Samsung, Munich Re Replace Static BI Dashboards with AI-Driven Actions

三星与慕尼黑再保险以AI驱动行动取代静态BI看板 商业智能投资难以获得回报，除非产品团队预先设计数据架构，并将预测性AI洞察直接接入自动化业务流程。 专家小组探讨了从静态、历史性的BI报告向主动、规范性AI Copilot的转变，后者能够发掘隐藏的行为模式，并将洞察生成时间从数周压缩至数秒。结合消费硬件、金融科技、再保险和客户服务等领域的企业案例，演讲嘉宾强调，AI价值的实现高度依赖于由产品经理主导的上游数据设计。他们着重指出了多项关键运营实践，包括跟踪洞察转化率、使非结构化对话数据结构化，以及应用动态持续风险监控。在整个讨论中，专家们重申AI旨在增强而非取代人类的领域专业知识，因此所有自动化建议都必须遵循人在回路（human-in-the-loop）的治理机制。 现代 AI 正在架起定性与定量数据之间的桥梁，将非结构化客户反馈与通话记录直接转化为可操作的运营洞察。 (1:24) 商业智能正从被动回顾的历史仪表盘，转向能够预测风险并提供具体避险行动的规范性（prescriptive）智能副驾。 (4:16) 现代 BI 的核心运营指标已从单纯的洞察生成转向“价值变现”，重点衡量洞察实际落地与规模化执行的比例及落地周期。 (11:51) 生成式 AI 与 LLM 正在将客服中心传统的质检延迟，从长达数周或数月的抽检大幅压缩至近乎实时的自动化评估。 (17:50) AI 的长远投资回报率依赖于上游的数据设计，要求产品经理在功能构思之初就明确遥测指标，并按用户意图对对话数据进行标记与切分。 (13:07)

Industry panelists argue that artificial intelligence is transforming business intelligence from static descriptive reporting into proactive, prescriptive execution by converting unstructured data into operational workflows. However, realizing measurable business value requires intentional upstream data design, cross-functional alignment, and strict human oversight. Converting Unstructured Conversations into Measurable Action (1:24) Traditional business intelligence historically relied on structured, historical data, leaving valuable qualitative customer touchpoints unexploited. Archana Acharya cites an under-the-desk elliptical startup that achieved a $100 million exit after an AI analytics engine parsed qualitative reviews to reveal an unaddressed customer base among senior citizens seeking joint-friendly exercise. In financial services and telecommunications, natural language processing turns live call audio into structured text to monitor prospective borrower sentiment during mortgage origination and uncover operational friction in enterprise contact centers. Transitioning from Historical Dashboards to Prescriptive Simulation (4:16) Panelists contrast legacy dashboards that merely report past metrics with AI systems that forecast anomalies and prescribe operational fixes. Nilay Atmaca illustrates this distinction through a Munich Re case study: portfolio analysis surfaced a 5% performance drop among top-tier life insurance agents, which AI diagnosed as high 30- to 60-day policy lapse rates tied to rushed Friday afternoon submissions. Banning Friday late-day submissions and requiring Monday morning reviews directly curtailed client remorse. Upfront Data Architecture Governs Implementation Success (11:51) A critical bottleneck in moving from analytical insight to production execution is the gap between insight generation and implementation. Atmaca notes that early initiatives suffered from low implementation rates of only 5% to 10% because projects lacked upfront alignment with business sponsors, leading to discarded work and damaged team morale. Establishing clear target outcomes improved implementation rates to 30%–40%, while non-deployed experiments were cataloged as formal learning KPIs. Acharya adds that product managers must intentionally structure conversational data and telemetry at inception, warning that retrofitting retrieval-augmented generation (RAG) and fine-tuning onto poorly structured archives inevitably degrades model accuracy. Augmenting Human Expertise with Decision Safeguards (23:47) The panel emphasizes that AI tools augment rather than replace specialized human judgment. In life reinsurance, Atmaca explains that automated underwriting clears standard, low-risk policies without chronic illnesses, freeing actuaries, underwriters, and medical professionals to analyze complex risks such as post-COVID long-term mortality. Angella Francis compares enterprise AI to a vehicle navigation system: the model calculates and presents multi-scenario options based on internal operational data, but human operators must weigh strategic tradeoffs and make the final determination. Dynamic Monitoring and Omnichannel Unification (34:24) Looking forward, the panelists outline shifts toward continuous portfolio surveillance and integrated customer intelligence. Rather than relying on point-of-sale snapshots, Munich Re incorporates third-party medical claims data alongside partnerships, such as discounted multi-cancer early detection testing, to monitor health trends dynamically. Venkata Pakkala explains that Samsung Electronics America is working to unify siloed metrics—including auto-scored quality assurance calls, web browsing behavior, Net Promoter Scores, and customer lifetime value—into coordinated AI agents that prescribe real-time, channel-specific actions under continuous expert review. While AI tooling substantially compresses analysis latency from weeks to seconds, panelists observe that enterprises still struggle with unstructured data governance and execution rates. Acharya notes that developing models capable of genuine cognitive reasoning remains an unsolved frontier for future system architectures.

本次研讨会的核心观点在于：现代商业智能（BI）正从被动的历史描述性报表演进为具备前瞻性的预测与规范型“智能副驾驶”，但要实现真正的业务价值变现，企业必须摆脱单纯追求洞察数量的误区，转而深耕上游数据意图设计、提高洞察落地执行率，并始终坚持人机协同的治理闭环。这一逻辑由三星电子（Samsung Electronics America）的 Venkata Pakkala、Heed AI 的 Archana Acharya、慕尼黑再保险（Munich Re）的 Nilay Atmaca 以及 Embrace AI Growth 的 Angella Watkins Francis 共同阐述。 范式跃迁：从静态报表到预测型副驾驶 (4:16) Nilay Atmaca 将传统商业智能与 AI 驱动型 BI 的差异比作飞机仪表盘与智能副驾驶：传统 BI 仅停留在提供高度、气温和航速等静态历史指标，回答“发生了什么及为何发生”；而 AI 驱动的 BI 则能实时运行预测模拟，提前发出颠簸预警并生成具体的规避路径。慕尼黑再保险在寿险再保险分析中便借此捕捉到了传统机器学习遗漏的微观异常：数据显示顶级代理人存在约 5% 的业绩反常下滑，AI 归因分析揭示周五下午匆忙提交的保单在 30 至 60 天内的失效退保率显著升高，原因在于急躁推销导致客户在周末产生购买反悔情绪。基于该发现，承保机构禁止在周五下午提交保单并推行周一晨会复核机制，大幅降低了失效风险。 Angella Watkins Francis 进一步指出，现代 AI 将企业从传统滞后数周或数月的周报、月报节奏中解放出来，将“洞察获取时间”（time-to-insight）大幅压缩至分钟乃至秒级。通过将数据分析从描述性回顾扩展至趋势预测与行动推荐，AI 成为指导管理层高效决策的企业级导航系统。 上游数据设计与非结构化会话结构化 (1:24) Archana Acharya 强调，跨层级的信息衰减是企业实现“从零到洞察”的主要阻碍。产品经理必须在产品设计首日就介入数据架构，明确界定用户行为产生的遥测数据及其下游分析意图，否则下游 AI 模型的输出将彻底沦为随机撞大运。在技术落地层面，微调专有模型与搭建检索增强生成（RAG）架构的最大技术瓶颈，正是缺乏清洗和标注的非结构化会话存档。企业必须将客服电话、邮件和在线聊天等交互解构为标准的主题语义单元，才能有效驱动自动化流程。 将非结构化反馈转化为量化收益已展现出显著商业价值。Archana 分享了一家生产桌下椭圆机硬件初创公司的案例：AI 引擎挖掘非结构化评价发现老年群体高度青睐该产品对老化关节低冲击的特性，企业据此调整营销策略，最终售出数百万台并达成 1 亿美元的退出规模；在金融科技与抵押贷款领域，实时语音量化技术能在通话中监测客户情绪并识别未满足需求，防止交易流失。Venkata Pakkala 也补充指出，三星电子利用大语言模型（LLM）将多模态通话与聊天记录进行自动化实时质检，彻底消除了传统人工审查长达数周至数月的滞后积压。 价值实现：考核落地执行率与组织复盘 (11:51) Archana Acharya 借用农业生长周期比喻从数据到价值的转化：播种与浇水代表数据的摄取与精炼，枝头开花结果代表形成“可落地洞察”的价值创造，而真正将果实采摘并食用才算完成“价值实现”。因此，现代商业智能的考核核心指标不再是系统生成了多少仪表板或报告，而是这些洞察实际被转化为业务行动并规模化推广的百分比，以及从发现到落地的周转周期。 针对执行落地难的问题，Nilay Atmaca 透露其团队早期仅有 5% 至 10% 的发掘洞察能够进入最终的生产实施，缺乏业务赞助商前期的明确价值定义导致大量项目被搁置，严重挫伤了技术团队士气。通过推行前期价值坦诚对齐，并将未落地实验中沉淀的行为与架构认知正式列为 BI 考核的关键绩效指标（KPI），团队的实施转化率成功提升至 30% 至 40%。 人机协同：赋能专业人员与端到端闭环 (23:47) 针对 AI 是否会淘汰专业人员的疑虑，Nilay Atmaca 指出 AI 在寿险核保中的定位是能力增强而非取代人力。系统自动处理无糖尿病或癌症等基础病的标准低风险“纯净案件”，使精算师、核保员与医生得以从繁重的行政审批中抽身，转而联合攻坚诸如新冠疫情长期死亡率影响等新型复杂医学课题；数据科学人员也将技能拓展至提示词工程与 RAG 准确性评测。Angella Francis 则提出 GPS 决策模型：AI 负责快速生成三条不同权衡的备选路线并起草分析报告，但只有人类管理者能结合企业战略与现实预算拍板最终路线。 Archana Acharya 强调，生成式模型本身存在偏见且缺乏真正的因果推理能力，必须由人类把关最终执行。例如 Heed 平台发现周五下午 5 点后客服漏接率达 30% 时，AI 仅负责提出部署即时聊天机器人或调整排班等结构化方案，最终由管理人员审核定夺。Venkata Pakkala 亦总结称，三星正致力于打破订单、社交网络反馈、客服记录及客户终身价值（CLV）等数据孤岛，由具备全触点视角的 AI 智能体生成个性化行动建议，但每一个自主推荐决策依然必须接受业务专家的持续监督与校验。 与会专家一致认为，商业智能的未来在于结合第三方动态理赔数据进行持续风险监控、与前沿技术企业合作，以及构建能够自动识别分析空白并生成图表的自学习仪表板。然而，受限于现有大语言模型在认知推理与因果判断上的短板，完全无人干预的自主运营仍难以实现。企业当前应避免盲目开展高风险的全面重构，而应立足现有客服录音等高密度内部资产从小处着手，通过快速验证单一可衡量的商业回报来逐步推进智能化升级。

## Background 该视频由人工智能行业平台 Ai4 于 2025 年 8 月 28 日在 YouTube 上发布。内容记录了一场以“从数据到洞察：AI 驱动商业智能的最佳实践”为主题的专家研讨会，由来自三星电子、Heed Inc、慕尼黑再保险以及 Embrace AI Growth 的多位行业专家共同探讨如何利用 AI 简化数据工作流、构建智能仪表板并克服数据质量与遗留系统集成的现实挑战。 ## TL;DR 研讨会围绕商业智能（BI）从传统“静态事实记录”向 AI 驱动的“预测与处方型决策”演进展开深入探讨。嘉宾们结合消费硬件逆袭、人身险退保微观模式侦测以及三星实时客服质检等实战案例，强调衡量 BI 价值的核心指标已转向“已落地洞察的转化率”与“洞察周转周期”。专家们一致建议企业在产品初始阶段即开展意图驱动的数据设计，从中小型高频场景与单一可量化 KPI 务实切入，并始终在全渠道智能闭环中坚持人机协同（Human-in-the-Loop）。 ## Atmosphere & style 全场研讨氛围务实严谨且极具行业洞见，主持与三位嘉宾之间的互动流畅自然、节奏紧凑。嘉宾们不仅分享了技术落地的成功范式，更坦诚复盘了早期实验中仅有 5%–10% 洞察得以落地的挫败与试错沉淀，配合“业务智能 GPS”、“驾驶舱副驾驶”等生动比喻，使整场技术讨论既有宏观战略高度，又充满一线作战的实操质感。 ## Key takeaways - [01:24](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=84s) 定性客户反馈（如电商长评、通话录音）通过 AI 转译为定量数据，能挖掘出关键业务洞察，例如帮助硬件初创公司转向老年客群并达成 1 亿美元收购退出。 - [04:16](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=256s) 传统 BI 仅能被动记录历史事实，而 AI 驱动的 BI 充当处方型“副驾驶”，能够通过预测推演与仿真模拟给出纠偏操作方案。 - [04:16](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=256s) 再保险案例显示，AI 能够侦测出人工难以察觉的微观模式，例如发现“周五下午”递交的新保单存在显著的冲动退保现象，推动企业建立周一审核机制。 - [09:37](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=577s) 企业拥抱 AI 的核心痛点在于认知雄心与工程落地之间的割裂，战略布局的第一步必须紧密围绕组织的核心价值驱动因素（Value Drivers）。 - [11:51](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=711s) BI 引擎的核心价值衡量标准已从单纯的“可落地的洞察”转变为“已落地且规模化的洞察比例”以及洞察落地的周转速度。 - [11:51](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=711s) 高质量数据必须在产品体验与功能设计源头就进行规划，而不是在系统建设尾声被动补救。 - [14:36](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=876s) 未落地的探索与试错沉淀同样应计入团队 KPI，将其转化为后续产品设计和算法迭代的可复用认知资产。 - [17:50](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=1070s) 三星利用大语言模型与生成式 AI 将语音实时转写，实现全自动客服质检，将过去耗费数周的复盘周期压缩并自动生成处方化改进建议。 - [21:36](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=1296s) 在落地 RAG 与大模型微调时，最大技术难点在于底层对话数据的结构化设计，必须按业务主题有效隔离与分类。 - [23:47](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=1427s) AI 在寿险核保等专业领域的角色是能力增强（Augmentation）而非取代人工，旨在自动化处理“干净案例”以释放专家精力投入前沿复杂风险评估。 - [27:46](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=1666s) 中小企业应从明确投资回报率（ROI）的高频场景切入，像车载 GPS 一样让 AI 提供多条路径建议，但始终保持人类在决策回路中（Human in the loop）。 - [34:24](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=2064s) 通过引入第三方医疗理赔数据打破静态核保快照，结合多癌早筛等前沿健康管理技术，保司可实现动态风险监控与控费控赔双赢。 - [36:34](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=2194s) 未来智能 BI 仪表盘将具备自学习与处方型能力，能主动追踪用户交互习惯并自动补齐缺失的可视化组件与分析视图。 - [38:13](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=2293s) 企业落地 AI 应务实起步，优先聚焦单一可量化的核心 KPI 与客服机器人等沉淀数据丰富的低门槛场景以快速验证成效。 - [39:30](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=2370s) 利用处方型 AI Agent 打破电商、社交、客服与 NPS 评价之间的数据孤岛，建立全触点统一画像，并在人工合规审计下输出个性化服务策略。 ## Detailed walkthrough ### [00:03](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=3s) 会议开场与嘉宾背景介绍 来自三星电子美国公司（Samsung Electronics America）的业务战略与运营分析总监 Venkata Pakkala 担任本次研讨会的主持人。在简短的热场并向到场观众致谢后，他邀请各位来自不同领域的行业专家依次进行自我介绍。 Heed AI 的数据科学与人工智能高级总监 Archana Acharya 首先发言，介绍自己在数据科学与人工智能领域的研发与落地经验。随后，慕尼黑再保险美国公司（Munich Re America）的数字平台产品管理助理副总裁（AVP）Nilay Atmaca 发言，表明自己的核心业务专注于人寿再保险领域的数字化产品与风控。最后，Embrace AI Growth 的创始人兼首席执行官 Angella Watkins Francis 进行介绍，她指出自己创办的机构致力于帮助有远见的企业高管通过战略性数据转型赋能组织，进而在商业和社会层面创造实际影响力。主持人随后点明本场讨论的主题：聚焦各行业中“AI 在商业智能（BI）中的应用”，深入探讨其价值定位、落地工具的实战执行方案，以及未来 AI 驱动型 BI 的演变趋势。 --- ### [01:24](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=84s) 消费品与金融科技实战：定性反馈量化驱动数亿美元增长 主持人 Venkata 向 Archana 抛出第一个问题，邀请她结合业务实际谈谈如何从“数据”跨越到“洞察”，并分享具体的落地案例。Archana 重点列举了消费品（CPG）与金融科技（Fintech）两个截然不同但皆具代表性的应用场景。 在消费品电商领域，Archana 分享了一家硬件初创公司的真实突围故事：该公司最初研发了一款放在办公桌下使用的迷你椭圆机（under-the-desk elliptical），市场定位原本是面向像与会者一样的普通办公室久坐白领。然而，针对白领人群的营销收效平平，业务增长一度陷入瓶颈。随后，该团队借助其搭载的“AI + BI”分析引擎挖掘反馈，竟意外发现“老年人群体”对这款椭圆机的反响异常热烈且购买意愿极高。背后的根本原因在于：老年人渴望找到一款体积极小、挪动方便，且使用时不会对关节造成剧烈负担的轻量化居家锻炼设备。 这一被传统统计所忽略的关键洞察，彻底改变了公司的战略方向。这家初创公司迅速调整定位与营销渠道，最终达成了数百万台的惊人销量，并成功以 1 亿美元的估值完成收购退出（$100 million exit）。Archana 由此指出，这正是商业智能与生成式 AI 深度结合的威力所在——“倾听客户的心声，将定性反馈转化为定量数据，并迅速转化为业务行动”。 随后，Archana 转向金融科技领域的房贷（Mortgage）业务场景。传统房贷机构的客户每天都会频繁拨打销售与银行顾问的电话咨询购房贷款。如果仅靠人工质检，海量对话录音往往沦为沉睡资产；而引入具备自然语言量化能力的 AI 系统后，算法能在销售人员通话的过程中进行实时对话转录与多维情绪、意向分析。系统可以实时侦测出哪些对话正在偏离轨道（走向负面）、哪些潜在客户的转化率极低，以及销售人员在通话中遗漏了客户的核心需求。通过这一“自然语言定性数据转定量指标”的实时干预机制，该房贷机构同样实现了显著的业绩增长。 ```mermaid flowchart LR accTitle: 从定性客户数据到定量商业智能行动流程图 subgraph Input ["非结构化定性数据"] A["电商用户长文本评价 / 调研反馈"] B["房贷销售实时通话语音与转录文本"] end subgraph Engine ["AI + BI 分析与量化引擎"] C["生成式 AI 意图识别与语义量化"] D["实时对话质量与流失风险检测"] end subgraph Action ["业务决策与成果落地"] E["重构产品定位：老人轻量健身器材<br/>（最终实现 1 亿美元退出）"] F["销售人员通话实时干预与挽回策略<br/>（驱动业务转化与规模增长）"] end A --> C B --> D C --> E D --> F ``` - **初创企业突围案例**：白领桌面下椭圆机转向老年人群体，凭借轻便易移动与关节零压迫的核心卖点，促成数百万台销量并以 1 亿美元成功退出。 - **房贷实时风控案例**：对房贷顾问的海量电话语音流进行实时转录与 NLP 量化，精准拦截濒临破裂的沟通对话，实时向业务员提示未满足的客户痛点。 --- ### [04:16](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=256s) 再保险视角下的预测性 BI：“周五下午递交”的微观模式侦测 主持人 Venkata 转向再保险专家 Nilay，询问在保险领域，尤其是核保（Underwriting）与理赔（Claims）环节中，“AI + BI”究竟意味着什么。Nilay 从根本上剖析了传统 BI 与 AI 驱动型 BI 的本质差异。 Nilay 提出，传统 BI 就像飞机驾驶舱中冰冷的静态仪表板，它只负责客观呈现“已经发生了什么”以及“为什么会发生”——显示当前的高度、航速、机舱温度等静态历史读数。然而，AI 赋能的现代 BI 则更像是一位拥有预测与规划能力的“副驾驶”（Copilot）。它不仅能说明过往事实，更能运行大规模仿真模拟，精准预测“未来可能会发生什么、何时会发生，以及你应当采取何种纠偏行动”。以飞行场景打比方，AI 仪表系统能提前 10 分钟发出“前方即将出现颠簸”的气象预警，并直接向机长推送绕开或应对颠簸的具体操作规程。 在人寿保险（Life Insurance）场景中，所有从业人员最核心关切的指标是死亡风险（Mortality Risk）。作为为直接保险公司兜底的再保险商（Reinsurer），慕尼黑再保的业务模式是直接承保各个直保客户送来的业务资产组合（Portfolios）。早期团队构建的传统 BI 仪表板，仅仅用于追踪各投资组合的平均死亡风险指标，并给出行内横向基准对标（Benchmarking），以此向保司客户反馈其在整个市场中的风险定位与致死风险驱动因子。 而当他们将 AI 深度注入 BI 分析看板后，监控系统从被动记录演变为主动提供预防性指标（Preventive Metrics）。Nilay 列举了一个帮助直保客户优化其保险代理人（Agent）绩效的真实案例：直保公司的销售主力是遍布各地的保险代理人及理财顾问。在海量代理人数据中，历来有业绩拔尖的高绩效代表和表现平平的低绩效人员。但在长期跟踪中，数据团队发现一个反常现象：即便是最优秀的一批顶尖代理人，偶发性出现保单失效（Lapse）的比例也会占到 5% 左右。传统业务逻辑往往认为这属于正常波动而选择忽略。 然而，AI 模型在对上百万条投保交易日志进行细粒度模式挖掘时，捕捉到了一个人工排查极难发现的“微观模式”（Micro-pattern）：**凡是由高绩效代理人在“周五下午”递交的寿险保单，后期的失效断保概率显著偏高**。 直保客户与代理人进一步复盘后发现，周五临近周末放假，代理人往往急于下班冲业绩，在向客户推介条款时仓促草率、沟通并不充分。客户在冲动签字后，经过周末冷静期反复琢磨，往往会产生“买家悔意”（Weekend Remorse），认为该保险产品既昂贵又不适合自身家庭规划，于是在短短一个月内迅速退保失效。掌握这一 AI 挖掘出的硬核洞察后，该直保客户迅速出台管理规定：**明令禁止周五下午递交新单，并强制推行周一上午复核机制**。这一微观业务流的微调，立竿见影地遏制了优质客户由于冲动退保造成的资产流失。 - **BI 理念升维对比**：传统 BI 是记录事实的仪表板（“发生了什么与为什么”）；AI 赋能的 BI 则是预测推演与处方型副驾驶（“将发生什么、何时发生、该怎么做”）。 - **人寿保险微观模式**：AI 发现周五下午草率提交的新保单退保率极高，由于“周末后悔情绪”，客户往往在一月内放弃保单；随后保司据此建立了“禁止周五下午交单，周一上午统一审核”的刚性机制。 --- ### [09:37](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=577s) 企业级落地思考：弥合技术与战略间隙，将数据变为动态“业务 GPS” 针对企业如何真正拥抱这一技术，主持人 Venkata 询问战略咨询专家 Angella Francis：AI 正在以何种方式重塑企业的商业智能实践？ Angella 形象地将企业数据在 AI 赋能前后的角色比作“纸质地图”与“智能 GPS”：在过去的传统商业时代，使用数据就像查阅老旧复杂的折叠纸质地图，管理者必须耗费大量时间精力在纵横交错的图表和指标中艰难摸索路线；而一旦将 AI 应用于企业底层数据，这些沉睡的数据就会瞬间转化为组织的“业务智能 GPS”，能动态根据路况指引方向，使决策变得更快、更准、更具确定性。 Angella 从跨越大型企业和私营领域的咨询视角指出，当前企业落地 AI 最大的痛点并非缺乏工具，而在于**认知意图与工程落地之间的巨大割裂（Gap）**——许多高层领导深知 AI 变革的必要性并抱有极大雄心，但在具体的工程技术手段、底层数据治理及业务场景匹配上，往往不知该如何从第一步切入。 她以会前与一位参会者的交流为例：该参会者所在的企业正面临严重的客服掉队（Customer drop-off）危机，由于坐席人力匮乏、等待响应时间过长，导致大量客户中途挂断流失。Angella 为其开出的解法并非盲目上马庞杂系统，而是通过轻量级的 AI 赋能模块迅速介入核心痛点，从而立竿见影地提升该客服部门的吞吐效率。她强调，企业战略性布局 AI 的第一法则，必须是审视组织的核心价值驱动因素（Value Drivers），确保技术应用能够分毫不差地咬合在业务价值创造的关键链条上。 随着讨论深入，主持人 Venkata 重新将话题转向 Archana，提出更进阶的技术落地拷问：从数据到洞察的整个管道中，过去究竟是如何运作的，而 AI 又是如何在加速流程、提升数据可见性（Visibility）以及增强可解释性（Explainability）等核心维度上掀起颠覆性变革的？Archana 随即准备对此展开深入剖析。 ### [11:51](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=711s) 从“可落地的洞察”到“已落地的价值实现”：数据设计的源头思维 行业内正在发生一个明显的变化：从数据到真正实现价值之间的距离正在大幅缩短。Archana Acharya 用了一个生动的比喻来阐述这一过程：把数据比作浇灌种子的水，当植物开花结果、果实挂在枝头时，这仅仅是“价值创造”（Value Creation），代表洞察已经产生；但只有真正把果实摘下来吃进嘴里，才是“价值实现”（Value Realization）。如今，企业追求的目标已经从过去的“可落地的洞察”（Actionable Insight）演变为“已落地且规模化的洞察”（Actioned and Scaled Insight）。因此，评估一套商业智能（BI）引擎有效性的核心指标，正是其所生成的洞察中有多少比例被真正执行并推广，以及完成这一过程所需的时间。 ```mermaid flowchart LR accTitle: 从数据到价值实现的全流程 subgraph S1["数据与构建阶段"] A["数据灌溉与设计"] --> B["BI引擎分析挖掘"] end subgraph S2["价值产出阶段"] B --> C["果实挂枝：洞察生成<br/>(价值创造)"] C --> D["采摘食用：落地与规模化<br/>(价值实现)"] end ``` 当被问及如何实现“零距离获取洞察”（Zero to Insights）时，Archana 指出大型企业最常见的问题是信息在各个层级之间严重流失，且在数据设计时缺乏以用户为中心的考量。在现场互动中，她询问台下有多少产品经理，结果举手者寥寥无几。她直言：“产品经理需要对数据设计投入更多关注。”在规划用户体验和产品功能时，必须在一开始就思考：这项体验会产生什么数据？未来打算用这些数据寻找什么洞察？又将如何据此采取行动？如果不在初始阶段就做好数据设计，最终产出的数据结果就只能碰运气。 - **核心指标转变**：BI 引擎评估标准不仅在于生成了多少洞察，更在于有多少比例的洞察被“落地且规模化”，以及落地的周转速度。 - **数据源头设计**：AI 能否带来回报，根本取决于数据质量与数据设计，这一工作必须始于产品设计之初，而非系统建设的尾声。 --- ### [14:36](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=876s) 衡量 BI 价值的关键 KPI 与试错学习的沉淀 在谈及验证 AI 赋能商业智能的关键绩效指标（KPI）时，Nilay Atmaca 呼应了 Archana 的观点，提出核心衡量标准在于：在所有被一致认可具有商业价值的洞察中，实际被落地执行的比例究竟有多高。他坦诚分享道，团队多年前刚起步时曾经历过许多惨痛的失败，从数据中提炼出的洞察最终能够进入实验、试点并落地的比例非常低，仅有 5% 到 10%。 团队随后进行了深刻的反思，对自身的做事方式保持“极其坦诚”（Brutally honest）的态度。他们意识到，业务发起人与产品经理必须在项目伊始就将商业价值界定得一清二楚。如果团队投入巨大心血进行实验，最终洞察却无法落地，团队士气和信心会遭受重创。通过持续调整，目前该落地转化率已提升至 30% 到 40%。尽管仍有很大提升空间，但团队如今也将“过程中的学习沉淀”计入 KPI：哪怕某些探索最终没有形成具体的系统上线，但只要能为后续产品设计提供宝贵经验，这种试错学习本身就是无价的产出。 - **落地转化率演进**：从最初仅有 5%–10% 的洞察能被试点和执行，逐步提升至目前的 30%–40%。 - **团队士气与业务对齐**：尽早明确业务价值，避免因成果无法落地而打击团队积极性。 - **学习价值纳入考核**：未落地的实验经验被提炼为可复用的认知资产，同样作为衡量产出的 KPI 组成部分。 --- ### [17:50](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=1070s) 三星实战：利用大模型与生成式 AI 加速客服质检与处方化行动 Venkata Pakkala 结合自身在三星（Samsung）负责客户服务运营优化的实战经历，阐述了 AI 如何极大加速业务流程。日常拨打客服热线时常听到的提示语“为了保证服务质量，您的通话可能会被录音”，在业务后台对应着大量的质检工作。传统模式下，人工复盘、抽检通话录音与在线聊天记录往往需要花费数周乃至数月的时间。 借助大语言模型（LLM）与生成式 AI 技术，三星将语音实时转写为文本，并结合聊天记录进行自动化的实时质检（Automatic QA）。这不仅使获取洞察的速度实现了量级跃升，更推动业务迈向下一个台阶——系统能够自动生成针对性的改进建议与“处方化行动”（Prescriptions），从而更迅速地改善终端客户的服务体验。 - **传统质检周期**：人工复盘录音与文本记录通常需要“数周至数月”。 - **AI 赋能质检**：基于大语言模型实现多模态（语音转文字、聊天记录）的实时自动化质检。 - **业务进阶目标**：从单纯发现问题的“洞察获取”，升级为由系统自动输出行动方案的“处方化决策”。 --- ### [19:46](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=1186s) 战略对齐与缩短洞察周期（Time to Insights） Angella Francis 从企业战略咨询的视角指出，AI 是一块极具变革性的“战略拼图”，必须清晰明确它在整个组织战略框架与价值驱动链条中的落点。企业在日常运营中产生的所有痕迹本质上都是数据，而要将海量数据转化为有效洞察，首要任务就是数据清洗——“优质数据输入，才能带来优质结果产出”（Good data in, good results, good output）。 在衡量 AI 成效的核心 KPI 方面，Angella 强调“洞察周期”（Time to Insights）是关键。过去企业依赖周报、月报来获取数据并做出分析决策，整个链路动辄拖延数周；而战略性融入组织的 AI 工具能够在数分钟甚至数秒内完成数据分析、模式识别与趋势预测。这种高效的响应能力帮助企业以更低成本、更高速度与更优质量做出业务决策，直接驱动组织业务增长。 - **战略拼图定位**：AI 工具的引入必须紧密围绕企业的核心价值驱动因素（Value Drivers）。 - **洞察周期压缩**：将传统基于周报、月报的滞后分析压缩为分钟级的实时预测与模式识别。 - **业务赋能效益**：通过更高质量的清洗数据与 AI 提取的有价值洞察，实现“更快、更便宜、更优质、更高效”的运营决策。 --- ### [21:36](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=1296s) RAG 与大模型落地中的绊脚石：面向业务意图的对话数据设计 在探讨落地 AI 与检索增强生成（RAG）等技术的具体工具与实操难点时，Archana Acharya 回顾道，团队在早期进行模型微调（Fine-tuning）和搭建 RAG 架构时，遇到的最大绊脚石依然是“数据设计”。实践证明，企业必须“带着明确的使用意图去设计数据”（Design with the intent of use in mind）。 特别是在处理客服等对话型数据（Conversational Data）时，数据设计的复杂度尤为突出。无论是短信、电子邮件还是电话语音，服务提供方与服务诉求方（无论是投诉还是咨询购买）之间的交互都可以被结构化和体系化组织。只有在数据构建之初就精准识别出业务主题（Themes）并进行有效隔离分类，才能让三星等企业所涉及的各类规模化客户服务场景真正释放出 AI 的潜力。 - **核心技术障碍**：在模型微调与 RAG 实施中最容易遭遇瓶颈的环节是底层数据设计。 - **意图驱动设计**：必须在数据产生和存储阶段就考虑最终的消费意图与使用场景。 - **对话数据结构化**：针对全渠道（文本、邮件、语音）的交互数据，通过提取并隔离核心业务主题实现标准化重构。 ### [23:13](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=1393s) 转化对话数据：从工作流编排到推理设计 当企业能够识别并针对性设计底层数据架构时，这些数据就具备了真正的可用性，因为业务团队能够围绕其实现全面的运营化（Operationalization）。