Mastercard Deploys Agent Pay as Banks Use AI to Protect Lending Moats

万事达卡推出 Agent Pay，银行借 AI 筑牢信贷护城河 金融机构若仅将人工智能用于基础工作流自动化，而非定性商业核保与零息稳定币利差，将面临估值被压低至公用事业股水平的风险。 专家小组探讨了金融机构在客户支持、后台文档处理及商业信贷决策中落地生成式 AI 和自主智能体的实用路径。与会发言者强调了一种“人在回路”的增强模式，该模式在管控风险的同时释放了显著的经济价值，并列举了万事达卡的产品入驻助手（Product Onboarding Assistant）和 Coinbase GPT 等实践案例。他们还探讨了宏观趋势，包括 AI 缓解逆周期信贷溢价的实证证据、银行发行稳定币的经济效益、依托万事达卡 Agent Pay 实现的自主商业，以及“了解你的智能体”（KYA）等新兴治理框架。最终，专家建议机构从解决急迫的内部摩擦点入手，同时利用 AI 巩固其关系型银行业务的竞争护城河，以防陷入同质化。 万事达卡高管指出，在客服和一线客户关系智能等高频、低风险的非结构化任务中引入人机协同（human-in-the-loop），能捕获生成式 AI 80% 至 90% 的经济价值，同时多模态智能体正大幅简化原本高度依赖人工的 KYC/KYB 与发票对账流程。 (4:20) 万事达卡通过严谨的产品化 MVP 路径落地多智能体入驻助手（Product Onboarding Assistant），通过辅助内部技术团队将单次银行技术集成的工时平均缩减了 28 小时。 (12:28) 实证研究显示，AI 信贷建模能基于客观数据定价以平抑信贷顺周期性，在经济扩张期降低贷款利息并在下行期维持信贷供给，从而将定制化关系型信贷扩展至中小企业（SME）。 (20:18) 随着消费发现场景向对话式 LLM 迁移，万事达卡推出 Agent Pay 框架以打通传统银行卡及稳定币的原生智能体结算，行业也正从 KYC/KYB 拓展至建立“Know Your Agent”（KYA）身份核验与安全治理机制。 (23:17)

Financial institutions are transitioning generative AI from experimental pilots into production by deploying autonomous agents across repetitive workflows while using artificial intelligence to augment, rather than replace, human judgment in relationship lending. Industry leaders and academic experts emphasize that long-term enterprise value depends on disciplined data orchestration, emerging governance frameworks like 'Know Your Agent,' and shielding relationship banking from commoditization. High-Volume Support and Back-Office Multimodal Workflows (4:20) Federico Cohen Freue of Mastercard identifies repeatable, low-risk tasks involving unstructured data as the most effective starting points for generative AI. In customer support, where identical queries recur hundreds of thousands of times daily, retaining a human in the loop captures 80% to 90% of economic value while mitigating operational risk. On the internal side, multimodal agents parse unstructured PDFs, invoices, and corporate filings for Know Your Customer (KYC) and Know Your Business (KYB) compliance, eliminating brittle template engineering and redirecting human scrutiny toward ambiguous, high-risk boundary cases. Cohen Freue advocates scaling through a disciplined product management approach. Mastercard's Product Onboarding Assistant exemplifies this: rather than automating complex integrations entirely, the multi-agent system indexes technical documentation and specialist feedback, eliminating an average of 28 hours of technical implementation time per bank onboarding. Similarly, Coinbase expanded its CBGPT tool from basic authentication support workflows into company-wide document querying and product specification drafting. Objective Underwriting and the Qualitative Lending Moat (7:42) Daniel Chi of the University of Nevada, Las Vegas, deconstructs commercial banking into two foundational tasks: information processing and credit decisioning. While legacy computing resolved information aggregation, credit decisions depend on the 'Five C's of credit' (Character, Capital, Capacity, Collateral, and Conditions). Evaluating borrower character historically required human discernment that rule-based systems lacked. Generative AI bridges this divide by contextualizing qualitative signals in unstructured records, allowing lenders to extend tailored relationship underwriting to small and medium-sized enterprises (SMEs) with sparse documentation. Chi highlights empirical research from Italian banks indicating that AI credit models attenuate traditional countercyclical rate structures. Historically, banks charged higher rates in expansions to subsidize distressed cycles. Granular AI modeling allows institutions to price loans at true underlying risk, lowering borrowing costs during expansions while maintaining credit availability during downturns. However, Chi and Bank for International Settlements research stress that commercial credit still hinges on qualitative trust and risk subsidization, making AI a force multiplier for human underwriters rather than a replacement. Agentic Commerce and the Economics of Stablecoins (23:17) As consumer product discovery migrates from search engines directly into conversational AI agents, payment networks are developing agentic payment rails. Cohen Freue notes that current autonomous purchasing relies on brittle, insecure screen scraping. Mastercard's 'Agent Pay' initiative aims to embed native transaction capabilities, identity safeguards, and multi-rail settlement across traditional cards and digital assets. Chi draws a historical parallel between emerging stablecoins and the pre-Federal Reserve era when private commercial banks issued individual gold-backed bank notes. Under proposed legislative frameworks such as the GENIUS Act, banks issuing dollar-backed stablecoins are not required to pay interest on those liabilities. This allows banks to collect customer funds at 0% interest and place the reserves at the Federal Reserve to earn risk-free yields—currently 4.25%. Chi explains that this spread represents a primary economic incentive for commercial banks embracing stablecoin issuance, alongside their utility as programmable settlement rails for autonomous agents. Enterprise Governance, KYA, and Strategic Differentiation (35:04) Deploying autonomous agents requires internal policy alignment before technical engineering begins. Chi observes that banking leaders prioritize establishing board-level ethical AI guardrails prior to implementation. As software agents gain autonomy to move funds and evaluate credit, both speakers anticipate a 'Know Your Agent' (KYA) compliance standard. Analogous to enterprise vendor risk management, KYA verifies agent identity, authorized permissions, and systemic safety controls. Strategically, Chi cautions banks against adopting AI merely for cost reduction, warning against a commoditization trap that mirrors the post-2008 financial crisis, when bank valuations dropped to price-to-earnings multiples of 10x compared to 20x for utility companies. Instead, Cohen Freue and Chi recommend directing AI toward acute operational friction while using the technology to fortify high-margin, judgment-heavy domains like commercial underwriting and relationship banking. While generative AI and agentic architectures offer measurable efficiencies across back-office operations and risk-adjusted lending, their enterprise viability depends on establishing industry-wide agent verification standards and navigating evolving regulatory frameworks governing bank-issued digital liabilities and autonomous transactions.

