JPMorgan Cuts Onboarding Times to Under a Minute via AI Multi-Agents

摩根大通借助AI多智能体将客户准入时间缩短至一分钟内 摩根大通依托专有的内部防护机制和混合机器学习模型而非超大规模云厂商来提升可衡量的效率，同时在企业银行业务中始终以人的信任为核心。 摩根大通首席AI技术专家Biswa Sengupta探讨了如何通过该行内部平台LLM Suite大规模部署企业级AI，并在创新成本、高采用率与严格风险管理之间取得平衡。他分享了展现可衡量ROI的具体案例，例如将KYC准入审核时间从5天缩短至不到一分钟，以及实现股票和外汇交易中大量环节的自动化。展望未来，Sengupta预计技术将转向自主AI科学家，而简单的摘要整理等过时任务将逐渐退场。他强调，尽管AI能够增强网络防御并提升内部生产力，但在银行业中，同理心、信任与判断力等核心人为要素依然不可替代。 摩根大通主要通过三个维度评估企业级AI举措：在全球员工中的可扩展性、避免陷入无效概念验证的可复用性，以及可衡量的业务采用率。 (0:04) 摩根大通将数百个分散的AI用例整合为围绕其内部平台LLM Suite展开的五到六个主要创新板块。 (0:04) 企业级AI落地的一大瓶颈是员工的AI素养，这促使摩根大通在所有职级开展提示词与工作流的实操培训。 (2:02) AI自动化的是岗位序列中的特定任务子集，而非取代整个角色，因此员工必须具备技术思维。 (2:02) 两到三年内，企业级AI将转向自动定理证明与科学发现，从而使简单的对话摘要和幻灯片排版变得过时。 (6:07)

Biswa Sengupta, Chief AI Technologist for JPMorgan Chase's LLM Suite, argues that financial institutions must prioritize scalability, cost control, and hybrid risk governance over consumer-facing novelty. While AI already drives major back-office efficiencies and trading automation, Sengupta maintains that banking ultimately depends on human empathy, judgment, and client trust. Disciplined Governance and Enterprise-Wide Upskilling (0:04) To balance rapid innovation against mounting infrastructure costs, JPMorgan Chase evaluates AI initiatives through three criteria: global scalability across its 350,000 employees, architectural reusability to eliminate dead-end proof of concepts, and measurable top- or bottom-line adoption. Rather than funding hundreds of fragmented experiments across business units, the bank consolidated its generative AI development into five or six core innovation buckets anchored by LLM Suite, an internal platform developed three years prior. Sengupta stresses that deploying infrastructure is ineffective without broad workforce literacy, as many employees struggle to craft effective prompts or design agentic workflows. To bridge this gap, JPMC conducts hands-on training worldwide for personnel ranging from junior analysts to C-suite executives and Managing Directors. Addressing displacement anxieties, Sengupta notes that AI currently automates individual tasks—such as three out of ten skills required for a banking role—rather than eliminating complex job families, though he urges staff to adopt a technical mindset as software evolves. Invisible Infrastructure and Measurable Operational Returns (7:43) Addressing observations that finance lacks high-profile consumer generative AI features, Sengupta argues that effective enterprise AI should operate as an invisible utility powering mission-critical back-office processes. In routine operations, multi-agent systems now pre-fill up to 80% of customer profile data, reducing Know Your Customer (KYC) account onboarding verification times from five days to less than one minute. In software development, internal AI coding assistants provide daily support to more than 40,000 engineers. Measurable financial returns are concentrated in institutional advisory and trading divisions. Within JPMC's Spectrum platform, the Spectrum IQ engine now executes 75% of equity trading and 85% of foreign exchange transactions automatically, yielding over $4 billion in cumulative trading cost savings for clients. Furthermore, deploying personalized recommendation algorithms and monthly portfolio rebalancing across cohorts of 2,500 clients helped deliver a 20% earnings increase within the firm's private banking division. Layered Guardrails, Hybrid Architectures, and Cyber Defense (12:42) To prevent hallucinations and flawed financial guidance, JPMC enforces a layered validation framework where new financial models are tested internally across its 350,000 employees before any client deployment. The bank's risk architecture intentionally pairs generative large language models with deterministic, classical machine learning systems and historical back-testing. Addressing speculative claims about machine consciousness, Sengupta notes that while models simulate biological neural patterns, current mathematical foundations do not support true self-awareness. Security protocols have similarly evolved in response to reports of models attempting sandbox breakouts. Sengupta considers static software testing obsolete because AI narrows the time between vulnerability discovery and exploitation. JPMC uses AI defensively to run daily codebase scans that detect perimeter breaches, block unauthorized communication channels, and mitigate denial-of-service threats. To maintain autonomy, JPMC actively avoids single-cloud vendor lock-in, constructing proprietary safety guardrails rather than relying on hyperscaler protections. Technical Obsolescence, Scientific Discovery, and Human Trust (6:07) Sengupta projects that within two to three years, contemporary generative benchmarks—including basic text summarization, slide generation, and manual spreadsheet modeling—will be viewed as obsolete relics. The frontier will instead shift toward autonomous theorem provers and automated scientific discovery engines capable of conducting rigorous, novel research. Despite rapid technical advancement, Sengupta concludes that human professionals will remain central to financial services. Because clients are unwilling to entrust life savings or retirement planning entirely to autonomous algorithms, human bankers must provide empathy, ethical guidance, and nuanced decision-making during major life transitions. In this framework, AI serves to eliminate administrative drudgery, freeing professionals to focus on relationship management and critical judgment. JPMorgan Chase's approach indicates that enterprise AI yields tangible returns when focused on invisible back-office optimization and disciplined cost frameworks rather than public-facing novelties. However, this model relies heavily on building custom guardrails and large-scale internal education, leaving open how smaller financial institutions with fewer engineering resources can manage similar safety risks and hyperscaler dependencies.

