Banks and Hospitals Limit AI Deployment Over Governance and API Volatility Risks

因治理与API波动风险，银行和医院限制AI部署 金融与医疗机构正放缓AI集成步伐，优先应对合规风险、供应商模型不稳定性及严格的验证要求，而非盲目追求快速迭代。 来自银行、医疗、销售工程和AI集成领域的行业专家小组探讨了企业如何在应对架构瓶颈、数据治理和技术快速更迭的同时，从数据中释放真正的商业价值。专家们强调了高风险敏捷方式与强监管企业环境之间的差异，指出立竿见影的投资回报率往往来自于提升人员效率与简化工作流程，而非替代员工或增加工具复杂度。他们还探讨了受监管领域持续进行模型重新验证的挑战，提倡将数据作为产品对待，并强调了建立统一企业数据文化的重要性。 在企业 AI 落地中，初创公司通常追求快速见效而忽视合规，而银行与联邦机构则要求在数据开放前完成严格的数据治理与风险把控。 (0:32) 数据源库的频繁变动与数据工程积压构成了核心架构瓶颈；同时，供应商频繁强制更新 API 模型在医疗等强监管领域引发高昂的重新验证成本。 (0:53) 企业常因缺乏统一数据治理而出现指标混乱（如不同部门存在 12 种“营收”定义），专家建议必须尽早从数据源头对齐业务指标。 (1:15) 在医疗等高成本领域，AI 的首要 ROI 在于辅助医护人员提高工作效率而非完全替代人力；同时，大模型直接访问数据源正推动中间分析工具的精简。 (22:10) 强监管行业普遍难以适应频繁的模型重训，医疗领域需经历 6 个月后台静默运行与小范围试点，而银行也倾向在内部沙盒充分监控后才考虑面向公众发布。 (37:20)

Enterprise leaders across banking, technology, and healthcare argue that unlocking data value depends on disciplined governance, pipeline reliability, and augmenting human workflows rather than pursuing frequent model updates or total workforce automation. Data Pipeline Bottlenecks and Metric Alignment (0:53) Operational friction in enterprise data initiatives rarely stems from an inability to access data; rather, data remains locked in rapidly changing source databases. Kelly Parker notes that moving this information into medallion architectures or analytics hubs creates severe engineering backlogs. Furthermore, basic data hygiene failures compound these technical hurdles. Parker cites an enterprise SaaS example where separate departments used 12 conflicting definitions of 'revenue,' requiring a complete audit of analytics pipelines and metrics at the source. Pragmatic Efficiency Over Total Workforce Replacement (22:10) In high-fixed-cost sectors like healthcare, Christopher Bartlett argues that practical return on investment comes from enhancing human efficiency rather than attempting full role replacement. Assisting medical staff—such as increasing the speed and accuracy of radiologists—provides immediate operational and financial gains, whereas moonshot automation goals carry high failure risks. This operational emphasis contrasts with executive perceptions. Parker points to a CIO Magazine survey revealing that while corporate executives often report widespread AI mastery, frontline implementers face execution hurdles, confusion, and fear of job displacement. Addressing this gap requires streamlining workflows; Raghu Vatte notes that as AI directly interfaces with source data, redundant intermediate tooling—including spreadsheets and middle-tier business intelligence platforms—faces obsolescence. Model Volatility and Regulatory Validation Runways (37:20) Rapid vendor release schedules create substantial friction for regulated enterprises. Bartlett emphasizes that deploying AI in clinical settings requires lengthy validation runways, such as running a model silently in the background for six months followed by small physician pilots. When external vendors deprecate APIs or update models every six to twelve months, enterprises must conduct expensive re-validation cycles that halt deployments and erode ROI. For regulated workflows, a fixed, 'fit-for-purpose' model with easily detectable errors offers far greater stability than constant upgrades. Financial institutions face similar constraints. Frank Kartmann explains that due to regulatory oversight and uncertainty, Deutsche Bank tests large language models strictly within internal sandboxes and monitors employee usage over extended periods before contemplating public-facing deployments. Cloud Architecture and Proactive Agentic Systems (1:37) To extract value from historical silos, Kartmann advocates treating data as a product governed by clear ownership, catalogs, and development roadmaps, supported by containerized cloud-native platforms like Kubernetes. These architectures support real-time requirements, including 24/7 customer service bots and continuous anti-financial crime stream processing. Technical selection criteria are shifting as well. Bartlett observes that engineering teams are moving away from continuous large language model benchmarking due to rapid release cycles. Vatte highlights the rise of agentic AI systems that proactively monitor underlying data to flag anomalies and security threats without user prompting, while Parker notes that AI-assisted 'vibe coding' enables internal teams to produce custom applications in roughly 40 hours, shifting traditional build-versus-buy procurement calculations. The panel underscores that enterprise data maturity is primarily an organizational and pipeline challenge rather than a race to adopt the newest algorithm. An ongoing tension remains between third-party vendors forcing rapid model updates and the protracted, high-liability validation cycles required in regulated sectors.

来自金融、医疗与企业软件领域的专家指出，释放企业专有数据价值的关键，并非盲目追逐前沿模型的频繁更迭或试图实现全盘自动化，而在于构建标准化的底层数据治理、以辅助人类员工效率换取务实投资回报，并在严格的监管验证与技术敏捷之间实现平衡。 数据指标统一与治理双轨制的平衡 (1:15) Fivetran 的 Kelly Parker 指出，企业将数据转化为运营实效的主要阻碍在于 IT 与数据工程的积压，数据常年被封闭在不断变化的源数据库中。他以自身在 SaaS 企业的经历说明缺乏标准化指标治理的代价：该公司各部门曾对“收入”（Revenue）存在多达 12 种互不兼容的定义（如月度经常性收入、年度合同价值、实际用量与未开票消费），导致分析管线产生严重偏差，迫使团队进行漫长且耗资巨大的重新审计与对齐。他强调数据负责人必须在源头尽早统一定义。 AI Squared 的 Raghu Vatte 进一步分析了企业在数据治理上的两极分化：小型初创团队倾向于激进推进，往往将治理和数据质量视为出成果后的次要问题；而大型银行与联邦机构则极其保守，必须在完全解决治理、合规与风险控制后才开放数据。他认为，核心战略挑战在于平衡两端，既要通过灵活架构让团队尽早提取业务洞察，又不能破坏治理底线。德意志银行的 Frank Kartmann 亦强调，企业必须将数据视作具备明确负责人、质量指标与发布周期的产品进行全生命周期管理，避免单纯为了收集数据而投资技术。 强监管行业中的模型验证与版本波动 (37:20) 针对高监管环境下的 AI 落地，全国儿童医院（Nationwide Children's Hospital）的 Christopher Bartlett 指出，服务商频繁推送新模型所带来的“模型波动”给企业造成了沉重的合规负担。在医疗等行业，临床 AI 工作流需要经过极其严苛的验证，若服务商每 6 至 12 个月就强制废弃旧接口，企业将被迫陷入昂贵且耗时的重新验证循环，从而吞食项目利润。Bartlett 解释称，医疗场景的模型落地存在极长的评估周期：典型流程包括让模型在后台静默运行 6 个月以收集性能数据，并仅向一至两位医生开放试点，确认其输出与医学专业知识一致后才可能在全院铺开。 因此，Bartlett 认为，频繁更新模型并不适合高风险临床决策，只要模型达到人类水平精度且错误易于被人类识别（即“适用且有效”），就极具商业与临床价值且无需频繁更替。针对这一问题，德意志银行的 Frank Kartmann 阐述了银行业的应对方式：由于外部不确定性与严格的金融监管要求，德意志银行将大语言模型（LLM）严格隔离在内部沙箱中，首先在内部员工中监测模型行为，只有经过长时间的内部稳定性验证后，才会评估面向公众提供工具的可行性。 务实赋能人类效率与中间软件的精简 (22:10) 在探讨人工智能的投资回报率（ROI）时，Christopher Bartlett 强调医疗机构具有庞大的固定成本结构并高度依赖专业医护人员。技术界常宣称要彻底取代放射科医生或临床医生，但他认为，当前最务实且即时的商业回报来自于提升人类员工的效率与准确率，而非消除岗位。通过协助放射科医生更快完成诊断，能够在不引发岗位替代阻力的情况下，迅速创造财务与运营价值，并维持管理层对多年期技术投资的信任。 Kelly Parker 引用《CIO 杂志》的调查指出，管理层对 AI 广泛落地的乐观判断与一线员工的怀疑、困惑及失业焦虑之间存在巨大鸿沟。Raghu Vatte 补充认为，当前企业转型的焦点正在从“取代人工”转向“淘汰不必要的中间软件工具”。在传统数据架构中，数据需要经过电子表格或中间商业智能（BI）工具的多层格式转换与抽取；而现代 AI 系统开始直接接入底层原始数据源生成洞察，使得大量中间层软件面临被淘汰的风险，从而实现工具链的大幅精炼。 云原生架构演进与主动型智能的兴起 (30:56) 为最大化释放数据价值，德意志银行的 Frank Kartmann 指出金融行业正加速向云原生与容器化架构（如 Kubernetes）演进，以打破部门间的数据孤岛并根据市场动态弹性扩展计算资源。现代银行架构依托实时流处理与全天候监控，已能实现秒级反金融犯罪（AFC）欺诈识别，并部署能够进行情绪与行为分析的 24/7 智能对话助手，在移动端、网页端与实体网点之间保持全渠道一致的个性化服务体验。 与此同时，技术选型范式也在发生深刻转变。Bartlett 提到技术主管因模型迭代过快已停止对每一个新大模型进行基准测试，转而依赖容器化隔离环境由安全团队控制风险。Raghu Vatte 指出企业正迈向“主动型智能体”（Agentic AI）架构，系统不再被动等待用户输入精确提示词，而是自主扫描底层数据以主动发现异常、业务偏差及安全漏洞。Kelly Parker 进一步指出，AI 辅助的快速编程（Vibe Coding）使得内部团队在约 40 小时内即可构建出定制化业务原型，深刻改变了传统的“自研还是购买”的软件采购决策。 讨论表明，企业释放数据价值的最大障碍通常不在于算法算力，而在于跨部门核心指标定义的统一、严格监管环境下漫长且刚性的模型重验成本，以及高管预期与一线执行之间的现实差距。如何在不触碰合规红线的前提下精简冗余的中间工具链、以增强人类工作效率获取立竿见影的投资回报，并在保持数据底座稳固的同时引入主动型智能体架构，是高监管行业在数字化转型中亟待解决的核心问题。