这意味着，无论客户交互来自哪一个沟通渠道，只要系统检测到特定的主题模式，就能立即触发对应的业务工作流。通过这种方式，企业成功将大语言模型（LLM）与生成式 AI 在自然语言处理和对话解析上的强大能力，与传统成熟的自动化工作流运营体系完美结合。对于企业智能化的下一步演进，架构设计不仅要满足信息提取与流程触发，更需要开始针对“推理能力”（Reasoning）进行系统性构建。 ```mermaid flowchart LR accTitle: 对话数据驱动业务工作流 subgraph Data["输入层"] A["多渠道对话数据<br/>(聊天/邮件/通话)"] end subgraph AI["AI 解析层"] B["大语言模型与标签体系<br/>(主题识别 & 意图提取)"] end subgraph Action["运营执行层"] C["自动化业务工作流<br/>(规则触发 & 流程编排)"] end A --> B B --> C ``` --- ### [23:47](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=1427s) 人身险核保革新：不是“替代人工”，而是“人机协同赋能” 在回应关于 AI 是否正在取代传统机器学习与数据科学的疑问时，来自慕尼黑再保险（Munich Re）的 Nilay Atmaca 直截了当地指出：“事实上根本不存在替代（Actually there is no replacing）。”在人身寿险领域，当外界普遍担忧某些职业是否会被 AI 淘汰时，保险科技团队启动“加速自动化核保”（Accelerated Automated Underwriting）项目的初衷，绝不是为了削减核保员，而是为了找出核保人员日常工作中那些“低价值、高耗时”的机械性任务，并借助 AI 予以简化。 最典型的低垂果实（Low-hanging fruit）是标准健康风险的申请人案例。如果投保人没有癌症、糖尿病等重大既往病史，属于非常干净的“标准风险”，其实并不需要资深核保师耗费大量精力去逐页核对厚重的医疗披露材料并计算标准保费。为此，产品经理、精算师、核保员与技术团队（数据科学家、机器学习工程师、数据工程师）紧密协作，梳理出机器可以全权处理的“干净案例”（Clean cases）。这样一来，专业核保员便能抽身投入到更具战略意义、更前沿的复杂课题中——例如与医学专家合作，深入研究新冠长期后遗症（Long-term effects of COVID）对人体健康及长远死亡率预测（Mortality predictions）的精算影响。 - **核保人员角色的重塑**：从繁琐的表单审核中解脱出来，专注于充满未知变量的前沿医学风险评估与复杂裁决。 - **数据科学家的技能演进**：传统数据科学家正将理论与统计经验拓展至提示词工程（Prompt Engineering）和检索增强生成（RAG）模型，其核心职责转变为验证和把控 AI 输出结果的准确性。 - **人机协同（Human in the loop）**：在整个人身险生态中，技术的本质是能力增强（Augmentation）与专业进阶，而非岗位替代。 --- ### [27:46](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=1666s) 中小企业的 AI 战略：从关注 ROI 到“GPS 决策模型” 针对中小企业如何从传统的商业智能（BI）图表汇报转向 AI 驱动决策，Angella Francis 强调，战略沟通的起点必须是明确哪些核心业务职能与任务能够真正拉动投资回报率（ROI）。企业在日常运营中往往忽略了一点：员工每天产生的大量业务交互本身就是极其宝贵的数据资产。中小企业完全可以利用自身沉淀的数据构建专属的大语言模型，并结合 AI Agent（智能体）来提升运转效率，实现“更快、更好、更省成本”（Faster, better, cheaper），从而更早识别业务规律并预测经营走向。 在具体落地路径上，不同企业的业务模式各异，因此集成的工具链也大相径庭。有些场景仅需部署一个简单的轻量化 Agent 去读取数据、提炼分析摘要并自动生成报告，随后交由人工进行复核。Angella 形象地将 AI 与人类的关系比作车载 GPS 导航系统：当你在 GPS 中输入目的地后，算法会规划出三条可行路线，但究竟是选择“用时最短”、还是“避开收费站省钱”、亦或是“风景最佳”，最终的权衡与决断权始终掌握在人类手中。中小企业必须深度拆解每一道日常工序，甄别出琐碎耗时的环节并交给 AI 处理，但在战略方向上永远要保持“人类在回路中”（Human in the loop）。 --- ### [31:09](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=1869s) 消费者大脑与对话标签：AI 决策建议与人类最终裁定 Archana Acharya 补充说明了对话数据处理的底层工具类别——“对话标签工具”（Conversational Tagging Tools）。她提到，人类之间的所有沟通（文字聊天、短信、电子邮件、电话）本质上都是对话，而人与大语言模型之间的交互同样是对话。对话标签在整个 AI 验证管线中占据核心地位，她去年在此类会议上结识的一位开发对话标签平台的初创公司创始人，其公司就在两个月前被 OpenAI 正式收购，用于强化后者的模型验证流程。 谈及 Heed Inc 的定位与人机协同理念，Archana 解释称 Heed 的核心定位是打造一个“消费者大脑”（Consumer's brain），旨在让组织内的每一位品牌经理、产品经理或业务分析师，都能拥有如同亚马逊创始人杰夫·贝索斯（Jeff Bezos）一般极致痴迷于客户洞察的能力。然而，当前的 AI 在本质上仍是我们自身认知偏差的镜像映射，且在大规模因果推理（Causal inference）上尚未成熟。因此，当系统从海量洞察中生成行动建议时，必须由人类进行最终审批。 ```mermaid flowchart TD accTitle: 人机协同建议审批决策流 subgraph Insight["洞察发现"] D["AI 发现业务痛点<br/>(如: 周五 17:00 后 30% 客服不在线)"] end subgraph Options["AI 方案生成"] Opt1["方案 A: 上线销售导向 Chatbot<br/>(满足即时响应需求)"] Opt2["方案 B: 动态调整销售班次<br/>(延长 2 小时覆盖流量窗口)"] end subgraph Human["人类决断 (Human-in-the-Loop)"] H{"业务负责人决策"} C1["采纳方案 A"] C2["采纳方案 B"] C3["组合并存执行"] end D --> Opt1 D --> Opt2 Opt1 --> H Opt2 --> H H --> C1 H --> C2 H --> C3 ``` - **洞察示例**：AI 分析发现，周五下午 5 点之后有约 30% 的客服人员未能及时响应进线咨询。 - **AI 提供的应对方案**： 1. *自动化替代方案*：快速调配一个具有销售转化能力的客服聊天机器人，承接追求即时反馈的潜在客户。 2. *排班优化方案*：根据来电流量在晚间特定时段显著衰减的规律，精准延长销售团队 2 小时的轮班时间。 - **人类决策者的价值**：人类管理者根据当期的预算、人力负荷及服务体验指标，自主决定采纳其中一种方案，或两者结合并行推进。 ### [34:24](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=2064s) 强化保险核保：第三方数据集成与早筛预防合作 在谈及 AI 与商业智能（BI）在保险及再保险领域的未来演进时，Nilay Atmaca 分享了慕尼黑再保险（Munich Re）行之有效的实践经验——购买并深度嵌入第三方数据以革新承保与风险监控。在传统模式下，保险公司的投资组合分析、风险监控与基准测试完全依赖于核保发生时对投保人状况的“单点快照”（Snapshot），而该数据在保单生效后便不再更新。然而，随着投保人年龄增长，健康衰退是客观存在的必然规律。 为了打破这种静态评估的局限，公司引入了第三方医疗理赔数据（Medical Claims Data）以持续动态监控投资组合的健康风险。通过统计方法对“年龄增长”这一固有变量进行标准化剥离后，分析系统能够精准识别出哪些特定客群或资产组合的健康衰退速度显著快于预期，从而及时发现潜在的新兴健康趋势。基于这些洞察，保险公司可以与客户展开协同，采取针对性的预防性健康管理措施。 ```mermaid flowchart TD accTitle: 保险动态健康监控与早筛协同流程 subgraph S1["动态风险监控"] A["第三方医疗理赔数据"] --> B["剥离/标准化年龄变量"] B --> C["识别超预期衰退的异常客群"] end subgraph S2["预防性干预合作"] C --> D["联合初创企业 (如 Galleri 癌症早筛)"] D --> E["向投保人提供低成本早筛服务"] E --> F["降低重疾理赔风险与赔付率"] end ``` 这种预防性协同的典型案例是与硅谷前沿生物技术初创公司合作推广多癌早检项目（如 Galleri 血液检测技术）。该技术仅凭单次抽血即可早期筛查 55 至 60 种不同类型的癌症。虽然该检测公开市场售价高达 900 多美元，但寿险公司通过战略合作以大幅折扣的方式向保单持有人提供，实现了投保人健康防护与保险机构控费控赔的双赢。 - **动态承保监控**：通过引入动态医疗理赔数据，摆脱投保时静态快照的束缚，实时追踪投资组合的群体健康变化。 - **癌症早筛合作**：联合硅谷初创企业推广 Galleri 血液检测技术，以保单优惠价覆盖 55 至 60 种癌症的早期筛查。 --- ### [36:34](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=2194s) 迈向自学习与处方型 BI 仪表盘的未来演进 面对保险行业普遍存在的创新周期偏慢、难以像初创企业一样高速迭代的痛点，Nilay Atmaca 强调，结合战略合作、高效的技术团队以及 AI 工具是实现降本增效的核心路径。例如，保险费率等公开市场信息过去依赖高成本的人工收集，如今借助 AI 驱动的网络抓取（Web Scraping）与数据整合技术，能够以极具性价比的方式将公开第三方数据无缝并入业务流程。 展望未来一至两年，商业智能系统将经历由“被动报表”向“自适应与处方型系统”的深刻蜕变。当前由数据分析师手动提取业务洞察、摸索应对策略的繁琐流程将全面走向自动化： ```mermaid flowchart LR accTitle: 商业智能仪表盘智能化演进路径 A["传统模式: 手动提取洞察"] --> B["自适应阶段: 追踪用户操作缺口"] B --> C["自学习阶段: 自动生成缺失图表/视图"] C --> D["未来模式: 预测型与处方型 (Prescriptive) BI"] ``` 未来的智能仪表盘将具备主动感知与自学习（Self-learning）能力，能够实时追踪终端用户的交互轨迹与点击行为。一旦系统察觉到业务人员在频繁通过线下表格进行某种交叉比对，仪表盘会自动提示甚至自主生成缺失的饼图、表格或分析维度。寿险与商业智能领域将迎来兼具预测性（Predictive）与处方性（Prescriptive）的下一代智能 BI 体验。 - **公开费率智能抓取**：利用 AI 降低公开市场保费数据的爬取与清洗成本，加速竞品基准分析。 - **自适应仪表盘交互**：BI 系统将主动学习用户的使用习惯，自动补充缺失的指标视图与可视化组件，消除人工手工制表环节。 --- ### [38:13](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=2293s) 务实落地：从单一可量化 KPI 与高频场景切入 AI 针对企业如何在未来竞争中抢占先机，Angella Francis 指出，AI 是一场极具颠覆性的底层变革，任何尚未利用 AI 产生具有实质性、可量化业务成果的企业，本质上已经在行业竞争中落后于对手。 面对扑面而来的技术浪潮，企业切忌盲目追求庞大而空洞的转型规划，最关键的在于“务实起步”： 1. **聚焦单一核心指标**：管理者下周返回工作岗位后，应首先寻找一个具体且可衡量的 KPI 作为突破口。 2. **低门槛场景优先**：智能客服问答机器人（FAQ Chatbot）是极佳的试点项目，因为企业内部早已沉淀了大量的历史通话录音与服务工单记录，底层数据现成可用。 3. **闭环验证成效**：通过在局部流程中引入轻量级 AI 工具，直观观察指标的改善趋势（"seeing the needle move forward"），以切实的数据回报为后续更大规模的 AI 部署奠定信任基础。 > "AI 是一项如此具有颠覆性的创新，任何还没在用 AI 做出实质性、可量化业务成果以驱动业绩的公司，实际上已经落后于竞争对手了。" --- ### [39:30](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=2370s) 全渠道数据归一与处方型 AI Agent 的业务闭环 Venkata Pakkala 结合三星电子（Samsung Electronics America）在客户服务与对话分析（Conversational Analytics）领域的实操经验，阐述了数据从离散洞察走向全生命周期赋能的演进过程。对话分析不仅能用于坐席层面的自动化质量保证（Automated QA），更重要的是从中沉淀多维度的客户声音（Voice of Customer）。 以往企业内部构建了大量基于传统机器学习的数据科学模型，例如细分客群模型、客户生命周期价值（LTV）预测模型以及购买意向倾向（Purchase Propensity）模型。但这些系统往往各自为政、深陷数据孤岛。一名典型的现代消费者会经历完整的跨渠道旅程：在官网浏览下单、在社交媒体分享体验、因使用问题联络客服解决，最后提交净推荐值（NPS）评分与评价。过去这些触点数据分散在电商、社交、售后等各个独立系统中。 ```mermaid flowchart TD accTitle: 全渠道数据整合与 AI Agent 闭环 subgraph Channels["全触点交互数据"] A1["官网订单与行为"] A2["社交媒体互动与分享"] A3["客服通话/对话分析 (QA+洞察)"] A4["NPS 评分与售后反馈"] end subgraph Core["AI Agent 智能整合引擎"] B["LTV 预测模型 + 购买倾向模型"] Channels --> B B --> C["高价值客户精准画像识别"] end subgraph Action["处方型个性化赋能与人机协同"] C --> D["跨渠道个性化体验建议 (Prescription)"] D --> E["人机协同 (Human-in-the-Loop) 审核监控"] end ``` 借助新型 AI Agent（人工智能体），企业终于能够将分散的触点数据打通归一，精准识别出最具商业价值的核心客户，并推导出跨渠道个性化服务的“处方型行动策略”（Prescriptions）。在展望未来广阔前景的同时，Venkata 特别强调了“人机协同”（Human-in-the-Loop）的底线原则：AI 所生成的行动建议和业务策略，依然需要人工的严密监控与审计，以确保商业决策的安全可控。至此，全场讨论圆满结束并正式进入问答环节。 - **多模态触点整合**：打破电商、社交口碑、客服录音与 NPS 问卷之间的数据孤岛，建立统一的客户价值全景图。 - **人机协同合规治理**：在 AI 自动输出处方型个性化策略的同时，设置人工监控节点以保障决策稳健度。 ## Notable quotes > [01:24](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=84s) 倾听客户的心声，将定性反馈转化为定量数据，并迅速转化为业务行动 — Archana Acharya > [04:16](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=256s) 传统 BI 就像飞机驾驶舱中冰冷的静态仪表板，它只负责客观呈现“已经发生了什么”以及“为什么会发生”……而 AI 赋能的现代 BI 则更像是一位拥有预测与规划能力的“副驾驶”（Copilot）。 — Nilay Atmaca > [04:16](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=256s) 凡是由高绩效代理人在“周五下午”递交的寿险保单，后期的失效断保概率显著偏高 — Nilay Atmaca > [11:51](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=711s) 把数据比作浇灌种子的水，当植物开花结果、果实挂在枝头时，这仅仅是“价值创造”，代表洞察已经产生；但只有真正把果实摘下来吃进嘴里，才是“价值实现”。 — Archana Acharya > [11:51](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=711s) 产品经理需要对数据设计投入更多关注。 — Archana Acharya > [19:46](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=1186s) 优质数据输入，才能带来优质结果产出（Good data in, good results, good output） — Angella Francis > [21:36](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=1296s) 带着明确的使用意图去设计数据（Design with the intent of use in mind） — Archana Acharya > [23:47](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=1427s) 事实上根本不存在替代（Actually there is no replacing） — Nilay Atmaca > [38:13](https://www.youtube.com/watch?v=VPX2uUZhRPQ&t=2293s) AI 是一项如此具有颠覆性的创新，任何还没在用 AI 做出实质性、可量化业务成果以驱动业绩的公司，实际上已经落后于竞争对手了。 — Angella Francis ## Mentioned Venkata Pakkala, Archana Acharya, Nilay Atmaca, Angella Watkins Francis, Samsung Electronics America, Heed AI, Munich Re America, Embrace AI Growth, OpenAI, Galleri, Jeff Bezos ## Video description Archana Acharya, Senior Director, Data Science, Heed Inc Nilay Atmaca, AVP, Digital Platform Product Management, Munich American Reinsurance Co. Venkata Pakkala, Director - Business Strategy and Operations Analytics, Samsung Electronics America Angella Francis, Founder / CEO, Embrace AI Growth Abstracts This session will focus on the practical steps needed to harness AI for enhanced business intelligence, from data preparation and integration to building AI-powered dashboards and reports. Panelists will discuss how organizations are leveraging AI to streamline data workflows, improve data visualization, and provide actionable insights that drive business outcomes. Learn best practices for adopting AI-driven BI tools and how to address common challenges, such as data quality and integration with legacy systems, to create a seamless and effective BI environment. [Source](https://www.youtube.com/watch?v=VPX2uUZhRPQ)

[1:24] Archana Acharya: 倾听客户的心声，将定性反馈转化为定量数据，并迅速转化为业务行动 [4:16] Nilay Atmaca: 传统 BI 就像飞机驾驶舱中冰冷的静态仪表板，它只负责客观呈现“已经发生了什么”以及“为什么会发生”……而 AI 赋能的现代 BI 则更像是一位拥有预测与规划能力的“副驾驶”（Copilot）。 [4:16] Nilay Atmaca: 凡是由高绩效代理人在“周五下午”递交的寿险保单，后期的失效断保概率显著偏高 [11:51] Archana Acharya: 把数据比作浇灌种子的水，当植物开花结果、果实挂在枝头时，这仅仅是“价值创造”，代表洞察已经产生；但只有真正把果实摘下来吃进嘴里，才是“价值实现”。 [11:51] Archana Acharya: 产品经理需要对数据设计投入更多关注。 [19:46] Angella Francis: 优质数据输入，才能带来优质结果产出（Good data in, good results, good output） [21:36] Archana Acharya: 带着明确的使用意图去设计数据（Design with the intent of use in mind） [23:47] Nilay Atmaca: 事实上根本不存在替代（Actually there is no replacing） [38:13] Angella Francis: AI 是一项如此具有颠覆性的创新，任何还没在用 AI 做出实质性、可量化业务成果以驱动业绩的公司，实际上已经落后于竞争对手了。