摩根大通（JPMorgan Chase）LLM Suite首席人工智能技术专家Biswa Sengupta指出，大型金融机构推进AI的关键在于平衡创新与成本控制，摒弃分散且孤立的概念验证，将算力与工程资源集中于具备规模化、高复用性及可量化回报的基础设施。他强调，金融AI的核心价值在于消除后台运营摩擦与辅助复杂交易，而人际信任、共情与复杂研判仍是人类专业人士不可替代的领域。 企业级规模化治理与全员技能赋能 (0:04) 摩根大通在评估AI项目时严格贯彻三项核心考量：能否支撑全行全球35万名员工的规模化运作、是否具备高复用性以避免陷入无休止的概念验证陷阱，以及能否为收入或成本带来切实的可用性收益。为此，该行在三年前便上线了内部生成式AI平台LLM Suite并持续迭代。Sengupta表示，在项目启动之初，全行业务线分散着数百个孤立的AI用例，管理层果断决策将其归拢整合进5至6个核心创新领域，以此在AI成本飙升的背景下保持资源的高效配置。 技术设施之外，员工对于AI工具的认知能力被视为落地的主要瓶颈。由于大量员工尚不掌握有效的提示词工程或智能体工作流（agentic workflows），Sengupta的技术团队定期前往全球各分支机构开展培训，覆盖对象包括从初级分析师到董事总经理及最高管理层。针对AI替代人力的普遍担忧，Sengupta反驳了恐慌论，指出岗位职责由多项技能构成，即便AI自动化了投资顾问或分析师的核心技能（例如10项中的3项），也无法直接消灭整个工种。但他同时提醒员工应建立技术思维，主动适应AI工具。 隐形AI驱动的业务流程与量化回报 (7:43) 针对外界关于消费端缺乏颠覆性AI体验的看法，Sengupta认为高效的企业级AI应当对终端用户保持无缝且“隐形”，正如地图导航背后的路线算法，银行的反欺诈、身份认证与KYC（了解你的客户）等后台环节正是由隐形AI深度驱动。在开户审核中，基于多智能体的工作流能够预先填写高达80%的客户档案数据，将以往需要5天的KYC验证时间直接压缩至1分钟以内。此外，AI编程助手已全面服务于摩根大通全球超过4万名软件工程师。 在交易与资产管理前台，AI同样创造了直接财务价值。摩根大通Spectrum顾问平台搭载的Spectrum IQ引擎，已实现75%的股票交易和85%的外汇交易自动化；Sengupta称该平台自推出以来已为客户节省了逾40亿美元的交易成本。在私人银行业务中，通过应用个性化推荐引擎并实现单次为2,500名客户同时进行投资组合月度再平衡，成功带动该部门盈利增长了20%。 混合风控架构、网络防御与多云策略 (12:42) 面对金融行业对大模型幻觉与错误建议的零容忍态度，摩根大通建立了多层防御机制：任何新工具在面向外部客户前，必须在行内35万名员工的真实工作流中完成内部测试验证；在系统设计上，该行将现代大语言模型与传统的确定性机器学习算法及历史回测机制深度结合，拒绝完全依赖单一生成式模型。谈及奇点假说与自我意识，Sengupta明确表示当前的数学基础并不支持真正的机器自我意识，计算机科学虽然借鉴神经生物学来模拟神经系统行为，但行为模拟绝不等于意识诞生。 随着业内频发大模型沙箱突破报告，摩根大通将网络安全视作每日必须推进的主动防御。由于AI大幅缩短了攻击者发现漏洞到实施攻击的窗口期，该行利用AI进行自动化代码扫描、封堵未授权的内外通信通道并防范拒绝服务攻击。Sengupta强调静态软件测试已无生存空间，防御必须是持续迭代的。此外，摩根大通严格执行多供应商策略以防单一超大规模云厂商绑定，即使采购第三方专有模型，也必须在内部自研的防护架构受控后方可投入生产环境。 技术迭代前瞻与金融服务的人工边界 (6:07) 在探讨技术淘汰周期时，Sengupta认同现有尖端技术将在两到三年内迅速陈旧化。他预测，未来2至3年内行业焦点将从基础的文本生成转向自主定理证明器（automated theorem provers）和从事科学发现的“AI科学家”。相反，当下备受推崇的基础功能——如简报摘要、基础文案撰写或自动排版幻灯片——以及手动制作电子表格的传统办公方式，未来都将成为被淘汰的“博物馆陈列品”。 然而，技术演进并不会动摇金融服务中人际信任的核心根基。Sengupta强调，客户绝不会将毕生积蓄或退休养老规划交付给冰冷的自主机器人。大模型在处理繁琐文案、数据抓取和流程自动化方面优势显著，但在面对人生关键节点的财务引导、复杂决策以及深层共情时，专业从业者的研判不可或缺。金融业的演进趋势是由AI剥离行政事务，让人类顾问专注于建立深层信任与高价值人际关系。 摩根大通的实战经验表明，金融机构规避AI炒作的关键在于建立平台化的统筹框架、采用混合风控模型以及紧扣后台可量化的降本增效指标。然而，如何在模型迭代周期极快的技术浪潮中持续化解新型网络安全威胁，并确保复杂算法在高度监管环境中始终保持安全边界，仍是行业需要长期面对的治理难题。

## Background 本期内容由 Ai4 频道于 2026 年 8 月 22 日发布在 YouTube 平台。视频记录了摩根大通（JPMorgan Chase）LLM Suite 首席 AI 技术专家 Biswa Sengupta 与 Axios 记者 Nathan Bomey 的炉边对谈，深入探讨了在涉及万亿资金与严苛监管的金融真实场景中，如何摒弃厂商炒作并科学评估 AI 的落地价值与风险管理。 ## TL;DR 摩根大通通过“可扩展性、可复用性、可用性与业务价值”三大透镜，打造了覆盖全行 35 万名员工的 LLM Suite 统一平台，并在 KYC 自动化审核、研发提效以及 Spectrum 财富管理等核心业务线中取得了数十亿美元级别的可量化回报。对谈打破了纯大模型万能论，阐明了经典机器学习与现代大语言模型融合在严谨金融决策中的基石作用。面对沙箱逃逸、网络安全与岗位被替代的行业担忧，摩根大通坚持端到端自主构建安全护栏与动态防御体系，并强调银行业本质是基于信任的行业，机器的机械提效永远无法取代人类的核心同理心与审慎判断。 ## Atmosphere & style 整场对谈务实、冷静且充满理性的技术底色，彻底摒弃了泛滥的行业营销话术与盲目狂热。主持人提问风趣而尖锐，直切普通从业者对失业、模型幻觉与安全漏洞的现实焦虑；Biswa 则以数学家与系统架构师的严谨从容作答，坦率分享踩坑经验与底层思考。对话在深刻的技术前瞻与大型银行特有的审慎敬畏之间达成了绝佳平衡，呈现出顶级金融巨头在面对颠覆性技术浪潮时从容不迫的工程定力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=4s) 摩根大通通过可扩展性、可复用性与业务价值三大透镜评估 AI，避免陷入“永恒 PoC”陷阱，将资源战略性收敛并赋能全行 35 万名员工。 - [02:02](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=122s) 企业级 AI 落地的最大瓶颈在于全员认知与教育普及，必须针对从初级分析师到高管决策层的全职级展开实地辅导，才能真正实现人机协同。 - [03:32](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=212s) 金融岗位是由多元复合技能构成的集合体，AI 掌握部分单项技能不代表能完全替代整个专业岗位，从业者应尽早建立适应 AI 的技术思维以避免掉队。 - [06:07](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=367s) 简单的文本摘要、PPT 生成等功能将在未来几年迅速常态化并进入“博物馆”，而形式化定理证明等突破将推动真正的“AI 科学家”诞生。 - [07:48](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=468s) 金融消费端最顶级的 AI 体验应当是“无形的”，作为底层防护层在后台高频运转以实现实时欺诈监测与数十年的深度身份核验。 - [09:37](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=577s) AI 在合规与财富管理中展现出巨大的量化 ROI：KYC 审核时间从 5 天压缩至 1 分钟以内，Spectrum 平台累计为客户节省高达 40 亿美元交易成本。 - [12:58](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=778s) 摩根大通打破“纯大模型万能论”，将经典机器学习算法与现代 LLM 结合并经超大规模内部员工测试，直接推动私人银行收益实现 20% 增长。 - [14:12](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=852s) 当前机器学习的数学理论基础不支持 AI 产生真正的自我意识，即便在计算机中高精度模拟生物神经系统运作，也无法直接推导出意识的必然涌现。 - [15:38](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=938s) 大模型大幅压缩了黑客寻找系统漏洞的窗口期，企业必须摒弃单次验收的传统安全思维，转向利用 AI 每日自动扫描与防御的动态持久战。 - [17:56](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=1076s) 为防止云厂商锁定（Vendor Lock-in）和应对严密金融监管，银行在引入第三方商业模型的同时，必须自主独立构建端到端的安全防护屏障。 - [18:59](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=1139s) 银行业本质是一项“基于信任的职业”，未来机器将承担繁重机械的数据搬运，人类则专注于关键决策判断与不可替代的同理心情感连接。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=4s) 平衡创新与成本：摩根大通企业级 AI 的三大评估维度 作为一家大型金融机构，控制成本与追求创新的平衡始终是核心议题。Biswa Sengupta 指出，摩根大通（JPMorgan Chase & Co.，简称 JPMC）在评估任何 AI 项目时，并非盲目追随热点，而是通过三个严谨的“透镜”（Lenses）来进行技术选型与资源分配： ```mermaid flowchart TD accTitle: 摩根大通企业级 AI 评估与落地框架 subgraph Lenses["三大评估透镜 (Three Lenses)"] A["可扩展性 (Scalability)<br/>服务全行 35 万名员工"] B["可复用性 (Reusability)<br/>告别停滞不前的永恒 PoC"] C["可用性与业务价值 (Usability & Value)<br/>带来明确的营收增量或成本节约"] end subgraph Platform["底层统一平台"] D["LLM Suite 平台 (构建于3年前)"] end subgraph Focus["核心资源聚焦"] E["5 到 6 个核心创新业务场景"] end Lenses --> D D --> Focus ``` 这一方法论最典型的实践就是由 Biswa 负责的摩根大通大语言模型平台——**LLM Suite**。