Ex-Cover Whale CAIO Sets 6-Step Plan to Halt Insurance AI PoC Failures

Agents, Actuaries & Agentic ai — A Fractional CAIO’s 6‑step Insurance Playbook With Cover Whale — Ai4

Ai4

Outsourcing generative models leaves insurers up to two years behind the technology frontier, requiring carriers to deploy lean in-house teams anchored to strict loss-ratio targets to survive tightening margins.

AI-assisted summary of Agents, Actuaries & Agentic ai — A Fractional CAIO’s 6‑step Insurance Playbook With Cover Whale — original source. Check the original for full context.

Source published: 2025-08-28

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 28, 2025, this presentation features Darien Acosta, Fractional Chief AI Officer and former executive at Cover Whale. According to the video description, the session shares a practical six-step insurance AI blueprint—Align, Govern, Build, Integrate, Scale, and Measure—designed to move initiatives beyond proof-of-concept slide decks while balancing governance and ROI.

## TL;DR Addressing the industry-wide stagnation where insurance AI proofs of concept wither on the vine, Darien Acosta outlines a practical execution framework for deploying production-grade AI. He compares the strategic trade-offs of building in-house versus outsourcing before detailing a six-step lifecycle anchored in hard financial metrics, human-in-the-loop governance, and a cross-functional five-person 'Tiger Team.' The talk maps current enterprise readiness against OpenAI's five levels of AI evolution, cautioning against premature agentic autonomy, and finishes with an analysis of latency bottlenecks in voice-driven customer service systems.

## Atmosphere & style The presentation carries the brisk, no-nonsense urgency of an executive briefing designed to cut through industry hype and regulatory paralysis. Acosta speaks with pragmatic authority and candor, framing technological adoption as a high-stakes operational necessity rather than a speculative experiment. It feels like an insightful, real-world masterclass tailored for technical and business leaders who need actionable guardrails to survive shifting market pressures.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=4s) Insurance AI proofs of concept frequently stall due to a lack of real-world deployment experience, which fuels intense regulatory fear and leads to costly inaction. - [02:16](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=136s) Building in-house offers complete tech stack control and proprietary data leverage, whereas outsourcing provides lower-risk OpEx solutions that lag the AI innovation frontier by one to two years. - [05:10](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=310s) A successful AI implementation lifecycle follows a structured six-step blueprint: Align, Govern, Build, Integrate, Scale, and Measure. - [05:28](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=328s) Step 1 (Align) requires anchoring initiatives to a specific quantitative financial metric paired with an operational guardrail, rather than pursuing vague efficiency goals. - [06:49](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=409s) Step 2 (Govern) necessitates early legal and compliance involvement, rigorous data mapping, and automated Human-in-the-Loop thresholds to prevent project shutdowns. - [08:20](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=500s) Step 3 (Build) is most effectively executed by a dedicated, cross-functional five-person Tiger Team composed of a data scientist, software engineer, insurance SME, product manager, and operations lead. - [09:19](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=559s) Step 4 (Integrate) emphasizes eliminating manual copy-paste handoffs by connecting AI models directly to core policy and claims management systems via APIs. - [11:10](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=670s) Step 5 (Scale) demands cautious, incremental expansion using A/B testing because technology and organizational culture must scale together to avoid internal rejection. - [12:01](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=721s) Step 6 (Measure) balances three core ROI drivers—loss ratio improvement, back-office efficiency, and end-to-end automation—without ever sacrificing auditability. - [12:53](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=773s) Major adoption pitfalls include leading with technology instead of business problems, regulatory paralysis, and building tools in isolation from operational staff. - [14:19](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=859s) According to OpenAI's five levels toward AGI, action-oriented agents are only around 50% mature; attempting agentic autonomy before mastering chatbots and reasoners creates severe operational risks. - [14:19](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=859s) Executive leadership structures will evolve as enterprise software and AI merge, eventually blending the Chief AI Officer (CAIO) and Chief Technology Officer (CTO) roles into one. - [17:44](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=1064s) Replacing legacy phone trees with fully voice-driven agentic IVR systems is currently constrained by speech-to-text and text-to-speech latency bottlenecks and high compute costs.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=4s) The Experience Gap: Why Insurance AI PoCs "Die on the Vine"

In the insurance sector, artificial intelligence is currently on everyone's lips, yet widespread adoption remains paralyzed by inaction. Most AI proofs of concept (PoCs) wither away because the industry suffers from a classic "chicken or the egg" dilemma: there is a profound lack of real-world deployment experience, which breeds intense anxiety around regulatory compliance and potential deployment failures. However, maintaining the status quo or bungling an AI initiative poses the greatest risk of all, leading directly to wasted capital, lost market opportunities, and a loss of competitive standing. Paraphrasing Thomas Edison, the speaker emphasizes that having an AI vision without actual execution amounts to nothing more than hallucination.

```mermaid flowchart LR accTitle: The AI PoC Failure Cycle subgraph Stagnation["The PoC Dilemma"] A["Lack of Deployment Experience"] --> B["Heightened Regulatory Fear"] B --> C["Indecision & Inaction"] C --> D["Lost Market Competitiveness"] D --> A end ```

[02:16](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=136s) Build In-House vs. Outsource: Navigating the Trade-Offs

When embarking on a generative AI initiative, organizations face a fundamental strategic fork in the road: building an internal capability versus hiring an external vendor. An in-house approach is generally preferable because it grants full architectural control, leverages proprietary data nuances, upskills internal talent, and delivers true cutting-edge capabilities. The risk, however, is that building internally demands strict top-down alignment, placing the manager's reputation squarely on the line if the project fails.

Conversely, outsourcing offers a standardized, turnkey playbook with clear SLAs and operational expenditure (OpEx) treatment instead of capital expenditure (CapEx), serving as a safe mechanism to buy time when boards demand progress updates. Yet, this safety comes at the cost of trailing the AI innovation frontier by one to two years. Making this choice is urgent because insurance operates on a "burning platform" characterized by nuclear verdicts and property and casualty (P&C) combined ratios hovering near 100%.

| Strategy | Advantages | Trade-Offs & Disadvantages | |:--- |:--- |:--- | | **Build In-House** | • Complete control over modern tech stack • Leverages proprietary internal data quirks • Invests directly in company talent pool | • Requires robust top-down executive alignment • High personal reputational risk for project leads | | **Outsource to Vendor** | • Turnkey implementation with standard playbooks • Shifts financial model to OpEx • Derisks failure; provides clear SLAs | • Leaves the company 1–2 years behind the AI frontier • Less customization to unique internal data |

[05:10](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=310s) The 6-Step Blueprint: Align, Govern, Build, Integrate, Scale, Measure

To successfully navigate AI deployments within insurance, organizations should follow a structured, six-step lifecycle framework designed to move projects from conception into production without getting derailed by corporate friction.

