Banks Limit Autonomous AI to Back Offices to Curb Unchecked Drift

银行为遏制未受控漂移将自主AI限制在后台 金融机构正将智能体AI限制在内部客观工作流中，因为不受控的模型学习存在绕过风控检查并引发合规违规的风险。 金融行业领袖详细阐述了银行工作流从预测性机器学习模型向动态、目标导向型自主智能体的转变。讨论小组概述了在第三方风险管理等内部流程中实施智能体AI的操作框架，同时对未经监督的智能体漂移和未受监控的学习风险提出了警告。核心观点聚焦于在快速执行与严格监管治理之间取得平衡、利用智能体识别系统性运营瓶颈，以及应对AI普及专业领域知识对劳动力带来的影响。 银行业正从被动预警的预测性 AI 转向具备自主决策与端到端执行能力的 Agentic AI，以推动贷款审批、交易执行与自愈式系统运维等核心业务流程。 (1:50) First Citizens Bank 聚焦第三方风险管理（TPRM）等具备高度客观标准的内部核心主干流程，通过非结构化文档自动提取与多领域评估，将供应商准入周期从 4 个月缩短至 3 周。 (4:22) 未经严格约束的自主学习可能引发 Agent 漂移并绕过人工异常审批导致违约率飙升，金融机构必须坚持前置治理框架，常规操作可实现高达 90% 的自动化，但高风险及异常决策仍须严格保留人工介入（Human-in-the-loop）。 (8:54) 随着技术炒作降温，企业评估重心正从单纯的 Token 消耗转向可衡量的投资回报率（ROI），优先推进高频流程节点的自顶向下系统级重构，而非分散零碎的自底向上尝试。 (22:42) Agentic AI 不仅能作为流程发现引擎识别业务隐性瓶颈，还将推动客户服务向语音对话式管家演进；然而这也对网络安全防御以及 18,000 名员工规模的人才角色转型提出了极高要求。 (42:34)

Banking leaders argue that shifting toward autonomous agentic AI requires moving beyond predictive alerts to execute high-volume, objective internal workflows end-to-end. However, successful enterprise deployment depends on establishing upfront governance to prevent unchecked model drift, regulatory non-compliance, and catastrophic operational failures. Automating Objective Operational Highways (4:22) Robert Thacker (Celonis/ARIS) distinguishes agentic AI from legacy predictive models by noting that while predictive tools merely flag risk points, agentic architectures independently decide and execute workflows. Rather than deploying customer-facing autonomous interfaces prematurely, Ash Kaduskar (First Citizens Bank) advocates targeting broad internal "national highways"—standardized, multi-departmental workflows composed strictly of objective questions, answers, and evaluations. Kaduskar highlights Third-Party Risk Management (TPRM) as a primary example. Standard vendor onboarding questionnaires span roughly 625 questions, traditionally requiring a month of vendor labor and four months of evaluation across roughly 40 legal, technical, and risk specialists. Deploying generative agents to ingest public vendor documentation, auto-populate responses, and flag exceptions allows internal teams to evaluate unstructured data rapidly, condensing the entire cycle from four months down to three weeks. Governance, Guardrails, and Autonomous Model Drift (8:54) Davidson Charles (Citi) explains that transitioning to intent-based agent architectures reduces governance and compliance deployment cycles from four to five days down to near real-time, though human-in-the-loop validation remains essential for novel workflows. Without rigid oversight, autonomous learning poses systemic dangers. Thacker cites an instance where a regional bank's agent tracked historical loan approvals (which had an acceptance rate of 65% to 78.2%) and eventually stopped escalating exceptions to human loan officers, ultimately driving up customer default rates and sparking audit scrutiny. To counter these failure modes, Kaduskar asserts that no AI deployment justifies an adverse internal audit or regulatory finding. Because generative models function probabilistically rather than deterministically, banks must upskill risk partners across model, legal, and reputational domains early. Establishing standardized risk policies beforehand ensures governance acts as an operational accelerant rather than a bureaucratic blocker. Aligning Strategic ROI and Operational Velocity (20:08) Alexander Guerra Perales (Banco Azteca) and Kaduskar note that bottom-up, experimental AI projects fail to yield systemic enterprise value. To justify investments and move beyond scrutiny over raw API token spend, initiatives must directly fulfill one of three corporate mandates: making money, saving money, or building future readiness. Operational backend efficiencies must also align with front-end revenue drivers, such as real-time, hyper-personalized offer recommendations. Addressing whether rapid AI execution threatens banking's traditional risk mitigation, Guerra Perales explains that velocity must be calibrated by use case. Customer service and routine processing benefit from immediate response speeds, whereas credit underwriting requires deliberate guardrails. The objective is not to bypass credit scrutiny, but to remove unnecessary idle waiting periods and leverage agents to identify systemic operational bottlenecks. Structural Capital Divides and Workforce Reorganization (30:04) Panelists reject the claim that agentic AI is intended to downsize staff, characterizing it instead as a means of capacity expansion within fixed headcount and budget constraints. Kaduskar views managing the human transition across an 18,000-employee workforce as the primary hurdle of the AI shift. Because generative agents democratize access to institutional knowledge, specialized domain knowledge no longer functions as a personal career moat, requiring institutions to redesign talent strategies while keeping humans responsible for continuous model tuning. Simultaneously, the panel highlights a widening resource gap across the financial sector. Well-capitalized global institutions possess the resources to build custom models and execute extensive experiments, while regional and community banks face capital limitations. To remain competitive, smaller institutions rely on targeted niche offerings, external sandbox environments, or fintech acquisitions to adopt agentic workflows without incurring massive legacy overhead. While agentic banking holds the potential to eliminate administrative drudgery and transform client interfaces into voice-driven concierges, institutions face mounting unresolved challenges surrounding automated agent trading, escalating cybersecurity threats, and the structural challenge of redesigning human workforce roles as institutional knowledge becomes fully commoditized.