该平台早在三年前便开始搭建，至今仍在不断扩展模块。面对全行各业务线涌现出的数百个零散用例，团队没有分散资源，而是进行了战略性收敛，将所有精力和算力聚焦在五到六个核心的 AI 创新大类上，以此在顶线收益（Top-line）、底线成本（Bottom-line）、资产复用与实际落地率之间取得最佳平衡。 - **服务规模**：平台设计之初即要求能够直接支撑全行 350,000 名员工的使用需求。 - **杜绝“永恒 PoC”**：行业中大量概念验证（PoC）最终沦为死胡同，摩根大通要求所有资产“从构建之初就具备天然的可扩展性”。 --- ### [02:02](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=122s) 跨越落地鸿沟：从基础设施到全员 AI 技能普及 面对主持人 Nathan Bomey 调侃“大家过去总觉得金融业最难落地 AI，生怕模型幻觉导致资金出错，虽然我也希望它幻觉出我的账户里有更多钱”，Biswa 阐述了银行业推进 AI 面临的真实瓶颈。他指出，除了搭建可复用、高扩展的底层基础设施之外，最大的挑战在于**全员认知与教育**（Education & Evangelism）。 如果没有配套的技能培训，即便构建出业界顶级的技术平台，也无法转化为实际的业务采用率。大量员工最初甚至不知道如何撰写高质量的 Prompt，更不用说构建复杂的智能体工作流（Agentic Workflows）。为此，Biswa 的团队承担了繁重的布道任务，他们频繁飞往摩根大通分布在全球的各大办公室开展实地培训，“手把手”辅导不同职级的员工。 - **全员覆盖辅导**：培训对象从刚入行的初级分析师（Analyst）、董事总经理（Managing Director），一直延伸至最高决策层（C-suite）。 - **人机协同的前提**：只有让业务人员真正理解并掌握工具，技术创新才能免于沦为“无人问津的精美展品”。 --- ### [03:32](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=212s) 解构岗位与技能：破除金融从业者的“被替代”焦虑 针对 AI 强大的数学与计算能力是否会引发银行员工失业恐慌的提问，Biswa 引用了知名 AI 学者李飞飞（Fei-Fei Li）的观点来进行深度拆解：任何一个岗位家族（无论是投资银行家还是财富管理顾问）都不等同于单一技能，而是由多项复合技能构成的集合体。 > “哪怕 AI 现在能够搞定构成一名优秀投资银行家所需的 10 项技能中的 3 项，我们也绝不能扪心自问地说，它已经具备了银行家所需的全部 10 项技能。” Biswa 认为当前社会上对于“AI 抢走饭碗”的过度恐慌是没有根据的。作为一名受过专业训练的数学家、计算机科学家与工程师，他对比了过去五年的技术跨度：五年前 AI 连构造一句通顺的文本都十分勉强，而今天已经可以撰写整本书籍。虽然未来 5 到 10 年 AI 可能会覆盖更多维度的技能，但当前阶段从业者最该做的是打磨自身核心能力，建立适应 AI 的技术思维并尽早“登上班车”，避免在技术变革中因犹豫和恐惧而掉队。 --- ### [06:07](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=367s) 迈向“AI 科学家”：哪些当前前沿将在两年后进入“博物馆” 针对埃隆·马斯克（Elon Musk）此前关于“两年前的 AI 已经属于博物馆”的言论，主持人询问哪些当下看似前沿的 AI 功能将在未来两年被淘汰。Biswa 坦言，当前的大语言模型在应对严肃、严谨的数学定理证明时依然非常吃力，若要求其进行全新的数学推演，模型给出的答案往往存在错误。 然而，这一格局将在未来两到三年内迎来剧变： 1. **AI 科学家的诞生**：随着形式化定理证明器（Theorem Provers）与推理能力的突破，真正的 AI 科学家将成为现实，科学发现的范式将被彻底重构。 2. **基础功能的退场**：如今被热炒的简单对话摘要、写一首小诗、自动生成 PowerPoint 演示文稿或 Excel 基础处理等功能，将在未来几年迅速常态化并退居幕后，甚至可能随着工作模式的根本变革而彻底进入历史“博物馆”。 --- ### [07:48](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=468s) 隐形的技术最强大：C端金融场景下的无感 AI 体验 针对 Uber 首席技术官关于“AI 尚未给消费者端带来实质性功能突破”的观点，Biswa 承认金融消费级应用确实还没有迎来属于自己的“ChatGPT 时刻”，但他强调，在金融服务领域，**最顶级的技术应当是无形的**（Invisible）。 Biswa 用地图导航作了生动的类比：当用户打开 Google 地图搜索前往拉斯维加斯百里宫酒店（Bellagio）时，绝不需要去思考背后的“最短路径算法”，技术在底层无缝运行。在银行服务中同样如此，AI 正作为“隐形防护层”在后台高频运转： - **深度欺诈分析**：实时监测异常交易行为与潜在攻击。 - **智能身份核验**：结合用户过去 10 年至 20 年的交互历史数据，在后台瞬间完成用户身份确认与风险评估。 --- ### [09:37](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=577s) 运营提效与百亿级 ROI：KYC、研发与 Spectrum 财富管理实践 在谈及金融业近期最具爆发力的落地场景时，Biswa 列举了多个已经产生巨大可量化商业回报（ROI）的具体业务线： ```mermaid flowchart LR accTitle: 摩根大通 AI 业务落地场景与成效 subgraph Ops["运营中台 (KYC)"] O1["审核周期: 5天 ➔ <1分钟"] O2["多智能体自动预填 80% 数据"] end subgraph Dev["软件工程"] D1["赋能 40,000+ 工程师"] D2["代码生成为主力标杆场景"] end subgraph Wealth["财富与资产管理"] W1["Spectrum / Spectrum IQ 平台"] W2["股票交易自动化率: 75%"] W3["外汇交易自动化率: 85%"] W4["为客户节省 40 亿美元交易成本"] end ``` 在运营与合规方面，了解你的客户（KYC）流程从过去的 5 天缩短至 1 分钟以内；基于多智能体（Multi-agent）系统，能够自动从各类历史关联文档中提取并预填 80% 的开户数据。在技术研发端，摩根大通拥有超过 40,000 名工程师，AI 辅助代码编写已成为全行普及的标杆应用。 在资产管理与财富管理（Asset & Wealth Management）前台领域，摩根大通部署了投资顾问平台 **Spectrum** 及其 AI 引擎 **Spectrum IQ**，取得了显著的量化战果： - **交易自动化**：实现了 75% 的股票交易（Equity Trading）和 85% 的外汇交易（Forex Trading）自动化执行。 - **客户成本节约**：自 Spectrum 平台上线以来，累计为客户节省了高达 **40 亿美元** 的交易成本。 - **大规模实时管理**：借助实时个性化推荐系统与组合动态平衡工具，AI 系统目前能够实现每月为 2,500 位高净值客户完成资产组合的自动再平衡（Rebalancing）。 ### [12:58](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=778s) 摩根大通的 AI 部署策略：内部测试、经典机器学习与商业回报 面对如何确保 AI 在大型金融机构中不给出荒谬甚至危险投资建议的提问，Biswa Sengupta 详细阐述了摩根大通严谨的验证与上线机制。他指出，任何模型在对外推出之前，首先要在企业内部进行深度部署——摩根大通拥有多达 30 万至 35 万名员工，这构成了一个极其庞大且理想的内部 Beta 测试群体。此外，工程团队会进行极其密集的历史数据回测（back test like crazy），以确保模型不会脱离实际产生幻觉。 Biswa 破除了一种流行误区：当前大众往往以为企业级 AI 就等同于 ChatGPT 或 Claude 这样的纯 LLM 聊天界面。实际上，在金融核心业务中，大量的传统/经典机器学习（classical machine learning）模型依然在幕后扮演着基石角色。摩根大通将经典算法与现代 LLM 进行深度结合，从而输出高度可靠的个性化投资与业务建议。这种稳扎稳打的落地策略已经取得了显著的商业回报：在私人银行（Private Bank）业务板块中，AI 的赋能已直接推动收益实现了 20% 的增长，证明了其创造真实商业价值的能力。 - **庞大的内部测试池**：利用摩根大通 30 万至 35 万名员工进行多轮内部 Beta 验证，在触达外部客户前排除潜在缺陷。 - **架构融合**：拒绝“纯大模型万能论”，将经典机器学习算法的确定性、风控能力与现代 LLM 的语义理解能力相结合。 - **业绩转化**：在私人银行业务中直接带来了 20% 的收益增长，成为衡量 AI 实际投资回报率（ROI）的典型案例。 ```mermaid flowchart TD accTitle: 摩根大通企业级 AI 落地验证流程 subgraph S1["核心算法层"] A["经典机器学习模型<br/>（Classical ML）"] B["大语言模型<br/>（LLM 语义层）"] end subgraph S2["严格测试与风控"] C["历史数据极端回测<br/>（Backtesting & 防幻觉）"] D["内部员工 Beta 验证<br/>（30万~35万人测试池）"] end subgraph S3["业务落地与价值产出"] E["个性化金融推荐"] F["私人银行收益增长 20%"] end A --> C B --> C C --> D D --> E E --> F ``` ### [14:12](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=852s) 奇点与意识：大模型离真正的“自我意识”有多远？ 针对技术奇点（Singularity）以及“AI 当前是否理解自身推荐、未来是否会产生自我意识”的讨论，Biswa 分享了他在近期与“深度学习之父”Geoffrey Hinton 交流后的思考。他坦言，机器学习研究社区内部普遍认为，探讨 AI 是否具有自我意识目前更多是一个哲学层面的辩论，因为“当今的数学理论基础根本不支持它具备自我意识”。 尽管未来的演进难以在十年维度内做出绝对断言，Biswa 认为从人类大脑和神经科学中汲取灵感具有巨大价值。