## Background 本内容由 Ai4 于 2025 年 8 月 27 日发布在 YouTube 平台上。视频记录了一场汇集来自 Fivetran、全国儿童医院、AI Squared、Wilson Bank & Trust 以及德意志银行等多位专家的圆桌研讨，旨在探讨如何将原始数据转化为驱动商业智能与 AI 落地的高价值业务洞察。 ## TL;DR 专家小组深入探讨了不同规模企业在数据治理与创新速度之间的权衡，并揭示了数据管道需求积压、指标定义混乱以及模型频繁强制升级带来的高昂重新验证成本等核心痛点。针对医疗与金融等强监管行业，嘉宾们强调了务实增效、合规优先以及将数据作为独立产品（Data as a Product）运营的重要性。最后，研讨展望了云原生架构、Agentic 主动智能与 Vibe Coding 时代的自建变革，并指出了直连内外部数据的业务分析 Chatbot 等高 ROI 落地场景。 ## Atmosphere & style 整场研讨会专业务实且充满行业一线实战的真知灼见，嘉宾们以幽默自嘲的口吻揭露了高管幻象与一线断层、把 Excel 当数据库等行业乱象。现场互动轻松活跃，不仅有主讲人分发糖果的趣味插曲，更在严肃讨论强监管合规与工程痛点时展现出严谨的思辨深度。听众能直观感受到技术前沿狂热与行业落地现实碰撞出的理性火花。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=V22p0JQN2_o&t=4s) 跨国数据管理的底层技术逻辑大同小异，核心差异主要在于不同国家和地区的语言及法律合规要求。 - [00:32](https://www.youtube.com/watch?v=V22p0JQN2_o&t=32s) 企业在数据价值化路径上呈现两极分化：初创公司急于产出而往往忽视治理与安全，而大型银行等受监管机构则坚持合规先行。 - [00:52](https://www.youtube.com/watch?v=V22p0JQN2_o&t=52s) 大模型服务商频繁强制废弃旧 API 与升级模型，迫使强监管行业的企业进行高成本的重复验证，严重侵蚀项目 ROI。 - [01:15](https://www.youtube.com/watch?v=V22p0JQN2_o&t=75s) 缺少源头数据治理会导致同一指标出现口径分裂（如一家公司内存在 12 种不同的收入定义），数据问题必须尽早追溯至源头纠正。 - [01:46](https://www.youtube.com/watch?v=V22p0JQN2_o&t=106s) 金融与传统行业释放数据红利的关键在于通过云原生与 Kubernetes 容器化架构，打破各部门之间的数据孤岛。 - [19:28](https://www.youtube.com/watch?v=V22p0JQN2_o&t=1168s) 金融业正利用实时数据与预测分析重塑全渠道客户体验，并通过情绪感知与反欺诈技术提升风控和交互质量。 - [22:10](https://www.youtube.com/watch?v=V22p0JQN2_o&t=1330s) 医疗 AI 的高投资回报率来自于切实提升一线医护的日常处理速度，而非盲目追求用技术完全替代医生。 - [23:38](https://www.youtube.com/watch?v=V22p0JQN2_o&t=1418s) 企业在推动 AI 落地时普遍存在断层：管理层乐观高估内部成熟度，而一线员工往往充满疑虑甚至感到被替代的威胁。 - [25:30](https://www.youtube.com/watch?v=V22p0JQN2_o&t=1530s) AI 的发展正在取代多层中间数据加工工具，未来用户可直接向底层数据提问并获得端到端洞察，无需复杂的中转流程。 - [28:18](https://www.youtube.com/watch?v=V22p0JQN2_o&t=1698s) 释放数据价值需要将数据视作具备明确 Owner、发布周期与质量指标的“数据产品”，并全面培养业务人员的数据思维。 - [30:56](https://www.youtube.com/watch?v=V22p0JQN2_o&t=1856s) 基础模型更新迭代过于迅猛，技术团队已从无休止的横向基准测试转向迅速选定可用方案并在安全边界内推进落地。 - [32:54](https://www.youtube.com/watch?v=V22p0JQN2_o&t=1974s) 数据交互正演进为智能体（Agent）主动感知并推送异常，同时 Vibe Coding 将内部系统开发周期压缩到数十小时，重构自建与采购的决策逻辑。 - [37:20](https://www.youtube.com/watch?v=V22p0JQN2_o&t=2240s) 医疗机构新模型上线前需经历数月后台静默运行和严苛审核，这种长周期的合规验证与外部模型的高频更新形成了尖锐矛盾。 - [41:56](https://www.youtube.com/watch?v=V22p0JQN2_o&t=2516s) 德意志银行等金融机构对大语言模型采取极端审慎策略，目前严格限制在内部环境测试监控，短期内不对外部客户开放。 - [41:56](https://www.youtube.com/watch?v=V22p0JQN2_o&t=2516s) 当前企业中投入产出比最高的落地场景包括：自动化会议纪要工具，以及能够跨内外部数据进行秒级损益和进销存分析的高管问答助手。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=V22p0JQN2_o&t=4s) 专家小组成员自我介绍与行业背景 会议由主持兼研讨者 Eric Ferguson 开场。他有着超过30年的银行从业经验，曾在渣打银行（Standard Chartered）、美国银行（Bank of America）工作，目前担任 Wilson Bank & Trust 的数字、数据与分析总监。他曾管理跨 35 个国家的数据，并指出跨国数据管理的本质差异仅在于语言和法律合规，底层的“比特与字节”无甚差别。随后，四位来自不同领域的专家依次分享了各自的背景： * **Raghu Vatte**（AI Squared 现场首席技术官）：AI Squared 旨在帮助企业将 AI 融入业务流程（涵盖传统工作流与新兴工作流），通过激活企业固有数据来挖掘深度洞察。 * **Kelly Parker**（Fivetran 销售工程高级总监）：专注于数据流动（Data Movement）。Fivetran 帮助企业将数据从各类数据源大规模迁移至数据仓库和数据湖，为后续的 AI 与分析项目打下基础。 * **Frank Kartmann**（德意志银行 大数据与云架构副总裁）：曾在德国中央银行等大型金融机构主导大数据平台建设，目前在法兰克福总部负责构建德意志银行的数据平台，整合各部门数据并赋能业务端用例。 * **Chris Bartlett**（全国儿童医院 联席首席数据科学官 / 俄亥俄州立大学 儿科教授）：兼具企业管理与科研创新双重视角。他在学术领域深耕视觉模型与前沿基础模型，在医院则主要负责数据治理与企业级数据应用。 --- > “在 35 个国家管理数据，真正的区别只有语言和法律……底层的比特与字节其实大同小异。” —— Eric Ferguson - Chris Bartlett 现场带来了俄亥俄州立大学著名的“七叶树”糖果（Buckeye candies）分发给嘉宾，并幽默承诺向现场提问的观众赠送糖果。 --- ### [00:32](https://www.youtube.com/watch?v=V22p0JQN2_o&t=32s) 将原始数据转化为业务价值：不同规模企业的策略差异 针对“如何将原始数据转化为实际业务或组织影响力”的问题，专家们从不同类型客户的痛点与路径展开了探讨： Raghu Vatte 指出，数据是企业最核心的竞争壁垒（Moat）。然而，不同规模与类型的机构在追求数据价值时，其时间线、风险偏好与治理策略呈现出极端的“双端摆动”现象： ```mermaid flowchart LR accTitle: 企业数据治理与创新策略演进路径 subgraph Small ["小型企业 / 激进派"] direction TB A["快速获取数据成果"] --> B["事后补救 数据治理与安全"] end subgraph Enterprise ["大型/联邦机构 / 保守派"] direction TB C["严格的数据治理与合规保障"] --> D["稳步推导 业务产出"] end Small -- "跨越治理与速度的鸿沟" --> Enterprise ``` * **初创与小型企业**：追求极致速度，往往将数据安全与治理视为“事后补救措施”，急于先看到结果； * **大型银行与联邦机构**：处于光谱的另一端，坚持“治理先行”，只有在彻底解决合规与安全隐患后才推进具体业务产出。 针对这一现状，关键在于提供平衡的工具链，避免激进者“画地为牢”，也防止保守者因过于推迟而错失时机。Frank Kartmann 则补充道，金融机构的根本落脚点是**提升客户满意度**。随着技术演进，银行正在通过两类典型场景释放数据价值： 1. **实时个性化服务**：构建能够实时调取客户私有数据的智能客服 Chatbot，打破传统分行营业时间的限制，实现 24/7 互动与精准产品推荐； 2. **反金融犯罪（AFC）**：通过实时数据分析与市场监控，即时识别并拦截诈骗交易，全面提升风控与运营效率。 --- ### [00:52](https://www.youtube.com/watch?v=V22p0JQN2_o&t=52s) 提取与运营数据价值的核心挑战 谈及当前在提取与运营数据价值时面临的最大障碍，专家们从技术流水线、模型生命周期和数据交付三个维度进行了剖析： Kelly Parker 首先指出了 IT 与数据工程团队普遍面临的**需求积压（Backlog）**问题。数据往往被锁在底层数据库中，而源头数据库又在不断变动，导致数据难以快速顺畅地流入金字塔/分层架构（Medallion Architecture）以供业务端运营使用。 Chris Bartlett 则提出了一个困扰企业端用户的痛点：**服务商强制更新 API / 模型引发的重复验证成本**。 ```mermaid flowchart TD accTitle: 模型强制升级导致企业成本增加流程 A["企业完成 API/模型验证并投入生产"] --> B["服务商 6 个月后废弃旧 API / 模型"] B --> C{"新模型在特定垂直场景表现"} C -- "可能退化 / 不稳定" --> D["企业必须重新进行合规与业务验证"] D --> E["侵蚀项目利润，导致 AI 方案被放弃"] ``` 在严苛的合规行业，一个模型经过验证后，企业希望其能稳定运行 3 年；但上游厂商每半年就强行废弃旧 API 并推送新模型。虽然新模型整体能力更强，但在特定垂直细分领域性能可能反而下降，迫使企业频繁重新验证，极大地侵蚀了项目 ROI。 Raghu Vatte 对此表示高度认同，并补充了企业适配数据的三大核心门槛： 1. **数据治理**：保障数据合法、合规与安全使用的底线； 2. **持续数据流与推理时上下文**：AI 不仅关乎模型训练，更关乎实时推理。必须确保数据管道能够持续输送最新的上下文数据，否则“垃圾进，垃圾出”； 3. **洞察的呈现与交互界面**：数据专家与业务人员存在认知鸿沟，如何将复杂的数据分析以直观、可交互的形式展示给普通业务人员，仍然是行业未完全解决的难题。 --- ### [01:15](https://www.youtube.com/watch?v=V22p0JQN2_o&t=75s) 数据治理缺失引发的“12种收入定义”惨案 在被问及“过往经历中最惨痛的数据失败案例及教训”时，Kelly Parker 分享了一个来自销售与营收分析领域的真实故事： 在一家基于 SaaS 与按量计费（Usage-based）模式的公司中，团队在进行早期数据分析时，惊讶地发现公司内部对于“收入（Revenue）”这一基础指标竟然存在**至少 12 种截然不同的计算定义**（如月度营收、年度营收、合同总价值、实际用量、未结清用量等）。 这一数据治理的缺失直接导致不同部门之间的分析结果产生严重偏差，最终不得不花费巨大精力去追溯源头、统一口径，并重塑团队的数据思维方式。 - **核心教训**：面对数据质量与治理问题，“越早追溯至数据源头（Source）越好”；一旦发现口径错误或数据异常，必须立即采取纠正行动。 - Eric Ferguson 也补充道，自己曾经历过因继承底层数据错误，导致相关指标直接影响到上市公司股票分析师评估的极端严重事件。 --- ### [01:46](https://www.youtube.com/watch?v=V22p0JQN2_o&t=106s) 未来展望：云原生架构与打破数据孤岛 展望未来，Frank Kartmann 认为金融及相关传统行业解锁数据最大价值的突破口在于**技术架构的云原生转型与组织数据孤岛的破除**： 要支撑敏捷的业务决策与个性化体验，企业必须构建具备高度弹性与可拓展性的云原生应用架构。通过采用容器化技术（如 Kubernetes），系统能够根据市场需求实时动态扩缩容。同时，技术架构的升级还必须服务于“打破业务部门间的数据孤岛（Data Silos）”，让跨部门的数据能够自由流动与融合，从而真正释放企业级的数据红利。 - **金融行业三大价值高地**：实时客户体验升级、精准风险管理与市场决策、高度自动化的运营效率提升。 - **技术落地关键词**：云原生（Cloud-native）、Kubernetes 容器化、动态扩缩容、跨部门数据融合。 <ElicitationsGroup message="要深入探讨本场研讨会的后续内容或技术细节，您可以选择："> <Elicitation label="获取下半部分剧透与数据平台构建指南" query="请继续提供该研讨会后续转录内容的详细解读，重点关注数据平台架构与企业 AI 落地最佳实践。"/> <Elicitation label="深入分析 Fivetran 与 AI Squared 的技术方案" query="请详细对比 Fivetran 的数据管道技术与 AI Squared 在企业 AI 流程集成上的技术架构与优势。"/> </ElicitationsGroup> ### [19:28](https://www.youtube.com/watch?v=V22p0JQN2_o&t=1168s) 银行业的实时数据价值与全渠道一致性 在金融服务领域，高质量与实时数据是驱动智能化转型的核心燃料。银行不仅需要利用高质量数据来训练模型，更需要通过分析实时数据，在第一时间捕捉客户诉求并提供量身定制的产品与服务。通过在业务流程中引入增强分析（Augmented Analytics）与预测分析等前沿技术，银行业迎来了重塑客户互动的巨大机遇。 行业内甚至开始部署具备情绪感知与行为分析能力的智能 Bot。虽然并非所有机构都已全面上线此类方案，但前沿实践表明，系统可以根据客户当前的即时情绪反应和特定场景需求，动态推送最匹配的金融产品，从而大幅提升客户满意度。与此同时，多渠道触点的增加也带来了全渠道一致性（Omni-channel Consistency）的挑战：无论是移动端 App、线下实体网点还是官方网站，客户在所有触点接收到的沟通与体验必须保持高度一致。此外，完善的数据安全与数据治理体系是这一切的基石，唯有守住安全底线，才能赢得客户长久的信任。 - **实时响应与定制化**：利用实时数据分析与 AI 预测能力，在客户产生需求的瞬间提供精准产品匹配。 - **情绪与行为分析**：前沿 Bot 能够识别客户情绪波动与交互行为，实现高度拟人化和有温度的推荐。 - **全渠道一致体验**：确保移动端、网页端与线下网点之间的信息与服务步调一致。 - **安全与治理护航**：通过健全的治理框架保障合规，筑牢客户信任防线。 --- ### [22:10](https://www.youtube.com/watch?v=V22p0JQN2_o&t=1330s) 医疗场景下的务实 AI：提升效率而非替代医生 医疗机构在本质上是重资产且高度依赖人力运转的实体，这使得技术赋能的潜在空间极为广阔，全行业对此也抱有高度共识。然而，计算机科学界往往沉迷于宏大叙事，动辄高喊“用 AI 彻底取代放射科医生”或颠覆某个医疗领域。这种设想固然能作为激励口号，但在现实中却很难快速兑现投资回报率（ROI）。 真正的医疗 ROI 来自于极其务实的“地面战”：只要一项技术能够达到接近人类专家的水平，同时大幅提高处理速度，回报就会立刻显现。根本不需要解雇放射科医生，只要让他们阅片和出报告的速度加快，投资回报就已经成立。尽管未来或许会出现大量机器人直接进行伤口护理等临床操作，但在当下，最理性的策略是立足现有流程，聚焦于提升一线医生与护士的人效，实现即时可见的收益，而不是盲目追求宏大目标最终落得失败退场。 ```mermaid flowchart LR accTitle: 医疗 AI 的落地路径对比 subgraph S1["宏大叙事路径 (高风险 / 难见效)"] A["目标: 完全替代医生/护士"] --> B["研发周期漫长且面临伦理审查"] --> C["难以短期兑现 ROI"] end subgraph S2["务实增效路径 (高 ROI / 易落地)"] D["目标: 赋能现有医护人员"] --> E["AI 辅助处理加速日常流程"] --> F["医生工作效率大幅提升"] --> G["即时实现投资回报 (ROI)"] end ``` --- ### [23:38](https://www.youtube.com/watch?v=V22p0JQN2_o&t=1418s) 管理层幻象与一线断层：对数据与 AI 认知的巨大鸿沟 当前企业在推进数据与 AI 战略时，管理层与执行层之间存在着一道巨大的认知鸿沟。《CIO》杂志近期的一项调查生动地揭示了这种脱节：当被问及组织内部应用 AI 的成熟度与效率时，高管们信心满满地宣称“我们做得棒极了，业务各环节都已经完全掌握并深度应用了 AI”；然而一线员工的真实反馈却是“绝无可能（No way, dude），我们根本没在用，而且对这套东西充满疑虑甚至一无所知”。 管理层往往被外界同行的宣传所裹挟，急切地质问“为什么别人能做我们不能做”，而真正负责落地、使用技术甚至担心被技术抢走饭碗的一线员工却处于恐惧与迷茫之中。这种局面宛如一百多年前经历的农业革命，当机械化浪潮席卷而来时，身处一线的劳动者切身感受到了替代压力。推进数据和 AI 建设的首要任务，应当是弥合这种管理层梦想与一线现实之间的断层，让全员真正理解这场技术变革的愿景与落脚点。 --- ### [25:30](https://www.youtube.com/watch?v=V22p0JQN2_o&t=1530s) 工具链的精简与数据视角重塑：从中间层到直接洞察 过去几十年的数据挖掘过程中，企业沉淀了海量沉睡的数据，却常常困惑于如何挖掘其最大价值。数据工程团队在治理和准备数据时，思维往往局限于“如何让下游的第一款工具使用这些数据”，导致数据在最终转化为业务价值前，不得不串联流转于两到三款中间工具之间——例如从底层导出到 Excel 做透视表，再导入 BI 工具生成图表以获取洞察。 当下的宏观趋势正在从“AI 取代人类”转向“AI 取代繁琐的中间工具”。如果用户可以直接向底层数据发出自然语言提问并瞬间获得精准洞察，那么夹在中间的诸多数据加工与展示工具就失去了存在的必要，这些工具若不主动进化，必将面临淘汰。因此，企业必须将数据本身视为直接提供洞察的源泉，而非服务于特定工具链的原材料；从这个视角出发，跨国数据治理（如新加坡与美国不同法规下的数据管辖）以及数据资产的维护和合规管理才会变得更加清晰与轻量化。 - **消除冗余中间层**：打破“底层数据 $\to$ 电子表格 $\to$ 报表软件 $\to$ 业务洞察”的长链条，推动从数据到底层洞察的直接交互。 - **工具生态优胜劣汰**：无法直接带来端到端价值的单点工具将加速被边缘化或被原生 AI 能力吞噬。 - **全场共识的调侃**：现场嘉宾一致强调，“必须向所有人再次声明：Excel 绝不是数据库”。 --- ### [28:18](https://www.youtube.com/watch?v=V22p0JQN2_o&t=1698s) 全民数据思维与以业务为中心的数据产品化 为了在组织底层彻底释放数据红利，必须让每一位一线业务人员意识到自己其实就是“数据工作者”。一线员工在日常工作中观察什么有效、什么无效，并逐步形成业务判断，这本质上就是收集与分析数据的过程。如果能够赋予他们基础的数据词汇和思维框架，当数据团队询问“如何帮你提升工作效率”时，他们便能提出清晰具体、可直接转化为数据科学任务的业务命题，而不是停留在抽象模糊的抱怨上，从而极大降低需求挖掘的沟通成本。 与培养全员数据素养相配合的，是企业必须打破“以技术为中心”的盲目狂热。很多公司在没有明确业务目标的情况下，仅仅为了拥有技术而投入大量资金采集数据。正确的路径应当是自顶向下：首先明确对外提供的产品与服务、确立核心业务目标，随后制定紧密对齐业务战略的数据战略。在此基础上，必须将数据视为持续运营与迭代的“数据产品”（Data as a Product），配备清晰的数据目录、明确的数据所有者（Data Owner）、严苛的质量指标、规范的发布周期以及长期的演进目标。 --- ### [30:56](https://www.youtube.com/watch?v=V22p0JQN2_o&t=1856s) 技术与模型选型困境：告别无休止测试与安全红线 在当前技术爆炸的环境下，企业在技术与模型选型上面临前所未有的压力。以往 CTO 和 CIO 们热衷于为每一个具体业务场景横向对比测试市面上所有的基础大模型（LLMs），但如今大多数技术领袖已经不得不叫停这种全面评测——模型迭代速度实在太快，无休止的基准测试只会导致业务停滞，团队只能迅速选定一款可用的模型并推进落地。 在医疗等强监管领域，模型选型最大的博弈往往发生在技术团队与安全隐私合规部门之间。安全团队对模型的数据来源审查极为严苛，通常要求将模型放置在严格隔离的容器（Container）中运行，甚至极难批准其进入生产环境。技术与学术团队普遍秉持“无论是开源还是闭源，只要发表且能解决实际问题就设法应用”的开放心态，而首席信息安全官（CISO）则时刻处于高度紧张的防守状态，以至于技术负责人开玩笑称“平时得定期给 CISO 喂降压药，生怕他被新技术的安全风险吓坏”。 --- ### [32:54](https://www.youtube.com/watch?v=V22p0JQN2_o&t=1974s) 从主动提问到 Agentic 智能：数据主动告警与 Vibe Coding 时代的自建变革 未来的数据交互形态正迎来根本性跃迁：从“人类向数据提问”演进为“智能体（Agent）驱动的数据主动感知”。过去，业务获取洞察的质量完全受限于提问者提出好问题的能力；而在 Agentic 趋势下，底层模型直接驻留在经过清洗的高质量数据之上，自主巡检、发现异常并主动发出警报（例如在安全攻防场景中主动提示潜在漏洞）。届时，人类的核心职责将转变为维护底层数据的真实性与可信度，而大量传统的查询与监控工具将彻底退出历史舞台。 与此同时，AI 辅助开发（Vibe Coding）正在彻底颠覆软件工程的“自建还是购买”（Build vs. Buy）决策。以企业内部工具开发为例，过去团队为了满足特定业务需求，从零手写代码搭建一套内部原型往往需要耗费数月时间；而在明确业务需求后，借助 Vibe Coding 仅用约 40 个小时就完成了功能完善的第二代版本。这种极低门槛、极高速度的构建能力，意味着开发者只要深刻理解业务场景即可快速交付工具，使得企业面对昂贵且不完全适配的商业软件时，天平正前所未有地向“自主定制构建”倾斜。 - **Agentic 主动智能**：系统不再被动等待 Query，而是自主在底层数据中挖掘异常、模式与商业洞察并向业务推送。 - **自建效率发生质变**：传统耗时数月的内部系统原型，在 Vibe Coding 赋能下仅需约 40 工时即可高质量交付。 - **软件采购逻辑重构**：业务人员无需精通底层编码细节即可构建贴合场景的专用工具，商业软件采买门槛面临重估。 ### [37:20](https://www.youtube.com/watch?v=V22p0JQN2_o&t=2240s) 强监管行业中的模型频繁迭代与验证困境 进入观众提问环节，主持人开玩笑提到提问者可能会分到糖果（Buckeyes 巧克力糖）。一位观众向 Nationwide Children's Hospital 的 Christopher Bartlett 以及德意志银行（Deutsche Bank）的 Frank Kartmann 提出了一个现实而尖锐的工程与合规难题：当前许多模型（包括连续训练模型或生命周期极短的模型）需要高频迭代，而在医疗和金融等高度受监管的行业中，监管机构要求实体必须对其自身或供应商使用的模型进行严格的验证与再验证（Model Re-validation）。在需要设置精准的精确率（Precision）、召回率（Recall）等阈值来防范欺诈或恶意请求的场景下，企业应当如何加速、优化或自动化这一验证流程，以便在面对频繁变更的模型时依然能保持合规与安心？ ```mermaid flowchart TD accTitle: 医疗行业模型验证与上线流程 subgraph S1["阶段一：静默运行与早期验证"] A["新模型/微调模型"] --> B["后台静默运行 6 个月"] B --> C["少数核心医生（1-2人）接入测试评估"] end subgraph S2["阶段二：合规评估与准入"] C --> D{"质量与合规评估<br/>(符合适用标准且错误易辨识?)"} D -- "未达标/风险高" --> B D -- "达标" --> E["进入全院/多医生正式部署阶段"] end subgraph S3["矛盾根源"] F["外部模型每 6 个月大幅迭代/更新"] -.->|周期脱节| B end ``` 针对这一问题，Christopher 从医疗健康的视角深入剖析了行业现实。在医疗领域，任何直接面向患者或临床医生的工具都必须经历非常漫长的准备跑道。例如，新模型通常需要在后台静默运行整整六个月以积累真实数据，并由一到两位医生专门进行试用评估，确认其输出确实符合预期后，才敢推向全院更多医生。Christopher 强调，出于合规、法律责任以及对患者安全负责的常识底线，这套耗时耗力的验证机制绝不能随意削减或取消。然而这也带来了一个根本性矛盾：如果外部模型厂商每六个月就推陈出新，这种高频变动与医疗机构的落地节奏完全脱节。 - **医疗质检的核心标准**：质量的定义是“符合适用目的（Fit for purpose）”；只要模型表现与人类相当，且出现的失误是显而易见、容易被人工挑选修正的，它就已具备极高价值，无需盲目追求频繁升级。 - **不同场景的容忍度差异**：在临床决策支持（Decision Support）方面，激进的持续集成与部署（CI/CD）并不适用；但在手术机器人（Robotic Surgery）等特定垂直场景下，自动化持续集成的可行性相对更高。 --- ### [41:56](https://www.youtube.com/watch?v=V22p0JQN2_o&t=2516s) 金融行业的审慎应对与高 ROI AI 应用落地 Frank Kartmann 简要补充了银行业的做法。金融行业同样身处极其严格的监管体系之下，面对大语言模型（LLM）目前仍存在的不确定性与风险，德意志银行在推进研发时采取了极为保守的策略：所有自研或测试的模型严格限制在内部环境运行，并由专门的监管/合规部门进行持续行为监控，短期内绝不直接向公众或外部客户开放，唯有经过长期验证无误后才会考虑推向市场。 随后，现场观众提出了最后一个问题：尽管大家经常调侃 Excel 的局限性，但在高管层面，Excel 依然是他们最习惯、能最快获取价值的工具；那么在各团队的实际落地中，究竟哪些 AI 工具或策略展现出了最高的“投入产出比”（Cost-to-ROI Ratio），真正释放了数据价值？ ```mermaid flowchart LR accTitle: 高管智能分析助手的数据流与价值转化 subgraph Input["数据源层"] D1["企业内部运营/财务数据"] D2["竞品/市场公开数据"] end subgraph AI["AI 智能分析引擎"] NLP["自然语言查询与推理"] end subgraph Output["业务洞察与决策"] Ans["损益预测 / 营收趋势 / 进销存对比"] Human["管理层人工决策"] end D1 --> NLP D2 --> NLP NLP --> Ans Ans --> Human ``` AI Squared 的 Raghu Vatte 对此分享了两大最具立竿见影效果的高性价比方向： 1. **会议纪要与行动项自动化工具**：基于 NLP 技术的会议记录、纪要提炼及 Action Item 跟踪工具已经非常成熟且被广泛采纳。过去这项工作需要专门的助理人员，而现在 AI 能以极低成本高效完成，为企业节省了大量组织协同成本。 2. **深度连接内部与外部数据的业务问答机器人（Chatbot）**：不同于“查询天气”的基础对话机器人，新一代分析型 Chatbot 能够直连企业内部数据库并结合公开竞争对手信息。高管可以直接用自然语言提问，例如“给出未来3个月或过去1年的损益表（P&L）”、“分析当前库存趋势”、“预测未来6个月销售额并评估库存是否匹配”。过去需要 3 名专业数据分析人员耗时整整一周才能整理出结果的问题，现在几秒内即可生成洞察，最后交由管理层拍板决策。这种极致的提效为企业带来了最直接的投资回报。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=V22p0JQN2_o&t=4s) 在 35 个国家管理数据，真正的区别只有语言和法律……底层的比特与字节其实大同小异。 — Eric Ferguson > [01:15](https://www.youtube.com/watch?v=V22p0JQN2_o&t=75s) 越早追溯至数据源头（Source）越好 — Kelly Parker > [23:38](https://www.youtube.com/watch?v=V22p0JQN2_o&t=1418s) 我们做得棒极了，业务各环节都已经完全掌握并深度应用了 AI — 高管调查反馈 > [23:38](https://www.youtube.com/watch?v=V22p0JQN2_o&t=1418s) 绝无可能（No way, dude），我们根本没在用，而且对这套东西充满疑虑甚至一无所知 — 一线员工调查反馈 > [25:30](https://www.youtube.com/watch?v=V22p0JQN2_o&t=1530s) 必须向所有人再次声明：Excel 绝不是数据库 — 现场研讨嘉宾 > [30:56](https://www.youtube.com/watch?v=V22p0JQN2_o&t=1856s) 平时得定期给 CISO 喂降压药，生怕他被新技术的安全风险吓坏 — 技术负责人 > [37:20](https://www.youtube.com/watch?v=V22p0JQN2_o&t=2240s) 符合适用目的（Fit for purpose） — Christopher Bartlett > [41:56](https://www.youtube.com/watch?v=V22p0JQN2_o&t=2516s) 给出未来3个月或过去1年的损益表（P&L） — Raghu Vatte > [41:56](https://www.youtube.com/watch?v=V22p0JQN2_o&t=2516s) 分析当前库存趋势 — Raghu Vatte > [41:56](https://www.youtube.com/watch?v=V22p0JQN2_o&t=2516s) 预测未来6个月销售额并评估库存是否匹配 — Raghu Vatte ## Mentioned Kelly Parker, Christopher Bartlett, Raghu Vatte, Eric Ferguson, Frank Kartmann, Fivetran, AI Squared, Nationwide Children's Hospital, Wilson Bank & Trust, Deutsche Bank, Standard Chartered, Bank of America, The Ohio State University, CIO Magazine, Kubernetes, Excel, Medallion Architecture, Vibe Coding ## Video description Kelly Parker, Senior Director, Sales Engineering, Fivetran Christopher Bartlett, Associate Chief Data Sciences Officer, Nationwide Children's Hospital Raghu Vatte, Field CTO, AI Squared Eric Ferguson, Director, Digital, Data & Analytics, Wilson Bank and Trust Frank Kartmann, Vice President of Big Data & Cloud Architecture, Deutsche Bank Abstract In an era where data is often the key to competitive advantage, knowing how to extract meaningful value from it is essential for success. This panel will explore practical strategies for transforming raw data into actionable insights that drive business intelligence and fuel AI initiatives. Through real-world examples and expert insights, you’ll discover effective tools, processes, and best practices for harnessing data’s full potential. Learn how to move beyond data collection to creating true business impact with data-driven decision-making. [Source](https://www.youtube.com/watch?v=V22p0JQN2_o)

[0:04] Eric Ferguson: 在 35 个国家管理数据，真正的区别只有语言和法律……底层的比特与字节其实大同小异。 [1:15] Kelly Parker: 越早追溯至数据源头（Source）越好 [23:38] 高管调查反馈: 我们做得棒极了，业务各环节都已经完全掌握并深度应用了 AI [23:38] 一线员工调查反馈: 绝无可能（No way, dude），我们根本没在用，而且对这套东西充满疑虑甚至一无所知 [25:30] 现场研讨嘉宾: 必须向所有人再次声明：Excel 绝不是数据库 [30:56] 技术负责人: 平时得定期给 CISO 喂降压药，生怕他被新技术的安全风险吓坏 [37:20] Christopher Bartlett: 符合适用目的（Fit for purpose） [41:56] Raghu Vatte: 给出未来3个月或过去1年的损益表（P&L） [41:56] Raghu Vatte: 分析当前库存趋势 [41:56] Raghu Vatte: 预测未来6个月销售额并评估库存是否匹配

[0:04] Eric Ferguson: Whether managing data across global boundaries or locally, the bits and bytes remain identical—the only real variables are language and legal frameworks. [0:32] Raghu Vatte: Proprietary data is an enterprise's true competitive moat. [0:53] Christopher Bartlett: Forcing an API upgrade every six to twelve months forces costly re-validation cycles that eat into project margins and halt AI deployment. [1:15] Kelly Parker: Resolve metric definitions at the source as early as possible, and act immediately when structural data inconsistencies are uncovered. [19:28] Frank Kartmann: By reading customer reactions in real time, these systems automatically surface specialized product offers to maximize satisfaction. [22:10] Christopher Bartlett: Immediate return on investment (ROI) comes from boosting human efficiency rather than replacing personnel. [23:38] Raghu Vatte: Historically, data pipeline architectures required multiple intermediate processing layers—such as extracting data into spreadsheets or intermediate business intelligence (BI) tools before extracting actionable insights. As AI systems gain direct access to raw data, these middle-layer tools risk obsolescence. [23:38] Raghu Vatte: Excel is an analysis tool, not a database. [28:18] Christopher Bartlett: Giving non-technical staff the vocabulary to articulate operational challenges in data-centric terms significantly accelerates the data science team's ability to scope and execute solutions. [28:18] Frank Kartmann: Enterprise data requires active lifecycle management, including data catalogs, dedicated data owners, clear quality metrics, defined release cycles, and long-term development roadmaps. [30:56] Kelly Parker: Prototypes that previously required months of manual software development can now be iterated into production-ready software in a fraction of the time. [37:20] Christopher Bartlett: If a model is 'fit for purpose'—meaning it achieves human-level accuracy and its remaining errors are obvious and easily spotted—it is 'worth its weight in gold,' and constantly updating it offers little marginal ROI.