内华达大学拉斯维加斯分校的Daniel Chi教授与万事达卡人工智能与数据运营高级副总裁Federico Cohen Freue指出，银行业引入生成式AI与智能体并非要取代人类专家，而是通过增强定性信贷决策与化解运营摩擦来沉淀竞争壁垒。他们认为，金融机构应坚持以人机协同与敏捷产品思维推进落地，依托非结构化数据处理拓展信贷与商业场景，同时防范业务被工具化、公用事业化所导致的估值滑坡。 业务落地：人机协同与痛点突破 (4:20) Federico Cohen Freue指出，当前生成式AI在银行高价值场景的落地主要集中于客户交互与后台运营，且普遍契合“高重复性、低风险与非结构化数据复杂”三大特征。在客户服务中，机构日常需处理海量同质化咨询，采用“人机协同”（Human-in-the-loop）机制即可捕获80%至90%的经济价值，同时防范系统性风险；在前台，系统能为客户经理在拜访客户前实时汇总客户背景与交叉销售线索，消除漫长的人工案头准备。在后台运营中，多模态模型正突破发票核对及KYC/KYB合规审核的瓶颈，直接解析非标准公司凭据与扫描PDF，免除了脆弱的模板配置流程，让合规专家得以集中精力审核高风险边缘案例。 在落地推广路径上，机构普遍采取由浅入深的渐进框架。以万事达卡为例，其工程团队研发了多智能体系统“产品入驻助手”（Product Onboarding Assistant），通过解析索引技术文档并建立技术专家的实时反馈闭环，协助解决银行接入网络时的复杂跨部门流程。万事达卡并未冒进追求端到端全流程无人化，而是先以单一产品的最小可行性产品（MVP）切入，最终实现单次入驻平均节省28小时技术实施时间。Cohen Freue强调，先针对员工日常工作中最耗时、存在严重瓶颈的具体流程寻求单点突破，是规模化铺开企业级AI最稳健的路径。 信贷重塑：破解定性瓶颈与小微普惠 (7:42) Daniel Chi从金融学术角度剖析，商业银行的核心中介职能依赖信息汇总处理与信贷决策两大支柱。传统计算机早已解决了信息处理环节，但在信贷决策“5C原则”（品行Character、资本Capital、能力Capacity、抵押品Collateral、环境Conditions）中，“品行”评估高度依赖对借款人声誉及还款意愿的定性判断，过往算法系统由于缺乏上下文综合理解力而无法介入。Chi认为，生成式AI能够阅读、理解并情境化分析非结构化文本，使得算法首次能够真正辅助商业信贷中最核心的人性化与定性分析。 这种技术演进正在重塑信贷供给与风险定价机制。Chi援引一项针对采用AI信贷分析工具的意大利银行的实证研究指出，AI工具削弱了传统信贷的顺周期性：传统模式下银行为抵御道德风险会在经济繁荣期加收保护性溢价，而AI模型依托客观数据分析真实风险，使得受试银行在繁荣期下调了贷款利率，并在下行期依然维持常规信贷供应，从而降低了借款人的融资成本。Cohen Freue补充指出，该能力对全球中小企业（SME）尤为关键。小微企业常因财务数据不完备而遭遇融资瓶颈，AI解析非传统与非结构化数据的能力，让银行得以将以往仅面向大型企业客户的精细化关系型信贷服务，低成本延伸覆盖至长尾小微企业。 智能体商业与稳定币：支付演进与利差机制 (23:17) 随着消费者的产品搜索意图逐渐从传统搜索引擎迁移至对话式大模型，自主智能体代理交易的需求加速显现。Cohen Freue指出，现有的交易闭环严重受阻，大多依赖脆弱且不安全的屏幕抓取技术，为此万事达卡于4月下旬公布了“Agent Pay”框架。该方案旨在建立原生网络标准，使自主智能体能在获得授权后代客安全付款，并提供与线下刷卡同等水平的身份认证、反欺诈与信任保障。该框架支持多元化支付轨道，允许智能体根据用户指令调用传统法币银行卡，或是连接受监管的合规稳定币钱包。 在探讨稳定币监管与商业银行的联动时，Daniel Chi提出了两点关键见解。在历史演化维度上，他将当前立法框架下商业银行以1:1美元储备发行代币的设想，类比为19世纪至20世纪初美联储成立前以黄金为储备的私人银行券时代，指出银行发行稳定币虽然创造了全天候可编程支付通道，但其购买力与稳定性根本上仍锚定于美元本身。在经济驱动维度上，Chi强调拟议的监管框架（如GENIUS法案）允许商业银行向稳定币持有者支付0%的活期利息，而银行吸收资金后可将其存入美联储并赚取当前约4.25%的无风险利率。这种零利息成本负债带来的高额净息差，成为传统银行竞相涉足稳定币领域的核心动力。 落地战略与治理：数据编排与KYA演进 (31:49) 针对生成式AI非确定性特征带来的部署挑战，Cohen Freue强调企业不应等待数据变得“完美”，而应建立常态化的数据治理纪律。万事达卡在基础设施层面的重心已从被动搭建静态数据湖转向跨系统的数据编排（Data Orchestration），确保核心数据集能够被敏捷调用。在业务落地层面，技术团队严格秉持产品管理思维，从可用数据、预期业务价值、构建复杂度以及风险敞口四个维度量化评估业务痛点，确保每一项智能体方案都具备可衡量的投资回报率。 在合规与治理领域，Daniel Chi指出银行业已达成高度共识，即必须由董事会和管理层自上而下确立负责任的人工智能伦理政策，先筑牢防线再铺开工具研发。当前各银行主要将AI作为合规工具，用于辅助审计流程与监管报告生成。随着智能体逐步拥有独立转移资金和评估信用的能力，主持人Jonathan King与Cohen Freue均认为，金融行业必须建立专门的“了解你的智能体”（KYA, Know Your Agent）框架。这一体系将参照成熟的第三方供应商风险管理机制，针对自动化智能体的身份识别、授权权限及底层工具链展开系统性验证，成为KYC/KYB在软件自主交互时代的必要演进。 Daniel Chi警示银行业切忌将自身降格为按公用事业定价的纯撮合通道，若仅在交易价格上陷入价格战，会导致估值倍数严重滑坡；AI的核心使命应当是巩固商业信贷、客户信任与复杂决策的护城河。不过，讨论也留下了尚未解决的治理挑战：在成百上千种银行发行的稳定币共存、自主智能体跨网络自主交易的未来图景中，如何在跨代币清算互通、非确定性AI决策责任归属以及全天候合规监管之间求取平衡，仍是传统银行步入生产级AI时代必须持续探索的难题。