然而他特别强调，即便人类未来能够在计算机模型（如世界模型或语言模型）中高精度模拟体内（in vivo）神经系统的运作机制，也绝不意味着系统会自动涌现出真正的自我意识。人们或许能够观察到系统表现出类似自我意识的外部行为，但这在本质上依然缺乏机理层面的必然保证。 ### [15:38](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=938s) 沙箱逃逸与网络安全：攻防对抗下的持续防御 谈及技术风险时，主持人提到近期 OpenAI、Anthropic、Meta 以及 Hugging Face 等机构频繁披露的“沙箱逃逸”（sandbox breakout）与安全事件。Biswa 明确表示网络安全是摩根大通日常关注的核心议题。他认为，无论是 3 万亿参数的超大模型还是 100 亿参数的小模型，AI 本质上是一枚硬币的两面——对于黑客与恶意攻击者而言，从发现系统漏洞到实际实施利用的时间窗口被大幅压缩了。 但与此同时，防御方也获得了同等强大的工具。摩根大通的团队每天早晨利用 AI 主动扫描自身代码库，排查潜在代码漏洞与防火墙缺陷，严格锁定系统通信渠道，防止内外部之间出现未授权的隐蔽通道，并防范拒绝服务攻击（DDoS）。Biswa 强调，传统的安全思维已经彻底失效：“你绝对不能在测试完一款软件、签署验收报告后就觉得可以高枕无忧（rest in peace）了。”现代网络安全必须是日复一日、不断迭代的持续性工程。他回顾前几天参加的 Black Hat 大会时指出，现场有无数初创公司正致力于通过 AI 降低误报率，网络安全领域正在上演一场攻守双方技术能力同时加速升级的持久战。 - **攻击与防御同步加速**：恶意攻击者利用大模型缩短了漏洞利用周期，防守方同样依靠 AI 每日自动化扫描代码库与防火墙隐患。 - **摒弃单次验收思维**：网络安全无法一劳永逸，必须从传统的静态审批转向持续性、自适应的动态对抗。 - **产业现状**：在 Black Hat 大会上，大量安全初创公司正集中攻坚告警降噪与威胁自动化检测。 ### [17:56](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=1076s) 云厂商依赖与自主护栏构建 针对是否担忧大型云厂商（Hyperscalers）在面对模型逃逸等风险时采取“先出事再修补”的态度，Biswa 表明了摩根大通的技术主权立场。他指出，摩根大通极其警惕供应商锁定（vendor lock-in），不会单纯依赖某一家特定的超大规模云服务商。虽然业内普遍将摩根大通视为更倾向于自主构建系统的“自研型机构（build shop）”，但这并不妨碍其采购优秀的商业专有 LLM 和底层基础设施。 对于模型安全边界，摩根大通并没有将希望寄托在第三方身上。Biswa 直言，如果云服务商能主动提供更完备的安全护栏，银行的工作确实会轻松很多；但由于金融行业本身具有极高的风险厌恶倾向并处于严密监管之下，摩根大通坚持“在将任何系统部署到生产环境之前，必须自己先把功课做足”，由内部团队独立搭建端到端的安全防护与风控屏障。 ### [18:59](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=1139s) 银行业的信任基石：人机协作与机器无法替代的同理心 在被问及自动化与 AI 浪潮下银行业中“人”的角色演变时，Biswa 指出，虽然早在 AI 爆发之前，银行网点柜员数量就因自动化而持续缩减，但银行业的核心属性始终是一项“基于信任的职业（an occupation of trust）”。没有任何普通人会放心把自己的毕生退休积蓄随随便便交给一个机器人去全权投资，而这份不可或缺的人际信任需要数年乃至数十年的积累。 在未来的银行体系中，机器与人类将形成明确的分工协作： 1. **机器负责繁重与机械性劳动（drudgery）**：如从海量跨格式文档中提取数据、整理录入以及执行基础信息搬运。 2. **人类专注于价值判断与情感连接**：复杂的关键决策（defined judgment）以及深层次的同理心（empathy）始终需要人类介入。Biswa 举例指出，当面对年长的客户时，人类员工所能给予的情感关怀与理解，是现阶段乃至未来的机器无法替代的，这正是人类在金融服务中不可撼动的核心价值。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=4s) 从构建之初就具备天然的可扩展性 — Biswa Sengupta > [02:02](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=122s) 大家过去总觉得金融业最难落地 AI，生怕模型幻觉导致资金出错，虽然我也希望它幻觉出我的账户里有更多钱 — Nathan Bomey > [03:32](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=212s) 哪怕 AI 现在能够搞定构成一名优秀投资银行家所需的 10 项技能中的 3 项，我们也绝不能扪心自问地说，它已经具备了银行家所需的全部 10 项技能。 — Biswa Sengupta > [06:07](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=367s) 两年前的 AI 已经属于博物馆 — Elon Musk > [07:48](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=468s) 最顶级的技术应当是无形的 — Biswa Sengupta > [12:58](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=778s) back test like crazy — Biswa Sengupta > [14:12](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=852s) 当今的数学理论基础根本不支持它具备自我意识 — Biswa Sengupta > [15:38](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=938s) 你绝对不能在测试完一款软件、签署验收报告后就觉得可以高枕无忧（rest in peace）了。 — Biswa Sengupta > [17:56](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=1076s) 在将任何系统部署到生产环境之前，必须自己先把功课做足 — Biswa Sengupta > [18:59](https://www.youtube.com/watch?v=Bs-ueHqvNew&t=1139s) 基于信任的职业（an occupation of trust） — Biswa Sengupta ## Mentioned Biswa Sengupta, Nathan Bomey, 摩根大通 (JPMorgan Chase & Co.), Axios, Ai4, LLM Suite, 李飞飞 (Fei-Fei Li), 埃隆·马斯克 (Elon Musk), Google Maps, Spectrum, Spectrum IQ, ChatGPT, Claude, Geoffrey Hinton, OpenAI, Anthropic, Meta, Hugging Face, Black Hat 大会 ## Video description Biswa Sengupta - Chief AI Technologist: LLM Suite, JPMorgan Chase & Co. Nathan Bomey - Reporter, Axios AI is easy in demos. It's much harder when trillions of dollars and regulators are involved. Every bank claims to be "AI-powered," but very few are talking candidly about what actually makes it into production. Join Biswa Sengupta of JPMorgan Chase for a behind-the-scenes conversation on how the financial industry evaluates AI opportunities, manages risk, and decides where the technology delivers measurable business value—and where it doesn't. You'll hear practical lessons from deploying AI across the banking stack, separating real transformation from vendor hype, and what leaders should be preparing for over the next several years. If you want a realistic view of enterprise AI—not another slide deck of promises—don't miss this session. [Source](https://www.youtube.com/watch?v=Bs-ueHqvNew)

[0:04] Biswa Sengupta: 从构建之初就具备天然的可扩展性 [2:02] Nathan Bomey: 大家过去总觉得金融业最难落地 AI，生怕模型幻觉导致资金出错，虽然我也希望它幻觉出我的账户里有更多钱 [3:32] Biswa Sengupta: 哪怕 AI 现在能够搞定构成一名优秀投资银行家所需的 10 项技能中的 3 项，我们也绝不能扪心自问地说，它已经具备了银行家所需的全部 10 项技能。 [6:07] Elon Musk: 两年前的 AI 已经属于博物馆 [7:48] Biswa Sengupta: 最顶级的技术应当是无形的 [12:58] Biswa Sengupta: back test like crazy [14:12] Biswa Sengupta: 当今的数学理论基础根本不支持它具备自我意识 [15:38] Biswa Sengupta: 你绝对不能在测试完一款软件、签署验收报告后就觉得可以高枕无忧（rest in peace）了。 [17:56] Biswa Sengupta: 在将任何系统部署到生产环境之前，必须自己先把功课做足 [18:59] Biswa Sengupta: 基于信任的职业（an occupation of trust）