[4:16] Nilay Atmaca: what has happened and maybe why [4:16] Nilay Atmaca: turbulence is going to happen in 10 minutes [11:51] Archana Acharya: actioned and scaled insight [14:36] Nilay Atmaca: brutally honest [17:50] Venkata Pakkala: this call is recorded for quality assurance purposes [19:46] Angella Francis: good data in, good results, good output [21:36] Archana Acharya: came back and bit us [27:46] Angella Francis: faster, better, cheaper [31:09] Archana Acharya: as consumer obsessed as Jeff Bezos [38:13] Angella Francis: every company that's not already doing something meaningful, something measurable with AI to drive their business results is already falling behind their competitor.

[0:03] Hi, my name is Wanketa Palakala. So I'm [0:06] from Samsung Electronics America. So [0:08] today I will be moderating this session. [0:10] Um thank you everyone for showing up [0:13] here. So awesome. So [0:17] um let's go let's start with some [0:19] introductions. So Archa. Yeah. [0:22] >> Hi my name is Archa. I'm with Heed AI [0:25] Inc. and I'm a senior director of data [0:28] science and AI there. [0:30] Hello everyone, thanks for coming. I am [0:32] Nili from uh Mini Insurance. My role is [0:36] AVP Pro digital product management in [0:39] life insurance. Thank you. [0:42] >> Hi, my name is Angela Francis, Angela [0:44] Watkins Francis, and I am the founder [0:47] and CEO of Embrace AI growth where we [0:50] help visionary leaders actually [0:52] transform the organization with data [0:54] strategically so they can make an impact [0:56] in society. [0:58] All right, thank you. So today we are [1:01] going to keep it very simple. So we have [1:03] folks from different domains here. So [1:06] the idea here for today is to be able to [1:09] understand h what AI in BI is, how is it [1:13] helpful and how are we executing it [1:16] today in terms of tools and how is the [1:19] future looking like for BI in terms of [1:21] AI. So let's go ahead and start it. So [1:24] Archa with you. So which domain uh [1:27] currently uh you are working with and [1:30] how uh let's talk about the business [1:33] intelligence for sure. So when we talk [1:34] about data right so from data to [1:38] insights can you talk us talk to us [1:40] about some of the uh ideas from there [1:43] and as well as use cases maybe for [1:45] starters. [1:45] >> Yeah. So I'll speak to two domains. I'll [1:47] speak to both fintech and CPG today. Um [1:51] I'll give you an example of a startup. [1:53] Uh not very long time ago there was once [1:55] a startup that uh was selling under the [1:58] desk ellipticals to human beings and [2:01] they were targeting officegoing workers [2:03] much like ourselves and they kept doing [2:06] that they were getting some growth but [2:08] not a whole lot. um through their you [2:11] know AI and BI engine they were able to [2:14] identify that seniors for some reason [2:17] were responding really well to their [2:19] product and the reason was it turns out [2:22] that they were essentially looking for [2:24] an easy exercise machine something small [2:28] that's easy to move around and wouldn't [2:30] be too strainous on their joints and lo [2:33] and behold with this insight this little [2:36] startup ended up selling multiple [2:38] millions of those under the desk [2:39] ellipticals and ended up with a $100 [2:42] million exit. And such is the power of [2:45] business intelligence when you combine [2:47] it with the power of you know AI [2:51] especially on the gen AI side [2:53] understanding what your customers are [2:55] telling you hearing them out converting [2:58] it into quantitative taking the [3:00] qualitative converting it into [3:01] quantitative and putting it into action. [3:04] And so what we are seeing now in the CPG [3:08] e-commerce service space is all this [3:11] collection of qualitative feedback data [3:14] being used in combination with [3:16] quantitative through your BI dashboards [3:19] and interfaces and decisions being [3:22] powered at speed. That's one example. [3:25] Another example from fintech um imagine [3:28] you know a mortgage company. a mortgage [3:31] company that essentially has homeowners [3:34] and people looking to be homeowners [3:35] calling their salespeople, bankers day [3:38] in and day out. And these conversations, [3:41] these transcripts as they happen, you're [3:44] able to intervene and make sure that you [3:47] are able to write track them in real [3:49] time in the moment. That happens when [3:52] you take all that information that's [3:55] natural language and you're able to [3:57] quantify it. you're able to understand [3:59] which conversations are going bad, which [4:02] conversations are less likely to convert [4:04] and what is the customer looking for [4:06] that your salesperson isn't really [4:08] giving them on the call and with that [4:11] that organization has seen growth as [4:13] well. Um I'm going to pass it back to [4:15] you. [4:16] >> Okay, thank you so much Nai. So from the [4:18] insurance uh domain standpoint, right? [4:21] So what does it mean by AI in BI in [4:24] terms of um claims or underwriting etc. [4:29] >> Yeah, I think um maybe we can make the [4:32] distinction between when we're talking [4:34] about traditional business intelligence, [4:36] we're talking about static dashboards [4:39] that can tell you what has happened um [4:42] and maybe why. So, think of it like a um [4:46] you're you're a pilot and you have a [4:49] dashboard, a panel in front of you. It [4:51] tells you where you are, what's the [4:53] temperature, um and uh the altitude and [4:57] speed and everything. So when we're [4:59] talking about AI power business [5:01] intelligence, we're not only talking [5:03] about what happened and why, but we're [5:05] talking about what can happen and when [5:09] and what are the actions that you can do [5:13] something about it. So in that copilot [5:16] example analogy, [5:18] um you can the AI can run a lot of [5:21] simulations and predict that there's a [5:24] turb turbulence going to happen in 10 [5:27] minutes. And here are the instructions [5:29] how you can uh manage that. So in [5:32] insurance we are um using a similar [5:35] approach. So in life insurance uh [5:38] mortality risk is the main risk that [5:41] everybody cares about. Um in our earlier [5:45] in our days we had our dashboards [5:48] business intelligence and we were just [5:50] tracking say average mortality risk and [5:54] benchmarking. As a reinsurer, our job is [5:58] to insure insurance companies. So, we do [6:00] have portfolios um that are um pockets [6:04] of business from each uh clients who are [6:08] our um c um who are our portfolio uh [6:13] providers. So our dashboards not only [6:17] monitored um risk over time but we also [6:20] had a chance to see a benchmarking so [6:24] that we can provide it to them and they [6:27] can understand their positioning in the [6:29] market uh what are their main drivers of [6:33] mortality and what they can do about it. [6:35] So when we powered AI into our BI [6:38] dashboards, uh we were able to provide a [6:42] lot more actionable [6:44] um insights to our clients and we were [6:47] able to not only monitor the risk but [6:49] also we have some preventive um metrics. [6:54] For example, um [6:57] uh for for one of our clients, we were [6:59] monitoring their agency uh performance. [7:03] We were helping them. Of course it's [7:04] their role to do that but as a as their [7:07] partner we uh we help them to provide [7:11] different um insights from a reinsur [7:14] perspective as well as from benchmarking [7:17] perspective. So when we identify [7:20] high-erforming and low performing [7:23] insurance agents, those are salespeople [7:26] talking like financial advisors talking [7:28] to people and providing them the most um [7:32] useful uh financial instruments. So one [7:35] of them is life insurance. So when um [7:40] when we looked at the data, we have [7:42] always seen that there are nicely [7:45] performing insurance agents and bad [7:47] ones. Um