```mermaid flowchart LR accTitle: The 6-Step Insurance AI Playbook subgraph Phase1["Strategy & Compliance"] S1["1. Align"] --> S2["2. Govern"] end subgraph Phase2["Development & Rollout"] S2 --> S3["3. Build"] S3 --> S4["4. Integrate"] end subgraph Phase3["Expansion & Tracking"] S4 --> S5["5. Scale"] S5 --> S6["6. Measure"] end ```

[05:28](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=328s) Step 1: Align — Anchoring AI to Financial and Guardrail Metrics

Alignment is the single most critical step in the entire playbook. Developing technology in isolation without stakeholder buy-in is the primary cause of project death. Teams must map key stakeholders and abandon vague aspirations like "improving efficiency." Instead, initiatives must target a specific quantitative financial metric paired with an explicit operational guardrail. - Replace generic objectives with hard targets, such as cutting loss adjustment expenses (LAE) by 8%. - Pair financial goals with operational constraints, such as improving loss ratios without degrading customer retention.

[06:49](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=409s) Step 2: Govern — Compliance, Data Mapping, and Human-in-the-Loop Controls

Omitting the legal and compliance teams early in the process invites an immediate shutdown of the initiative later on. Governance requires conducting comprehensive data mapping to establish where sensitive data resides, assessing the downstream legal repercussions of generative model hallucinations, and establishing strict Human-in-the-Loop (HITL) policies. - Map out data lineage, exposure levels, and liability risks for model errors. - Implement explicit automated boundaries—for instance, claims that exceed a designated dollar threshold must automatically route to a human adjuster rather than an automated decisioning engine.

[08:20](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=500s) Step 3: Build — Assembling the 5-Person In-House "Tiger Team"

When building proprietary capabilities in-house, execution speed and quality depend on maintaining a lean, cross-functional team structure. The ideal setup is a dedicated five-person "Tiger Team" with clearly separated responsibilities: - **1 Data Scientist:** Builds, evaluates, and tunes models. - **1 Software Engineer:** Handles traditional software architecture, code quality, and robustness. - **1 Insurance Subject Matter Expert (SME):** Provides domain context, underwriting logic, and workflow expertise. - **1 Product Manager:** Maintains development velocity, roadmaps, and alignment. - **1 Operations Lead:** Understands internal ecosystem communications, workflow handoffs, and system dependencies.

[09:19](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=559s) Step 4: Integrate — Connecting Core Systems via APIs to Eliminate Copy-Paste

A recurring failure mode in insurance automation is partial execution, where an automated output is manually copied and pasted into the next application by an employee. Because manual handoffs inevitably introduce errors, AI capabilities must be tightly coupled to enterprise infrastructure—such as policy administration systems (PAS) and claims management systems (CMS)—via robust APIs. Crucially, software engineers must not be left alone to design the end-user interface; interfaces designed solely by engineers rarely match how operational personnel work.

[11:10](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=670s) Step 5: Scale — Safe A/B Testing and Cultural Alignment

Scaling must be managed cautiously rather than through a forced top-down overhaul. Deployments should expand incrementally across small business segments using controlled A/B testing frameworks. As the maxim goes, "technology and culture scale together or not at all"; attempting to steamroll legacy human processes with raw technology triggers organizational rejection.

[12:01](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=721s) Step 6: Measure — Balancing Core ROI Drivers with Auditability

The final phase evaluates performance against executive priorities while preserving institutional safety. Successful deployments focus on driving three primary business outcomes: 1. **Loss Ratio Improvement:** Sharper underwriting and risk selection. 2. **Back-Office Efficiency:** Reduced administrative friction and operational cost. 3. **End-to-End Automation:** Higher throughput without manual interventions.

These performance gains must never come at the expense of regulatory standing. Sacrificing auditability for short-term operational wins represents dangerous short-term thinking.

[12:53](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=773s) Common Pitfalls and Remediation Strategies

Navigating enterprise AI adoption requires overcoming three chronic organizational pitfalls: - **Tech-First Rather Than Problem-First:** Jumping straight into shiny AI capabilities without anchoring them to a measurable business problem. *Remedy:* Start with the quantitative business outcome, then select the minimal viable technology. - **Regulatory Paralysis:** Paralyzing fear over compliance risk. *Remedy:* Adopt a "crawl, walk, run" deployment strategy that scores early, low-risk wins. - **The Adoption Gap:** Internal resistance caused by building tools in a vacuum. *Remedy:* Engage operational stakeholders continuously throughout the development journey to foster ownership.

[14:19](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=859s) The 5 Levels to AGI and the Convergence of CTO and CAIO Roles