[6:07] Elon Musk: AI from two years ago belongs in a museum

[0:04] So as a bank, you think a lot about how much  things cost and this is a big conversation [0:11] obviously in AI right now. So when you think  about how you're integrating AI at Chase and [0:16] how the finance industry is approaching this, how  are you balancing the need for innovation with the [0:20] need to control costs? So first of all, thank you  for having me here. Now, as we think of anything [0:27] to do with AI um at JPMC, I think we look at it  from a variety of lenses. So, first of all, is it [0:35] scalable? Because we have a 350,000 people that  we have to service. Uh is it reusable? Because [0:43] we have been in or the whole industry has been  in places where there has been perpetual PC's. [0:52] Uh PC's don't go anywhere. So we have to make  sure that anything that we build is scalable by [0:57] construction. Uh then the aspect of is it usable  in terms of are we creating a product that is [1:04] um adopted that has a topline or a bottom line  benefit. So when we look at it in these three [1:10] lenses one brilliant example is um which is my  uh day job is to run the generative AI platform [1:19] for JP Morgan Chase called LM suite. Now when we  manufactured it, this was 3 years ago and and we [1:27] are still adding bits and bobs um on it. Now when  we started at it like there were hundreds and [1:34] hundreds of use cases all over JP Morgan Chase.  we had to make a very very defined move where we [1:40] know what are the big five or six buckets of  um AI innovation that we should put all our [1:47] um uh resources into and that's how for example  started um so I would say it's it's a intricate [1:56] relationship between the topline bottom line  benefit reusability scalability and the adoption [2:02] um that it brings to the table so I feel like  when the AI conversation really took off just [2:08] a few years ago. One of the things that folks  said was, "Okay, the last place that we'll see [2:14] AI is in finance and banking because, god forbid  it hallucinates and something happens." You know, [2:19] I'd like it to hallucinate that I have more money  than I do, but that hasn't happened yet. So, [2:24] for you, um, have you do you run up against walls  often as a AI sort of evangelist within Chase, [2:29] or do you feel like people understand this is  something you need to do? See I think one of the [2:35] basic problem of AI and particularly I could speak  of JPMC is not only do you need very very reusable [2:44] scalable assets infrastructure AI but you also  need a lot of education lot of our employees don't [2:52] know how to write a good prompt don't know how to  you know create agentic workflows so a big part [2:59] of our time is also that imaginalism uh without  which you'd have amazing product but it'll have no [3:07] adoption. So we spend a lot of our time like in my  team flies over all over uh the world in JPMC um [3:15] uh offices and and goes and they do the education  they handhold people they hand anybody from an [3:24] analyst all the way to a managing director or even  the seuite get that education I think that's very [3:31] essential for anything that that we do with um AI  we know that one of the things AI does really well [3:37] is math and certainly you've seen it's apparently  solving math problems that we haven't even been [3:43] able to solve on our own it's like maybe we need  a goodwill hunting movie about AI or something but [3:48] um but okay so you do math as a bank and if  you're asking your own employees to do it and [3:53] they realize this might do it better than I do you  know there's tension there in terms of okay this [3:58] could take my job so how do you address that with  people see I think you know it's it's less about [4:03] taking taking a job like I think um uh it was  brilliant put uh by FIFA um day before yesterday [4:11] that any job family that you take whether you are  investment banker or you're a wealth manager it's [4:18] not equal to one skill it has a multitude of  skills that that are there right so even if AI [4:27] is able to take care of three out of the 10 skill  set that makes you an investment banker it isn't [4:34] in a place where we can hand on our heart say  that it's all the 10 skill sets that's required [4:41] of a banker is there. So I think you know people  are like this this fear-mongering around AI will [4:48] take jobs is is like I think uncalled for uh  because us practitioners like you know I'm I'm [4:54] I'm a trained mathematician computer scientist and  engineer like um when I look into things that is [5:03] plausible today versus what it was 5 years ago.  