, it's I have a [25:29] history in that space. So, I'm kind of [25:31] the one person internally that knows how [25:34] to design programs for crypto exchanges [25:37] and Bitcoin miners and AI [25:39] infrastructure, high performance [25:40] compute, like that kind of stuff. Um, I [25:42] am actively trying to download my brain [25:45] into the system so that I'm not needed [25:48] on every single call. And people are [25:50] like, "Aren't you worried like you're [25:51] building your own replacement?" I'm [25:53] like, "No, because it's going to need [25:54] refinement." And I'm the only one that [25:56] can refine it, but I can democratize my [25:59] brain so that people are not relying on [26:01] me. I'm not a bottleneck. I'm not a [26:03] Tracy. That is the one of that is the [26:07] reason that we do or we don't win a [26:09] piece of business based on availability. [26:11] Um, so yeah, we're trying to do that. [26:13] >> Yeah. [26:13] >> Right now. Then Gary, you can come to [26:15] Vegas anytime you want to because you've [26:16] got the [26:18] >> someone else is working there. [26:19] >> That's right. [26:20] >> Okay. So, touching on this underwriting [26:23] part, Andrew, you did mention that uh [26:25] you do assess risk and then price them. [26:29] So, how do you quantify with these new [26:30] tools? Did anything change for your [26:33] existing risk? And is it teaching you [26:35] new moves like how chess players are [26:37] learning new moves from those parts? [26:40] Yeah, it's it's really a fun [26:42] conversation to have when when you look [26:44] at at a lot of large insurance companies [26:46] and the Hartford is definitely one of [26:47] them. We have we have expertise [26:49] everywhere. So, you think about small [26:51] businesses, middle businesses, large [26:53] businesses across all almost all lines [26:55] of business that we have in all [26:57] industries, but I think where AI is [26:59] really coming is kind of what I was [27:00] explaining before. This idea that we're [27:01] going to be able to get our clients and [27:03] our customers to the right places fast, [27:05] I think is a really differentiating [27:07] opportunity. we can solve almost any [27:10] problem. We just need to get the right [27:12] people and we need to get it done [27:13] quickly. And I I think about [27:15] opportunities all over our organization [27:18] where we have, like I said, the Tracy, [27:20] but we also have individuals that [27:22] specialize in so many different types of [27:24] insurance. Uh but you need to be able to [27:26] get them to the right to the right point [27:27] of contact. And that's really what we're [27:29] aiming to do. Um the example that I give [27:31] oftentimes on more of a basic level is [27:33] that if we receive let's say a [27:35] submission that has 10 separate [27:38] locations. So it's a company that owns [27:40] 10 different locations, right? Nine of [27:42] them might be in uh in Illinois uh not [27:46] in a catastrophe area with relatively [27:48] low insurance limits, right? We don't [27:51] need to spend a lot of time there. We [27:53] want to get through that very quickly. [27:54] Now the 10th location might be in Miami. [27:57] It might have a hurricane risk. Maybe [28:00] there's a a tornado hail risk. That's [28:02] the one that we want to spend a lot of [28:03] time on. The pro the process which most [28:06] insurance carriers deal with today. All [28:08] 10 of those locations need time and [28:11] attention. And I think that we're our [28:13] customers are going to be better served [28:15] if we spend less time on the nine [28:17] locations that are just not that not [28:19] that critical, not that complex, and we [28:21] spend more of our time on the more [28:22] complexity to be able to provide more [28:24] solutions and solve more problems. [28:26] you talk about new risks. Um, so we get [28:28] a lot of questions from um, commercial [28:31] clients about AI and how does that [28:34] change their risk profile? Um, and and [28:37] it does. So, in a lot of cases, carriers [28:40] are coming out with AI exclusions for [28:42] policies that are saying like, look, if [28:44] the claim originates out of someone [28:45] else's LLM, we're not covering it. We [28:48] don't care what the claim is. even if [28:50] it's traditional bodily injury, property [28:52] damage, things that are right down the [28:53] middle, if it's from AI, we're not [28:56] covering it. Um, so that's that's [28:57] something everybody needs to pay [28:58] attention to. And then they're trying to [29:00] figure out how do we shift that burden [29:02] back onto the AI companies that we're [29:04] adopting because we don't know how to [29:06] assess whether this AI is safe or not. [29:08] The sales guy says it is, but like you [29:10] can't trust them. Um, so we're doing a [29:13] lot of work. We we've done this for [29:15] years in the cyber space where cyber [29:18] security companies are out there selling [29:19] security. We can protect you. We can [29:21] defend. Um well these companies are are [29:24] adopting them but they don't have the [29:26] internal muscle to know does this thing [29:28] actually work. Same thing with AI. Um so [29:32] the solution for that is the AI [29:34] companies also need their own AI [29:36] specific insurance. And again, my age [29:39] showing um it's kind of like the Tommy [29:41] boy thing if you know the movie, right? [29:43] Like put a guarantee on the box like it [29:45] makes people feel good seeing that on [29:46] the box, right? Um that's what the AI [29:49] companies need to compete. I look at it [29:51] more as like a marketing expense. Um it [29:54] is protection. It is uh covering in the [29:57] event the AI causes some kind of injury. [30:00] This is really important if you're [30:02] talking about health care, right? If [30:04] we're doing some imaging scanning and we [30:07] miss something and it results in someone [30:09] getting sick or dying, you can bet the [30:12] lawsuits are going to fly and that AI [30:14] company's name is going to be top of the [30:15] list, right? So, how do I know as a [30:18] company that's kind of thinking about [30:20] investing in your platform that if that [30:23] happens, you're going to be it's not all [30:25] my dollar, you're going to play, too. AI [30:28] companies are now b buying AI specific [30:30] insurance. So, it's kind of it's created [30:32] a whole new cottage industry for [30:34] insurance products for us to sell, which [30:35] we love.