but out of the good ones that [7:52] sometimes uh their performance was low [7:55] and we never thought that there could be [7:58] something behind it because the [7:59] percentage wasn't a lot, right? It's [8:01] like 5% or something. Um and uh when we [8:07] powered our AI to understand the reasons [8:10] why what we noticed was that and that's [8:13] the data we would probably it's not our [8:16] go-to data to look at the reasons but we [8:19] noticed that um Friday afternoon [8:23] submissions of agents their lapse rates [8:27] uh are higher. [8:29] So um we provided this insight to our [8:33] clients and then they implemented uh no [8:36] Friday afternoon submissions. What [8:38] happens is that uh they rush and they [8:40] don't really inform the clients. So with [8:43] the weekend remorse the clients think [8:45] about okay probably this was not the [8:48] best decision I made or it doesn't look [8:50] like it fits to me personally perfectly. [8:53] So then within a month or so these [8:55] policies lapsed. So after talking to [8:58] clients and identifying this reason um [9:02] they implemented that solution no Friday [9:05] submissions and then uh they required [9:08] also Monday morning reviews. So this is [9:11] one of the thing that um I think it's a [9:13] a micro pattern among the millions of [9:16] data that we wouldn't we would not have [9:19] thought of it um using just traditional [9:22] machine learning techniques. So with the [9:24] help of AI um this BI dashboards really [9:28] provided very good insight. [9:29] >> Yes, thank you. So I would love to know [9:32] more about how you guys have done it. So [9:33] but I will uh come to you in a bit. So [9:37] Angela, so you help uh businesses, [9:40] right? So how is AI impacting um the [9:44] businesses in terms of business [9:45] intelligence? Well, in terms of what I [9:47] want to want to start off with about um [9:49] the data and business intelligence, your [9:52] data in your organization [9:54] becomes business intelligence like a [9:57] GPS, your business GPS. So regular when [10:00] you think about the map we used in the [10:02] olden days. I don't know if anybody [10:03] remember using the map where we had to [10:05] take time to read through that. Now AI [10:07] applied to your data becomes your [10:09] business intelligence. So you can make [10:11] uh decisions so much quicker, so much [10:12] faster, so much better. Um and so in [10:16] terms of my domain, I come from a a [10:18] different perspective in terms of um uh [10:21] several corporate and both private [10:23] sector. And so I see that there's this [10:25] gap in between the knowledge and the [10:28] intention of wanting to implement AI and [10:31] then also the [10:34] actual technical aspect of actually [10:36] implementing it. So as we look at and [10:38] see how AI can actually change the [10:42] meaning of the data that you have in [10:45] your organization and that's where we [10:47] see the so many different use cases. I [10:49] was speaking to one the participant who [10:52] came early yesterday and she was saying [10:53] they're looking at their customer [10:55] service see how many uh clients are [10:57] being dropped off and not being able to [10:59] helped because either they're waiting [11:00] too long or they don't have enough [11:01] agents in that. And one of the quick [11:03] things that we can suggest is [11:04] implementing some AI into that to really [11:07] turn that um and improve that efficiency [11:10] within that sector for them [11:12] organization. So in my domain, it's [11:14] really about strategically looking at AI [11:16] in really terms of how it fits into your [11:18] organization and the best way it fits [11:20] into what your value drivers are for [11:22] your organization. And I'll leave it at [11:24] that. [11:24] >> Yeah, thank you. So Archa, so coming [11:27] back to you. So let's pick a use case, [11:31] right? So from data to insights and [11:33] let's talk about KPIs and what we really [11:35] would like to understand is from data to [11:37] insights how it used to be and how AI is [11:40] making an impact in terms of [11:42] accelerating or in terms of making more [11:45] visible explainable etc. So how AI is [11:48] making an impact? [11:49] >> I think the biggest thing that we are [11:51] starting to see across the industry now [11:53] is the distance between your data to [11:57] value real realization is reducing and [12:00] if I had to explain it in a simple term [12:02] like imagine that you know you you put a [12:04] seed um and you're watering it and every [12:07] time you water the seed and it's growing [12:09] that waters the data. When this thing [12:12] brings out a flower and a fruit, right, [12:15] um, and the fruit's there on the plant, [12:17] you've not plucked it yet. That's value [12:20] creation. That's your insight. It's [12:21] there. You've not plucked it yet. If you [12:24] pluck the fruit and you eat it, that's [12:26] value realization. That's essentially, [12:28] you know, you've gone from what we used [12:30] to call actionable insight to actioned [12:33] and scaled insight. And so now that [12:36] distance gone down significantly. And so [12:39] the pursuit of actionable insight is now [12:43] the pursuit of actioned and scaled upon [12:45] insight. And therefore the metric to [12:48] understand what's working and what's not [12:50] is to understand the percent of insights [12:52] that your BI engine is creating that's [12:55] actioned and scaled and how in what time [12:58] are you able to do it. That's going to [13:00] be key. So value realization that makes [13:04] all the difference. You can create as [13:05] many insights as you want. Can I have a [13:07] follow-up question on that? So, what uh [13:11] what makes zero to insights really [13:13] possible? What are some key things that [13:16] uh organizations and leaders need to [13:18] really keep in mind to make that [13:20] possible? I think a lot of times [13:23] especially in the larger enterprises [13:25] what I see happen is uh there is a lot [13:28] of loss of information across the layers [13:31] and the intention of designing data with [13:33] the user in mind that doesn't come [13:36] across so if you if there are there [13:38] product managers in the room today are [13:40] you product any product I see a hand I [13:43] see a hand I see a hand all right there [13:45] are very few product managers in this [13:47] room and I can tell you just by that [13:49] fact that product managers need to be [13:51] paying more attention to data design. [13:54] When you're designing an experience, [13:56] when you're designing features, you need [13:58] to think about data from a design [14:00] perspective as well. What data is that [14:03] experience going to generate? And what [14:05] are you going to do with it? What [14:07] insights are you going to look for? And [14:09] how are you going to act on it? If you [14:11] don't do that right from the beginning, [14:14] what you get out is, you know, a [14:16] randomized output that's just based off [14:19] of luck. So those five product managers [14:21] who raised their hands, I'm really happy [14:24] you guys are all on the right track cuz [14:26] guess what makes AI return possible? [14:30] It's your data quality and data design [14:32] and it starts at the very beginning not [14:34] the end. [14:36] >> Thank you. Thank you Eli. So from your [14:39] KPIs perspective, right? So which KPIs [14:43] do you consider uh to validate the fact [14:46] that AI is really really helping in [14:48] terms of business intelligence uh and [14:50] also bringing the data to insights [14:52] right? Yeah, sure. I think uh one of the [14:54] KPIs is really very similar to what [14:57] Archana mentioned uh out of all the [15:00] actions that um from the insight that we [15:04] gather and agree upon that these could [15:07] create business value, how many [15:10] percentage of those that we are actually [15:12] implementing? And a lot of the times [15:15] with my experience is that the ones that [15:17] you're not implementing and [15:19] understanding reasons why um with [15:22] continuous learnings is also possible [15:25] for us as well right because we're [15:27] trying to make the machines learn right [15:29] machine learning and self-learning AI. [15:32] Um but I think uh there is a big [15:35] opportunity with the leaders and um and [15:38] business people to understand what is [15:40] failing and of course we have to we did [15:44] have a lot of um big failures and um [15:46] initially years ago when we started um [15:50] seeing certain insights that I mentioned [15:52] and implementing them our rate was very [15:54] very low like five to 10% of the things [15:58] that we learned from data we were able [16:00] to experiment [16:02] um pilot and then implement. Uh but [16:05] moving forward uh our learning was that [16:08] we analyzed we came together and we [16:11] became brutally honest with ourselves [16:13] and how we're doing things and we [16:16] understood that um [16:19] similar points that you mentioned about [16:21] the role of a product manager or I would [16:23] say that the business sponsors uh the [16:26] value of the business business value has [16:29] to be really very uh clear in the [16:31] beginning. So um this affects um teams [16:36] morale as well. At the end of the day [16:38] when they experiment something when they [16:40] put a lot of effort in it and then that [16:42] experiment or insight is not being [16:44] actualized or or implemented they're [16:47] also losing some some faith along the [16:50] way. Um so right now we are not at that [16:54] highest level still from insights to an [16:57] action and impact. I would say that 30 [17:00] to 40%. It is still a lot of improvement [17:03] but there is I think still a lot of um [17:05] way to go. uh but I think being really [17:10] honest uh with our work and ourselves [17:13] and clarifying that business value in [17:17] the beginning and coming up with a [17:19] couple of options right so you can say [17:22] that okay this is the value of business [17:24] there's nothing to be implemented here [17:27] however the learning we're going to get [17:29] out of that that experience is priceless [17:31] and we will use it say designing other [17:34] products and stuff like that to me uh we [17:36] are also counting that in uh as part of [17:40] our KPIs. Um there is nothing um very [17:44] concretely implemented at the end of the [17:46] day but learnings are also uh part of [17:49] the KPIs now. [17:50] >> Got it. Got it. So uh at Samsung so [17:53] currently I'm uh working with the [17:55] customer service operations optimization [17:58] especially understanding how AI can help [18:00] right. So I will tell one example. So [18:03] every every one of you would have heard [18:05] when you call the customer service hey [18:08] thank you today we are going to record [18:10] this call for quality and assurance [18:12] purposes right so it's actually true so [18:15] we do uh conduct a lot of quality [18:18] assurance on the calls recorded as well [18:20] as the chat transcripts but for us for [18:23] me the KPI that we are targeting is the [18:27] acceleration now for us traditionally [18:30] for us to be able to go through the [18:32] transcripts and everything, right? So, [18:34] it used to takes anywhere from weeks to [18:37] months. Um, now with AI, so we are able [18:41] to trans create transcripts, voice [18:45] transcripts, convert that to text and [18:48] also we have chat transcripts. So, we [18:50] can use LLM as well as genai to enable [18:54] automatic QA um to do the quality [18:59] assurance in real time. So for us um our [19:05] customers matter. So providing the best [19:07] service to the customers matter. So now [19:09] with one of these use cases that I am [19:11] talking about. So AI is really really [19:14] helping us to accelerate the uh insights [19:17] not only insights but the next leap here [19:20] is to be able to generate actions and [19:22] prescriptions automatically so as to [19:24] bring better service to our customer. So [19:27] that's about me but Angela. So from uh [19:31] of course you from your domain [19:33] standpoint helping out the businesses [19:34] right. So in terms of uh KPI what KPI [19:38] are you considering like how do you know [19:40] if the AI is actually helping your um uh [19:44] businesses? [19:46] >> Okay. So each business we like to start [19:49] out their basic strategy strategically [19:52] because AI is so transformative and it's [19:55] a puzzle piece and you have to know [19:57] where it fits into the whole [19:59] organization of your strategy. What are [20:02] your value drivers that drives your [20:04] business results that you want and so [20:07] when you basically everything you do in [20:09] your company every single thing is data. [20:12] And when you get the all collect all [20:15] that data, the first thing you want to [20:17] do is basically in order to turn it into [20:19] insight is to basically clean the data. [20:21] And in cleaning the data because as I [20:23] trying to say data in, good data in, [20:25] good results, good output. So the data [20:28] in, get the insights with the AI and the [20:31] one of the key um key KPIs is as [20:35] following up on what you said too is [20:37] that the time to insights, right? so [20:40] much. You have reports, weekly reports, [20:43] monthly reports. So those decisions on [20:45] based on those reports or those data in [20:47] those reports will take weeks, months [20:49] for you to actually analyze. [20:51] With an AI tool can do a quick analysis [20:54] in minutes, less than minutes to give [20:56] you insights, predict what's happening, [20:59] detect patterns, and being able for you [21:01] to make decisions so much faster than [21:03] you could on a normal basis. So some AI [21:06] tools strategically [21:09] integrated into your existing [21:10] organization can be valuable in terms of [21:13] the outcome you get. Starting with the [21:15] data that you have in your organization, [21:16] clean good data and then using the AI [21:19] tool to extract the insights that's [21:21] meaningful, measurable insights that can [21:23] really help your organization grow and [21:25] then the outcome we know will be faster, [21:27] cheaper, better and more efficient [21:29] operations of your organization leading [21:30] to your growth and able for companies to [21:33] go to make um better decisions. [21:36] >> All right. So Archa coming back to you. [21:40] So [21:42] um in terms of the tooling right so uh [21:47] have you guys implemented uh any use [21:50] sorry use cases with AI but then in [21:52] terms of tooling not going into names or [21:56] but techniques and tips standpoint have [21:58] you guys faced any obstacles if so how [22:01] you overcome them [22:03] >> I think the biggest obstacle uh [22:05] especially early on when we started [22:07] implementing and fine-tuning our models [22:09] and implementing rag [22:11] came back and bit us was the data design [22:13] aspect. Um and so it became very clear [22:16] to us that you needed to design with the [22:19] intent of use in mind. Now when you [22:22] think about conversational data um and [22:25] designing of data around conversational [22:29] data that's a very very interesting [22:31] exercise. Um if you do that well enough [22:35] then all those use cases that we've all [22:36] talked about you know the servicing [22:38] based use cases such as yours from [22:40] Samsung and uh yours from uh as well [22:45] they're all getting unlocked at scale [22:47] now you imagine that you know [22:49] irrespective of the medium whether it's [22:51] text whether it's email whether it's [22:53] phone call if two individuals are [22:55] interacting one's a service provider and [22:57] one's a service seeker uh whether it's [23:00] complain or buying buying something and [23:02] just seeking information about buying [23:04] something, you can structure and [23:06] organize that information in a certain [23:08] way. If you start identifying themes and [23:11] isolating what themes you want to [23:13] identify and design your data [23:14] accordingly, that becomes utilizable [23:17] because you can operationalize around [23:19] it. anytime irrespective of the channel [23:21] I see theme XY Z pop up start workflow X [23:26] and so now you've combined the whole [23:29] language based conversational data power [23:32] of genai LLMs with power of what we've [23:35] already you know done very well at for [23:37] years which is operationalization of [23:40] workflows and up next we'll have to [23:42] start designing for reasoning as well [23:44] but that's probably a conversation for [23:46] the next AI for [23:47] >> thank you Archa so Neil Right. So uh [23:51] especially in terms of like uh life life [23:54] insurance, right? So uh how AI is [23:57] replacing the traditional machine [23:58] learning algorithms or um traditional [24:01] data science that you used to that we [24:03] used to do to uncover insights, right? [24:06] >> Actually there is no replacing. [24:08] >> Okay. Yeah. So [24:09] >> um so um this was one of the things. So [24:12] in life insurance when um when AI is [24:16] everybody's top of mind subject and [24:19] we're talking about certain occupations [24:22] are going to be replaced by AI or [24:24] anything like that. So when we started [24:26] first um experimenting this project [24:29] accelerated automated underwriting um [24:33] our purpose was not really um replace [24:36] underwriters. Our purpose was okay what [24:39] are the uh routine low value higheffort [24:44] tasks that underwriters are doing in [24:47] their day-to-day jobs that AI can help. [24:50] So one of the uh lowhanging fruit was [24:54] for example [24:56] um an applicant um an individual if they [25:00] have a very standard risk if they don't [25:02] have say cancer diabetes all these uh [25:05] risk factors that are very important uh [25:08] for life insurance underwriting. uh you [25:11] wouldn't necessarily need an underwriter [25:13] to go back all those papers or um [25:16] disclosures um and say that okay this is [25:20] a standard risk and here is the standard [25:22] premium that the person needs to pay. So [25:26] um we worked with underwriters again the [25:28] product manager actuaries underwriters [25:31] and then the tech people um who are data [25:35] scientists, machine learning engineers, [25:37] data engineers. We came together and [25:39] said that what are those cases that are [25:41] clean case that uh a machine can do for [25:45] you um so that you can put more effort [25:48] into how AI can even help you more. Um [25:53] what are the more strategic part of your [25:56] job that you need more time uh to spend [25:59] on? Um and uh also new cases, right? [26:05] There's a lot of advancements uh in the [26:08] medicinal area that uh helps a lot for [26:11] our um life insurance underwriting. Um [26:15] and acturies and medical doctors [26:17] literally work together. Um uh for [26:20] example uh we are currently studying the [26:23] longerterm effects of COVID and how it [26:27] affects of course person's health um and [26:30] what that means in actuarial sense uh [26:34] longerterm mortality predictions as well [26:36] as um underwriting. [26:39] So this is a black box right now right [26:42] the subject of research in a lot of [26:45] fields and we don't know and [26:47] underwriting uh opinion and aspect is so [26:51] important. So when we automatize those [26:53] clean cases underwriters have more time [26:56] to work on more important and futuristic [27:00] innovative uh projects. So that's [27:02] something it's like human in the loop [27:04] aspect in the life insurance um um area [27:08] for data scientists as well right they [27:11] are evolving their skills now into AI [27:14] prompt engineering [27:16] um and they are uh studying literally [27:19] with drag models and how can uh how can [27:23] they uh use their day-to-day statistical [27:26] theoretical experience so that they can [27:29] check the accuracy of AI outputs. [27:33] Um so there is really like at least uh [27:36] from our perspective in the life [27:38] insurance area with all these different [27:40] types of roles there is no replacement [27:42] there is augumentation and there is [27:44] advancements. [27:46] >> Got it. So Angela so from in terms of [27:49] small businesses right so now um from [27:52] strategy standpoint [27:54] um how do you how do you approach [27:56] talking to your business partners? So [27:58] for example, hey traditionally BI has [28:00] been these beautiful charts reporting [28:02] and then we need to do more work to be [28:04] able to understand what's driving why [28:06] why it is happening what is happening [28:08] from the AI standpoint right so now with [28:10] the advent of AI so we have more [28:11] techniques more tools so strategically [28:15] uh what do you how do you approach [28:17] talking to your business partners [28:19] >> so strategically we start out I as I [28:21] mentioned before we have to companies [28:24] have to think about what are the task, [28:28] the opportunities, the the the functions [28:31] in their organization that are really [28:33] driving their ROIs, their return on [28:36] their investments and try to focus some [28:38] of those first so they can see [28:40] measurable measurable outcomes based on [28:43] implementing the AI. you're [28:44] strategically going to a business, you [28:46] want to look at what is your mission, [28:48] what are goals you want to achieve and [28:50] how can AI be used to augment what we do [28:54] to do it faster, to do it better, to do [28:56] much better uh and efficiently. So [28:59] really you want to start out with [29:00] basically what are the and you talk to [29:01] different people in the organization it [29:03] comes and one of the things that [29:04] sometimes we miss out on when we go into [29:06] a work on a daily basis we sometime [29:08] because we have just so used to doing a [29:10] work over and over again we miss out on [29:12] some of the things that are really in [29:14] really data because everything we do in [29:17] your organization is data it can be used [29:19] to create your own large language model [29:24] your own LLM in your organization with [29:26] the data that you have in your [29:28] employees, your team can pull from that [29:31] database, pull from that large language [29:33] model and be used as in into agents to [29:36] really uh make things I guess and I like [29:39] to use the word faster, better, cheaper, [29:41] right? Create more predictable uh um [29:43] outcomes, more patterns to see more [29:45] patterns and be able to drive us to [29:47] where we really need to go in order to [29:49] be uh profitable and be successful in [29:52] business. So as again we really start [29:54] out each company the tools that they use [29:57] will be different because their [29:58] functions are different. The way that AI [30:01] is integrated will be different simple [30:03] and sometimes we just need to automate [30:05] some things and just put an agent in [30:07] agent tool in there to read the data do [30:10] the u read the information present some [30:13] analysis and really uh create some [30:15] reports and then move the report on to [30:18] human. Of course, we want to keep human [30:20] human human in in the loop. Human in the [30:22] loop. Keep humans because machines will [30:26] just like when you have your GPS, you'll [30:28] go in there and you you put in your [30:30] destination. They'll give you three [30:31] options. How do you choose the the three [30:34] options? It has to be do I want to get [30:35] there faster? Do I want to get there [30:37] cheaper to avoid the tolls? Or how do [30:39] you want to get there? So, it's the [30:40] human who will help you to decide that [30:42] with the with the with the AI. So [30:44] strategically for each organization will [30:46] be different because you would just [30:48] basically have to actually do a deep [30:51] analysis of the tasks, the roles, the [30:54] functions, all the little steps that are [30:57] done each and every day that makes your [30:59] organization run efficiently and see how [31:02] you can automate and AI those mundane [31:06] tasks to become more efficient. [31:09] >> Thank you Angela, you brought up a very [31:11] good point, human in the loop. So Archa [31:13] we we were talking about the human in [31:14] the loop especially in the work that you [31:17] are doing especially bringing the [31:18] consumer brain uh near to the um humans [31:22] right so can would you like to talk [31:24] about [31:26] >> Sure um I'll finish answering your last [31:28] question first because I forgot to [31:30] actually give the tool the category of [31:32] tools that I was talking about is called [31:34] the conversational tagging tool um and [31:37] so if you think about it all your data [31:39] exchanges between humans chat chat, [31:42] text, emails, calls, they're all [31:44] conversational. But guess what else is [31:46] conversational? The exchanges that [31:49] people have with LLMs. Uh there was a [31:51] founder here last year that I met who [31:54] had a conversational tagging platform. [31:56] His company got acquired by OpenAI 2 [31:59] months ago and they're using it for [32:02] their own validation pipelines now. So [32:04] you can understand how important [32:06] conversational tagging is to the whole [32:08] AI pipeline. Now with that I move to the [32:10] next question human in loop. Um so heed [32:15] is essentially a consumer's brain that [32:17] powers your most consumer obsessed team [32:21] member agent that you can imagine. So [32:23] whether you're a brand manager, whether [32:26] you're a product manager, whether you're [32:27] a business analyst, it doesn't matter. [32:30] But imagine that every team member [32:32] within your organization, every employee [32:34] of your company gets as consumer [32:36] obsessed as Jeff Bezos. And there you [32:39] have it. That's what he does for you. [32:42] But AI at the end of the day is a [32:45] reflection of our own biases. Um, it's [32:49] not very good with reasoning just yet. [32:52] New models are coming out. They're [32:54] getting better and better with certain [32:56] reasoning. And as causal inference makes [32:59] their way with these machine learning [33:01] large language models, you'll start [33:03] seeing them getting even more better. [33:05] Um, but they're not there yet. And so [33:07] for that reason, you need a human in [33:09] loop for all the practical aspects that [33:13] the AI is either knowingly or not [33:16] realizing it has taken in the data set [33:19] and reflecting back on