## Background 该视频由行业会议频道 Ai4 于 2025 年 8 月 28 日发布在 YouTube 平台上。内容为万事达卡（Mastercard）AI 与数据运营高级副总裁 Federico Cohen Freue、Coinbase Ventures 首席投资人 Jonathan King 与内华达大学拉斯维加斯分校金融系主任 Daniel Chi 教授共同参与的圆桌研讨，探讨银行业如何跨越试点阶段，安全且可扩展地在前后台生产环境中部署生成式 AI 与自主 AI 智能体。 ## TL;DR 圆桌深入剖析了生成式 AI 与 AI 智能体在银行业的实际落地路径，涵盖前台客户经理赋能与后台多模态 KYC/KYB 自动化等典型范式。嘉宾从金融学核心“5C 原则”与信贷中介本质出发，阐述了 AI 如何通过精准风险量化平抑传统信贷的反周期波动，并分享了万事达卡在产品接入助手（Product Onboarding Assistant）与 Agent Pay 支付协议上的实战经验。讨论还前瞻性地探讨了银行发行稳定币的历史镜鉴与无息套利模式，并提出了面向自主智能体交易的“了解你的智能体”（KYA）合规框架，强调银行应利用 AI 强化人际信任与复杂决策以避免同质化陷阱。 ## Atmosphere & style 整场研讨氛围兼具严谨的学术深度与务实的产业实战感，展现出跨界思维碰撞的高效与张力。来自顶级商学院的金融学者、加密风投投资人与万事达卡核心高管形成了极佳的互补视角，既有对百年银行史与信贷理论的深层解构，又有关于多智能体架构与支付协议的硬核落地剖析。对谈节奏紧凑、观点犀利，既仰望智能体商业与稳定币结合的未来蓝图，又始终聚焦于一线痛点解决与 MVP 渐进式交付的务实底色。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=4s) 万事达卡等金融机构的 AI 战略涵盖顶层业务押注、数据基础设施建设、员工素养赋能及全生态能力输出四大维度，推动 AI 从概念验证走向核心生产环节。 - [04:20](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=260s) 银行 AI 落地呈现双重范式：前台聚焦高频、低风险、非结构化数据的客户经理赋能与人在回路；后台多模态 Agent 则彻底重塑 KYC/KYB、发票核对等繁琐合规与核保流程。 - [07:42](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=462s) 商业银行本质是信贷中介，生成式 AI 的核心突破在于从单纯的数据聚合统计，升级为能够辅助评估信贷“5C 原则”（尤其是主观且复杂的借款人品德 Character）的智能决策大脑。 - [12:28](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=748s) 万事达卡“产品接入助手”等标杆案例表明，AI 落地应坚持务实的 MVP 策略与知识动态闭环，将人机协同定位为赋能专家提效，而非盲目追求完全无人化的端到端自动化。 - [16:33](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=993s) 商业信贷的基石是依赖人际信任的“关系型银行”（Relationship Banking），AI 的定位在于增强并维系长期的微妙信用博弈，而非取代客户经理的真实纽带。 - [20:18](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=1218s) 实证研究显示，引入 AI 信用分析能够基于客观真实风险定价，消除道德风险带来的高额溢价并平抑信贷反周期波动，使长尾小微企业获得低成本流动性支持。 - [23:17](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=1397s) 万事达卡推出 Agent Pay 解决 AI 商业支付断点，在网络层建立严格安全护栏以替代脆弱的屏幕抓取，同时全面支持传统法币与稳定币自托管钱包结算。 - [28:22](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=1702s) 未来银行广泛发行 1:1 美元稳定币的模式在历史结构上高度类似于百年前的“自由银行时代”，清算网络将面临兼容数千种不同发行主体代币互通与路由的挑战。 - [30:45](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=1845s) 监管立法允许银行向储户发行稳定币无需支付利息，使银行能够将零成本资金存入美联储赚取无风险净息差，形成极具吸引力的新套利商业模式。 - [31:49](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=1909s) 企业级 AI 基础设施的核心在于系统间的连接与编排而非单体数据湖，落地需采用产品管理视角，围绕数据可用性、业务回报、实现难度与风险层级进行四维评估。 - [35:04](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=2104s) 随着自主 Agent 代表实体发起交易和信贷决策，行业正在从传统 KYC/KYB 向“KYA（了解你的智能体）”框架延伸，强化对 Agent 技术架构、底层模型与授权逻辑的穿透式合规认证。 - [38:19](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=2299s) 银行业应警惕沦为低估值的公用事业化管道服务商，须将 AI 作为复杂评估与关系型借贷的“决策增强器”，优先从业务一线的痛点环节切入以构建差异化护城河。 - [41:14](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=2474s) 从易于上手的“低垂果实”场景切入积累实战经验，能够产生飞轮效应，推动全组织建立统一的 AI 战略思维并深入探索差异化业务创新。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=4s) 研讨会开场：来自学术界与产业界的背景介绍 研讨会伊始，来自内华达大学拉斯维加斯分校（UNLV）李商学院（Lee Business School）的金融系主任兼教授 Daniel Chi 首先向在场观众致意。他热情地提到，会场就在他的家乡拉斯维加斯，从他办公室窗户就能远眺著名的拉斯维加斯大道（the Strip），距离现场还不到一英里。今年恰逢 Daniel Chi 从教二十周年，他在金融学领域深耕多年，近期正主导在全校范围内构建涵盖金融科技（FinTech）与人工智能（AI）的综合金融课程体系。他自豪地提到，UNLV 金融系在过去三年实现了惊人的增长，学生选课规模从约 1,000 人激增至本学期的 2,700 人，增幅超过 150%。学院设有商业银行专业项目并组建了商业银行咨询委员会（Commercial Banking Advisory Board），委员会成员在日常沟通中，最频繁且高度聚焦的核心话题正是人工智能在银行体系中的实际落地与影响。 紧接着，万事达卡（Mastercard）AI 与数据运营高级副总裁 Federico Cohen Freue 介绍了自己的职责。他主要负责引领万事达卡利用 AI 技术与前沿数据能力推动核心战略倡议，涵盖三大工作支柱：其一是企业整体的数据与 AI 战略，明确公司的战略押注方向并推动实际执行；其二是内部 AI 赋能，构建支撑大规模模型落地的数据基础设施，确保员工享有安全可靠的 AI 工具；其三也是他最热衷的人员组织与培训，提升全员的 AI 素养并向产业生态传达正确认知。Federico 特别强调，这不仅需要从企业内部视角出发，还要求团队深入洞察整个银行业生态体系（从支付延伸至各类业务生命周期）中的真实痛点与实践得失，以便全面赋能合作伙伴。 - **UNLV 金融系规模扩张**：过去三年学生总人数从约 1,000 人快速增长至 2,700 人，增速突破 150%。 - **万事达卡 AI 战略维度**：横跨企业战略决策、大规模数据基础设施构建、内部员工赋能培训以及面向广泛银行生态系统的能力输出。 --- ### [03:15](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=195s) 从理论走向现实：生成式 AI 与智能体的演进历程 主持人 Jonathan King（Coinbase Ventures 首席投资人）接管对话并梳理了 AI 发展脉络。他指出，虽然 AI 理论与机器学习技术最早可追溯至 20世纪 50年代的图灵测试（Turing test），但在长达数十年的时间里主要局限于实验室与后台数据模型。真正的转折点发生在 2022 年底 ChatGPT 的横空出世，它在不到 6 个月内便聚集了突破 1 亿的活跃用户，成为历史上增长最快的消费级应用。 如今，技术形态正经历由单一大型语言模型（LLM）向具有自主执行能力的“AI 智能体”（AI Agents）演进。这些智能体属于专用软件实体，具备针对既定目标与任务自主调用工具、执行多步工作流的能力。行业正在以极快的速度从早期的概念验证与理论探讨，全面迈向银行业和金融服务业的核心实际生产环境。Jonathan 随之向 Federico 发问：万事达卡以及更广泛的金融体系中，生成式 AI 和 AI 智能体目前究竟在哪些场景切实释放了生产力价值？ --- ### [04:20](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=260s) 客户服务赋能与后台繁琐流程重塑的双重范式 Federico 将 GenAI 与 AI 智能体在银行服务中的价值创造归纳为两大模式：面向客户的前台交互（Customer-facing）与面向内部运营的后台协同（Internal-facing）。在前台客户交互层面，成功落地的 Agent 应用普遍具备三大特征：“可重复性（repeatability）”、“低风险（low-risk）”以及处理“繁杂的非结构化数据（messy unstructured data）”。客户服务之所以率先爆发，是因为客服每天面临成千上万次相似提问，依赖相对固定的知识库，且非常容易引入“人在回路”（Human-in-the-loop）机制——无需 100% 全自动接管即可释放 80% 至 90% 的业务价值。