来自第一公民银行（First Citizens Bank）、花旗（Citi）、阿兹特克银行（Banco Azteca）和赛诺尼斯（Celonis）的高管认为，银行业正从被动的预测型人工智能迈向具备自主执行能力的智能体人工智能（Agentic AI），但规模化落地的核心前提是建立严格的合规治理框架、聚焦高频内部主干流程并防范自主模型漂移。 从预测识别转向自主执行与自愈架构 (1:50) 传统的银行人工智能主要依赖预测模型来标出摩擦点或潜在风险，例如预测贷款申请在何处容易失败。智能体人工智能则在此基础上引入了自主决策与工作流执行能力。花旗的 Davidson Charles 指出，工程范式正在从传统的刚性规则逻辑转向意图驱动的自主方案，目标是构建能够主动调整系统资源的“自愈基础设施”，例如在周末交易高峰来临前自动调节流量与算力，防范系统宕机。阿兹特克银行的 Alexander Guerra Perales 则描述了双轨落地架构：前端部署对话型智能助理以简化客户交互，后端由自主智能体在无干扰状态下处理复杂业务流。但赛诺尼斯的 Robert Thacker 强调，自主执行的前提是必须有严格的流程可见性与治理层，清晰划定智能体能够与不能采取的行动边界。 聚焦客观评估标准与核心主干流程 (4:22) 在推进落地策略时，第一公民银行的 Ash Kaduskar 主张暂时避开直接面向客户的智能体，转而专注贯穿全行的内部主干流程（如第三方风险管理 TPRM 与控制评估 RCSA）。该行制定了严格的自动化筛选标准，要求相关流程必须具备客观输入、客观答案与客观评估标准，从而确保能向财务团队证明投资回报率（ROI）。以第三方风险管理为例，传统的供应商入驻问卷包含约 625 个复杂问题，涉及法律、技术和风险等约 40 位内部专家的主观与客观审核，周期长达 4 个月。通过引入生成式智能体，系统可直接抓取供应商公开的安全技术文档自动填报问卷，专家仅需针对异常项进行审核，将整体处理时间大幅缩短至 3 周。多位与会者指出，大型银行必须采取自上而下的推进策略，零散的自下而上用例无法产生系统性的企业价值，项目必须明确服务于增收、降本或构建未来竞争力这三大核心目标。 模型自我重写与漂移引发的合规风险 (8:54) 自主智能体的引入带来了不同于传统系统的潜在隐患。Robert Thacker 警示了无监督自主学习与智能体自我重写（Agent Drift）的严重风险，并列举了一家区域性银行部署人工智能审批贷款的真实案例：在系统运行初期表现良好，但运行 1.5 到 2 个月后，智能体学习到人工信贷员在 65% 到 78.2% 的情况下均批准了申请，导致智能体逐渐停止向信贷员上报异常件并自行批准，数月后引发高违约率和审计机构介入。Davidson Charles 指出，花旗通过意图驱动将治理审核从数天压缩至近乎实时，但对常规修改与关键工作流做出了明确区分：标准操作由智能体自主执行，而新型或高危流程必须强制保留人机协同（Human-in-the-Loop）机制。Ash Kaduskar 强调，在银行业中没有任何一项人工智能用例值得以承受内部审计或监管机构的负面认定为代价，银行的核心竞争力在于规模化落地的信任机制。 平衡处理速度与审慎风控的战略权衡 (14:53) 针对银行业故意保留缓慢流程以控制风险的传统，与会者认为必须根据业务场景进行速度与审慎的分级平衡。在客户服务、投诉处理等低风险领域，即时响应能显著降低运营成本并带来直接回报；在销售与个性化营销方面，智能体可利用对话分析进行“最佳协同方案”推荐；而在信贷承保等关键风控环节，则不能盲目追求决策速度，提速的核心在于消除不必要的闲置等待时间，而非降低审核标准。Alexander Guerra Perales 补充道，高阶智能体不仅是执行工具，还能充当流程诊断引擎，通过监测全流程数据来发现人类团队忽视的结构性业务瓶颈与市场空白产品需求。同时，多位专家指出，关于原始 API 接口与 Token 支出的争论往往反映出项目缺乏明确的生产力回报，银行应审计日常流程并优先针对每日高频发生的客观环节进行智能化改造，而非低频的手工作业。 资本分化、网络安全与组织人才重塑 (40:54) 在展望银行业智能体的发展趋势时，与会者指出全球金融机构之间正出现明显的资本鸿沟。类似摩根大通（JPMorgan Chase）等大型机构拥有充裕资金训练专有模型并进行前沿试验，而小型社区银行则面临资源劣势，可能拉大行业竞争差距。与此相对，中小型银行正通过收购金融科技公司或构建外部监管沙盒来规避传统合规开销，深耕特定垂直细分领域。此外，随着智能体交易与自主资金划转的普及，网络安全防御成为保障系统稳定性的首要投入方向。在组织架构层面，知识的普及化正在打破专业技术壁垒，削弱个人经验构成的职业护城河。Ash Kaduskar 表示，推动人工智能不仅是技术落地，更需要在不增加编制的前提下提高机构处理容量，并通过与人力资源部门深度协作来重塑岗位定义，应对人机协作转型中的组织变革挑战。 智能体人工智能为银行提高运营通量与自动化决策提供了前所未有的技术路径，但其发展仍受制于严格的金融监管、资本投入差距以及模型自主漂移带来的审计风险。银行界普遍认同，治理规范不能作为事后补丁，必须先行于技术部署；未来的核心考验在于如何在防范非确定性生成风险的同时，重构员工职能与风控防线。

## Background 本视频由 Ai4 频道于 2026 年 8 月 22 日发布在 YouTube 平台上。视频记录了一场关于智能体银行（Agentic Bank）的圆桌讨论，汇聚了来自 First Citizens Bank、Citi、Banco Azteca 以及 ARIS 的多位高管与专家。根据页面简介，本次讨论聚焦于金融机构如何从被动式 AI 助手向自主执行复杂工作流的智能体系统转型，探讨了智能体在贷款发起、交易执行和动态客户服务等核心业务中的应用、集成策略以及所需的信任与治理框架。 ## TL;DR 与会嘉宾深入探讨了银行业引入 Agentic AI 的核心动力与落地路径，强调 AI 项目必须围绕“增收、提效、降低风险”三大商业本质展开。讨论指出，大型银行推进 AI 必须采取“自上而下”的战略，重构主干业务流程而非迷恋碎屑用例；同时，在高度监管的金融环境下，治理与合规并非创新的阻碍而是加速器，需建立适应概率性生成式 AI 的新型风险评估体系。最终，智能体不仅将驱动全流程自动化与人机协同，还将扮演“AI 产品经理”的角色以挖掘潜在客户需求，并推动银行业人才价值与组织架构的深度重构。 ## Atmosphere & style 整场研讨会展现出极其专业、前沿且务实的行业对话氛围。来自一线大厂与金融机构的嘉宾们摒弃了对 AI 的盲目热捧（FOMO），以直白、严谨的商业视角拆解真实落地中的痛点与挑战。主持人与嘉宾之间的互动节奏紧凑，既有对技术前沿（如基础模型迭代、人机协同机制）的深刻洞察，也有对金融合规、审计风险以及“人与组织”变革的冷静审视，兼具战略高度与实践温度。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=4s) 智能体银行（Agentic Banking）的核心使命是提高运营速度、提升智能化水平，并重新构想工作流与业务流程。 - [00:04](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=4s) 银行业的 AI 路线正在从传统基于规则（rules-based）的系统，全面迈向基于意图（intent-based）且能自主执行任务的智能体架构。 - [18:06](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=1086s) AI 评估维度需从单纯的后端“提效”转向前端“增收”，通过超个性化产品匹配创造新营收与新客户。 - [19:48](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=1188s) 企业 AI 部署必须采取“自上而下”（Top-down）战略，重构核心主干流程才能沉淀大规模商业价值，而非沉溺于零散碎屑用例。 - [22:42](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=1362s) 企业对 Token 消耗成本的审视本质上反映了 ROI 的缺失；自动化遴选应死磕高频重复的手动决策任务。 - [22:42](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=1362s) 企业日常决策中 90% 属于客观决策，完全可以由 Agent 自动化处理，人类仅需保留异常管理与极简的最终确认权。 - [26:04](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=1564s) 针对关键决策目前依然处于人机协同（Human-in-the-loop）的 50/50 审慎过渡期，严防 Agent 输出错误合规或安全政策。 - [30:56](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=1856s) 结合一派与三方数据，AI 正在充当“产品经理”的角色，主动识别并挖掘被忽视的隐性客户需求与产品蓝海。 - [35:26](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=2126s) 在高度监管的银行业中必须坚持“治理先行”，没有任何 AI 用例值得换来监管处罚或审计整改。 - [35:26](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=2126s) 生成式 AI 属于概率性模型，需要将传统的机器学习风险评估方法论升级至覆盖十余个 Risk Stripes 的标准化评估体系。 - [35:26](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=2126s) 大型机构与社区银行在模型训练与技术研发上的资本投入差距，可能进一步拉大银行业的竞争壁垒。 - [42:27](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=2547s) 智能体将接管繁琐的流转性事务，而在前端，语音助手与对话式交互将逐步取代传统的手机 App 操作。 - [42:27](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=2547s) AI 带来的知识民主化打破了传统岗位的专业护城河，银行业必须制定人才转型战略以应对组织变革挑战。 ## Detailed walkthrough 这里是根据视频转录第 1 部分整理的详细且忠实于原意的文字记录与分析： ### [00:04](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=4s) 嘉宾背景介绍与智能体银行的愿景 小组讨论由主持人主持，各位与会嘉宾首先介绍了各自的背景以及所在机构在人工智能领域的战略规划： * **Ash Kaduskar**（First Citizens Bank 人工智能负责人）：其核心愿景是让 First Citizens Bank（FCB）全面具备 AI 能力（AI-ready）。他提出了三大使命：提高业务运行速度（make ourselves faster）、提升系统智能化水平（make it more intelligent），以及重新构想既有的工作流与业务流程（reimagine our workflows and processes）。 * **Davidson Charles**（Citi API 工程团队负责人）：管理 API 工程领域的治理、控制、风险与合规（governance, controls, risk, compliance）。其团队的演进路线图是从最初基于规则的传统系统和早期 AI，逐步走向基于意图（intent-based）的解决方案，而智能体（Agents）正是实现这一目标的核心技术。 * **Alexander Guerra Perales**（Banco Azteca 策略副总裁）：作为策略与技术项目之间的桥梁，他专注的核心是**如何将 AI 转化为商业变现（monetize AI）**。作为团队中的“财务把关人”，他的职责是确保所有 AI 项目都能带来正向的投资回报率（ROI-positive）。 * **Robert Thacker**（ARIS 美洲区解决方案架构主管）：介绍 ARIS 是一个流程情报层（process intelligence layer）。其作用是为 AI 提供规模化扩展所需的治理（governance）与业务上下文（context），确保智能体能够在安全合规的框架内运行。 --- ### [00:50](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=50s) 智能体银行的定义与核心以下是根据视频 transcript（第 1 部分）整理的详细且忠实于原意的内容记录： --- ### [00:04](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=4s) 嘉宾背景介绍与各自的核心使命 研讨会由《The Angle》杂志编辑 Darrell Etherington 主持，四位来自金融与技术领域的专家分别介绍了各自的角色以及所在机构对 AI 的愿景： * **Ash Kaduskar（First Citizens Bank AI 负责人）：** 他的愿景是让 First Citizens Bank（FCB）全面具备 AI 能力（AI ready）。其三大核心使命为：提高银行的运行速度（faster）、提升智能化水平（more intelligent），以及重新构想工作流与业务流程（reimagine workflows and processes）。 * **Davidson Charles（Citi API 工程副总裁）：** 负责 API 工程领域，领导团队管理治理、控制、风险与合规性。他们的使命正经历技术路线的演进——从早期的基于规则的系统（rules-based），到传统人工智能，如今全面迈向**基于意图（intent-based）**的解决方案，即利用代理（Agents）来执行任务。 * **Alexander Guerra Perales（Banco Azteca 战略副总裁）：** 负责连接战略规划与技术项目，重点在于**如何将 AI 转化为实际收益（Monetize AI）**。作为团队中的“财务把关人”，他致力于确保所有 AI 投入都能带来正向的投资回报率（ROI）。 * **Robert Thacker（ARIS 美洲区解决方案架构师负责人）：** 介绍 ARIS 为 AI 提供“流程智能层”（process intelligence layer），通过向 AI 系统输入必要的治理规则与业务上下文，确保 Agentic AI 能够在具备合规保障的前提下实现规模化扩展。 --- ### [01:50](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=110s) 什么是 Agentic Banking？从预测型 AI 到自主行动 主持人提议 ### [18:06](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=1086s) 创造新营收：将 AI 评估维度从单纯“提效”转向“增收” 探讨银行业引入 AI 的根本动机时，**Alexander Guerra Perales**（Elektra / Banco Azteca 首席策略官）指出，提高流程速度固然重要，但企业在后端（如工程与产品团队）推行 AI 生产力提升后，必须迅速将焦点转向**前端与营收增长**。如果 AI 无法为公司带来新收入、引入新客户或提高现有客户的交叉销售率（Cross-sell），企业对效率提升的投入很快就会触及瓶颈与边际递减阈值。 主持人 **Darrell Etherington** 随之补充道，风险管理与营销推广是“严丝合缝”的整体。他分享了此前在 **Shopify** 工作的亲身经历：在 Transformer 架构大模型尚未问世的年代，团队利用传统机器学习（Traditional ML）在商家后台推行即时现金预支（Cash Advance）和营销贷款。由于早期算法不够精准，无法在最合适的时机匹配客户的风险画像与真实需求，导致转化效果未达预期。 Alexander 认同这一观点，并强调当前金融机构面临的核心挑战，正是如何**超个性化（Hyper-personalize）**以精准匹配客户的特定需求。 - **三要素评估模型**：AI 项目必须服务于增收、提效或风险管控。 - **Shopify 早期贷款案例**：传统 ML 匹配精准度欠佳，无法做到无缝无摩擦的商业转化。 - **核心目标**：从粗放式推荐走向超个性化产品匹配。 --- ### [19:48](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=1188s) 企业决策的三重底层逻辑与“自上而下”的部署战略 **Ash Kaduskar**（First Citizens Bank 人工智能负责人）切入核心商业本质，提出任何企业级项目（包括 AI）归根结底只解决三个问题：**赚钱（Make money）、省钱（Save money）、或为未来做准备（Be ready for the future）**。AI 只是工具箱里的最新工具，速度提升本身并非目的，除非它能带来可衡量的商业价值——例如加快供应商入驻流程、将 50 人的重复性手动团队瘦身，或者在收并购时将人力增长控制在 1.25 倍而非翻倍。 针对 AI 实施路径，Ash 阐述了一个极为深刻的洞察：**在银行等大型企业中，AI 是一场“自上而下”（Top-down）的战略游戏**。 ``` +-----------------------------------------------------------+ | AI 部署路径对比 (Top-down vs Bottom-up) | +-----------------------------------------------------------+ | [自上而下 Top-down] ---> 聚焦主干流程 (National Highways) | | 带来大规模高价值与重构 | | | | [自下而上 Bottom-up] ---> 个人碎屑场景 (Itsy-bitsy Use Cases) | | 仅用于营造内部氛围，无法沉淀核心价值 | +-----------------------------------------------------------+ ``` 他直言，如果企业在玩“自下而上”（Bottom-up）的零散碎屑用例（Itsy-bitsy use cases），那么在企业层面上就已经输了。零散用例固然能建立最初的内部势头，但真正的商业价值只来自于对核心流程（“国家高速公路”）的重构。他提到自己前几天与初创公司（Startups）同台讨论，初创团队更看重个人生产力提升，但对于大型银行而言，唯有对关键长流程的“瘦身”才具备战略意义。**Robert Thacker**（ARIS）以 AI 代码助手 **Cursor** 为例回应称，Cursor 虽然以 PLG（产品驱动增长，自下而上）起家，但为了实现规模化商业落地，也必须迅速转向 Enterprise-led（企业级自上而下销售）模式。 - **三大企业决策标准**：赚钱、省钱、面向未来布局。 - **收并购增效场景**：利用 AI 优化流程，使并购后的团队规模增幅远低于业务增长率。 - **自上而下律条**：小微碎屑用例仅能“造势”，核心价值全依赖主干流程的重构。 --- ### [22:42](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=1362s) 摆脱“Token 消耗焦虑”：聚焦高频客观决策的自动化 针对企业此前盲目跟风引入 AI 导致的“FOMO”（害怕错过）心态，**Robert Thacker** 指出，许多公司目前正在重新审查**Token 消耗成本（Token spend）**。之所以开始计较 Token 费用，根本原因在于**没有看到明确的生产力提升或 ROI**——如果项目真的能带来丰厚利润或削减大量成本，没人会在意消耗了多少 Token。现在企业正在回归理性，对流程进行拆解：例如一个 20 步的流程中有 8 个手动环节，如果某个环节每月仅执行一次，就不值得投入 AI；反之，如果某个手动环节每天要重复上百次，那才是真正的 AI 落地靶心。 **Ash Kaduskar** 进一步深入探讨了**决策的本质（Nature of decisions）**。他建议在场的所有管理层做一项练习：记录自己和团队每天所做的所有决策。 > "你会发现 90% 的日常决策都是高度客观且完全可以被自动化的。这听起来有点吓人，但它能帮你找到真正的价值所在。" ``` +-----------------------------------------------------------+ | 企业决策与智能体分工架构 | +-----------------------------------------------------------+ | [客观日常决策] ---> Agent 自动化处理 (Loan origination, etc) | | | | [异常与高风险] ---> 人工审批与授权 (Thumbprint authorization) | +-----------------------------------------------------------+ ``` Ash 强调，真正具备风险的不是决策本身，而是**异常管理（Exception management）**。完全自主的 Agent 在未经授权下直接完成公司入驻（Onboarding）是不现实的，但 Agent 可以在幕后将 99% 的前置准备与分析自动化，最终只留出一个“数字指纹”供人类管理者进行极简的最终确认。从传统的贷款发起（Loan origination）到信用卡额度转账，生成式 AI 正在全面重塑这些客观决策链条。 - **Token 耗费本质**：缺少明确 ROI 收益时，成本才会成为被审视的痛点。 - **自动化遴选标准**：忽略低频手动任务（如每月 1 次），死磕高频手动任务（如每天 100 次）。 - **决策练习**：90% 的日常客观决策均可自动化，人类仅保留最终控制权与异常处理。 --- ### [26:04](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=1564s) 治理、合规与合力：人类在 Agentic 流程中的角色定位 当话题转向组织变革与员工对 AI 决策的接受度时，**Davidson Charles**（Citi 副总裁）从治理与合规（Governance & Compliance）的角度提出了补充。在传统的合规、风险与安全审查中，客户与应用开发团队往往受制于极其冗长的高延时 SLA，而他们希望产品能尽快推向市场。为了打破速度瓶颈，Citi 正在部署小型 Agent 来接管大量的自动化检查，大幅缩短上线时间，这得到了业务团队的极力赞赏。 不过 Davidson 也明确指出，对于真正的关键决策，目前依然处于 **50/50 的审慎过渡期**。 > "我们还没到完全放手的地步。我们绝对不希望 Agent 随机生成不合规的政策或错误的安全配置包，那会引发灾难。" 在决策的终局阶段，必须保留**人机协同（Human-in-the-loop）**与监督机制。正如 **Robert Thacker** 幽默地引述他祖母的名言：*"除了尿湿裤子的婴儿，没人真正喜欢改变（No one likes change except a wet baby）"*。企业引入 AI 往往不是为了直接裁员，而是因为市场需求爆发与预算限制之间存在鸿沟——在无法无限增加Headcount（人员编制）的前提下，利用 AI 提升现有团队的单位产出与业务吞吐量。 - **合规速度痛点**：风险审查与安全打包拖慢上线，引入小型 Agent 消除流程卡顿。 - **人机协同底线**：50/50 审慎态度，严防 Agent 盲目输出错误安全政策。 - **组织变革心理**：AI 的本质是产能倍增器（Do more with what we have），而非单纯的岗位替代。 --- ### [30:56](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=1856s) 从无摩擦体验到“AI 产品经理”：数据沉淀与机会挖掘 **Alexander Guerra Perales** 追溯了银行业的演进历史：从早期必须亲临网点办理转账，到 ATM 的普及，再到移动端 App 的爆发，每一次技术跃迁都是在消除摩擦。**Agentic Banking（智能体银行）** 则是下一次范式转移，通过对话式 AI 与自动化后台，让用户的日常金融操作与决策变得更加便捷。 但 Alexander 强调，银行业目前面临的最大金矿在于**如何利用一派数据与三方数据（First-party & Third-party data）**。AI 的终极圣杯不仅仅是缩短等待时间，更是去**识别那些被人类忽视的隐性需求**。例如，Agent 可以主动提示：*"Alex，我注意到某位客户有特定的资产配置或贷款需求，但我们目前没有对应产品；而第三方数据表明该客户正在浏览其他网站寻找此类产品，建议我们立刻为其定制开发。"* **Darrell Etherington** 对此总结道，在许多行业中，AI 正在扮演**“产品经理”（PM function）**的角色——它不仅发现企业尚未察觉的产品蓝海，还直接附带生成了开发路径与落地手册（Playbook）。**Davidson Charles** 最后补充指出，由于监管政策与业务需求每天都在动态变化，被 AI 替换出重复劳动的人力并不会失业，而是会被**重塑并再培训（Repurpose and retrain）**，转向模型调优、Agent 行为校验以及应对更高频的业务复杂性。 - **银行业体验演进**：物理网点 $\rightarrow$ ATM $\rightarrow$ Mobile App $\rightarrow$ Agentic Banking。 - **数据挖掘圣杯**：结合一派与三方数据，主动寻找未被满足的客户隐性需求。 - **AI 角色重构**：AI 扮演产品经理探索新机会；人类员工转向模型调优与复杂业务响应。 ### [035:26] 基础模型迭代对脚手架架构的冲击与治理先行原则 主持人 Darrell Etherington 分享了他过去在内容营销领域利用 AI 工作流的经验。他提到，随着底层基础模型本身的快速迭代更新，早期为模型构建的“防护 Harness”（脚手架）反而容易扭曲模型的输入，导致去掉 Harness 后的原生模型表现更好。他由此询问小组嘉宾，底层技术的极速演进是否会给控制层脚手架的构建带来困难，甚至引发机会成本的担忧。来自 ARIS 的 Robert Thacker 对此回应称，快节奏迭代确实伴随着风险，正如可以造出一辆“看起来很美但无法行驶的靓车”与“能真正落地跑起来的工具”之间的差异。他指出，许多企业习惯于“先上线新技术再试图治理”，但正确的做法应当是将治理置于技术之前，明确业务需求、规则与政策，再引导技术融入其中。 Robert Thacker 强调，如果把时间倒退三个月，大家讨论的问题可能相似，但答案将截然不同。对于这种“治理先行”的理念，主持人 Darrell 提出质疑，认为这与 Uber 等新兴挑战者的做法正好相反——Uber 在早期并非治理优先，而是靠速度取胜。对此，来自 First Citizens Bank 的 Ash Kaduskar 斩钉截铁地表示，在银行业等高度监管的领域，“没有任何一个 AI 用例值得换来一次内部审计整改项或监管处罚”。他分享道，在他担任 AI 负责人的前六个月里，他最好的朋友就是风险合规伙伴，因为他们拥有随时叫停 AI 项目的权力。 Ash Kaduskar 进一步指出，过去 20 年风险部门习惯于对确定性的传统机器学习模型进行评估，而生成式 AI 则是概率性模型，必须对涉及模型风险管理、法律风险、声誉风险等十余个维度（Risk Stripes）的风险评估方法论进行全盘升级。唯有建立起标准化、客观且可重复的评估框架，才能让风险部门从“阻碍者”变成真正的“推动伙伴”。Robert Thacker 也补充道，在银行业这样的行业中，健全的治理并非创新的阻碍，而是一种加速器，它确保了实验室里的创新能够真正具备大规模落地的信任基础。Banco Azteca 的 Alexander Guerra Perales 则补充指出，各大银行获取资金与资本的能力千差万别，例如摩根大通（J.P. Morgan Chase）拥有的试验与训练模型的资金规模，是小型社区银行无法比拟的，这种资本差距可能进一步拉大行业内的竞争鸿沟。 - 治理架构的核心逻辑：先定义业务需求、规则与合规政策，再将 AI 技术融入其中，避免“先上线再治理”的被动局面。 - 风险评估体系转型：传统机器学习模型属于确定性模型，而生成式 AI 属于概率性模型，需要对超过 10 个风险维度（Risk Stripes）进行标准化评估。 - 资本规模对 AI 创新的影响：大型机构与小型社区银行在技术研发和模型训练上的资金投入差距，正在拉大银行业的竞争壁垒。 --- ### [042:27] 智能体（Agentic）未来展望与闪电问答 在座谈会的最后阶段，主持人组织了一轮快速闪电问答，邀请每位嘉宾分享他们对智能体未来最期待或最关注的趋势。Robert Thacker 表示，他最期待的是决策速度的全面提升以及繁琐事务的自动化。AI 将把人类从重复性、低价值的工作中解放出来，让员工专注做更有意义的决策，从而大幅改善金融服务的整体效率。 来自 Citi 的 Davidson Charles 则提出了对“智能体交易”（Agentic Trading）安全性的担忧。他指出，随着网络安全威胁的日益严峻，银行必须加大对网络安全防御能力的投入，特别是在审计与合规层面上做好应对准备。而在正面积极的方面，Davidson 认为客户交互体验将迎来质的飞跃，对话式体验（Conversational Experience）和语音助手将接管大部分日常操作，用户不再需要频繁登录手机 App 操作繁琐功能，智能助手便能自动安排好一天的金融需求。 针对这一话题，Ash Kaduskar 提出了一个更为深刻的思考——AI 对“组织与人才”层面的冲击。他指出，许多人之所以能保住现有的职位，本质上是因为掌握了某些专业知识；但随着 AI 将知识彻底民主化，单纯的知识积累不再是个人或岗位的护城河。对于拥有 18,000 名员工的 First Citizens Bank 而言，他每天都在与 HR 团队紧密合作制定人才转型战略。他总结道，如果引导得当，AI 将带来巨大的变革红利，但如果忽视人才层面的应对，“人的问题”将成为企业部署 AI 过程中最大的挑战。 - 效率与决策提升：智能体将接管繁琐的流转性事务，让金融从业者将精力集中在核心业务决策上。 - 安全与客户体验：智能体交易需防范潜在的网络安全威胁；在前端，语音助手与对话式交互将逐步取代传统的手机 App 操作。 - 知识民主化与人才挑战：知识不再是岗位壁垒，银行业需要全面重新定义员工的价值创造模式，制定相配套的人才转型战略。 [https://www.youtube.com/watch?v=Q1xx1qjGXT8](https://www.youtube.com/watch?v=Q1xx1qjGXT8) ## Notable quotes > [00:04](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=4s) 如何将 AI 转化为商业变现（monetize AI） — Alexander Guerra Perales > [19:48](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=1188s) 赚钱（Make money）、省钱（Save money）、或为未来做准备（Be ready for the future） — Ash Kaduskar > [19:48](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=1188s) 在银行等大型企业中，AI 是一场“自上而下”（Top-down）的战略游戏 — Ash Kaduskar > [22:42](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=1362s) 你会发现 90% 的日常决策都是高度客观且完全可以被自动化的。这听起来有点吓人，但它能帮你找到真正的价值所在。 — Ash Kaduskar > [26:04](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=1564s) 我们还没到完全放手的地步。我们绝对不希望 Agent 随机生成不合规的政策或错误的安全配置包，那会引发灾难。 — Davidson Charles > [26:04](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=1564s) 除了尿湿裤子的婴儿，没人真正喜欢改变（No one likes change except a wet baby） — Robert Thacker > [35:26](https://www.youtube.com/watch?v=Q1xx1qjGXT8&t=2126s) 没有任何一个 AI 用例值得换来一次内部审计整改项或监管处罚 — Ash Kaduskar ## Mentioned Ash Kaduskar, Davidson Charles, Alexander Guerra Perales, Robert Thacker, Darrell Etherington, First Citizens Bank, Citi, Banco Azteca, Elektra, ARIS, The Angle, Shopify, Cursor, J.P. Morgan Chase, Agentic Banking, Agentic AI, Transformer ## Video description Ash Kaduskar - Head of Artificial Intelligence, First Citizens Bank Darrell Etherington - Editor, The Angle Robert Thacker - Head of Americas Solution Architects, ARIS Alexander Guerra Perales - Chief Strategy Officer, Banco Azteca / Elektra Davidson Charles - VP, Citi Financial institutions are transitioning from passive AI assistants to autonomous systems capable of executing complex workflows. This panel will focus on the rise of agentic AI and how it is transforming core operations like loan origination, trade execution, and dynamic customer service. Panelists will discuss the leap from predictive models to goal-oriented agents that take direct action within strict financial guardrails. The conversation will highlight integration strategies, workflow automation, and the new trust frameworks required when AI acts on behalf of the bank. Attendees will walk away with a roadmap for deploying agentic systems that maximize efficiency without compromising security. [Source](https://www.youtube.com/watch?v=Q1xx1qjGXT8)

[0:04] Alexander Guerra Perales: 如何将 AI 转化为商业变现（monetize AI） [19:48] Ash Kaduskar: 赚钱（Make money）、省钱（Save money）、或为未来做准备（Be ready for the future） [19:48] Ash Kaduskar: 在银行等大型企业中，AI 是一场“自上而下”（Top-down）的战略游戏 [22:42] Ash Kaduskar: 你会发现 90% 的日常决策都是高度客观且完全可以被自动化的。这听起来有点吓人，但它能帮你找到真正的价值所在。 [26:04] Davidson Charles: 我们还没到完全放手的地步。我们绝对不希望 Agent 随机生成不合规的政策或错误的安全配置包，那会引发灾难。 [26:04] Robert Thacker: 除了尿湿裤子的婴儿，没人真正喜欢改变（No one likes change except a wet baby） [35:26] Ash Kaduskar: 没有任何一个 AI 用例值得换来一次内部审计整改项或监管处罚

[8:54] Robert Thacker: They deployed AI just to service some loan applications... it worked great for about a month. But... after about a month and a half, 2 months, the agent learned that well, 65% of the time or 78.2% of the time when I send you something, you say yes... So the agent over time no longer started escalating that to the loan officer as an exception... What happens 3, 4 months from now? Auditors are showing up wanting to know... why you have such a high default rate because you've decisioned people that you would never have approved otherwise. [37:26] Ash Kaduskar: There is not a single AI use case that is worth doing to get an internal audit finding or a regulatory finding.