[0:04] Thank you. Thank you. So, a little bit [0:07] about myself. I've been at banking a [0:09] little over 30 years. Uh, manage data in [0:12] 35 countries. Um, the only real [0:14] differences are language and legal. Um, [0:17] the bits and bites are about the same. [0:19] Um, so given that, you know, I've been [0:21] in Standard Charter, Bank of America, [0:24] now more recently, Wilson Bank of Trust. [0:26] Um, to where we are working on some of [0:27] the same problems that you're probably [0:28] working through. So given that I'll give [0:31] everybody an opportunity to introduce [0:32] themselves. [0:36] >> Okay. Thanks Eric. I'm Ragu. I am the [0:39] field CTO for AI squared. Uh we at AI [0:42] Squared help our customers on their AI [0:45] journey bringing AI into of their [0:48] enterprise workflows. Now these [0:50] enterprise workflows can be legacy [0:52] workflows or like new age workflows. But [0:55] most of them start with the legacy [0:56] workflows and then they start to like [0:58] create more and more insights that [1:00] becomes their mo from the data that they [1:03] have. Okay. And uh that is the reason I [1:06] think like today's uh discussion is [1:08] going to be very relevant. Hopefully [1:10] like we'll have some questions towards [1:12] the end. Looking forward to it. [1:14] >> Hello everybody. Hope you're having a [1:15] great time in Las Vegas. Is it hot [1:17] enough for you? [1:19] Have they asked that? Did they ask that [1:20] in the keynote? I was wasn't in there, [1:22] but figured the heat is the story, [1:24] right? Um I'm Kelly Parker. I am from [1:27] Fiverr. Fiverr, if if you don't know [1:29] Fiverr, we're all about moving data. I [1:31] wore my data movement shirt. Um we help [1:34] organizations get data for AI and [1:37] analytics projects into data warehouses [1:39] and data lakes. And we do it at very [1:42] large scale with from any source into [1:45] any great analytics or AI destination. [1:47] Um I'm really looking forward to talking [1:49] to you today. [1:51] Yeah, my name is Frank Katman. Uh I am [1:55] working for Deutsche Bank. Um the [1:58] biggest uh German bank. Um we are [2:00] located in in Franfor um in Germany. [2:05] There I am uh working with a team um [2:08] just building up the the data platform [2:12] um just for collecting and managing data [2:15] of the different departments making it [2:17] available for the business um users um [2:22] yeah to elaborate the own use cases. Um [2:25] before uh this position I was working [2:29] also for other bigger GM banks including [2:33] also the the central bank uh building um [2:37] the big data platforms. [2:39] Yeah, that's it. [2:42] My name is Chris Bartlett. I am the [2:44] associate chief data sciences officer [2:45] for Nationwide Children's Hospital. I'm [2:47] in the healthcare industry. That's my [2:48] enterprise hat. I'm also a faculty at [2:51] the Ohio State University. I'm professor [2:52] of pediatrics and so I have the [2:54] innovation hat. We I do a lot of work on [2:56] vision models and foundational frontier [2:58] models in in that space. Uh but here to [3:00] I'm here today to talk mostly about the [3:02] governance and the enterprise hat hat [3:04] that I wear. Uh as a as a faculty at [3:06] Ohio State, I do have to say that I [3:07] brought some Ohio State Buckeye candies [3:09] with me. I passed them out to the panel [3:11] earlier and if you ask a question this [3:13] panel, you can come up here and grab [3:14] yourself a Buckeye. So So please do so. [3:17] >> No comment. [3:19] >> All right. [3:20] >> All right. All right, let's start off [3:21] with uh the first question. So, how are [3:23] you or your customers successfully [3:25] transforming raw data into real business [3:27] or organizational impact in no [3:29] particular order because some of you are [3:30] helping and some of you are using. So, [3:33] who would like to start with that? [3:36] >> I can go. [3:36] >> Go ahead please. [3:37] >> So, it [3:40] it depends on um the customer. So what [3:43] we are seeing is that like everyone has [3:46] a different uh requirement and the [3:50] different goal and more importantly a [3:52] different timeline in terms of when they [3:54] want to achieve these goals. So when we [3:57] look at uh our federal customers uh [4:00] versus our very large banks versus [4:03] mid-size banks or financial [4:05] organizations or going down one tier and [4:08] looking at uh um smaller organizations, [4:11] smaller companies, what have we seen is [4:14] that like everyone has come to realize [4:18] that their real mo is the data that they [4:20] own. Okay, it is about the data that [4:22] they own and they want to be able to [4:24] like first of all get the most value out [4:27] of the data but also making sure that [4:31] they have something unique that their [4:34] competitors cannot uh provide or they [4:36] have something unique that their [4:38] customers are going to look towards. And [4:42] what we are seeing is that when you go [4:46] on that journey there is everything from [4:50] the quality of the data to the [4:52] consumption layer where your insights [4:54] from the data are shown to the your [4:56] customers right like your end users and [4:59] when you go on that journey you have to [5:01] like uh they are depending on the size [5:04] of the company their risk profile risk [5:07] management all of those things also [5:08] change. Some of them are very very [5:10] aggressive. What we find is like uh the [5:12] smaller the organization they are very [5:14] aggressive and they are thinking about [5:16] like security they're thinking about [5:18] data governance like what we were [5:19] talking about earlier uh they're [5:21] thinking about it as an after thought [5:24] and uh let me first get something and [5:26] then I'll think about it and as we grow [5:28] go up the chain and federal being like [5:31] the other extreme they are like okay I [5:34] need to solve all of these issues first [5:36] before I can get to the outcome and what [5:39] we are seeing is when we it it's both [5:42] sides are the two ends of the pendulum, [5:45] two ends of the spectrum and uh the it [5:49] it is very hard to convince somebody who [5:51] is very very conservative to like okay [5:53] take like very aggressive actions at the [5:55] same time like it's also very hard to [5:57] convince somebody who wants a answer [5:59] tomorrow to tell them that like okay you [6:03] have to slow down this is not going this [6:05] is going to hurt you in the long term [6:06] you have to do five different things [6:07] before you get to that outcome right So [6:10] when balancing this and like giving them [6:14] enough tools so that they don't paint [6:16] themselves into a corner while [6:18] traversing this is the challenge uh and [6:21] I see that like whoever is able to like [6:23] provide a good solution uh for that is [6:26] going to be able to like help these [6:28] customers or these solutions or these [6:31] scenarios come to life. [6:34] >> Yeah. So you're building the foundation [6:36] before you actually can start to [6:37] monetize the data. you know obviously a [6:39] very important step. Anybody else want [6:41] to add anything to that or [6:44] >> I could emphasize in a financial [6:46] institution the the customer is always [6:50] in in focus. Yeah. Everything what is [6:53] developed is just um with the target of [6:56] getting customer satisfaction. Of course [6:58] they are private companies and want to [7:00] to earn money but uh through um yeah [7:05] offering through customer satisfaction [7:06] offering the best products and and [7:10] services and we we are seeing in the [7:13] last years um technology is evolving [7:16] very quickly. Um we are we get new um [7:21] methods of collecting data of processing [7:23] data and uh also new means of [7:27] interacting with the customer and uh the [7:30] banks are just utilizing these new [7:32] technologies to approach the customer [7:34] the customer and offer always better [7:38] services um to yeah [7:43] to to get the customer um yeah satisfied [7:46] to get [7:48] um to achieve customer satisfaction and [7:50] the use cases that are developed uh yes [7:55] utilize these new technologies. Um [7:59] for instance um two examples developing [8:03] chatbots that have access to in real [8:07] time to customer um [8:11] tested private customer data when [8:13] interacting with a with a customer. um [8:17] is more freely because that traditional [8:23] um forms of dealing with customers in [8:26] branches because it's always available [8:29] 24/7 and not only open uh during branch [8:33] opening hours and also through realtime [8:37] analysis of customer data. the bank is [8:41] able to to provide a more better [8:45] tailored a more user friendly product [8:49] that um that it was before just with [8:52] traditional segmentation of the market. [8:55] That could be an example. And another [8:57] one I think very quickly is um [9:02] um antifinancial crimes for instance [9:04] just analyzing in in real time data [9:07] monitoring the market for um to avoid um [9:13] fraudulent transactions. So we are [9:16] utilizing technology just to um [9:21] to enhance um [9:25] customer experience [9:27] um better decision making um better [9:31] market risk and um [9:35] um an operational uh efficiency. So [9:39] these are two examples of we how we are [9:43] utilizing new technology to bet to get [9:46] better results and um yeah from data. [9:50] >> Great. Thank you for that. Uh next [9:52] question was uh what's the biggest [9:54] challenge you're facing right now in [9:56] extracting or operation operationalizing [9:58] data value? You want to start? [10:01] >> Yeah. So, um, one of the things that we [10:04] see all the time as it as it relates to [10:06] this is that there's a long backlog in [10:12] IT and data engineering to get to the [10:14] data. It's not that it's impossible. [10:17] It's not that it's even super [10:19] challenging. The problem is that it's [10:22] locked away into a database and the [10:24] database is constantly changing. you're [10:26] getting new data in there and to get [10:28] that data into whatever medallion [10:31] architecture or or the method that's [10:34] going to be used for getting access to [10:36] that data and doing something [10:38] operationally with it and [10:41] overcoming that challenge of speed and [10:45] uh creating inertia within the data team [10:48] to to get that data that they need. [10:51] >> Anybody want to add anything to that or [10:54] >> I can go for it. So double plus good on [10:57] that comment. I think getting to the [10:58] data is hugely important. Um but I want [11:00] to throw something out because this is [11:01] the right room to hear it. Uh one thing [11:04] that hinders adoption of AI in in [11:06] getting value is something that some of [11:08] the service companies are doing to [11:10] enterprise people such as myself. So we [11:12] get access to the data. Big hurdle. We [11:14] finally got it. We validate a process. [11:16] Great. I'm using an API. I'm using a [11:18] model that's fixed. It's API. Then in [11:20] six months they say, "Well, we're not [11:21] going to allow you to keep using that [11:22] API anymore. we should use a better [11:24] model. We all know AI is getting better [11:26] over time, but for any niche purpose, it [11:28] might be getting worse. And this is [11:29] exactly what we've seen. Giving a [11:32] continued access to a model that has [11:34] been validated for the next three years [11:36] saves me a lot of money. If you force me [11:38] to revalidate every six months or a [11:40] year, you might be eating it in my [11:42] profits, in which case it's no longer a [11:43] useful project. And so, we just wouldn't [11:45] adopt AI for that. So, I think that's [11:47] something that I find a challenge. It's [11:49] not the biggest challenge admittedly, [11:50] but it's something that, you know, folks [11:52] in this room, somebody has influence [11:53] over this. This is something enterprise [11:56] needs and it isn't really being [11:58] addressed fully at the moment. [12:00] >> Can I add something to that please? [12:01] >> Okay. uh before I u add something I did [12:04] not seed that question like I did not [12:06] ask him to say the things but uh what I [12:08] mean is like