OpenAI's five-stage framework illustrates industry progress toward Artificial General Intelligence (AGI) alongside current organizational readiness: 1. **Chatbots (Conversational AI):** Roughly 98% mature and widely deployed. 2. **Reasoners (Problem Solving):** Roughly 80% mature, represented by advanced thinking models like OpenAI o3 and GPT-5 tier systems. 3. **Agents (Action-Oriented Systems):** Currently at roughly 50% maturity; these systems carry operational risk when executing unauthorized or flawed actions autonomously. 4. **Innovators (Invention & Strategy):** Currently limited to PhD and advanced academic research environments. 5. **Organizations (Fully Autonomous AI Enterprises):** Roughly 1% mature.

```mermaid flowchart LR accTitle: 5 Levels of AI Evolution L1["Level 1: Chatbots (~98% Mature)"] --> L2["Level 2: Reasoners (~80% Mature)"] L2 --> L3["Level 3: Agents (~50% Mature)"] L3 --> L4["Level 4: Innovators (Research Level)"] L4 --> L5["Level 5: Organizations (~1% Mature)"] ```

Attempting to build fully autonomous agentic workflows before an organization understands chatbots and reasoning engines is a hazardous leap over necessary evolutionary steps. Looking forward, executive leadership structures will shift: the Chief AI Officer (CAIO) and Chief Technology Officer (CTO) roles will progressively merge over the coming years as enterprise technology and AI become indistinguishable.

[17:44](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=1064s) Q&A: Voice-Driven Interactive Voice Response (IVR) Latency Bottlenecks

An audience member asks why insurance customer service lines still rely on frustrating phone tree menus ("press one for claims, press two for underwriting, press three if you hate us") rather than an intelligent, human-sounding voice bot capable of handling basic administrative tasks like address changes or claims routing.

The speaker explains that building an intelligent conversational IVR requires integrating three AI layers: chatbots, reasoning engines, and agentic execution. While technically viable today, widespread deployment is held back by voice-to-text and text-to-voice latency over telephony systems. Delivering low-latency, real-time audio conversations requires massive computational infrastructure and high compute expenditures, making widespread, cost-effective rollout a work in progress that requires further engineering refinement.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=4s) Die on the vine — Darien Acosta

> [00:04](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=4s) chicken or the egg — Darien Acosta

> [00:04](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=4s) vision without actual execution amounts to nothing more than hallucination — Darien Acosta

> [02:16](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=136s) burning platform — Darien Acosta

> [05:28](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=328s) improving efficiency — Darien Acosta

> [08:20](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=500s) Tiger Team — Darien Acosta

> [11:10](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=670s) technology and culture scale together or not at all — Darien Acosta

> [12:53](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=773s) Tech-First Rather Than Problem-First — Darien Acosta

> [12:53](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=773s) crawl, walk, run — Darien Acosta

> [17:44](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=1064s) press one for claims, press two for underwriting, press three if you hate us

## Mentioned Darien Acosta, Cover Whale, Ai4, Thomas Edison, OpenAI, OpenAI o3, GPT-5

## Video description Darien Acosta, Fractional Chief AI Officer, ex - Cover Whale Abstracts

What good are LLMs and vision models if they never move beyond the proof‑of‑concept slide deck? In insurance, agentic AI promises sharper underwriting and faster claims, yet governance fears and siloed teams keep most initiatives on ice. In 15 brisk minutes I’ll share the six‑step blueprint I developed as a Chief AI Officer - Align, Govern, Build, Integrate, Scale, Measure - using anonymized field stories to make each step concrete. I’ll highlight the three KPIs that really predict ROI and the guard‑rails that satisfy both actuaries and the business. Join me for the Q&A and leave with a QR‑linked 100‑day audit checklist you can apply the moment you’re back at your desk.

[Source](https://www.youtube.com/watch?v=bhfXJ7vvTDU)

中文简报

前 Cover Whale CAIO 提出六步方案，化解保险业 AI PoC 失败困局外包生成式模型使保险公司落后技术前沿长达两年，承保机构须组建锚定严格赔付率指标的精益自研团队，以在日益收窄的利润空间中求生存。针对保险行业 AI 概念验证（PoC）半途夭折、普遍陷入停滞的现状，Darien Acosta 提出了一套用于部署生产级 AI 的实操框架。他对比了自研与外包的战略权衡，随后详细阐述了以硬性财务指标、人机协同（Human-in-the-Loop）治理和五人跨职能“攻坚团队”（Tiger Team）为核心的六步生命周期。该演讲对照 OpenAI 的 AI 演进五阶段评估了当前企业的就绪度，告诫切勿过早追求智能体自主化，并在最后深入分析了语音客服系统中的延迟瓶颈问题。保险行业的 AI 概念验证（PoC）常因缺乏实际部署经验及对合规风险的担忧而停滞，但在综合成本率逼近 100% 的行业压力下，裹足不前反而是最大的竞争风险。 (0:04) 企业在路径选择上面临权衡：自研可掌控核心技术栈并释放专有数据价值，但对管理层声誉风险高；外包虽能转为 OpEx 模式降低试错风险，却会让企业在 AI 创新前沿落后一至两年。 (2:16) 推进落地的六步蓝图涵盖对齐、治理、构建、集成、扩展与衡量；建议组建由数据科学家、工程师、业务专家、产品经理和运营组成的 5 人突击队，并通过 API 直连核心系统以消除手动复制粘贴。 (5:10) 目前智能体（Agents）成熟度仅约 50%，盲目跨越推理阶段构建全自主系统风险巨大；随着技术与业务深度交织，未来首席 AI 官（CAIO）与首席技术官（CTO）的职能将逐步走向融合。 (14:19)

Reading summary