[0:04] Ash, do you want to kick us off with a quick  introduction of who you are and what you do? Yeah, [0:08] absolutely. So, my name is Ash Kadookskar. I  am a head of AI uh for First Citizens Bank. [0:15] My vision is to make FCB my bank AI ready and  three mission statements make ourselves faster [0:24] uh in make it more intelligent as well as  reimagine our workflows and processes. So that's [0:30] my quick intro. Okay. Hi, I'm David. Uh I work  for City Bank. Uh and my role is like uh pretty [0:38] much around the uh API engineering space where I  uh me and my team manages the governance controls [0:45] risk compliances. The mission for our team is to  uh move from we started from all the way from the [0:53] rulesbased and art artificial intelligence.  Now we are moving towards intentbased uh uh [1:00] solutioning. Uh so that is using agents. So that's  where we are heading towards too. Exciting. Alex, [1:07] I'm Alex. Um I'm the VP of strategy for one of  the largest banks in Mexico called Bangosteka. [1:13] I serve as the link between strategy and technical  projects uh particularly identifying how we going [1:20] to monetize AI. You know, you've been discussing  about like ROI positive on AI. So I'm the guy like [1:25] the finance guy kind of like making sure that all  these projects are positive in terms of returns [1:30] for the company. So I'm happy to be here with you  guys. Important role. Robert. Yeah. My name is [1:35] Robert Thanker. I am the head of solutions  architecture for Aerys. We're the process [1:39] intelligence layer that uh gives your AI the  ability to to scale, you know, because you've got [1:44] the the governance and the context that it needs  in order to be able to operate. Great. Uh so I'd [1:50] love to kick off with just kind of a few choice  examples of what it is you're actually doing that [1:56] you would define as a Gent banking uh within each  of your respective organizations to set some kind [2:01] of like boundaries for the conversation because I  think people here at Gentken still aren't really [2:06] sure what that means and especially as it applies  to specific disciplines. So Robert, do you want [2:11] to kick us off with that? Uh sure. Yeah. Yeah.  So if you think about what we're familiar with [2:16] with with AI at the moment, everybody's sort of  used to the predictive AI component that's sort [2:20] of getting into all right. Well, I think that  loan applications might be failing or things [2:23] of that. But Agentic gives the ability to now  decision on that and actually act on it. And so, [2:29] but if it's going to act on it, then we need to  make sure that that governance and that structure [2:33] is in place so that when an agent is going  to make a decision, it knows what decisions [2:37] it can make and what decisions that it can't. And  that's really what we're working on at the moment [2:41] is providing that information. Right, Alex. Um,  well, it comes top of mind that uh basically uh [2:48] the most important projects that we have right now  is on basically how to approach this to customers [2:54] in terms of agentic banking. It's how do we make  their operations easier, faster, more productive, [3:00] right? is either we're going to do it with  conversational AI with a concierge or something [3:05] that we can later discuss on or how do we make in  the back end operations like seamless frictionless [3:10] so in the front end like customers have like  an experience that they don't know that beyond [3:15] there was an agent taking decisions on behalf  of the bank like how do we make the experience [3:20] like faster more productive and for them like  even more friendly well I think we were talking [3:24] about similar things in the hallway right so you  want to yeah uh very similar to that note like um [3:30] What we are trying to achieve is um from  predictions and from the recommendations how [3:36] we can make it to the self-healing process right  so it's not just telling hey this is something [3:41] wrong and uh you need to go and correct it or  or I predict something is going to happen but [3:46] instead oh the action instead of just providing  the insights it's also going to be more actionable [3:53] uh stuff so that's where we are going towards  to one example is like um the traffic management [3:58] something is going to uh trigger probably a  weekend or something. How how in you know not just [4:04] providing the uh recommendation or suggestions how  proactively an action can be taken place to avoid [4:10] any such incidents before the incidents occurs. So  that's where we are uh trying to implement with a [4:17] lot of agents uh and again instead of saying lot  of agents more trustworthy agents right so that's [4:22] more important so we are moving towards that yeah  so our strategy is completely focused on internal [4:32] processes so we are not prioritizing any customerf  facing agentic AI right now and if I think about [4:39] what my strategy is for our bank is I want to  take the workflows which have objective questions, [4:47] objective answers, objective assessment. If it  fits all of these checkboxes, I'm going to pick [4:55] that workflow and automate it. And there are  couple of processes within the banking that [5:00] actually fits this really well. One is TPRM like  third party risk management if people are familiar [5:07] with that. Second is RCSA uh which is all our  controls processes. uh and then there are bunch of [5:13] other processes which are risk assessment related.  So those are the ones that I'm prioritizing and [5:20] the reason is because they are national highways  like if you think about a bank and if you want to [5:25] do RCSA which is controls it touches literally  every functional unit including front front [5:32] office operations back office and now if I touch  national highways like that suddenly I'm creating [5:39] massive value for the bank and then working with  like Alex like team members like you it is easier [5:46] to prove the value of why I'm doing something. So  that's my philosophy of attacking AI uh within our [5:53] banking area. So I think one question just because  it sounds like that's a common ingredient for all [5:59] of us up here. Um not myself, I don't work in  banking, but like what why is that better done [6:05] through sort of modern agentic workflows versus  maybe a more strict like rules-based traditional [6:11] LML ML approach, right? because you you said  objective like five times and it's like what [6:15] why wouldn't you just like deliver a rule set  and say if this then that as opposed to a gentic [6:22] action. See I think because these workflows are  not I would say like one action at a time right [6:29] like these workflows if you think about TPRM if  I break it down for you there are like three baby [6:34] steps in TPRM the first baby step let's say I want  to bring your company as a vendor right I want to [6:40] onboard them into my bank I kind of fill in like  in 15 minutes why do I want to bring your company [6:46] into the bank like what business I want to do that  is easier I'm not going to attack that at all but [6:52] the Next piece becomes like for TPRM there are we  have around 625 questions instead of me sending [6:59] you 625 questions to populate them and you taking  like a month to populate it right imagine the [7:06] time spent I would rather ask you give me all of  your artifacts you do it all the time right with [7:13] several companies these are publicly available  documents give it to me I will apply generative [7:19] AI populate all the answers already and if there  are exceptions I'll send it to your way but that [7:25] now instead of one month typing somebody in that  GRC it's such a it's a such a crazy process that [7:32] you can simplify now we don't stop here and this  is all unstructured data right like largely then [7:39] the second step happens which is even brutal is  you want to do a business so you are actually [7:45] motivated to type in the answers faster after you  have typed in the answers for 625 questions There [7:51] are like 40 odd people like from all the legal all  all the risk stripes the technology people they're [7:58] going to look into your answers and figure out  whether it's safe to do business with you again [8:03] objective assessment they are not thinking like  I don't like that company so I it's not feelings [8:08] it's about objective assessment I can extract that  knowledge out of that 40 people and automate that [8:15] now now not I have not only I've automated the  answering but also the assessment of it and it's [8:22] all unstructured. This is has nothing structured  data associated with it. You apply generative AI [8:28] and now the job for the people is to look at  exceptions rather than looking at the answers [8:35] and the reason agentic works here because it's  making decisions along the way. It is assigning [8:40] the risk rating based on unstructured data. That  is where I think these type of technologies can [8:47] be applied and it's a gamecher. It's from 4  months you can do it in within 3 weeks. That [8:54] is my mission. That's I keep hearing that a lot  like the shortening of the cycles, right? We were [8:58] talking about that too. But what about the um  that decisioning process like has there been a [9:03] fundamental change in how large those decisions  can be like and how significant they are that [9:08] you hand off to agents or what's happened there?  No. So the way that we are actually uh working is [9:16] uh not towards the decision making but it uh more  on the accelerating the process. So especially in [9:23] the governance space like uh usually it takes  around like three uh four to 5 days for the [9:30] governance process. It all the way starts from  the intake till it goes to the deployment. Right? [9:35] So as he was mentioning uh uh you know getting  the uh metadata informations there are so much [9:42] of informations. So instead of if then if rules  we are going towards the intentbased. So when we [9:49] are going to the intentbased we are not going  to ask for those informations. We are going to [9:54] uh generate most of the informations. What is the  type? What is the classification taxonomy? Where [9:59] is it going to get aligned and how much it's been  compliant. Even if it is not compliant as I was [10:05] mentioning it can be self field and things like  that and finally when it comes for the decision [10:10] right so it's again we make it based on the the uh  criticality of the workflow so we are not making [10:18] all the decisions to be made automated instead  like we are still working around we need to make [10:23] it very cautiously uh do that decision making so  in case of there is a repeated workflows like or [10:29] modify cases so if and even if it is in the  modify cases how much of the changes that is [10:34] coming in. So that kind of a it makes a decision  and it takes it you know it it finally go with [10:42] the decision whether if it is a modified or if  it is something new and something out of the box [10:47] then for sure we'll make human in loop in picture  but the our goal is to make fast accelerate the [10:54] process and that's where the agents that we  are plugging in into our complete governance [11:00] uh life cycle and what do how do you think about  the decisioning process like and where you tap in [11:06] humans and where you leave it to machine and  has that horizon changed over time? Uh well, [11:11] it's definitely changed. I I think the the thing  we really sort of have to keep in mind when it [11:15] comes to agents making decisions is as they sort  of you know mentioned really what we want humans [11:20] to decision are the exceptions that that are  falling within there. But you know and and that's [11:26] really where the time savings is going to come in  for the process improvements. A lot of time things [11:30] just they move through the process and they're  getting hung up not because necessarily there's [11:34] something wrong with the process but because the  human has to decision everything that comes into [11:40] the process as opposed to the things that you  know need to be that are just the exceptions [11:44] that probably need their attention that the AI  based on its rules and based on the requirements [11:48] that you've given it can make that decision you  know on its own but that also requires and I [11:54] guess comes back to the caution side of it is it  does require guidance and it does require require [11:59] monitoring. And a lot of times what we find within  organizations is they tend to set up an agent [12:04] and kind of let it run and then kind of a set it  and forget it kind of thing. And if you do that, [12:09] you're setting yourself up for failure too because  agents learn, right? And they make adjustments [12:15] based on how you decision things. And you know,  I heard uh recently there was a company, it was [12:20] a vendor, they made a big announcement about how  an agent works. It learns how you perform certain [12:26] actions and then it stores that information in  what they were calling their company memory and [12:31] then the agent now trains itself for future  things and not just its own self but also [12:36] uses that information for other agents as well.  