this is a real important [12:11] problem to solve but for the viability [12:14] and the return of investment of [12:16] especially in the enterprise space and [12:18] uh I don't like to like uh pitch uh my [12:21] own company when we are talking having [12:23] this panel but that is exactly the [12:25] reason why we started AI squared because [12:28] you cannot go on one hand and say like [12:31] okay give me a model that is static for [12:33] the next 3 years. In this age of like AI [12:35] innovation, it is going to change it. [12:37] You want it to change because otherwise [12:39] the next model is going to behave better [12:41] than this one or the next version is [12:43] going to behave. Now, how do you have [12:44] your cake and eat it too? Okay. So, [12:47] those are the kind of problems that uh [12:49] are being solved right now and like I do [12:51] think like there is a good chance that [12:53] like your wish is going to come to in a [12:56] near future too and we are not the only [12:58] ones working on it. There are a lot of [12:59] other companies who are solving this [13:01] problem because for me there are three [13:04] big hurdles uh that uh are for that are [13:08] in the way of adapting the data. The [13:11] first one is about data governance and [13:12] like I think like there are people who [13:14] are more qualified than me to talk about [13:16] data governance on this panel. I'll look [13:18] to you. The second one is about having [13:21] that continuous stream of data and being [13:24] able to like get the newer data not the [13:26] stale data because in AI is not only [13:30] about training the model but it is also [13:32] about inferencing right at the end of [13:34] the day you're doing all of that so that [13:35] you can get an inference and the [13:38] insights out of that and especially when [13:40] we're talking about enterprises you're [13:42] also talking about the new data that is [13:44] generated on a continuous basis you want [13:46] to use that data as a context text to [13:49] for your insights, right? So what that [13:53] means is that like if you don't have [13:54] that data pipeline, if you don't have [13:56] that same level of uh scrutiny and [13:59] discipline in those areas, then your [14:02] insights are not good enough or they are [14:04] only as good as the data and the data [14:06] quality that you provide. Right? So that [14:08] is one and the last one I would want to [14:10] say is like it's about how do you [14:12] surface those insights too because [14:15] somebody who is living in the data world [14:17] is very very can look at it I mean for [14:20] example I know like you live in the [14:21] quality and data world right you look at [14:23] I go and tell you like hey I have [14:25] implemented my stack like this you're [14:26] like oh this is great this is or this is [14:28] not good but somebody who is not living [14:30] in the uh data world and who just cares [14:33] about like how do I see those insights [14:36] is it relevant to Can I give feedback in [14:38] a right way? Those are the kind of [14:40] things that like are not there yet and [14:43] people look at chat GP it has captured [14:46] their imagination and they're going to [14:48] say hey if chhatgity can do it why can't [14:50] I do it right like okay so those are the [14:52] kind of hurdles that we need to cross [14:54] but I think that that is that is you [14:56] brought up a really good point that is a [14:58] real big hurdle right now but thankfully [15:00] there are like some real good companies [15:02] working uh to solve those problems. [15:06] Great. Thank you. Um, real quick, and I [15:09] guess this is the bigger question that [15:11] many of us have gone through. It's when [15:15] did you see something that went wrong? [15:18] How did you identify it? And then how [15:20] did you uh fix it? And then move on, [15:23] move forward. And I'm sure all of you [15:25] have had a bad experience of some sort. [15:27] So, uh, somebody want to try a bad [15:30] experience? [15:32] I've had many myself, but [15:36] Maybe I could start in just a [15:38] >> real simplistic one. Um, and the bad [15:41] experience may be just me that had a bad [15:44] experience because of a um what I'll [15:49] perceive as a lack of data governance [15:52] and it was around revenue. Uh, so I'm [15:55] I'm in sales and uh and revenue is is of [15:58] course one of the key things that we're [16:00] focused on. Um but as uh a SAS-based [16:05] company and a usage-based company, [16:07] revenue takes on a lot of different [16:09] forms in the way that it could be [16:10] calculated based on monthly revenue, [16:13] annual revenue, contract value, usage, [16:18] um uh outstanding usage and um having [16:23] done some early analysis on some of our [16:27] data and um I won't go through the the [16:31] exact finding, but um it was um learning [16:35] that there were at least 12 different [16:38] definitions of of revenue within our [16:41] company, which one was the appropriate [16:44] one, and then modifying not only the [16:48] analytics that we were putting together, [16:50] the uh outcome that we were going for, [16:54] and then um ultimately our mindset shift [16:57] and attitude on finding that were all [16:59] kinds of key things that I I found out [17:02] that that I had to experience. Um what [17:05] did I learn from it? As it relates to [17:08] governance and data, go to the source as [17:12] quickly as you can and when you find [17:15] that you've made a mistake, take action [17:17] as quick as you can. [17:20] >> Yeah, that sounds like a fun one. Um the [17:22] one I experienced was a little bit more [17:25] extreme because it ended up in the uh [17:27] stock analysts with a public company. Um [17:30] and it wasn't an error that we made but [17:32] we inherited and uh yeah hindsight's one [17:35] you know uh 2020. So uh so looking ahead [17:39] um where do you see the biggest value um [17:42] from unlocking uh data and from from [17:46] your sector in particular um as well as [17:48] potentially even others that you might [17:50] be looking at. [17:52] You want to start [17:54] >> Frank go ahead. [17:55] >> Yes. um in our case or in the financial [17:57] industry, I see the biggest um [18:01] opportunities just in um leveraging the [18:05] data analytics um technologies [18:08] and uh um [18:11] strategic data management just to yeah [18:16] to enhance um customer experience. Um I [18:20] mentioned it before um [18:24] uh decision making, risk management, uh [18:26] operational efficiency [18:30] and um [18:32] going to the yeah to the technology [18:34] itself. Um the opportunities are in yeah [18:38] leveraging um scalable [18:42] um [18:44] cloud native applications um for [18:47] instance that can um react quickly [18:50] flexible that uh um [18:54] yeah use containerized technologies [18:59] Kubernetes [19:01] um that can react flexible according to [19:05] market needs can scale quickly and also [19:10] um facilitate [19:13] um integration across departments and [19:16] that is um a challenge uh that I see I [19:20] didn't mention before but just breaking [19:22] the silos in the in the in the business [19:24] department than appear because um you [19:28] want to um to use data for um building [19:32] your model but you need quality data And [19:35] um yeah this this is a a measure and [19:39] other um uh point could be just um uh [19:45] utilizing re uh real time data analyzing [19:47] real time data um [19:51] just to to give the the the best um [19:57] offerings to the to the customer. um um [20:01] in real time to introduce to utilize AI [20:07] when um um in the in the in the in the [20:11] processes of the bank uh augmented [20:14] analytics um predictive analytics just [20:18] utilizing this new technologies in the [20:20] bank. This is uh um the opportunities [20:23] that that we have and also um creating [20:27] new new products that uh that are um [20:31] yeah specific for for the needs of the [20:34] of the customer just reacting in real [20:37] time to the yeah [20:40] to the request of the customer. Um banks [20:43] are even implementing [20:45] um yeah bots and and doing some kind of [20:50] emotional and behavioral analysis and uh [20:53] this is amazing not in my case but I am [20:57] reading it from the industry and uh [21:00] depending on how the customer reacts and [21:03] uh what yeah specific needs are are in [21:07] this any [21:10] day then he gets [21:12] Yeah, a specific product offer just to [21:16] to enhance customer satisfaction at the [21:18] end. Other points I [21:22] quickly um could be just um omni channel [21:27] consistency because new technologies [21:30] open new like new doors to um to to get [21:34] in contact uh with the customer to [21:38] implement um yeah the the the [21:44] yeah communication with the customer and [21:46] it's important that he gets always the [21:49] same um the same communication through [21:54] um [21:56] um mobile apps in the branch through [21:58] website. So omni channel consistency and [22:02] at the end of course um security and [22:04] data governance is very important just [22:06] to enhance uh trust from customers. [22:10] >> What about from the uh healthcare [22:12] perspective? Do you see anything? The [22:14] opportunity seems vast there obviously. [22:16] >> Oh yeah. uh hospitals are big bundles of [22:19] fixed costs and we're very reliant on [22:21] humans for for things. So there's a lot [22:23] of opportunities here and I think in the [22:25] healthcare space everybody recognizes [22:26] that opportunity which is great. So [22:28] there's a lot of general support for [22:30] something. Anything specific of course [22:31] needs to be discussed thoroughly. [22:33] Um, one thing I find that, you know, my [22:36] my computer science world that I I live [22:38] in, we love to talk about let's replace [22:41] radiologists and let's do something big. [22:43] And that's great as a motivator, but ROI [22:46] doesn't come from something like that [22:47] necessarily. If you can do something [22:49] about as well as a human and faster, [22:51] that's ROI. You don't have to get rid of [22:54] a radiologist. Just make them work [22:55] faster and suddenly you've got return on [22:58] investment. So I think from the health [22:59] care perspective, the talk in the tech [23:02] world really isn't as boots on the [23:05] ground as it as it ought to be in the [23:06] health care space because there's so [23:08] much we can do just to improve human [23:09] efficiency that's going to get us return [23:12] uh on investment and it doesn't [23:13] necessarily mean replace radiologists, [23:16] replace physicians, replace nurses, that [23:17] kind of thing. Hey, I'm looking forward [23:19] to the day when, you know, there's lots [23:20] of robots actually doing wound care and [23:22] and all that kind of stuff, but in the [23:24] short term, you know, we can do so much [23:26] with what's already there. Uh, and to [23:28] me, that's that's where we like to put [23:30] our focus on is let's get the RAI now so [23:32] my CEO is my friend in six years rather [23:35] than try to do something big and fail at [23:37] it. [23:38] >> Sure. Great. Uh, so moving on. Uh, if [23:41] you could change one thing about how [23:43] organizations approach data today, what [23:45] would it be? And you want to start from [23:47] the five trend perspective because the [23:50] ETL process is obviously the beginning. [23:53] >> So let me take a little different [23:54] approach though. Uh there was a recent [23:56] article in CIO magazine where um they [24:00] surveyed executives and they surveyed [24:04] doers on AI and they said um um what's [24:09] the how how efficient is your [24:10] organization using AI? And the [24:12] executives were like we are awesome. [24:14] We're doing it. we have AI mastered. [24:17] We're using it in all parts of our [24:19] business. And the uh the the rest of us [24:23] were like, "No way, dude. We're not [24:25] using it. I'm skeptical that I don't [24:26] know anything about this." And um [24:29] there's a a chasm here between what the [24:33] executive team dreams about. We were [24:35] talking about this the about the [24:37] executives um like why not more? I see [24:39] them down the street doing it. Why are [24:41] we doing it? um versus the people that [24:43] are getting it put into are responsible [24:45] to put it in place or responsible to use [24:47] it or have the fear of their job going [24:49] away and whatever uh realm that that [24:53] might manifest itself into. I think [24:55] there's in from data and and really [24:59] making the most of data and AI as we go [25:02] forward bridging that gap between the [25:05] executive perspective and the rest of [25:08] our perspective and having a vision for [25:12] what this AI revolution like the [25:16] agricultural uh revolution that we went [25:18] throughund and something years ago [25:20] manifests itself for people like us who [25:23] are are out here and and maybe replaced. [25:28] >> John Deere is gonna look very [25:29] differently soon. [25:30] >> Yeah. If I may add one thing, right? [25:33] >> Um so what uh we what I would like to [25:37] see is that like when whenever people [25:39] are talking about data, they are or [25:42] especially people who are uh curating [25:44] the data and everything they're looking [25:46] at how can my next application use this [25:49] data to get the best value out of it. [25:52] And usually it is not the end [25:54] application. There are like two other [25:56] three other tools that come into the [25:58] picture before the end result of that [26:01] data comes in. And over the past couple [26:03] of decades like we have been doing a lot [26:05] of data mining and there is a lot of [26:06] data sitting out there which people do [26:09] not really know what is the best value [26:11] of it. So, and again like to your [26:15] previous question too, the way the trend [26:18] the Uber trend right now is that people [26:21] are not thinking about replacing people [26:24] by AI yet at least. Uh but they're [26:28] definitely thinking about replacing [26:29] tools. They're saying it this this data [26:33] is not going to go into an Excel. I mean [26:35] I'm just picking Excel on but like I [26:37] don't really need an Excel to like [26:40] create a pivot chart for me so that some [26:42] other BI tool can look look at it and [26:44] then give me an insight and if I can ask [26:47] the data like or wherever my data is hey [26:50] give me that insight and it gives me [26:52] that insight then why do I need these [26:54] two tools in the middle right so we see [26:56] that trend quite a bit and it is just [26:59] beginning by the way but uh I do see [27:01] that like a lot of the tools either they [27:04] have to evolve or they will be it's not [27:07] even replaced they will just go out of [27:09] existence because there is no real need [27:11] for those tools anymore because you are [27:13] able to get the data but one thing I [27:15] would like to see for the data is that [27:18] people need to start thinking about data [27:20] as a source that can give them insights [27:23] not as a data as a source that would [27:25] help the first tool in that tool chain [27:28] and when you start thinking about it [27:30] then the way you curate data the way you [27:32] preserve it way way you govern it. All [27:35] of those things become different because [27:38] just like what you were talking about [27:40] yesterday, right? Like data is data but [27:42] now a data in Singapore is completely [27:44] different from data in US and the rules [27:47] that govern those are completely [27:48] different. Just because you use a tool [27:50] in Singapore doesn't mean you need to [27:52] use a different tool. But if you can [27:54] eliminate the need of those tools, then [27:56] all of a sudden the number of hurdles to [27:59] get to that outcome which is universal [28:01] becomes like smaller, right? So that is [28:04] something that is very important. [28:05] >> Yeah. And I'll just add one my own [28:07] personal hot button is that Excel is not [28:10] a database. Just so everybody knows [28:13] >> what [28:16] >> I think you have two more opinions. This [28:17] is a really good question. [28:18] >> Yes. Yes. Um, so I'd like to push things [28:21] down even to a more granular level [28:23] because I love everything that's been [28:24] said so far. For sure. I would like [28:26] everyone in my organization to realize [28:28] they are a data person, right? Because [28:30] right now folks, I mean, how how does a [28:32] frontline worker develop opinions? They [28:34] see what works and they see what doesn't [28:35] work and over time that becomes their [28:36] opinion of things. They're collecting [28:38] data whether they realize it or not. and [28:40] having more folks engaged in thinking [28:42] about, you know, how they're forming [28:44] their opinions and how they're [28:45] collecting data. When you speak to them, [28:47] how can I make your job more efficient, [28:49] if they're more data oriented, they can [28:50] come back and give me a question that I [28:52] can convert more easily into a data [28:54] science operation, right? If if it's too [28:56] abstract and too conceptual, we have to [28:58] spend a lot of time digging. And so I [29:00] think if we got more people in the [29:01] organization thinking in terms of you [29:03] know I'm collecting data this is I'm [29:04] learning about this because of this just [29:06] give them the vocabulary and sort of [29:07] that extra sort of change in viewpoint [29:10] and I think our interactions would be a [29:12] lot more uh specific early on because [29:15] getting finding a question into data [29:17] science terms is is is a really hard [29:19] problem and the folks that we interact [29:21] with have no experience with that. So I [29:23] would like to convert everyone to a data [29:24] person. I think that would just be [29:25] immensely helpful. I also could like to [29:28] to add something. Uh maybe to get the [29:31] most value of the data, the companies [29:32] should uh begin just defining what uh [29:35] they want to um to to reach with with [29:39] data. That's uh beginning with uh um [29:43] specifying products and services they [29:44] want to offer to the customer and [29:47] defining business goals, objectives and [29:49] then a data strategy and then align this [29:53] data strategy with the business goals [29:55] with when just to to build the products [29:57] AI or whatever they want to offer to the [30:00] customer because currently or in some [30:03] cases we have a like a um [30:06] technologycentric [30:08] um mindset. we some companies are just [30:10] investing in technology and collecting [30:12] data just for the sake of having it uh [30:15] without defining business goals. That's [30:17] some point and the other thing um I also [30:21] think is important is just um [30:24] considering data or treating data as a [30:28] continuously managed product. So with [30:31] with a catalog we should know which data [30:34] we have in the company just to use it [30:37] correctly. with with data owners um with [30:42] um quality metrics and uh release cycles [30:46] and um long-term um development targets. [30:52] Um yeah. Okay, great. So, one of the [30:56] biggest decisions that you're going to [30:58] have to make as an owner is to what [31:01] technologies are you going to use? And [31:05] some of that is foundational. Some of [31:07] that is centered around security. Some [31:10] of that is going to be specific to the [31:12] industry that you're in. So, how do you [31:14] go about deciding that um which ones you [31:17] choose? I mean, you could walk through [31:18] the the vendor showcase out here and [31:21] there's many many choices obviously out [31:23] there. Um how do you go through that [31:24] decision-making process and maybe maybe [31:27] start in the health care space and see, [31:29] you know, how do you think about that to [31:30] start? So, it's that's it's an [31:33] interesting question. Um, I was just on [31:36] a call not too long ago with a bunch of [31:37] uh CTO's and cos that we we get together [31:40] and have a have a discussion [31:41] periodically and uh people were asking, [31:44] do you still test all the different AI [31:46] models out there, all the different LLMs [31:48] for your application? And most people [31:50] admitted that they had to stop because [31:52] it's happening too fast. You can't [31:54] compare them all the time. You just kind [31:55] of have to make a decision and go with [31:57] it. And so I think when it comes to the [31:59] the the tool, I mean, beg, borrow, [32:01] steal, anything that works is what [32:03] works. Um, I think the security folks [32:07] are where in healthcare in the privacy [32:09] folks, that's where the really tough [32:10] conversations happen. And so they say, [32:12] "We like this model, but you know, it [32:14] comes from a source that we're not happy [32:15] with. We need to be super careful. Put [32:17] it in a container and, you know, you'll [32:18] probably never get permission to use it [32:20] in production." Um, that's fine. I'm I'm [32:22] happy to to live in that world. But uh [32:24] for the most part, you know, the [32:26] academic side of our organization is [32:28] very much into if it's published, if [32:30] it's out there, if it's open source, [32:31] even if it's closed source, let's let's [32:33] find a way to use it if it solves a [32:34] problem. [32:37] >> And uh you know, jokingly, we were [32:39] talking earlier about, you know, my [32:40] chief information security officer, I [32:42] have to feed him blood pressure [32:43] medication every once in a while because [32:44] I'm a little worried about, you know, [32:46] him being fearful of what we choose. [32:48] But, uh, anybody want to add to that? [32:50] like the way you approach the tool uh [32:53] decision [32:54] >> the the tool decision uh is uh again [32:59] I'll go back to the same thing about [33:01] your cir right like uh hopefully he gets [33:03] his build preure medication uh it it [33:05] also depends on the people who are [33:07] helping him out okay not everyone has [33:10] somebody who is as experienced as you to [33:12] like actually help them out and also the [33:16] people around them who are like skilled [33:18] enough to like make the right decisions [33:21] at the right time and it is about the [33:22] timing especially when it comes to [33:24] security especially when it gets to [33:26] governance. If you do, if you corrupt [33:28] the data, if you like tain the data, [33:29] then it takes so much more to like bring [33:31] it back to a pristine uh state. But more [33:35] importantly, it is also about like [33:37] giving these people enough tools to [33:40] actually [33:42] do their job. And uh to your point [33:45] earlier, it's very hard to expect [33:47] somebody to form a very good question [33:50] like almost like a data scientist so [33:52] that they can get the right insights, [33:54] right? like and your insights that you [33:57] are going to get are only going to be as [33:59] uh good as uh the question that you are [34:02] going to ask. Okay. And the next trend [34:05] uh which is again um I this is an AI [34:07] conference I have to talk about agentic [34:09] uh but uh uh the thing is like the next [34:12] trend that we are seeing which is [34:14] actually I do believe that this is the [34:15] right trend is that why should somebody [34:18] ask a question you have your data okay [34:22] and there are experts that are going to [34:24] like talk about uh how you can [34:28] categorize the data how you can like [34:29] actually like create curate the data but [34:32] then it is also about the models and the [34:34] models that are using that instead of [34:37] you going and asking tell me something [34:40] what if the model is able to go and to [34:42] the data look at it and say hey I see a [34:45] discrepancy I see an anomaly I see [34:46] something that like you need to look at [34:49] um especially in the security field [34:50] because it is there there is so much of [34:53] it that like you need a huge uh team to [34:57] like even keep up with what is happening [34:58] today and like not everyone can afford [35:00] it. So what you need is the data that is [35:04] able to tell you like what it is rather [35:07] than you asking that question to that [35:09] and I see that trend happening and it is [35:11] actually I mean it's still not there but [35:13] we are going in the right path towards [35:15] that and it's not far off too like I'm [35:17] pretty sure in a year