对于简单咨询，系统可实现完全自动化解答；而在价值更高、更具复杂度的私人财富管理或商业银行业务中，Agent 的核心价值在于为一线客户经理（Banker）提效： ```mermaid flowchart TD accTitle: 银行前后台 AI 智能体工作流 subgraph FrontOffice["前台业务赋能"] A["多源非结构化信息"] --> B["AI 检索与上下文提取 Agent"] B --> C["商业银行经理 / 财富顾问"] C --> D["高价值客户深度沟通与交叉销售"] end subgraph BackOffice["后台运营自动化"] E["发票 / 公司章程 / 扫描 PDF"] --> F["多模态理解与制裁筛查 Agent"] F --> G{"置信度评估"} G -->|"置信度高"| H["自动化归档与执行"] G -->|"疑难 / 异常"| I["人工合规官与核保人深度复核"] end ``` 以往商业银行面临的经典顽疾，就是客户经理难以快速获取客户全貌信息、推荐适配的交叉销售产品。现在，AI 能够在对话与拜访前将背景信息提炼到位，极大免除繁重的人工作业，使银行家能聚焦于高附加值的客户沟通。 在后台流程方面，多模态模型的兴起打破了长期以来依赖纯人工的桎梏。诸如发票核对（Invoice reconciliation）、了解你的客户（KYC）以及了解你的业务（KYB）等流程，涉及大量不同版式的发票、企业注册证明及扫描 PDF 文件，同时还需要反复跨系统执行制裁名单筛查。以往若想做自动化，必须针对每种版式专门开发规则模板；而现在利用多模态 Agent，系统能自主阅读任意格式文件并抓取关键实体信息，进而调度下级 Agent 自动完成合规筛查。这种变革将核保人（Underwriter）和合规审查官（Compliance officer）从繁重的“工作流编排者”角色中解脱出来，让他们将专业精力完全集中在置信度不足、需要批判性分析与调查的疑难边缘案例上。 - **前台 Agent 落地核心特征**：高重复频次、低执行风险、高度依赖非结构化数据且支持“人在回路”。 - **商业银行知识赋能**：秒级整合客户背景与产品目录，从繁杂材料检索转变为精准交叉销售辅助。 - **后台多模态作业革新**：无需预先定义表单格式，Agent 即可跨发票、公司文件和制裁名单完成自动化编排，大幅降低 KYC/KYB 的进入壁垒。 --- ### [07:42](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=462s) 学术视角：信贷中介本质与“5C 原则”的智能突破 面对同样的价值落地问题，Daniel Chi 教授转换视角，从金融学理论层面深度剖析了 AI 对商业银行本质职能的颠覆。他指出，尽管现代商业银行的业务版图已涵盖资产管理等多元版块，但其不可动摇的核心仍然是“信贷中介”（Credit Intermediation）——吸收一方的存款并将其转化为面向另一方的贷款。从职能上看，商业银行运作信贷中介本质上分为两大连续任务： 1. **信息的汇集与处理**（Aggregate and process information） 2. **制定信贷决策**（Make credit decisions） 在过去数十年中，计算机技术的普及和算力爆发已经有效解决了第一项任务，使得海量金融数据的存储、统计与计算高度自动化。然而，第二项任务——“信贷决策”——长期以来被视为人类智慧的专属领地。除了高度标准化、金额较小的简单个人消费类贷款可以通过自动化规则审批外，绝大部分信贷决策从未能够真正委托给机器。 Daniel 解释道，如果系统学习过银行信贷学，就必然知晓衡量贷款违约风险与承贷能力的经典“5C 原则”： - **品德（Character）**：借款人的声誉、诚意与还款意愿 - **资本（Capital）**：借款人的财务杠杆与资产净值 - **还款能力（Capacity）**：借款人的现金流与日常偿债水平 - **抵押品（Collateral）**：用以担保债务的资产价值 - **环境条件（Conditions）**：宏观经济走势与行业竞争环境 在 5C 框架中，首当其冲且最核心的就是借款人的“品德（Character）”。评判一个企业主或借款人的品格极其依赖人类经验的微妙洞察与直觉判断，以往传统机器与固定代码完全无能为力。但如今生成式 AI 的突破在于，它不仅具备汇总与理解文字及多源非结构化数据的能力，更能生成超越单一离散数据点的综合洞察，这标志着人工智能正在从单纯的信息搬运工具，真正转变为能够辅助银行专家完成深层信贷决策的智能大脑。 ### [10:25](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=625s) 银行业的本质差异与智能体商业的崛起 Daniel Chi 教授深入剖析了银行业与传统商品零售平台的本质区别。他指出，像 Amazon 或 Walmart 这样的巨头，核心业务是提供实体商品的交易平台；而银行提供的则是服务与信用（Credit）。信用是一种无形资产，高度依赖人类的智能判断与复杂决策，这也正是人工智能在银行领域能够释放真正价值的落脚点。 Jonathan King 对此深表赞同，并强调整个行业正迈入由 AI 驱动的“下一代自动化”新阶段。过去极其繁复的工作流，如今正在引入具备思考与推理能力的智能体（AI Agents）。他分享了自己前一天主持的一场关于“智能体商业”（Agentic Commerce）圆桌讨论的经历：现场被问得最多的问题便是“我们是否正在将人类完全踢出业务流程？”