Insurance artificial intelligence initiatives frequently stall due to an experience gap and regulatory anxiety, even as tight underwriting margins demand operational reform. A successful transition requires choosing deliberately between in-house builds and outsourcing while executing a six-step framework tied to measurable financial metrics and human-in-the-loop governance. The Deployment Gap and Strategic Sourcing Trade-Offs (0:04) Most insurance AI proofs of concept (PoCs) fail because a lack of real-world deployment experience triggers intense regulatory anxiety, resulting in corporate paralysis. The speaker notes that inaction poses severe competitive risks, especially as property and casualty (P&C) combined ratios hover near 100% and nuclear verdicts rise. When initiating generative AI projects, organizations face a strategic fork between internal development and third-party outsourcing. Building in-house provides full architectural control, leverages proprietary data nuances, and upskills personnel, though it exposes project leaders to substantial personal reputational risk if the initiative fails. Conversely, outsourcing to vendors derisks execution through turnkey playbooks, clear service level agreements (SLAs), and operational expenditure (OpEx) accounting. However, vendor solutions typically leave the insurer trailing the AI innovation frontier by one to two years. A Six-Step Framework for Production AI (5:10) To transition AI projects from conception into production, the speaker outlines a six-step lifecycle: Align, Govern, Build, Integrate, Scale, and Measure. Alignment mandates abandoning vague goals like 'improving efficiency' in favor of precise financial targets paired with operational guardrails, such as reducing loss adjustment expenses (LAE) by 8% without compromising policyholder retention. Governance requires early engagement with legal teams to map data lineage, evaluate hallucination risks, and enforce strict Human-in-the-Loop (HITL) policies, such as automatically routing claims over a designated monetary threshold to human adjusters. Proprietary execution centers on a dedicated five-person 'Tiger Team' consisting of a data scientist, software engineer, insurance subject matter expert, product manager, and operations lead. To avoid the common failure mode of employees manually copying and pasting automated outputs, systems must be integrated directly into policy administration and claims systems via APIs. Scaling should proceed gradually through controlled A/B testing, tracking improvements across loss ratios, back-office throughput, and regulatory auditability. Remediating Common Adoption and Governance Pitfalls (12:53) AI initiatives in insurance routinely suffer from three organizational pitfalls: technology-first prioritization, regulatory paralysis, and internal adoption friction. To counter technology-first impulses, organizations must begin with a quantitative business objective and deploy only the minimum viable technology necessary to solve it. Regulatory paralysis is best resolved through a phased 'crawl, walk, run' deployment cadence that establishes early, low-risk wins before taking on complex operational surfaces. Addressing the adoption gap requires continuously embedding operational stakeholders throughout the product development cycle, preventing tools from being developed in an engineering vacuum that human workflows ultimately reject. Agentic Maturity Levels and Latency Bottlenecks (14:19) Using OpenAI's five-tier framework toward Artificial General Intelligence (AGI), the speaker estimates conversational chatbots at roughly 98% maturity and problem-solving reasoners at 80% (exemplified by systems like OpenAI o3). Action-oriented agents stand at roughly 50% maturity, carrying material operational risk if allowed to execute flawed actions autonomously, while fully autonomous organizations remain at roughly 1% maturity. Attempting to deploy autonomous workflows before mastering conversational and reasoning systems represents a hazardous strategic leap. Technical friction also limits voice-driven interactive voice response (IVR) telephony systems. While conversational voice bots could theoretically replace touch-tone menus for administrative tasks, deployment is currently constrained by voice-to-text and text-to-voice latency bottlenecks and high computational expenses. As these technologies mature, the speaker forecasts that the roles of Chief AI Officer (CAIO) and Chief Technology Officer (CTO) will progressively merge as enterprise IT and AI systems become indistinguishable. While AI holds potential to sharpen loss ratios and streamline administrative overhead, practical execution depends on solving telephony latency constraints and resisting premature agentic autonomy. Long-term success requires treating compliance and auditability as essential operating constraints rather than secondary considerations.

精炼阅读摘要

演讲者指出，保险机构在引入生成式人工智能时常陷入因缺乏实操部署经验与担忧合规风险而导致的行动瘫痪；为打破概念验证难以投产的僵局，企业应在权衡自建与外包利弊的基础上，遵循涵盖财务对齐、风控治理、精益构建、系统集成、受控扩展与可审计衡量的六步法执行框架稳妥推进落地。概念验证陷阱与自建外包抉择 (0:04) 演讲者分析认为，保险行业对人工智能的探索大多受困于缺乏真实环境部署经验与担忧合规失败的恶性循环，导致大量概念验证（PoC）项目未及推广便半途夭折。然而，在巨额诉讼判决频发以及财产及意外险（P&C）综合成本率普遍徘徊于100%左右的紧迫背景下，维持现状或执行不当反而是造成资本浪费与竞争劣势的最大风险。演讲者援引爱迪生的观点强调，缺乏实际执行力的AI愿景本质上无异于“幻觉”。在技术路线选择上，企业面临自主构建与外包采购的战略抉择。自建内部能力能够赋予机构对现代技术栈的完整控制权、深度利用专有数据特性并培养内部技术人才，但需要自上而下的坚定支持，项目负责人也需承担直接的职业声誉风险。相反，外部供应商提供交钥匙方案与明确的服务等级协议（SLA），并在财务上计入运营支出（OpEx），是应对董事会短期汇报压力的避险方式；但该路径的代价在于将使企业在AI技术前沿上滞后一至两年，且难以契合专有数据的独特性。指标对齐、风控治理与五人团队 (5:10) 为确保AI项目顺利投产，演讲者提出了六步法生命周期框架，其前三个阶段重点解决战略定调与团队配置问题。第一步是对齐（Align），这是防止孤岛式开发致死的最关键环节；团队必须摒弃“提升效率”等宽泛表述，将目标锁定在量化财务指标与显性业务护栏上，例如在设定将理赔理算费用（LAE）削减8%的同时，明确要求在改善赔付率过程中不得损害客户留存率。第二步是治理（Govern），即尽早纳入法务合规团队，完成数据血缘梳理并评估模型幻觉造成的法律责任，同时必须设定严格的人机协同（HITL）机制，例如规定超出特定金额门槛的理赔必须自动流转至人工理赔员审核，禁止系统全自动决策。第三步是构建（Build），演讲者建议在内部组建精简高效的五人“老虎团队”（Tiger Team），以保障研发效率与敏捷性。该团队明确由五类角色构成：负责模型构建、微调与评估的1名数据科学家，负责底层软件架构与代码工程鲁棒性的1名软件工程师，提供承保逻辑与业务场景支持的1名保险业务专家（SME），负责推进开发节奏与路线图的1名产品经理，以及负责梳理内部系统交互与工作流依赖的1名运营负责人。深度系统集成与受控扩展策略 (9:19) 在实施的后半阶段，第四步是集成（Integrate）。演讲者指出，保险自动化的典型硬伤是“部分自动化”，即员工仍需手动在不同系统间复制粘贴数据，这极易引入人为差错；因此AI系统必须通过API深度嵌入保单管理系统（PAS）和理赔管理系统（CMS）等核心基建，且前端交互绝不能由软件工程师独立设计，必须贴合一线人员的实际工作习惯。第五步是扩展（Scale），推广过程必须借助受控的A/B测试在小范围业务单元渐进开展，演讲者强调“技术与组织文化必须协同扩展，否则寸步难行”，依靠行政命令强推技术往往会引发组织的排异反应。第六步是衡量（Measure），成效追踪应聚焦承保风险筛选优化带来的赔付率改善、后台管理摩擦减少带来的运营增效，以及无需人工干预的端到端自动化率提升三大核心回报杠杆。演讲者特别提醒，业务收益绝不能以牺牲监管合规性与可审计性为代价。针对“技术导向而非业务驱动”、“合规恐惧导致瘫痪”以及“脱离一线引发采纳抵触”这三大常见陷阱，机构应当从量化业务问题出发选择最小可行技术，采取小步快跑的“爬、走、跑”策略积累低风险成果，并在全流程中让运营人员深度参与以培育主导意识。智能体成熟度演化与语音延迟瓶颈 (14:19) 在技术成熟度与演进路径方面，演讲者引用OpenAI的五级框架评估指出：第一级对话式聊天机器人已达到约98%的成熟度并广泛部署；第二级以OpenAI o3与GPT-5梯队系统为代表的推理模型成熟度约为80%；第三级具备行动执行能力的智能体（Agents）成熟度仅约50%，自主行动容易引发越权与执行失误风险；第四级创新发明系统尚停留在顶尖科研阶段；第五级全自主AI组织成熟度仅约1%。演讲者警示，在尚未掌握前序能力前跳跃式推进全自动智能体工作流极具危险性；同时预测随着AI与基础架构的深度融合，首席人工智能官（CAIO）与首席技术官（CTO）的双轨职责将在未来数年内走向合一。针对保险客服电话至今仍普遍依赖按键式交互语音应答（IVR）系统、未能全面由拟人化语音Bot接管的疑问，演讲者解释称，打造智能对话IVR在技术架构上需深度整合聊天、推理与智能体执行三层能力。尽管技术链条已基本跑通，但在电话通信网络中，语音转文字与文字转语音的双向端到端延迟构成了主要物理瓶颈；要实现低延迟的实时对话体验需要极其庞大的计算基础设施与高昂的算力成本支持，大规模低成本落地仍有待工程层面的进一步优化。演讲者总结表明，生成式AI在保险业的渗透并非单纯的技术替换，而是一场涉及核心系统集成、合规风控边界与组织文化适应的系统工程。现阶段企业应客观看待智能体技术的成熟度缺陷，避免激进跃进，在严格恪守可审计性与人机协同边界的前提下，通过小步快跑的闭环验证实现具体的承保与运营效益。