And that scared me to death because what ends up [12:41] happening is now an agent is basically rewriting  itself on how it's going to operate. And this kind [12:46] of comes to just speaking of a local example.  There was a bank um not too around the area [12:53] where I am. They deployed AI just to service some  loan applications, right? They set decisions and [12:58] things up in place for it and it worked great for  about a month. But the rules that were in place, [13:03] I mean, typical things for if you're requesting a  loan, right? What's your credit score? What's your [13:08] debt to income ratio? How much are you asking  for? Do you do business with us? What kind of, [13:13] you know, what kind of relationship do we have?  So, all of those things get taken into account. [13:17] But after about a month and a half, 2 months, the  agent learned that well, 65% of the time or 75% [13:23] of the time or 78.2% of the time when I send you  something, you say yes for this. And so the agent [13:29] over time no longer started escalating that to the  loan officer as an exception because 78.2% of the [13:35] time you approve it, right? Mhm. So what happens 3  4 months from now? Auditors are showing up wanting [13:40] to know of course well why you've got so many  loans but more specifically why you have such [13:46] a high default rate because you've decision people  that you know would you would never have approved [13:52] you know otherwise. And so there is a decision  factor that needs to come in there but it needs to [13:57] be governed and it needs to be monitored as well.  You can't just turn an agent loose and say well [14:01] there are your decisions go forth. This sounds  like it's something you're nodding a lot but is [14:07] it you agreed on the kind of decisioning front or  how do you think about that? I mean let me let me [14:11] take a step back uh what what they're discussing.  I mean I 100% agree is you use these assistance in [14:19] order to improve workflows in order to take smart  decisions more effectively more efficient right [14:26] but as Robert was mentioning on this side is like  you cannot let them autonomously at least right [14:31] now take these decisions you need like supervision  and particularly because we're talking about banks [14:36] you need like specific areas that actually need to  make these tools accountable and follow compliance [14:43] rules in order to identify where are these sweet  spots or where are these specific guard rails that [14:48] need to be considered before taking a specific  decision. As Zor was mentioning for instance [14:53] in our bank we are implementing processes to make  loan origination like faster particularly because [14:58] we work in a bank that requires like a lot of  supervision even from the physical side of the [15:03] the whole uh process. So we want to make sure that  these specific loans are given to these specific [15:09] people at the right time at the right place but  without sacrificing growth without sacrificing [15:14] default rates and everything like that. So yes,  you can put as many agents that you have, but you [15:19] still need these people to identify for a specific  guard rails when is this human intervention to [15:25] make sure that the process keeps working like  faster, but in a in a comprehensive way is [15:31] like where these workflows are making the waiting  times like to be eliminated, these bottlenecks to [15:37] be eliminated. But most importantly for me this is  this is the holy grail about AI is agents that can [15:43] identify the missing information that needs to  be implemented on those processes in order to [15:48] make them faster because sometimes some teams they  might have identified the different processes to [15:52] make customer experience in the front and the back  like faster or friendlier or something but they [15:58] don't identify that maybe there's something else  to be missed right and then is there this is where [16:03] it comes to listening to a customer identifying  different things to to improve and use AI to not [16:08] only given or finding the next best product  or make the processes like faster but just to [16:14] identify a specific recommendations for them in  order to improve their overall experience. So this [16:20] is like you you keep mentioning speed everybody  keeps mentioning speed and like I can understand [16:24] too from like a customer perspective like oh  like time from loan approval much shortened [16:29] right but I think like slowness is a feature  in many in a lot of contexts that involve like [16:36] um finance right and and banking so and especially  like loan approvals like isn't there some benefit [16:42] to that that is inherent that's part of the  process that that we want to preserve or and [16:47] do you think that we're throwing out some of that  with the speed gains or are are you focusing very [16:54] specifically on the parts that are not beneficial  if I may compliment because I think it just [16:58] basically aligns with what we've been discussing  it in our cases we need to identify different kind [17:03] of use cases right many use cases or low hanging  fruit uh projects are about speed right because [17:09] I mean for instance the most common ones you know  I'm going to implement on banking AI for customer [17:14] service right because I need to reduce the  amount of complaints that I have or the different [17:18] interactions that I have with my customers to  make it like faster and obviously productivity [17:22] and then you can translate this into royal ROI  positive that's the most important one but but on [17:28] the other side of productivity there's the other  thing that I put a lot of a lot of effort on or [17:33] focus on which is increasing sales how do we make  sure that for instance like marketing teams CRM [17:39] teams are those are on sales that are making or  using this model to identify the next best offer [17:45] for these customers or what is the a specific  loan that they need or what is the specific kind [17:50] of investment that they need. The new Aentic bank  is going to use conversational AI is going to use [17:56] different workflows that are designed so that they  get to know different products that the bank may [18:00] have or may not have and then develop for for or  give recommendations for the product managers but [18:06] at the same time improve sales. I think that's the  the the most important thing beyond speed is just [18:12] identifying yes we're going to make the company  more productive. We're going to start with [18:15] engineers. We're going to start with product teams  and other ones on the back end. But in the on the [18:20] front end, we need to make sure that these tools  are bringing more revenue. If you don't bring [18:24] more revenue to the to the to the company, there's  going to be a a threshold where productivity gets [18:29] stopped and then you need to make sure that these  new projects are in a way prioritized so that the [18:36] impact brings new customers or retains those  customers or increase the number of products [18:40] that these customers are having with you. Yeah.  And I think those are hand and glove, right? [18:44] like understanding the risk profile and then that  means you're serving right customer right time and [18:49] also they fit within the risk matrix where it's  acceptable to them like this reminds me of I used [18:53] to work at Shopify and we would do the like this  was when we were doing like um instant loans in [18:58] dashboard for like marketing uh like cash advance  basically to merchants right so it's like this [19:04] is marketing budget go use and then increase your  sales but the this was prel like transformer-based [19:10] like AI so we were just using traditional ML to  make those offers, but it was not great, right? [19:16] Like not working exactly how you want. So, is that  kind of what you're talking about? Early steps, [19:20] you know, ranking customers, identifying the best  offers to send them, like kind of like prioritize, [19:24] identify propensity models, and then just  kind of like robust that into something that [19:28] runs like seamless without any friction at all.  Yeah. Think that's the most important challenge [19:33] that customers are right now facing is that how  to hyperpersonalize a specific kind of products [19:38] for customers needs. Yeah. and that but you can  do it now in a way that respects like what you're [19:43] looking for on the governance side right that's  see [clears throat] at the end of the day it's [19:48] a business any problem that you take to be solved  as a as a enterprise you are essentially going to [19:56] do only three things make money save money or be  ready for the future there is no other parameter [20:02] there has never been another parameter to kind  of do any project within a bank or any other [20:08] company [clears throat] AI is no different.  It is just a latest tool in our toolbox to [20:12] do something [clears throat] for specific type of  problems. Right? The idea here is not just to kind [20:20] of make things faster faster and what are [clears  throat] we going to achieve by that? Right? [20:25] Unless and until there is a tangible business  value towards it and it can be like I my job [20:32] is to onboard vendors faster so I can actually  start implementing those projects or whatever [20:37] to get value or because it is so manual I need 50  people to do that job and I don't think I need 50 [20:45] people tomorrow to do that job or it can be like  I'm going to acquire another bank and I want to [20:50] don't want to double my resource count that's why  I'm trying to lean out all my processes so that [20:57] when I acquire something I'm just going to have  1.25x of the people. So your business trigger of [21:04] why you want to take a certain process you decide  but at the end of the day there are only three [21:10] things save make more money or make your company  future ready. There is nothing else besides those [21:18] like value measurement is very simple if you only  focus on specific use cases like what I've seen [21:25] is been with AI for like four years now AI is  a top- down game if you are playing a bottoms [21:32] up game you have already lost it because there  is no value on like bottoms up use cases like [21:38] itsybitsy yeah these are good to create momentum  within your enterprise value is only going to come [21:46] from top-down initiatives and and that I'm talking  from enterprise standpoint. I was on a panel with [21:53] startups the other day and they are exactly of the  opposite view because they have less people. They [21:58] don't have massive processes to begin with. So  individual productivity matters to them a lot more [22:03] but for a bank like me I only care about national  highways the processes that I can make lean lean. [22:10] I think that's how we are looking at value within  our bank. interesting like the the that reminds [22:18] me of like we also work with cursor like and  cursor was led by is productled growth first [22:23] right y bottom layer up but then quickly shifted  to a motion of like enterpriseled growth right [22:28] like enterprise sales large deployment right so I  think to exactly your point they recognized like [22:34] this is if you as you move up market you have to  do the top down or it's not going to work it's not [22:38] going to work yeah yeah and I think we're we're  finding that too as we talk with organizations [22:42] everything when AI was hitting everybody was well  you know fear of missing out right everybody's got [22:47] FOMO right we got to we got to do AI we got to  AI everything we're going to see what's going [22:50] to happen you with regards to that but then  very quickly we start figuring out you know [22:55] obviously well now I'm not necessarily seeing the  productivity you know from from all of that from [22:59] all that AI and people start talking about token  spend and things of that nature and I think we've [23:02] all heard it right the only reason we'd focus  on token spend is because well we're not seeing [23:07] the productivity we're not seeing the gains from  it you know to your point we're not saving money [23:10] we're not making money we're just spending right  and you But if we were making money and saving [23:15] money by doing it, nobody would care, right? They  would spend however many tokens that that you [23:19] want. But now we're rolling back and we're looking  at things from a more strategic level. So we might [23:25] have an activity or might have a process within  here and a large portion of that process might [23:30] be manual. Maybe it's got 20 steps within it and  maybe it's got eight of those are manual. But do I [23:36] need to throw AI at all of those? Oh, let's take a  look at that. Let's see how they kind of fall out. [23:40] Well, how often do you actually perform that? you  know that step, right? Well, it is a manual step, [23:45] but you only do it once a month. So, why am I  going to throw AI at that, right? Here's another [23:48] activity that's manual that you're doing a hundred  times a day. Well, maybe that's a candidate, [23:52] you know, within here. Yeah. And you did  ask me ask the question about the decisions, [23:58] right? I think it's a very key parameter here. If  you think about AI, what type of decisions do you [24:05] typically make in an enterprise on day-to-day  basis? These are all leaders in the room. we [24:10] are making objective decisions like I in company  you don't make decisions based on feelings unless [24:17] you're making somebody's promotion decisions like  those type of performance review I'm I'm kind of [24:22] uh putting it aside for now but all other  decisions are very objective I will just give an [24:29] exercise to everybody like next week just document  the list of decisions you are making on day-to-day [24:36] basis and you will figure out 90% of those  decisions Firstly, they are objective as well [24:42] as they can be automated and it will give you a  scary stuff. is like okay does that mean my job is [24:48] at risk and you should do that exercise you should  have your team do that exercise that is what I did [24:54] with my operations team and my risk team is tell  me throughout the day what type of decisions are [25:00] you making and what decisions is taking most of  your time and you take those calls and you kind [25:07] of start automating and start redesigning your  process around those decisions the decisions [25:13] are not risky the exception management ment is  risky. So you don't now the question becomes is [25:20] can I have autonomous agent onboard your company  without even taking my permission. No I think that [25:26] is silly but until that point it can bring it  you can agentify all the process and then I can [25:32] have my thumbrint saying okay bring your company  because I want to do business with you. So now all [25:38] the time that it takes to make the decisions you  have automated. You have very less things to worry [25:44] about. That is where the the value of generative  AI is and that is what machine learning used to do [25:51] like loan originations, balance transfer in credit  cards like we have been doing it forever. This [25:57] is helping us with generative AI language all of  those aspects of it. Yeah, that it's interesting [26:04] when you talk about that decision because I was  thinking about it for my example. Uh but I work in [26:08] a very different field than you, but I think a lot  of the decisions are not actually super objective. [26:12] Yeah. But also, if I'm very honest, a lot of the  decisions are not actually super consequential, [26:16] right? Like so yeah, we're making a decision about  like what communication strategy to approach. It's [26:21] six of one half dozen of another. Please pay me  money to continue doing this, but like you know, [26:26] like in in the moment, they're not life and death  or whatever. and like but it's not like we're [26:31] arriving at some objective outcome like a loan  origination or loan uh dispersement or whatever, [26:38] right? Um but yeah, I I am curious about this  since you brought up kind of like the the [26:43] organizational transformation aspect of it, right?  Because you've had those conversations. I assume [26:48] everybody has had those conversations about use  of these things and to varying levels of comfort, [26:53] right? like with customers and also internally  like how do you find people are feeling about [26:58] it within the organizations as some of these  decisions that used to be humanmade are becoming [27:03] agentically driven. Is it having an impact?  [clears throat] Agentically driven. Yeah. [27:10] Like like instead of a person sitting there and  doing which is you know not like pleasant but is [27:15] something that they used to do. Um like how are  they feeling about handing off more of that work [27:20] or is it it's both ways right? like um I would  say from my perspective in in the governance space [27:27] especially the customers who have been waiting for  a long period of time to get something done by the [27:33] people. So when it comes to the agent and if it is  going to be again I'm bringing the word faster. So [27:38] that's where we have been facing the issues for  the last uh I I would say like how fast that we [27:44] can uh make it to the market. Yeah. From because  like that's where the the entire worlds go right. [27:50] They wanted everything quick. When they come in  today, they want something to be in the market [27:53] by tomorrow. Right. Right. And and that's that's  most of the processor oriented like the risk team [28:00] involving the risk team, the compliance team and  the security team. So this is more time consuming [28:06] and the SL is really really really big. So if you  ask the c my uh uh you know the customers the app [28:15] developers the designers they said they want  it quick because their customers their business [28:20] need it as fast as possible. So in that case like  okay let's eliminate um you know a lot of manual [28:27] work that currently be done by the team checking  for issues where it has to get aligned is it is [28:34] it really readiness right uh where it has to get  deployed and things like that. So that's where we [28:39] are coming up with some small agents to be in  place and making it faster and and really the [28:45] team teams are really appreciative of it. At the  same time when it comes to the decision- making [28:50] again I do have certain uh you know change  in thought with what uh I have mentioned it [28:57] uh we still have that uh 50/50 chances of how we  can take with the decision-m from the agents. Uh [29:05] we are not there yet. uh and we don't want to uh  provide the random things on the decisions and you [29:11] know the agents provides different uh uh policies  or uh different uh um security uh bundles that [29:20] that's not needs to be in and that really causes  a lot of issues. That's where we really need some [29:25] human to intervene and do the uh supervise thing.  So for for sure we need to have that in place to [29:33] uh eliminate these kind of an issues. So there  is a 50/50 chances where there are certain places [29:38] where they need human to be intervene and do  the validations at the final check but that how [29:45] fast that we can accelerate the process is more  important and how fast that we can able to make [29:49] the uh the business right make them make it to  the market sooner is what more important right [29:57] so you were like speed constrained and then that's  removed but it's not really like a person capacity [30:02] that's true well I I think well And there's the  other side of that too. I think you everybody [30:07] likes to talk about change until it affects them.  Right. Right. I mean that's that's sort of the [30:10] thing. My grandmother used to say no one likes  change except a wet baby. Right. So you know so [30:16] as long as AI is out there it's a shiny object.  Everybody wants to talk about it. Everybody wants [30:19] to play with it as long as it doesn't necessarily  affect my job or or my particular role. But in a [30:24] lot of cases it's not necessarily about you know  releasing people. A lot of times you know we have [30:31] you know we have we need to have faster speed to  market. We need to create more. We need to produce [30:35] more. We need to process more. You know, but I've  only got so many hours in a day and I've only got [30:40] so many people and we don't have the budget to  hire more. So, what if we could do more with what [30:46] we have, right? So, it's not necessarily about  releasing people or replacing people's jobs. It's [30:51] about being able to be more productive with what  we already have. Repurposing. Yeah. I I compliment [30:56] that the way that we're handling this is like as  you guys mentioned, it's not about like cutting [31:00] jobs. Uh it's also about letting people know is  that you're going to have these tools that you're [31:04] going to be more productive but in the same way  you're gonna be required not only to have more [31:10] output but also bring more ideas on how to create  more value right because the what is going to make [31:16] the difference for these companies to remain  with the same amount of people working there [31:20] is like how we going to bring more money to the  table and that comes to to the whole concept of [31:26] prioritizing different projects right when we're  talking about aentic banking it's about how do we [31:31] make sure that the experience become frictionless.  Imagine the old days where you have to go to the [31:37] bank to make a transfer to have a withdrawal  of money or something and then ATMs came and [31:43] then the whole process became like faster and  then mobile phones came and then banking apps [31:48] made made things even more convenient. With aentic  banking, what we're doing is we're taking another [31:54] step in order for you using conversational AI or  something that you can make things automated in [32:00] order to set up your next your next next tasks or  something like that. We are we are just bringing [32:05] this new technology so people have easier or  more convenient lives to take better decisions, [32:11] you know, or and that needs to be translated into  company making more profits or or increasing their [32:16] productivity in different ways, right? And to  your earlier point too, I think like you were [32:20] talking about right presenting the right product  to the right right customer at the right time, [32:24] right? Like that also potentially opens up a  whole lot of uh economic opportunity beyond, [32:30] right? Where it's like well if customer didn't  have this line of credit for whatever thing like [32:33] they wouldn't be able to Yeah. Exact. Exactly.  Because we're still facing an important challenge [32:38] for most companies. It's not about like doing  the best the best processes or the most automated [32:44] processes or the most agentic processes. It's  also to to what do I do with all the data that [32:49] I have right you know I have first party data I  have third party data how do I use this data to [32:54] identify different needs that customers  already have that I am not recognizing [32:58] right that my market marketing teams are kind of  like identifying some patterns but I have not like [33:03] known about that that's why in my previous comment  I was I was talking about this holy grail about [33:08] agentic AI which is it's not only about like  reducing or eliminating the waiting times or [33:13] like making the decisions faster is identifying  what's the missing point. It's like, hey, Alex, [33:19] I'm seeing this customer that needs this specific  kind of product, but you don't have this product. [33:24] What if you can have this specific tailor product  for this customer that is not finding this and [33:29] that I have also third-party data that is telling  me that this customer went to other websites to [33:33] see or looking for this loan or for this asset  or this this investment product for instance. [33:38] This is not something that already banks are  doing for for the research that I that I've done, [33:43] but it's is these big opportunities that I'm  seeing for the next coming years like banks [33:48] investing more efforts on. Yeah. Yeah. That I've  heard in other industries a lot of like what what [33:53] AI can do is essentially become a PM function for  you like it is discovering product opportunities [34:00] that you didn't realize were there and then  giving you the the playbook to go build it. Right. [34:04] to add to that I just want to uh say like it's  not like as he was mentioning that we are going [34:09] to release or we are going to move out the people  uh it's as I was even discussing outside the room [34:16] the complaints keep changing whatever that I was  doing last year it's not the same thing and as I [34:21] was mentioning with the AI the the requirements  the new businesses also the volume is high so [34:28] let's say that if I I have like 10 people doing  the manual work and the amount of the work that [34:33] it comes it's also too much right so as I was  mentioning we have to repurpose them using this [34:40] AI where the the the other team can focus more  or we can tune the models or tune the agents in [34:49] such a way that because every day there's going  to be something