or two if we are [35:19] having the same discussion we are not [35:21] going to be talking about like hey what [35:22] is the question I need to ask what is [35:25] it's like okay how how is that better [35:27] than the other the tooling is going to [35:29] elevate to a different level where now [35:31] you your data is telling you stuff and [35:34] now your job is to like make sure that [35:36] the data that is telling you is the [35:37] correct data is the curated data is the [35:40] one that you trust uh and at that point [35:43] this tooling again I go back to the same [35:45] tooling that like some of the will just [35:48] go extinct uh for a good reason uh but [35:50] the new tools that are going to are [35:52] coming up are things that are going to [35:55] like tell you like what your data wants [35:56] to tell you rather than the other way [35:58] around. One one last thing that we're [36:00] we're seeing um is with AI and and kind [36:05] of thinking through vibe coding, will [36:07] there be as many business tools that are [36:11] available versus built? I kind of think [36:14] back to like the you were talking about [36:16] um Hadoop and big data and uh comm built [36:19] on commodity hardware and open- source [36:21] technologies and DIY. And if you were a [36:25] vendor in that space during that time, [36:27] you're going like, "Wow, we've we've [36:29] already gone through this in the data [36:31] warehouse revolution. Now we're going [36:32] through it again in the big data [36:34] revolution." Um, and I speak from [36:36] experience because we had a business [36:39] tool that we were using. It didn't meet [36:41] all of our needs. Like, well, let's [36:42] let's put together a prototype of what [36:44] we think that we would want so then we [36:47] know what we're looking for. That took [36:49] months to code. But then version two [36:53] through vibe coding when we had our [36:55] requirements was done in like 40 hours. [36:58] So there I think there is going to be [37:01] this this um decision point of build [37:04] verse buy when you can build through [37:06] vibe coding just so easily. You don't [37:09] have to have really any expertise except [37:11] in the application itself. [37:14] >> Great. So, we've exhausted the um [37:16] prepared questions um and so now there's [37:20] opportunity for you to ask the [37:21] questions. I think there was candy [37:22] involved um as well if you so choose to. [37:26] But uh do you have a microphone ready? [37:29] >> Yes, please. [37:32] And you can you direct the question to [37:34] who think besides myself? [37:42] This this question is I think probably [37:44] for Christopher and and Frank in [37:46] particular. Um [37:48] you talked a little bit about the [37:51] difficulty when um when you have new [37:53] versions of a model and having to then [37:57] revalidate those models. Um but in some [38:00] contexts it's very important to have [38:03] models that are either continuously [38:06] trained or um you know their their their [38:10] life cycle is really short for some [38:12] reason. Um and so in the context of a [38:15] highly regulated entity that is required [38:18] to validate the models that they're that [38:21] they're using or that their vendors are [38:22] using, what are the things that can be [38:25] done to uh enhance, speed up, automate [38:29] the process of a highly regulated entity [38:33] to be comfortable with these models that [38:35] may change more often. So just to give a [38:39] little bit more detail about some of [38:40] those things, right? You could, for [38:42] example, let's go beyond LLMs and let's [38:44] just talk about other models that may be [38:46] returning just true faults on a um uh on [38:50] categorizing fraud data or or or a [38:53] malicious request of some kind, right? [38:55] Um you could set things like precision, [38:58] recall, and whatnot in order to be able [39:00] to say, "Hey, these are our thresholds, [39:02] right?" Um, those are just some examples [39:05] of sort of what I'm curious about, but [39:07] but really this is intended to be [39:09] open-ended about what what you find [39:11] would be the most helpful or if it's [39:13] just not really an option. [39:18] >> Do you want to start or I can start? [39:20] >> Maybe you [39:22] I' I'd be interested in hearing the [39:23] contrast in finance. Um, fantastic [39:26] question. in in in healthcare we have a [39:29] long runway to get things in front of [39:32] patients or in front of in front of [39:34] doctors we call providers um that's a [39:37] good thing right so if we we have a [39:38] model we let it run silently in the [39:40] background for say six months so we can [39:42] collect enough data to make sure that [39:44] it's really functioning the way we want [39:45] it to we might have one or two [39:47] physicians who have access to that model [39:49] in those six months to make sure that [39:51] they're agreeing that it it seems to be [39:53] doing what it's supposed to be doing and [39:55] then we can make it go live in front of [39:56] addition physicians or the entire [39:57] hospital, something like that. That that [40:00] piece really shouldn't go away in [40:02] healthcare. It's this is part of [40:03] regulations, part of legal liability and [40:05] and just you common sense as to what you [40:07] want to put in front of a patient. So, I [40:09] think that's not going anywhere. But for [40:11] me, if I've got a six-month ramp to get [40:14] something new out there, then I have [40:16] validation that happens prior to that. [40:19] If models change every six months, it's [40:21] completely out of sync with what I need [40:22] to to deploy something routinely. And so [40:25] I agree with you. I think continuous [40:27] deployment, LLMs are a special case. Um, [40:30] but any model that I can fine-tune and [40:32] has a half-life that's longer, it helps [40:34] me be, it helps me get a more ROI [40:36] because good enough is, right? Fit for [40:39] purpose is what quality is, right? And [40:41] so once you hit that, it's about as good [40:43] as a human and it makes obvious mistakes [40:46] that are easy to to pick. That right [40:48] there is just worth its weight and gold. [40:50] and the need to mess with that. You can [40:52] say the model's better, but is it better [40:54] for what I need? No. And so I I think I [40:57] think that's that's kind of a mindset. [40:58] Now, I agree with what we're hearing [41:00] this panel. Other areas of the world [41:02] don't think the way we do because they [41:04] they have a little bit more freedom, but [41:05] I I've got a couple runways that I just [41:08] I can't make them shorter and uh to try [41:10] to make them shorter. I'm not sure [41:11] that's in everyone's best interest in [41:13] the healthcare space. And so we we tend [41:15] to to shy away from that. I think there [41:17] are niche places where you can get [41:18] around that a little bit more, [41:20] especially if it's a similar model. [41:22] There's already been a lot of trust [41:23] built up. I mean, there's probably ways [41:24] to adjust that a little bit, but it's [41:26] not going to go to zero. And so, I think [41:28] continuous improvement might not be [41:29] where we want to sit for certain [41:31] applications in terms of decision [41:33] support. In other places, like you talk [41:36] about robotic surgery or something like [41:37] that, I can see a lot more continuous [41:39] integration being completely viable and [41:41] and and completely helpful. But in other [41:44] places, I just I'm not quite sure it [41:46] works well with the sort of where we're [41:48] at from a regulatory perspective, but [41:50] open to new ideas and it sounds like AI [41:52] squared's got some some thoughts on [41:53] this, too. So, yeah, I I'd love to talk [41:55] to anybody who can help us with that. [41:56] >> Hey, real quick. Sorry. We have five [41:58] minutes. So, do you mind if we take one [42:00] more and we can actually like meet you [42:02] in the back as well if you want to go [42:04] into more detail on one of them [42:06] >> and we'll have Buckeyes [42:07] >> and he has he has candy. Yes. So, do you [42:10] mind uh if we [42:11] >> Yeah. very very quickly answer in the [42:12] financial industry you mentioned it um [42:14] is a very highly regulated industry and [42:17] because of this still existing [42:20] insecurities uh when using LLMs in my [42:25] specific case uh we are developing them [42:27] but still not offering these products to [42:31] the broad public outside they are being [42:33] like tested internally and um yeah we [42:37] are monitoring behavior and Um, of [42:41] course, there's a um an an um [42:46] institution for Yeah. Um, [42:49] yeah. It's regulated. Yeah. And um after [42:53] some time, if uh everything goes well, [42:56] we'll begin uh offering this these tools [42:59] to Yeah. to the public. But at the [43:01] moment, they are only used internally. [43:04] >> One last one. Go ahead. [43:05] >> Oh, yeah. Um, I know we dunked on Excel [43:07] a little bit, but sometimes I find, [43:09] especially for execs, the Excel is the [43:11] best database I can give them. It's the [43:13] one that they get the most value out of [43:14] sometimes. Speaking of which, uh, are [43:16] you guys anybody on the panel, can you [43:18] give us some or one AI tool that you [43:21] your your data teams found to be the [43:24] best on like a cost to ROI ratio that [43:27] you just integrated as a strategy or a [43:29] tool that really gave you the the true [43:32] value for your data? [43:36] an example of an AI tool that I mean I [43:40] can't I don't want to like pick one or [43:43] but there are a lot of uh things that [43:46] are becoming really really efficient [43:49] especially in terms of meeting minutes [43:52] in terms of taking notes in terms of [43:54] giving you action items and like going [43:56] back I mean this is something that like [43:59] you needed assistance you needed [44:00] somebody notetakers and they have to be [44:02] good at something right now because of [44:05] the NLP and [44:07] innovations in NLP and improvements in [44:09] NLP I think like that is something that [44:12] is going to stick that has already been [44:14] accepted throughout there are multiple [44:16] of them I don't want to pick one versus [44:18] another that comes to mind immediately [44:21] um and the other things are regular chat [44:25] bots not not like okay what is the [44:28] weather kind of a thing right like okay [44:30] so there are chatbot integrations that [44:32] can be done and they are gaining really [44:34] really good traction where like an exec [44:36] or like anyone in the company can go in [44:38] and ask hey what are the here is my [44:42] revenue or give me the P&L for the next [44:45] uh 3 months or like give me the P&L for [44:48] the past uh one year or they can just go [44:51] give me a trend of like what my [44:53] inventory is like or tell me what my [44:56] sales uh will look like in the next 6 [44:59] months and let me know if my inventory [45:01] is on par with that. So those are the [45:03] kind of questions now you can really ask [45:06] and these are going to be like looking [45:08] at your internal data. They are going to [45:09] go and like look at your competitor's [45:12] public data that is available and make a [45:15] real good decision out of it or an [45:18] insight and give you that insight. the [45:19] decision is still with people at this [45:22] point, right? So, but those are the kind [45:25] of things that those are the two things [45:26] that come to my mind where like there is [45:28] a lot of interest whenever people see [45:31] that whenever they see because they see [45:34] value immediately. That same question [45:36] used to take like a week to answer. They [45:38] needed like three different people to be [45:40] there and those three people have to be [45:42] on their staff to do that. They need to [45:44] trust them. All of those things are [45:45] going away and those are the kind of [45:47] integrations and those are the kind of [45:49] tools that are gaining a lot of traction [45:51] right now. All right, I think we're at [45:54] time. Uh thank you everyone and again if [45:56] you have more questions we will be happy [45:58] to uh meet you in the back. Uh but [46:01] thanks again.