Detailed notes — zh

## Background 该内容由 Ai4 于 2025 年 8 月 28 日发布在 YouTube 平台上。视频主讲人为前 Cover Whale 兼职首席 AI 官（Fractional CAIO）Darien Acosta，主要分享面向保险行业的生成式与智能体 AI 落地部署“六步法”实战框架。 ## TL;DR 很多保险业 AI 项目停留在概念验证（PoC）阶段，主因在于缺乏部署经验与对合规监管的恐惧。演讲者主张优先选择内部自研而非外包，并提出由对齐、治理、构建、集成、扩展及衡量构成的“六步法框架”。同时介绍了由 5 人组成的“老虎队”最佳配置，警示了过度追求 Agentic AI 而忽视底层推理能力的风险，并在 Q&A 环节探讨了语音 AI 在理赔场景中的延时与成本瓶颈。 ## Atmosphere & style 演讲整体呈现出专业、干练且极具务实导向的风格，语言节奏明快且直击痛点。演讲者凭借一线高管的丰富经验，将复杂的 AI 架构与严谨的保险精算逻辑融会贯通，在深刻剖析行业困局的同时提供了清晰可行的解题思路。问答环节则显得真实自然，展现出对前沿技术落地点滴细节的理性考量。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=4s) 大多数保险业 AI 项目停留在概念验证（PoC）阶段，本质上是因为缺乏落地经验与对监管风险的恐惧构成的“先有鸡还是先有蛋”难题。 - [00:04](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=4s) 引用爱迪生名言“没有执行力的愿景本质上只是幻觉”，警告企业在 PoC 上铩羽而归或保持现状会白白浪费资源并失去核心竞争优势。 - [00:21](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=21s) 企业面临自研（In-house）与外包（Vendor）的路径选择，演讲者更倾向于自研，因为自研能积累内部人才与长远控制力，避免技术滞后。 - [00:21](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=21s) 保险业处于核保理赔变难、损害赔偿金激增以及综合成本率接近盈亏平衡线的“正在燃烧的平台”上，保持技术前沿性至关重要。 - [00:51](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=51s) 提出了保险业 AI 落地的六步法框架：对齐目标（Align）、治理与合规（Govern）、构建与团队（Build）、系统集成（Integrate）、安全扩展（Scale）以及量化衡量（Measure）。 - [00:51](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=51s) 项目的首要环节（Align）切忌设定模糊目标，必须选定核心财务指标（如降低理赔调整费用/LAE）及护栏指标（Guardrail KPI）。 - [00:51](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=51s) 必须尽早引入合规部门（Govern），厘清数据存储与风险，并设置“人机协同/人在环路”（Human-in-the-loop）机制来复核高风险业务。 - [00:51](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=51s) 自研团队的最佳配置是 5 人“老虎队”（Tiger Team）：数据科学家、软件工程师、保险领域专家、产品经理和运营负责人。 - [00:51](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=51s) 系统集成（Integrate）应通过 API 将生成式 AI 链接至核心系统，避免断层和员工手动复制粘贴，且不能让工程师单独主导 UI/UX 设计。 - [00:51](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=51s) 安全扩展（Scale）主张技术与文化必须同步通过 A/B 测试稳健扩展；量化衡量（Measure）则需聚焦于赔付率、后勤效率与自动化程度三大高管最关注的 KPI。 - [01:13](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=73s) 企业应避免三个常见陷阱：过度关注 AI 技术而偏离业务问题、因害怕监管风险而止步不前、以及缺乏业务利益相关者的全程参与。 - [01:13](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=73s) 对照 OpenAI AGI 五级演进模型（Chatbots、Reasoners、Agents、Innovators、Organizations），警惕在未熟练掌握基础模型时盲目跃进至 Agentic AI 系统。 - [01:13](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=73s) 预测未来几年内首席技术官（CTO）与首席 AI 官（CAIO）的职责将走向重叠与融合，“所有技术终将 AI 化”。 - [01:17](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=77s) 在理赔电话中实现自然人声 AI 交互的主要瓶颈在于实时文本转语音的延迟问题（Latency Issues）与计算成本，但技术理论上已可行。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=4s) 为什么大多数保险业 AI 概念验证（PoC）会半途夭折？在保险行业中，许多生成式 AI（Gen AI）及 Agentic AI 项目往往停留在概念验证（PoC）阶段而无法真正落地。这种现象本质上是一个“先有鸡还是先有蛋”的难题：保险业普遍缺乏 AI 部署与实施的具体经验（Experience Gap），而这种经验断层又进一步加剧了对监管合规风险（Regulatory Issues）的恐惧，导致企业不敢轻易迈出关键一步。然而，保持现状或在 PoC 项目上铩羽而归才是最糟糕的场景。它不仅会白白浪费资源，更会导致企业错失市场先机。如果在部署尝试中没有吸取任何教训，企业将在日益激烈的竞争中失去核心优势。 * **关于执行力的警句**：演讲者引用托马斯·爱迪生（Thomas Edison）的名言警告道：“没有执行力的愿景本质上只是幻觉”（Vision without execution is hallucination）。目前全行业都在大谈特谈 AI，但多数只是停留在口头讨论，缺乏实质性的落地执行。 --- ### [00:21](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=21s) 自研（In-house）与外包（Vendor）的路径权衡企业在启动 AI 项目时面临的首要抉择是：究竟是在内部自主研发（In-house），还是将其外包给第三方供应商（Vendor）？演讲者明确表示更倾向于自研，因为自研带来的长远价值远远超出其潜在风险。 ``` ┌─────────────────────────┐ │ AI PoC 实施路径抉择 │ └────────────┬────────────┘ │ ┌─────────────────────┴─────────────────────┐ ▼ ▼ ┌──────────────────────┐ ┌──────────────────────┐ │ 内部自研 (In-house) │ │ 第三方外包 (Vendor) │ ├──────────────────────┤ ├──────────────────────┤ │• 掌控力强，部署前沿 │ │• 开箱即用，标准化流程│ │• 投资内部人才团队 │ │• 为董事会问询争取时间│ │• 熟悉自身数据与特性 │ │• 明确 SLA 与责任划分 │ │• 需高层对齐与承担责任│ │• 属于 OPEX 非 CAPEX │ │• 避免技术滞后 1-2 年 │ │• 技术的敏捷性与前沿较差│ └──────────────────────┘ └──────────────────────┘ ``` 演讲者强调，保险业当前正处于一个“正在燃烧的平台”（Burning Platform）上：核保与理赔变得越来越艰难，核定损害赔偿金（Nuclear Verdicts）大幅激增，财产及意外险（P&C Insurance）的综合成本率（Combined Ratio）已接近 100% 的盈亏平衡线甚至亏损。在这种严峻的市场环境下，保持技术的前沿性至关重要。 --- ### [00:51](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=51s) 保险业 AI 落地六步法框架为确保 AI 方案能够成功落地并创造价值，演讲者提出了一个由六个核心环节组成的“六步法框架”（Six-step Playbook）： ``` ┌───────┐ ┌───────┐ ┌───────┐ ┌───────────┐ ┌───────┐ ┌─────────┐ │ 1. Align│──▶│ 2. Govern│──▶│ 3. Build │──▶│ 4. Integrate│──▶│ 5. Scale │──▶│ 6. Measure│ │ 对齐 │ │ 治理 │ │ 构建 │ │ 集成 │ │ 扩展 │ │ 衡量 │ └───────┘ └───────┘ └───────┘ └───────────┘ └───────┘ └─────────┘ ``` * **1. Align（对齐目标）**：这是整个框架中最关键的一步。绝不能闭门造车或设定“提升 IT 效率”这类模糊笼统的目标，必须与利益相关者（Stakeholders）充分沟通，选定一个核心财务指标（如“降低 8% 的理赔调整费用/LAE”）以及一个护栏指标（Guardrail KPI，如“在降低赔付率的同时不损害客户留存率”）。 * **2. Govern（治理与合规）**：合规部门或法务总监（General Counsel）如果未被尽早引入，随时可能叫停项目。必须厘清数据存储位置、数据风险等级以及模型出错后的后果，并明确设置“人机协同/人在环路”（Human-in-the-loop）机制，例如高额理赔必须交由人工复核。 * **3. Build（构建与团队配置）**：自研团队的最佳规模是 5 人的“老虎队”（Tiger Team）： - **1 名数据科学家**（Data Scientist） - **1 名传统软件工程师**（Software Engineer） - **1 名保险领域专家**（Subject Matter Expert） - **1 名产品经理**（Product Manager） - **1 名运营负责人**（Operations Lead，熟悉各业务系统交互） * **4. Integrate（系统集成）**：避免出现“自动化了某一段流程，员工却还在手动复制粘贴输出结果到下一个系统”的断层现象。应积极利用 API 将生成式 AI 链接至核心系统（如保单管理系统 PAS 或理赔管理系统 CMS），切忌采用零碎搭积木的方式。同时不要让工程师完全主导 UI/UX 设计，因为工程师眼中的合理逻辑对普通用户而言往往难以理解。 * **5. Scale（安全扩展）**：通过小范围的 A/B 测试进行稳健扩展。“技术与文化必须同步扩展，否则两者都无法扩展”（Tech and culture scale together or not at all），决不能单靠技术蛮干破坏公司既有的业务流程。 * **6. Measure（量化衡量）**：聚焦于高管与董事会最关心的三大 KPI（赔付率改善、后勤部门效率提升、自动化程度增加），同时确保系统能够通过合规审计，不因追求 AI 效益而增加合规与监管风险。 --- ### [01:13](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=73s) 避坑指南与 OpenAI 的 AGI 五级演进模型演讲者指出了企业在推进 AI 时常见的三个陷阱：**偏离业务问题而过度关注 AI 技术本身**（应以业务目标倒推技术选型）、**因害怕监管风险而止步不前**（应采取“先爬、再走、后跑”的循序渐进策略，从小胜开始积攒信心）、**缺乏利益相关者的协同**（必须让业务方全程参与并保持积极性）。演讲者分享了 OpenAI 关于通往通用人工智能（AGI）的五级演进模型，并对保险业的技术现状与发展趋势做出了评估： ``` Level 1: Chatbots (聊天机器人) ───────▶ 进度 98% (非常成熟) Level 2: Reasoners (推理者) ───────▶ 进度 80% (如 o3 / GPT-5 思考模型) Level 3: Agents (智能体) ───────▶ 进度 50% (尚显不成熟，易出错) Level 4: Innovators (创新者) ───────▶ 仅限于科研/PhD 层面 Level 5: Organizations (组织级 AI) ───────▶ 进度仅 1% ``` * **跳级的风险**：如果企业在连 Chatbots 和 Reasoners 都还没用熟的情况下，就急于跃进至 Agentic AI 系统，就如同在楼梯上尝试“跳级跨越”，极易引发操作风险。 * **未来趋势预测**：演讲者预测，在未来几年内，首席技术官（CTO）与首席 AI 官（CAIO）的职责将重叠并最终走向融合，“所有技术终将 AI 化”。 --- ### [01:17](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=77s) 问答环节：语音 AI 在理赔中的应用与延时瓶颈在 Q&A 环节中，一位听众提问：为什么目前在理赔电话中很少看到完全由自然人声 AI 接听并处理简易业务（如变更地址、报案理赔等）的场景，而是依然需要听取传统的按键菜单（IVR）？演讲者回应道，这实际上融合了 Chatbots、Reasoners 和 Agents 三种技术的结合。当前未能大规模落地的主要瓶颈在于**延迟问题**（Latency Issues）。将文本处理实时转换为语音（尤其是在电话线路上）需要较高的计算成本，为了降低延迟，企业必须投入巨额资金运行专门的系统。不过从技术理论上讲，这一目标已经实现，未来只需要在工程上做进一步的优化与精细化打磨。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=4s) Vision without execution is hallucination — Darien Acosta > [01:53](https://www.youtube.com/watch?v=bhfXJ7vvTDU&t=113s) Tech and culture scale together or not at all — Darien Acosta ## Mentioned Darien Acosta, Cover Whale, Ai4, Thomas Edison, OpenAI, AGI, GPT-5, o3, PAS, CMS, Agentic AI, Gen AI ## Video description Darien Acosta, Fractional Chief AI Officer, ex - Cover Whale Abstracts What good are LLMs and vision models if they never move beyond the proof‑of‑concept slide deck? In insurance, agentic AI promises sharper underwriting and faster claims, yet governance fears and siloed teams keep most initiatives on ice. In 15 brisk minutes I’ll share the six‑step blueprint I developed as a Chief AI Officer - Align, Govern, Build, Integrate, Scale, Measure - using anonymized field stories to make each step concrete. I’ll highlight the three KPIs that really predict ROI and the guard‑rails that satisfy both actuaries and the business. Join me for the Q&A and leave with a QR‑linked 100‑day audit checklist you can apply the moment you’re back at your desk. [Source](https://www.youtube.com/watch?v=bhfXJ7vvTDU)

Source quotes

[0:04] Darien Acosta: Vision without execution is hallucination [1:53] Darien Acosta: Tech and culture scale together or not at all

Source quotes

[0:04] Darien Acosta: Die on the vine [0:04] Darien Acosta: chicken or the egg [0:04] Darien Acosta: vision without actual execution amounts to nothing more than hallucination [2:16] Darien Acosta: burning platform [5:28] Darien Acosta: improving efficiency [8:20] Darien Acosta: Tiger Team [11:10] Darien Acosta: technology and culture scale together or not at all [12:53] Darien Acosta: Tech-First Rather Than Problem-First [12:53] Darien Acosta: crawl, walk, run [17:44] press one for claims, press two for underwriting, press three if you hate us

Original transcript

[0:04] Okay. So, uh, next slide. [0:09] All right. So I'm going to talk about um [0:13] basically [0:15] part of my experience [0:17] um working with deploying Gen AI um [0:22] solutions uh within insurance companies [0:24] and the framework [0:26] uh that I uh created to basically ensure [0:30] success. Um so the the first question is [0:35] uh why do most AI uh proof of concepts [0:38] die on the vine? Um and to to me this is [0:42] uh partially due to a chicken or the egg [0:44] problem. Uh there's a lack of uh [0:48] experience in the insurance industry uh [0:51] in terms of deploying imple and [0:54] implementing uh uh AI. Um so there's an [0:58] experience gap and because of that [1:01] experience gap uh there's uh a fear uh [1:06] that uh there are uh regulatory [1:11] uh issues that uh might prevent a a [1:16] successful uh deployment. But the [1:19] reality is that the cost of doing [1:22] nothing or failing with an AI pro uh [1:25] proof of concept is uh the the worst [1:29] case scenario. uh it leads to uh lost [1:32] opportunities and uh wasted uh resources [1:37] and generally uh if you're not learning [1:40] anything from your deployment um you [1:43] you're basically um not being [1:46] competitive in in the market. I think [1:50] uh right now every AI is on the tip of [1:54] everyone's tongues. We're all talking [1:56] about it, but it's it's pretty much just [1:59] all talk. Um and uh there needs to be [2:02] more execution. So it like Thomas Edison [2:06] said uh vision without uh AI uh without [2:11] execution is basically hallucination. Uh [2:14] next slide. [2:16] So [2:18] uh the first [2:20] was there a slide before this or or no? [2:24] Okay. All right. So the the slides are [2:27] outdated but I I'll just speak from it. [2:29] Um the the the first uh thing that needs [2:34] to be done uh is you have to make a [2:36] decision. Is your proof of concept going [2:39] to be built in-house or are you going to [2:41] outsource it to a a vendor? Uh generally [2:45] uh I like in-house because I think the [2:48] the pros uh out outweigh the the cons. [2:52] uh if you build in-house uh you can uh [2:55] have more control and ensure that [2:58] there's uh cutting that you you are [3:00] deploying cutting edge solutions. You're [3:03] investing in your company's uh talent [3:05] pool uh and you're you'll move faster [3:08] because you're uh your people know your [3:12] data and the the quirks of working with [3:15] that data versus [3:17] uh outsourcing to a a vendor. Uh [3:21] generally [3:23] um it's it's turnkey. Uh there's a [3:27] standardized playbook. Uh it's maybe a [3:30] safer bet because it buys time. Uh if uh [3:34] your board is asking you where the [3:36] company is on uh AI integration, you can [3:40] just say well we have XYZ uh company [3:43] working on it. Uh there's clear SLAs's [3:46] and general accountability. And [3:48] generally when you outsource uh you uh [3:52] you're it's an operational expenditure [3:55] versus capex. Uh there's cons in either [4:00] uh path. Uh if you build in-house, you [4:04] need uh you really need top down [4:06] alignment. Uh and the manager's [4:09] reputation is on the line if it flops. [4:11] uh if you outsource you don't have to [4:13] deal you can derisk the probability of [4:16] failure but you're exchanging that risk [4:19] for uh being one to two years be behind [4:23] uh the the frontier uh of potential AI [4:27] applications. [4:29] Um but [4:32] generally uh this is still uh good to do [4:36] because [4:38] in insurance in general is a a burning [4:42] uh platform. It's getting harder to to [4:45] work in this space. There's an increase [4:47] in nuclear uh uh uh verdicts. um uh PNC [4:53] is almost uh operating at a combined [4:57] loss ratio of 100 um and uh it's just [5:02] becoming more and more of a difficult uh [5:05] space to operate in. Uh next slide. [5:10] So in in general this is the framework [5:13] um [5:15] you know step one align step two govern [5:19] three build four integrate uh five scale [5:22] and six measure I'll go into each one uh [5:26] next slide. [5:28] So al a line is the most important um [5:32] aspect of of this framework. uh you want [5:35] to set clear goals and know what you're [5:38] you're promising to your company either [5:41] uh using a vendor or building in-house. [5:45] Uh this is the number one thing that [5:48] will kill any initiative is sort of [5:51] going off on your own and just building [5:52] something without talking to anyone. Um [5:55] so you you will want to have [5:58] conversations map your stakeholders uh [6:01] and identify the the one key financial [6:05] metric uh that you're trying to target [6:08] and improve uh and then one guard rail [6:11] uh KPI uh so for example improve uh uh [6:15] loss ratio without harming retention [6:19] um [6:21] in in in general Um, [6:25] yeah, you don't want a generic you you [6:27] don't want to go into this with a [6:29] generic uh objective like you don't say, [6:33] well, I'm going to use AI to improve it [6:35] efficiency. You'll say, I want to use uh [6:38] uh AI to achieve a quantitative measure [6:41] like uh cut loss adjustment expense by [6:44] 8%. Uh, next slide. [6:49] Second is uh governance framework. um [6:53] your entire your internal compliance uh [6:57] department or team or general counsel [6:59] will shut down your project if you don't [7:02] include them. So this is the the second [7:05] this is part of alignment but it's it it [7:08] goes deeper because um you need to [7:11] understand [7:13] uh where your data lives what data your [7:17] um [7:18] what what what data you're using uh what [7:22] the potential risk level if that data is [7:24] misused uh if if a gen AI model makes a [7:29] mistake what what's the uh what are the [7:32] repercussions s. Um, so all of this [7:35] needs needs to be mapped out in [7:37] conjunction with uh the whoever is [7:41] running compliance at your company. In [7:43] general, you want to add in also human [7:46] in the loop policies where you're [7:48] saying, "Okay, uh we can automate uh [7:52] these specific steps, but these this [7:55] specific trigger will always uh need to [8:00] be reviewed by by a human. Uh for [8:03] example, uh if there's a specific claim [8:06] above a certain amount, uh you don't [8:08] want an AI automatically making a [8:10] judgment on that. you want to pass it to [8:13] the appropriate uh human employee. [8:17] Next slide. [8:20] Um this one is the actual uh fun part [8:24] assuming you're building uh in-house. Uh [8:27] there's a lot of uh different ways you [8:29] could approach uh working with a with a [8:32] tech stack. Um I'm assuming this is [8:38] you're you've chosen to build in-house. [8:41] uh what I find work works is a team of [8:44] uh five people I'm calling them a tiger [8:46] team uh so one data scientist one [8:50] traditionally trained software engineer [8:52] one subject matter expert in insurance [8:55] uh one product manager to to keep things [8:58] uh moving along and one operations uh [9:01] lead the person who understands how all [9:04] the different uh components of your [9:06] company systems uh work uh and [9:09] communicate with each there. [9:12] Um [9:14] I could go longer on this but running [9:16] out of time. Next slide. [9:19] Um in integrate. So this one is [9:24] basically if you automate one part of a [9:28] a a workflow [9:31] generally what I've seen at companies is [9:36] PE uh employees are looking are happy to [9:41] have one component of the workflow [9:44] automated but then they're just like [9:46] copying and pasting uh the output from [9:49] one process into uh uh the input of [9:53] another process. Generally, I like to [9:56] adopt the the the the phrase uh if [10:00] there's a manual process, you want to [10:01] avoid that because manual means [10:03] mistakes. Uh and we have technology here [10:07] that uh can be um organized in a way [10:12] where you uh link workflows together uh [10:16] using APIs. Uh and generally if you [10:20] build a uh a a Gen AI product, you want [10:25] to link it to uh core uh platforms like [10:29] the policy administrative uh policy [10:32] administrative system or a claims [10:35] management system. Most of these [10:37] platforms support APIs. Um but you you [10:40] really want to lean into it. Don't don't [10:43] uh um take uh peacemeal approaches. Um, [10:49] also, uh, don't let your engineers [10:51] design all of the, uh, components or or [10:55] the the general look and feel of the [10:58] product that you're building. Um, what [11:00] makes sense to an engineer probably [11:03] makes no sense to uh, a regular person. [11:07] Uh, next slide. [11:10] So this is once you've built a specific [11:15] uh proof of concept, you want to scale [11:17] it out or scale it up uh in um a a a [11:22] safe way, a limited way. You want to AB [11:25] test uh with your uh existing with with [11:29] small portions of of your of your [11:31] business. Um basically there's the [11:34] saying like tech and culture scale [11:37] together or not at all. um [11:40] you you need uh both those things to to [11:43] work in in in conjun in conjunction. You [11:46] don't just want to use technology to [11:48] bulldoze um all the uh existing [11:52] processes uh in your uh in your company. [11:57] Uh next slide. [12:01] Okay. So these are the three uh KPIs [12:05] that I think in general uh most uh [12:10] boards most executive um uh teams are at [12:15] insurance companies are looking to [12:17] achieve. Uh basically this is uh [12:20] improvements in loss ratio uh back [12:23] office efficiency and automation. [12:27] Um, and you want to achieve these two [12:31] things while also not