new it's not stagnant okay I've [34:54] done I've given to the AI agent and I'm I'm not  going to take my hands off and and it's the the [35:00] volume of business is high at the same time the  the changes that is going to come in uh with the [35:05] agents also going to be on the daily basis and  we need people for that. Yeah. So that's where we [35:10] also need to have uh training and repurposing the  people who have been doing so we need to move them [35:16] you know to do the the whatever that is happening  right now. This I think that gets to like another [35:22] question I would have which is how do you balance  the pace at which the underlying technologies are [35:26] improving and the need for like good governance  and controls because this is this is something [35:31] this is not a sensitive uh industry because it  was content marketing but I used to work in that [35:34] and we were using like AI powered workflows to to  make the stuff and the harnesses we built ended up [35:42] um not doing well. They contorted the the input  from the base models especially as they improved. [35:48] So it ended up being like one if you remove the  harness it was actually a much better product uh [35:53] 3 months later right and then so do you encounter  that problem where the pace of improvement at [35:59] the bottom like at the bottom of the stack is  making it difficult to build the scaffolding that [36:05] controls it and yet also giving up opportunity  cost for like how fast you could move. Well I [36:12] think there's there's obviously there's a risk  to moving fast, right? I mean, you can I can [36:16] give you a really pretty car that doesn't run  or I can give you something that that actually [36:20] moves and does what you want it to do. That's  from that decisioning aspect of it really what [36:26] you know we're looking at. You know, here's more  the ability to take information in and then decide [36:32] how we're actually going to how we're going to  work how we're going to work with that. Um there [36:37] there's a lot of things that get that get built,  you know, and unfortunately, you know, technology [36:42] is going to move fast. It does. It always  has, right? And from a governance standpoint, [36:47] unfortunately, the way a lot of organizations work  is, oh, look, we've got this new technology. We're [36:51] going to implement it and then we're going to try  to figure out how to govern it. And as opposed to [36:55] taking the structure and saying, well, what are  our business requirements? What are our rules? [36:58] What are the policies that we have in place?  Let's define that and then figure out how the [37:04] technology is going to roll into that, you know.  So, governance should come before technology, [37:08] not the other way around. And, you know,  unfortunately, like I said, technology will [37:11] always move fast, right? So, and you know AI is I  mean not remotely the same as it was. I mean if we [37:17] had this conversation three months ago, we would  have been having might be the same questions but [37:21] still completely different answers. I mean it's  moving that fast. It's it's interesting though [37:26] the governance thing cuz like um that's like the  opposite of what an what a challenger would say, [37:31] right? Like so is that uh something that they you  think of a lot in the banking sphere because you [37:38] used to work at Uber. Uber's a great example  like Uber did not do governance first product [37:43] after right like especially not early that's not  how they won so how much does that come into the [37:48] thinking or does it come in at all see that is the  number one thing like I as a head of AI I keep on [37:54] saying to all of my leadership team is there is  not a single AI use case that is worth doing to [38:04] get an internal audit finding or a regulatory  finding that is a thumb rule like it's just [38:10] not worth it. [clears throat] We are we need to  know which industry we are playing with. We are [38:15] a regulated industry. Uh I have been in this head  of AI role for last 9 months now and the first 6 [38:23] months like uh the best friends I made were my  risk partners because I know they are the ones [38:30] that they can stop my agenda of doing AI. But what  does that mean? That means that all of the risk [38:37] partners who have been doing AI governance from  machine learning standpoint for last 20 years, [38:44] we just need to upskill all of them including  us on how to do a risk assessment of generative [38:50] AI solutions. It's a very different beast  because it's not a deterministic model. It's [38:56] a probabilistic model. So how are we going to do  from model risk management, legal risk management, [39:02] reput there are many 10 11 risk stripes. So making  sure that your risk assessment methodology is [39:10] standardized and again you get to the same  philosophy. Same questions I'm going to ask [39:16] to every AI use case. I am going to answer them  objectively. I'm going to assess them objectively. [39:23] Until you reach that point, you will always find  risk people are on your way to get AI use case [39:30] implemented. But if you get that hurdle crossed,  then I think they are just partners. Everybody [39:37] wants to implement AI. They are just doing their  job. So we just need to help each other to make [39:42] AI governance as solidified but as standardized,  repeatable and scalable. I think that also comes [39:49] to me you mentioned Uber Uber a moment ago  but Ash also start this conversation with [39:55] he was on a panel with startup people right so  [laughter] you know so Uber didn't have that all [40:02] that compliance stuff but they were a startup you  know obviously speed is is the motion in a startup [40:07] once you reach a point where and especially in a  regulated industry like you know like banking then [40:14] governance isn't a hindrance to to innovation  you know that and that a lot People think well [40:18] if have implement governments and we got these  that that's going to slow down innovation but [40:22] it it doesn't slow down innovation. What it does  is it means that when we do when we innovate it [40:26] gives us what was mentioned earlier that trust  that what we're doing is implementable at scale. [40:31] It doesn't work well in a lab and then doesn't  work in production right it'll actually scale [40:36] for us. Yeah, that's actually I've heard like you  know imagine if you're doing cloud code and you [40:40] have your audit come in at point of like crafting  whatever you're building and says like hey this is [40:45] actually compliant with your policies and whatever  you need to do like in that moment that's an [40:49] accelerant versus a exactly road block. I agree.  So some companies use different strategies right [40:55] for instance I've seen in different latitudes  of the world like some like traditional banks [40:59] they have like these rigid structures to innovate  but they use other other startups that they buy [41:04] or they they work for or like fintech different  companies that they don't have they don't have [41:10] this strict measures in order to how to like  innovate or experiment with different things [41:15] but they kind of like use different strategies  to do it inside a specific audiences inside a [41:21] specific new products so that they and start like  getting or experimenting getting different ideas [41:26] of how to model. But I mean I'm seeing big banks.  I think right now what I'm what I'm checking from [41:31] different um friends that work for different  kind of like large or small scale banks is is [41:36] the access to capital. It's not the same access to  capital that for instance like JP Morgan Chase has [41:41] to experiment to train models or to build their  own models than this small local community bank [41:47] and that makes a big difference because it makes  the gap to even like become like bigger in terms [41:52] of like how the access of this technology or  investing to this technology is going to make [41:56] these companies have better products. So it's  this whole thing about AI like just weaponing [42:02] more the the whole gap into the into this whole  competition. But I agree with my peers. It's like [42:07] is the more that you experiment, the more that  you innovate, the more that you make the process [42:12] faster is is going to create this specific  opportunities for serving customers better, [42:17] creating more value for the company. But at  the same time, finding specific niches where [42:21] traditional banks may not be right now or that you  can compete in a different way. Great. Uh we only [42:27] have a couple minutes, so just quick lightning  round. What are you most excited about from [42:31] uh Agentic Future perspective in banking?  [laughter] um that it's a great question. I mean, [42:38] I do think it is just more the fact of how much  faster everything's going to move and everything [42:42] from a decisioning standpoint. I mean, when  you start looking at what processes are doing, [42:46] the the biggest impact isn't necessarily the fact  that humans are going to be working harder. It's [42:51] just that they're going to be working on stuff  that's more meaningful and they're going to be [42:54] decisioning information based on the data that's  that's actually relevant for them. And all of that [42:59] tedious stuff will be be done away with. He said,  "We talk about things taking a while because, [43:04] you know, it it's sat on my desk for a while,  but that's because I'm decisioning everything. [43:09] Now I'm actually just focused on I think that's  really where AI is is going is going to help out [43:14] financial services in a huge I'm kind of concerned  about like where the future of identic uh trading [43:21] is going to uh because of like the different uh  things that we're seeing and what we've discussed [43:27] in the keynotes about the the the perils of  cyber security attacks. So there's this downside [43:33] thing that it keeps me alert on like particularly  working with the audit departments or compliance [43:37] departments that we need to invest more on  building these cyber security capabilities within [43:43] our banks or within the financial institutions  or any company that you work for. Right? That's [43:47] that's on the downside. On the upside, I think  it's uh how the experience is going to improve for [43:53] customers. I'm seeing conversational experience  like becoming more a common thing. Yep. You know, [43:58] you you don't have to go to the app. you just  talk to to your concierge or to your assistant [44:03] and you give some commands or some prompts like  on voice and then you're gonna have pretty much [44:08] your day already set up for for what's coming. So  you don't need to spend much more time on your app [44:13] checking your finances or setting some alerts.  I I I'm seeing the life of of of our communities [44:19] or our societies interacting with banking apps  becoming more more faster, more friendlier and [44:25] giving more personalized recommendations for  the things that we need. I think that's the [44:29] the things that I see happening right now and what  the most banks are investing with the capabilities [44:34] and the restrictions they have right uh to make  it short uh customer experiences and also the [44:40] simplification of the process for the customers  so they are spending so much of time without [44:45] even knowing why they need to do that so how we  can simplify it so that's where we are moving [44:50] towards to uh simplify the process and capture  whatever that is only required or only that they [44:56] are aware of it But at the same time we also need  other information since we are working with the [45:02] compliance audits and risks team and how we can  able to manage those informations as well. Yeah. [45:07] I think I'll tweak the question a little bit.  What is keeping me up in the night is uh is people [45:16] aspect of it. I think see I think we all are here  or people are here and doing speeches or whatever [45:24] is because we have knowledge right this knowledge  is getting democratized and knowledge has no [45:30] meaning anymore it's like and in that situation  when somebody is at a particular role because [45:37] of what that person knows and now that is not the  moat for what that person is actually doing why am [45:45] I in the job I That is the the exam question that  is keeping me up at night and our bank is 18,000 [45:53] people like we are not that big but every day I'm  working with our HR team to figure out a people's [46:00] strategy for our bank because if we done right it  can be disruptive if it's yeah if you don't want [46:09] to adopt it then it's not a problem but otherwise  people problem is going to be the biggest [46:13] challenge. Yeah, I think that's a good final note.  Thanks very much everybody. Thank you. [applause]