事实恰恰相反，目前的实践是在强化人类的能力，大幅提升生产力。Jonathan 提到 Coinbase 去年发布的关于加密货币与 AI 交叉融合的研究报告，预测生成式 AI 与自然语言交互界面未来将成为人机交互及金融交易（如购买稳定币等）的首要模态，彻底重塑人类与机器在数字经济中的协作方式。 ### [12:28](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=748s) 万事达卡实战案例：产品接入助手的落地与演进 Federico Cohen Freue 详细分享了万事达卡（Mastercard）一个成功从概念验证走向规模化生产的应用——“产品接入助手”（Product Onboarding Assistant）。将银行客户接入万事达卡的全球网络是一项高度复杂的技术工程，不仅涉及底层的网络对接，还涵盖多种业务工具的启用、不同支付轨道的切换，甚至涉及法定货币与稳定币的交互。整个流程极为繁冗，需要万事达卡内部与客户的多支团队紧密协作。 ```mermaid flowchart TD accTitle: 万事达卡产品接入助手工作流 subgraph Knowledge["知识资产整合"] A["技术文档 / 产品说明"] --> C["多智能体系统<br/>Multi-Agent System"] B["产品管理笔记 / 规范"] --> C end subgraph HumanLoop["人机协同与闭环演进"] C --> D["赋能万事达卡技术人员"] D --> E["协助客户解决接入难题"] E -->|新最佳实践与反馈| F["持续知识回流"] F --> C end ``` 面对这一挑战，万事达卡并未试图追求完全无人化的端到端自动化。Federico 强调，面对面的客户沟通、需求拆解、问题预判依然需要深厚的人性化支持，大语言模型目前无法仅凭“旁听会议记录”就替代人际协作。因此，团队将 AI 的价值精准锚定在两个核心维度： 1. **海量知识的精准穿透**：智能体遍历全部技术文档、接入指南与产品管理笔记，实时推断客户当前阶段的具体诉求，并为万事达卡的支持人员提供下一步的精确操作指引。 2. **知识资产的动态闭环**：在服务不同客户的过程中，团队沉淀出的最佳实践与新洞察，会实时反哺给这套自研的多智能体系统（Multi-agent system），让系统在人机协同中持续演进，避免了传统模式下频繁人工推倒重写底层知识库的沉重负担。 - **显著成效**：通过在技术实施阶段为客户提供更直接、明确的指导，平均单次复杂的技术对接流程可削减约 28 小时的工作量。 - **产品化思维推进**：万事达卡拒绝将 AI 落地视为“要么停留在实验、要么一蹴而就实现大规模”的二元论选择，而是坚持软件工程的 MVP（最小可行产品）理念——先在一个产品线试点，挑选认同感高、执行力强的内部先锋团队反复验证并听取反馈，跑通后再逐个产品线横向拓展。即使第一版 MVP 仅能解决 50% 的痛点，也远比等待开发出完美系统能更快带来真实业务价值。 ### [16:33](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=993s) Coinbase GPT 的“爬、走、跑”路线与银行业的真实鸿沟 Jonathan King 呼应了万事达卡务实的产品化策略，指出 Coinbase 在 AI 落地中同样奉行“先爬、再走、后跑”（Crawl, Walk, Run）的递进准则。Coinbase 内部推出的 "Coinbase GPT"（CBGPT）最初仅切入极小的场景：辅助新入职的客户支持团队处理日常例行问题（如登录故障排查及相关工作流检索）。在验证了其实际成效后，CBGPT 的覆盖范围逐步推广至全公司所有员工，不仅能帮助产品经理撰写更高质量的 PRD（产品规格说明书），还承担起检索公司 Google Workspace 内部海量资产的智能企业搜索中枢。 转向学术界与更广泛的商业银行时，Daniel Chi 教授现场进行了一次趣味互动，询问台下观众有多少来自银行业，在看到大量银行从业者举手后，他坦言就当前银行界的实际情况来看，绝大多数应用仍停留在简单的初级自动化层面：许多被贴上 AI 标签的工具，与本轮大模型爆发前的既有系统并无本质差别。 Daniel 指出，商业贷款历来是银行业最核心的“现金牛”与盈利支柱，而商业贷款的核心基石是“关系型银行”（Relationship Banking）。银行绝不可能让冰冷的计算机取代人际信任纽带，其真正诉求是“增强”这种关系，但目前技术离该目标仍有相当漫长的距离。他援引国际清算银行（BIS）关于 AI 与关系型借贷的最新工作论文解释道：商业信贷的百年潜规则是“把钱借给不需要钱的人；而在客户真正缺钱时，银行反而不愿贷款以规避下行风险”。因此，传统客户经理通常在市场景气、风险较低的顺周期阶段向企业收取高于实际风险溢价的费用，以此储备空间，好在逆周期与危机时刻继续向真正需要资金的企业注资履约。AI 能否在理解并维系这种极为微妙的长期信用博弈中发挥作用，依然是整个行业面临的未解课题。 ### [20:18](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=1218s) AI 工具平抑信贷反周期性与解决信息不对称 Daniel Chi 教授深入探讨了银行业传统信贷周期的痛点。在传统的银行业逻辑中，信贷供给和利率具有明显的反周期性（counter-cyclicality）：银行通常以“我们在经济下行时为您提供下行保护”为由，在经济形势好时向借款企业收取更高的风险溢价；而当企业在危机中最需要资金时，往往面临高昂的资金成本。