us that needs to [33:21] be checked. And so that's why you need a [33:24] human in loop. So when you have a [33:25] suggestion or a recommendation of action [33:28] because of an insight, a human needs to [33:31] approve it and say, "Yep, this is the [33:32] action that I want to take." What AI can [33:35] offer you is, "Hey, you just told me [33:36] about 30% of your service agents, you [33:40] know, don't really respond after 5:00 [33:42] p.m. on Fridays. Here are some of the [33:45] actions I can take for all the incoming [33:47] pathways." Um, we can spin up an option [33:50] of saying, "Hey, do you want to chat [33:52] with a chatbot?" So instead of moving on [33:54] to an agent, I have a salesperson [33:56] chatbot that's essentially taking on and [33:59] you know taking care of people who are [34:00] need instant gratification. That's one [34:02] option. Another option is I can increase [34:05] the shift hours of your salespeople and [34:08] adjust it so that you can cover the [34:09] additional two hours that you need [34:11] because after that the call volume drops [34:13] significantly or the request volume [34:15] drops significantly. the human is going [34:17] to choose which option or if both to [34:20] implement and how to implement in the [34:22] parameters. [34:24] >> Okay, thank you Arjuna. So Nille, so [34:26] coming into the future, so what is your [34:29] perspective on AI in BI especially in [34:33] the insurance domain? [34:37] Um so what we have seen that really [34:40] worked for us is uh purchasing and [34:44] embedding third-party data into [34:47] insurance underwriting. [34:49] Um in the past when we are doing [34:51] portfolio analysis and monitoring and [34:53] benchmarking we were only using the [34:55] snapshot of a uh person and then the [34:58] portfolio at the time of underwriting [35:01] which never changes but we know that as [35:03] we get older our health declines. [35:07] So uh we purchased a third-party medical [35:11] claims data uh and we are using that to [35:15] monitor the risk of our portfolio health [35:18] of our portfolio. Um you can see that of [35:22] course we're not expecting health [35:24] improving over time for anyone right as [35:26] we get older but you can get rid of the [35:29] um age factor. You can standardize that [35:32] and then see what else is playing a [35:35] role. Are there any new health trends [35:37] emerging or a certain clients or uh or [35:41] portfolio is declining um in terms of [35:44] health a lot faster than the other ones. [35:48] So now we able to see that and we are [35:51] talking to our uh clients what can we do [35:54] together uh some sort of preventive [35:56] actions. For example, we um couple of [35:59] years ago we partnered with a startup um [36:03] gallery is the name of the test. Um so [36:06] uh they are they are in Silicon Valley [36:08] and what they are doing is with u with [36:11] the blood test they are identifying 55 [36:14] 60 different types of cancers early [36:17] detection. So we're pro for example we [36:20] partner with our clients life insurance [36:22] companies that they are providing this [36:24] test. It is I think $900 something [36:26] dollars right now if you want to [36:27] purchase but um if you have a if you're [36:30] a policy holder you can get it a lot [36:32] cheaper. So in terms of the future I [36:35] think uh embedding AI into your process [36:38] and getting web scraping for example uh [36:41] for business intelligence for premiums [36:44] that is public data that is available to [36:46] us and now with AI um it's a lot more [36:50] cost effective to integrate that uh [36:53] third party public data into your [36:56] processes um and considering um [36:59] third-party data purchases and [37:02] partnering with startups. So in the [37:05] insurance space, insurance is not like [37:07] startups that you can move faster and do [37:10] a lot of innovation. But with the [37:11] strategic partnership and um using your [37:15] tech team uh more effectively [37:19] and uh having the benefit of AI, I think [37:23] in in one or two years all these things [37:26] that we are doing right now trying to [37:28] manually extract insights um and trying [37:32] to figure out what we can do about it [37:34] will be more automated. So we will have [37:38] dashboards that will for example track [37:41] the user interaction and then tell you [37:43] that your dashboards are missing this [37:45] certain aspect or chart pie chart table [37:48] because your users are doing this uh [37:51] manually. So it will do it itself um and [37:55] self-learning [37:57] um more AI powered intelligent business [38:00] intelligent dashboard. So I think uh in [38:02] one or two years in the life insurance [38:04] space this is very likely that um we [38:07] will have a predictive and prescriptive [38:10] uh BI dashboards. [38:13] >> Thank you. So Angela, so what's your [38:15] perspective for AI uh in future for [38:19] especially for businesses? [38:22] AI again as I say is such a [38:25] transformative innovation that every [38:29] company that's not already doing some [38:33] with AI something meaningful something [38:36] measurable with AI to drive their [38:38] business results is already falling [38:40] behind their competitor [38:42] basically [38:44] go my thing if you do one thing starting [38:48] next week when you go back to work is [38:49] really find one measurable of a KPI, one [38:52] small thing that you can change in your [38:54] organization to be able to use [38:56] integrate, use a small AI to change it. [38:59] Whether it be for information, [39:01] frequently asked question, a chatbot is [39:03] one of the easiest things that you can [39:04] you can implement in your organization [39:06] because all the data is already there [39:08] from previous as you said in your [39:10] organization where you you call [39:12] recorded. So all that information is [39:14] there, all that data is there. So we can [39:16] start by doing just one thing and just [39:18] seeing the needle move forward and keep [39:21] in mind make it measurable so that you [39:23] can see the impact from integrating AI [39:27] and see the difference of what it makes [39:28] in your organization. [39:30] >> Thank you. So from my perspective so as [39:33] I mentioned before the customer service [39:35] conversational analytics use case. So um [39:38] what we are doing in terms of the [39:40] analytics right so the conversations [39:43] have both customers aspects as well as [39:46] agent aspects in terms of agents the use [39:48] case that I talked about is automated [39:51] quality assurance but at the same time [39:53] we can also generate insights from those [39:55] conversations about our customers. Now [39:57] what we are seeing is from data to [39:59] insight standpoint. So traditionally we [40:01] have developed a lot of intelligence in [40:03] terms of machine learning or pure data [40:05] science. So we have lots of segments [40:07] hyper segments about the customers and [40:09] now when we are looking at different [40:12] channels independently so we have [40:14] established enough analytics now it's [40:16] the time for us to be able to bring [40:18] everything together so um one use case [40:21] that I would like to talk about right so [40:22] some a customer who come to our website [40:25] place an order and then um they may like [40:30] the product the way the uh uh website [40:33] operates so they may have shared the [40:34] message socially Okay. And then they may [40:37] come to the customer service and they [40:40] had a problem, talked about it, solved [40:42] it and then they have given us an NPS [40:45] score or a promoter comments etc. So now [40:48] we should be able to put together all [40:50] the information at one place and [40:51] understand how valuable this customer [40:54] is. Traditionally we have customer [40:56] lifetime value models from the machine [40:57] learning standpoint and we have [40:59] propensity models from their purchase [41:01] propensity standpoint. So similarly [41:04] there are much siloed areas. Now with [41:07] the AI agents we should be able to put [41:09] everything together to be able to [41:12] identify who are our most valuable [41:15] customers and how should we treat them [41:16] at the same time what kind of [41:18] prescriptions that we need to generate [41:20] in terms of personalizing their [41:22] experiences in various channels. So this [41:24] is allowing us to do that at the moment. [41:26] So um anyway so from our future [41:29] standpoint it's looking bright. At the [41:31] same time what we have talked about [41:33] right so human in the loop element it's [41:35] really really important because the [41:37] actions and the prescriptions that the [41:40] AI is going to generate has to be [41:42] monitored as well. Uh now with that I [41:45] think we are almost at the end of the uh [41:47] session. Any any questions?