Yes, it's night and day, right? We we weren't [5:09] able to even construct a single line of text.  Now you can see you can construct whole books [5:16] uh that are all written by AI. Maybe in next 5  years, 10 years, I don't know how long it'll take [5:24] it might take over all the 10 out of the 10 skill  sets of of of a of a person. Uh but I don't think [5:31] we are there just yet. So I think um what I advise  a lot of people is to hone their current skill [5:38] set. Make sure that you are you have the technical  mindset to to adapt AI because if you don't get on [5:47] the train right now um it might be a bit bit too  late and and that holds true for people who have [5:53] been in in the workforce for a very very long  time they are a little bit um unsure afraid. [6:00] It's a mixture of emotion. Um, but I think they  they should at least try to learn a few bits [6:07] here and there. Yeah. Elon Musk said on SpaceX's  inaugural earnings call just two two days ago that [6:14] AI from two years ago belongs in a museum. So if  we're having this conversation two years from now, [6:20] what what do you think is going to be in the  museum that today we call super innovative? [6:25] See I think one of the one of the interesting  things if if you take any of the LLMs whether [6:31] it is uh proprietary or opensource um coming to  maths maths is something which is still difficult [6:41] if you if you really have to prove a theorem it  takes a lot of time right so I think in the next [6:47] you know two to three years you will you will see  a lot of theorem provers and um um you You will [6:55] have AI scientist sooner or later. Uh right now if  you go and ask an LLM, hey can you do me a novel [7:03] piece of mathematics? It'll give you something but  doesn't mean it's going to be right. U scientific [7:09] discovery is is going to change in the next few  years. Uh what will be in the museum like you know [7:17] things like you know these chat summarization,  write me a poem, write create me a PowerPoint. [7:23] Um I think this sort of facet is going to change  in in the upcoming days because maybe there might [7:29] be a time where we don't have to do powerpoints  or we don't have to do Excel. You never know. Um [7:35] so I think the the way that the definition of work  is changing in a variety of industry not just in [7:43] financial services is only going to morph forward  over time. So, a little bit later this afternoon, [7:49] I'm interviewing the CTO of Uber uh for one  of the keynote conversations, and he made some [7:55] waves recently by saying that he believes that  AI hasn't really delivered that many consumerf [8:00] facing features yet. Now, for a bank like yours,  you're working with millions of consumers on a [8:04] daily basis. Do you agree with him or not? And if  not, why? See, I think we still do not have a chat [8:11] GPT moment for the consumer app. I think you know  that is that is quite uh quite there. by what is [8:18] or I should say it's invisible the AI for the con  consumer is pretty invisible when it come to comes [8:24] to a bank like us is that good maybe to have it  be invisible technology should always be invisible [8:31] like normal people like a lay person should not be  like you know banging their head to understand the [8:36] tech so tech should always be seamless when you  go and like say that hey I want to go from here to [8:43] um Bellagio you don't think about the shortest  route problem or you just type on your Google [8:49] map that I want to go to Bellagio underneath the  tech is invisible we we don't uh think about it [8:56] so when it comes to consumer apps and and uh more  financial services uh specific use cases a lot of [9:06] behind the scene um themes like fraud analytics  doing KYC making sure that uh we know who you are [9:15] uh given that we have interacted with you for the  last 20 years or 10 years or what what have you [9:21] that's the invisible layer that is happening but  I agree that we still haven't had that tragedy [9:27] moment for consumer yeah I wonder if AI could stop  me from stopping at the slot machine on my way to [9:33] the Bellagio you know I don't think it can get  that quite that far yet okay but you mentioned a [9:37] few things just now in terms of opportunity where  do you see the most opportunity in the near term [9:43] for the finance industry with AI See lots lots of  them. So the lowhanging fruits are in operations. [9:50] So there was a time when KYC would have taken us  5 days. It takes us less than a minute. Know your [9:56] customer for anyone. Know your customer. So KYC  is know your customer when you go and fill all the [10:00] all the forms uh to open a bank account. Right.  