compromising [12:35] uh the ability of the company to pass [12:37] audits. Um, if you're if you're trading [12:41] uh if your success with AI uh increases [12:45] regulatory risk, that's just short-term [12:48] thinking. [12:50] Uh, next slide. [12:53] Okay. So these are the uh three three [12:58] pitfalls uh and how they can uh be [13:01] addressed. Uh generally [13:05] focusing on AI capabilities [13:09] uh rather than business problems. Uh [13:11] some some people they're like we're [13:13] going to do something with AI uh bec and [13:17] this is uh similar to uh the alignment [13:20] piece. Um you want to make sure that [13:23] everything is is quantitative. Start [13:26] with the business outcome then select [13:28] the appropri appropriate technology. Uh [13:31] and then uh governance. This is again uh [13:35] a hearken back to um step step two. Um [13:41] if if your organization is afraid of uh [13:46] regulatory risk, start with very small [13:49] small wins. Generally, uh you want to [13:52] walk uh crawl and then run. [13:55] And then uh back to to alignment um [14:00] there's the adoption gap. Uh you always [14:04] need to include uh your stakeholders in [14:08] the com in the conversations. Uh get [14:11] them excited and uh involve them in in [14:14] in the journey. [14:16] Uh next slide. [14:19] Okay. Um, [14:22] this is this is a bonus slide just to [14:24] kind of like spice things up because I I [14:26] do think the the framework um is is [14:30] generally uh self-explanatory once you [14:34] you work through the pro the the process [14:36] and I will include uh these slides as a [14:39] download uh in the application but uh [14:43] this is something uh that I think about [14:46] uh very often um Where is AI going? Um [14:52] what is the the the pathway to what they [14:56] call uh AGI, a a um a system that can do [15:01] anything. Uh and if you're unfamiliar, [15:03] there's five levels for what they [15:06] consider what OpenAI is considering uh [15:09] the pathway to AGI. [15:11] There's chatbots, which I think is we're [15:14] like 98% [15:15] uh there. There's reasoners which are uh [15:18] problem solvers. This is like your 03 [15:22] model or chat GBT5 with thinking. Uh [15:26] that's like 80% accurate or 80% of the [15:29] way there. And then there's agents which [15:32] everyone's talking about agents now. [15:34] These are systems that can take action [15:35] on your behalf. And I I think the [15:38] reality is that a the agentic systems [15:41] I'm seeing are a bit uh premature. Um, [15:45] we're like halfway through uh a as as a [15:51] as an industry or as a society figuring [15:55] out uh how we could uh improve [15:59] uh the [16:02] agents ability to not make mistakes [16:05] because there's risk, right? Um if it [16:08] what if what if it takes the wrong [16:10] actions? [16:12] The the next level beyond that uh is [16:15] innovators something that can help a [16:17] person uh invent things. We're we're [16:20] seeing some of this but mostly at the P [16:22] uh PhD and research level. This doesn't [16:25] really help people in business unless [16:28] they're uh using things unless they're [16:31] using AI to to strategize. And then the [16:34] the final level is like uh organizations [16:37] basically an AI that can do the work of [16:40] the entire company. Uh, and we're only [16:42] like 1% there. Um, I think trying to uh [16:49] get a company to build agentic systems [16:52] before your company has gained [16:54] familiarity with like how chat bots [16:56] work, how reasoners uh work is a little [16:59] bit like this image, right? It's like [17:01] you're you're trying to skip steps. Um, [17:04] and in general, you you want to start [17:06] small. uh again with the uh the walk [17:09] crawl walk crawl run. Uh but my [17:13] prediction is that in a few years uh [17:16] basically we'll see more and more more [17:18] overlap between CTO and chief AI uh [17:22] officer roles uh until they eventually [17:25] merge and everything that is uh is tech [17:27] will become AI. Uh next slide and then [17:32] um that's the end of my talk. I'll take [17:35] questions, but um in in general um yeah, [17:39] let's let's connect. Um and first [17:43] question, [17:44] >> something that seems to me to be within [17:47] the realm, but I haven't seen it [17:48] actually out there yet that instead of [17:50] getting the voice message, you're [17:53] calling about claims, hit one. If you're [17:55] calling about underrite, hit two. If [17:56] you're calling because you hate us, hit [17:57] three. [17:58] uh instead of that you actually have a [18:01] human voice come on and say this is so [18:03] and so insurance company can I help you [18:06] and this is this is a a bot doing this [18:08] and it's technically there it just I [18:12] haven't seen it actually appear [18:14] >> and then have it then do the simplest of [18:17] stuff that doesn't require thinking like [18:20] change of address I need to file a claim [18:22] and then it moves you there so you feel [18:24] like you haven't had to go through that [18:27] stupid cycle of listening to all the [18:29] alternatives. [18:29] >> I I think that's a combination of both [18:31] one, two, and three. So, chat bots, [18:33] reasoners, and and agents. Um the reason [18:36] why you're not seeing that now is [18:38] because of latency issues. It takes too [18:41] long to um basically have uh a a text [18:46] process be translated into a voice voice [18:49] process, especially over the phone. Um [18:53] it there's ways to get around it, but [18:55] generally it require it it makes things [18:57] computationally more expensive. So [18:59] basically you'd have to throw a lot of [19:00] money at running your own systems to [19:03] reduce this this latency uh component. U [19:07] but like theoretically like we're [19:09] already there. It just needs a little [19:12] bit more refinement uh in in my opinion. [19:15] Um and like that's that's the end of my [19:18] talk. Thank you. [19:20] Catch me after if you're if you have [19:22] questions.