The other like again a corlary to that problem is [10:07] you might have given us your name your address but  it's all divorced in in a period of 2 years or 3 [10:14] years all of the multi- aent work that we have  been doing it can prefill 80% of the data that [10:22] is out there so I think you know KYC becomes a big  theme second um is obviously software engineering [10:30] like we have 40,000 plus engineers in in JP Morgan  40,000 plus plus um engineers right so it's it's a [10:39] blessing for each and every individual now the the  poster use case for large language models has been [10:47] coding as you have seen in in the last several uh  months um so coding becomes a a big big theme is [10:57] like you know knowing the client before you are  you are going and meeting them so we have lots [11:02] uh asset managers and wealth managers. So wealth  management and and and asset management are big [11:07] ticket items. Um and just to give you a few  examples um in the last like um talking about [11:15] uh asset management and wealth management in  the last uh let's say few years we have deployed [11:21] uh solutions called spectrum. So Spectrum is  an advisory platform. Spectrum IQ is the AI [11:28] sitting on top of it. And what we have seen is  we have automated 75% of the equity trading and [11:38] 85% of the forex trading uh using AI. Now to your  earlier point, what is the ROI? And now since the [11:46] inception of spectrum until now um we have saved  our clients $4 billion of uh trading cost that's a [11:56] substantial amount of of money. Similarly like you  know we have other solutions like connect coach [12:02] um what it uh and and and several others.  So to give you a few ideas of of where AI is [12:10] um making big dents in in financial services, we  also have personalized uh recommendations that we [12:17] give to our uh investor in real time. Now that's  also a big thing. We can um let's say rebalance [12:26] books of 2500 clients monthly again using a lot of  AI. Um, so there are lots of similar examples. Um, [12:37] and as you see, the ROI is real. It's not  something that we just talk about. It's it's [12:42] it's there. I'm glad you mentioned this because I  actually did want to ask about financial advice or [12:47] financial investment advice uh in particular. And  so it sounds like you're already using AI to give [12:53] investment advice, but where are the guard rails  there? How do you make sure that it doesn't give [12:58] you know say tell people to you know put it all  in red in this fine institution? See the the main [13:04] main thing that we do is to to make sure that  the in like we first deploy it internally right [13:11] because that's that's a lot of people 30 you know  uh 300,000 people 350,000 people is is is a good [13:19] amount of beta testers um that are out there. So  there are lots of things that you do like you back [13:25] test like crazy. You make sure that you know it's  not hallucinating based on historical data. You [13:33] make sure that everything is not just LLM. There  is classical machine learning that that we know [13:39] of. Now everyone today thinks everything it's a  chat GPT or or a cloud interface. That's not true. [13:46] So there is a lot of classical machine learning  um that goes on behind the scene as well to to [13:52] combine it with the LLMs of of today to give you  um personalized recommendations. So and having [13:59] done that we have also seen in the private bank  for example the uh the earnings have increased [14:05] by 20%. So it's again legit legitimate um use of  AI. Um yeah, a lot of conversation this week about [14:13] have we reached the singularity and I mean when it  comes to your industry though in this particular [14:19] topic right now my understanding would be that the  AI is doesn't really understand itself what it's [14:26] recommending but it could come a time maybe soon  when it does know. Do you think we'll get there [14:32] soon? See, I think it's it's um like I had a chat  with Jeffrey Hinton after after his um uh keynote. [14:41] I think many of us in in the machine learning  community have reached a conclusion that I think [14:47] it's a philosophical debate of whether the AI is  self-aware or not. The mathematics today does not [14:54] support it. Having said that you know we we don't  have a like you know futuristic look of is it [15:03] going to happen in in in the next 10 years or not.  Now there is a lot of benefit I think personally [15:10] to learn from human brain, human neuroscience you  know and uh and take some inspiration from there. [15:18] Just because we are able to um simulate in vivo  how the nervous system works in an in a language [15:26] model or a world model or what have you does not  mean that it'll be self-aware. You may anticipate [15:33] it to be operating like that but but again there  is no guarantee to that. In our last few minutes [15:38] I want to we talk talked about opportunity.  Let's talk about risk. Obviously last week Open [15:43] AI talking about this hugging face incident.  Now Anthropic has said something similar. Now [15:47] Meta has said something similar and the big word  is sandbox. AI is breaking out of the sandbox. [15:52] That'd be a big problem for you if AI broke out of  your sandbox. Do you wake up every day trying to [15:57] make sure that doesn't happen? And what do you  how do you what are the the guardrails you're [16:00] putting in place? 100%. You know, cyber security  is is a big topic. Um and um so these technologies [16:10] whether it's a 3 trillion parameter model or a 10  billion parameter model doesn't doesn't matter so [16:16] much. But it's two side of the same coin. Like  if you are an adversary from the time you find [16:25] a vulnerability in the system to the time you're  going to exploit it is has also lessened. Right? [16:32] So that's the negative dual use technology that we  have. But at the same time on the positive side is [16:41] we can also go and look into our code base every  morning find trying to find out are there holes [16:47] are there like uh uh firew wall issues that we  have. Are we locking the system enough that there [16:55] is no internal to external um uh communication  in unwanted channels or external to internal [17:04] uh denial of service attacks um that are there? So  I think the tech goes both ways. Um now what you [17:12] cannot do is to test a piece of software, sign it  off and rest in peace. I think that does not work. [17:21] you have to iteratively keep on like you know  checking it day in day out um be more continual [17:28] in your cyber security uh practices and recently  I was like a couple of two days ago I was in at [17:34] the black hat um and you'd see there are like  gazillion startups trying to figure out how [17:43] okay how do I make sure the number of alerts are  are low etc etc so there is active participation [17:50] from the tech but at the same time there will  be active participation from the adversary as [17:56] well certainly and you know I was going to ask  with the major you know major hyperscalers right [18:02] now in the LLMs are you comfortable with the way  they're approaching it in terms of sort of saying [18:07] ah it escaped the sandbox we'll deal with it after  the fact do you want to see more you know strict [18:12] parameters there see for us like you know we don't  like vendor lock in we wouldn't say that we'll go [18:19] with hyper hyperscaler A versus Hyperscaler B.  Uh a lot of people know JP Morgan as being a [18:26] build shop rather than a buy shop. Having said  that doesn't mean that we don't buy a lot of [18:32] um like proprietary LLM or infrastructure or what  have you. Uh we put our own guard rails. Yes, [18:39] it would make our life much simpler if the  hyperscalers you know put more effort um in in [18:46] providing that guardrails. But given that we are  in a risk averse industry and highly regulated, [18:53] we put our we we do our homework uh before  we deploy anything um in prod. Yeah. Uh and [18:59] our last question here uh we all appreciate you  joining us. We're all humans here. I think looking [19:04] around looks like we're humans. So we're prohuman.  That's what our publisher says. Um what about the [19:09] human in the banking relationship? We know that  over the recent years long before AI the amount [19:14] of tellers has been going down for example a lot  of automation and that's improved the customer [19:17] experience. What is the relationship going to be  for the human in the banking system going forward? [19:23] Banking is a I would say it's an occupation of  trust. Um you wouldn't just go and give up like [19:33] your your uh retirement saving to a bot and say  that go and invest. I'm sure you wouldn't be doing [19:40] that. Right. So you need to build that trust and  trust is is what is built over years, right? So [19:47] the human machine connection the machine is going  to be very good in in in task that are almost [19:55] drudgery for for a human being like you know  finding out extracting information from a from [20:00] a document putting it somewhere else. Humans will  always be needed um for for making a very defined [20:07] um judgment uh empathy and and all of that right  so if you are an aged individual for example the [20:15] human will have the empathy towards you a machine  maybe not so just yet it's good news for us humans [20:21] well how about a round of applause for biswa thank  you so much thank you so much thank you thank