然而，一项基于意大利银行样本的最新学术研究显示，银行在发放贷款前引入 AI 信用分析工具后，有效平抑了这种反周期性——在经济向好时向企业收取更低的利率，而在经济下行时依然能以常规水平的合理利率向企业提供信贷。 这一改变从根本上重塑了借贷双方的关系。AI 信用分析依赖客观数据、能更精准地量化实际风险，使银行可以按照资产真实的风险水平定价，而无需额外附加高额溢价来对冲信息不对称或借款人的道德风险（moral hazard）。这不仅增强了银行的定价竞争力，也让企业借款人享受到了更低的融资成本。 Federico Cohen Freue 对此表示高度赞同，并结合万事达卡（Mastercard）在中小微企业（SME）领域的实践补充道，全球小微企业普遍面临“极度渴求快速、灵活获取资本，但自身财务信息往往不完备”的痛点。AI 能够解析大量非结构化、非传统数据（non-traditional data），将过去仅服务于大型企业的大额批发关系银行业务（wholesale commercial banking）能力延伸至长尾小微企业，大幅提升审批效率与资金流动性。 ```mermaid flowchart TD accTitle: AI in Credit Pricing and Information Asymmetry subgraph Traditional["传统信贷模式 (Traditional Lending)"] T1["信息不对称与道德风险<br/>(Information Asymmetry)"] --> T2["附加高额风险溢价<br/>(Risk Premium)"] T2 --> T3["信贷供给反周期波动<br/>(Counter-cyclical Credit)"] end subgraph AI_Enabled["AI 赋能信贷模式 (AI-Driven Lending)"] A1["多模态/非结构化数据输入<br/>(Unstructured Data)"] --> A2["AI 风险精准量化模型<br/>(Objective Risk Analysis)"] A2 --> A3["真实风险定价 + 利率平抑<br/>(Risk-Reflective Pricing)"] end subgraph Impact["市场影响 (Market Outcomes)"] A3 --> M1["银行端：提升定价竞争力"] A3 --> M2["小微企业端：快速获得低成本流动性"] end ``` - **研究样本**：意大利银行业部署前置 AI 信用分析工具的实证数据。 - **机制突破**：消除道德风险溢价，使定价完全基于资产实际风险。 - **中小企业赋能**：长尾小微企业通过非结构化数据评估，获得接近大型企业客户的信贷体验。 --- ### [23:17](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=1397s) AI Agent 与稳定币交汇：Agent Pay 与自动化商业 作为加密风投领域的主持人，Jonathan King 顺势引出了稳定币（Stablecoins）与 AI Agent 的融合议题。他将稳定币定义为可编程资金（programmable money）与 24/7 全天候即时跨境结算的数字美元，并提出前瞻性问题：当具备自主保管钱包（self-custody wallet）的 AI Agent 能够代表企业和个人直接进行价值交换时，银行与支付基础设施该如何演进？ Federico Cohen Freue 详细介绍了万事达卡在 4 月底推出的 **Agent Pay**。他指出，当前消费者的产品搜索与发现行为正加速向 AI 迁移——用户不再只是在搜索引擎输入“最好的跑鞋”，而是直接向 Agent 详述自己的跑步习惯、步态（gait）特征及偏好，并获取精准推荐。然而，现有的 LLM 体验在“支付结算”这一环节形成了断点。部分工具试图通过脆弱且存在安全隐患的屏幕抓取（screen scraping）代填卡号，这在大规模商业化场景下无法满足安全标准。 ```mermaid flowchart LR accTitle: Agent Pay Transaction Flow and Trust Architecture subgraph Discovery["AI 交互与发现"] User["消费者需求"] --> Agent["AI Agent (决策推荐)"] end subgraph Execution["安全支付执行 (Agent Pay)"] Agent --> Guard["Agent 级安全护栏"] Guard --> Net["万事达卡网络层验证<br/>(Mastercard Network)"] end subgraph Settlement["多元结算选择"] Net --> Card["传统刷卡/法币账户"] Net --> Stable["稳定币自托管钱包"] end ``` Federico 强调，万事达卡的底线是让“Agent 驱动的商业（Agentic Commerce）”达到与当下人们线下刷卡或手机轻触支付（Tap to Pay）相同的信任度与安全基准。Agent Pay 不仅在 Agent 层面和网络层面建立了严密的防护栏（guardrails），同时坚持“支付选择权”原则——既支持传统法币，也全面接入稳定币钱包链路，顺应多国正在成型的监管框架。 - **Agent Pay 定位**：万事达卡网络层能力的演进，允许 AI Agent 在获得安全授权后代表人类完成合规支付。 - **痛点替代**：淘汰安全性极差的屏幕抓取（screen scraping）支付模式。 - **多渠道共存**：AI 代理支付与实体卡线下 Tap 支付长期并行，稳定币作为底层受支持资产之一无缝嵌入。 --- ### [28:22](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=1702s) 稳定币与自由银行时代：学术视角的历史镜鉴 Daniel Chi 教授从金融历史和制度经济学角度对稳定币监管与发行做出了客观剖析。针对美国最新推进的 GENIUS Act（天才法案）等立法框架中关于“允许银行发行由美元 1:1 储备支持的稳定币”的规定，Chi 教授指出，这一场景在未来数年可能演变成“全美涌现数千种由不同银行发行的稳定币”，而万事达卡等清算网络将不得不处理数千种面值完全相同（均为 1 美元）但发行主体各异的代币。 Chi 教授将这一现象与 100 多年前美国“自由银行时代（Free Banking Era）”的金本位纸币体系进行了生动对比。在美联储统一发行联邦储备券（Federal Reserve Note）之前，富国银行（Wells Fargo）、摩根大通（J.P. Morgan）以及各地方银行均发行各自的银行券（Bank Notes），理论上均以黄金作为 100% 储备兑付支撑，但最终该分散体系因效率和挤兑风险被单一法币体系取代。 - **历史类比**：现在的“银行发行稳定币”在结构上高度相似于百年前各银行发行的黄金储备代金券。 - **锚定本质**：早期银行券锚定黄金，而现在的稳定币锚定已脱离金本位的法币（美元）；稳定币的价值稳定性本质上等同于美元自身的购买力稳定性。 - **清算挑战**：未来网络需兼容成百上千种由不同金融机构发行的 1:1 美元稳定币资产的互通与路由。 ### [30:45](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=1845s) 稳定币免息红利与银行的新套利模式 在探讨加密货币与传统金融的结合点时，关于稳定币监管法案中的关键条款引发了深入讨论。相关法案中存在一条对银行机构极为有利的规定：当银行向其客户或储户发行稳定币时，无需为其支付利息。相比之下，传统的支票账户（Checking Account）需要支付少量利息，储蓄账户（Savings Account）则需支付更高一些的利息；如果银行能够发行稳定币，即可实现“零利息支出”。 这种资金成本结构解释了为何银行业突然开始集体拥抱稳定币。银行发行稳定币吸收资金后，可以将这笔资金存入美联储系统，获取美联储当时约为 4.25% 的年化利率，而对储户端支付的利息却为零。这种零成本吸储与美联储无风险利率之间的净息差，构成了一个极具吸引力的套利商业模式。 - **法案免息优势**：银行向储户发行稳定币无需支付利息，打破了传统活期与储蓄账户的付息刚性成本。 - **利差套利空间**：银行将发行稳定币获得的资金存入美联储系统，可无风险赚取约 4.25% 的年化收益率。 --- ### [31:49](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=1909s) 万事达卡的数据底座建设与产品化 AI 推进路径 谈到生成式 AI 的落地实操与基础设施挑战时，万事达卡（Mastercard）分享了其推进企业级 AI 战略的核心支柱。数据战略是整个体系的基石，但这并不意味着所有底层数据必须在起步时就做到完美无缺，而是要建立起全局性的持续治理流程。万事达卡设立了专门的数据战略团队，深入各个业务部门梳理数据资产、规范存储方式，确保数据在合规与负责任的前提下被充分利用，并精准识别出哪些是最具业务价值、应优先接入的高优先级数据。 在技术设施方面，企业需要构建能够支撑关键数据汇聚的通用架构模式，但这绝非单纯搭建一个“数据湖”，核心在于实现底层系统之间的高效编排与连接。在具体执行层面，团队采用严格的产品管理（Product Management）视角，由产品经理深入各个业务场景，系统评估数据可用性、业务回报率、开发难度以及风险等级。尽管 AI 的非确定性（Nondeterministic）特征和新兴的智能体系统开发需要特定专长，但将成熟的传统技术落地方法论迁移至 AI 领域，依然是确保项目成功的关键准则。 ```mermaid flowchart TD accTitle: Mastercard Enterprise AI Implementation Architecture subgraph DataStrat["数据战略治理层 (Data Strategy)"] A["业务部门数据盘点"] --> B["合规与负责任使用规范"] B --> C["高价值数据优先级识别"] end subgraph InfraOrch["基础设施与编排层 (Infrastructure)"] C --> D["企业核心数据底座"] D --> E["数据连接与编排引擎 (Orchestration)"] end subgraph ProductLens["产品化落地评估层 (Product Lens)"] E --> F["业务用例深入调研"] F --> G["多维可行性评估<br>(价值 / 难度 / 风险)"] G --> H["AI 智能体与应用交付"] end ``` - **全生命周期数据治理**：数据战略并非一次性项目，而是跨部门协同的持续性规范建设。 - **重在编排而非单体数据湖**：构建 AI 基础设施的关键在于打通系统间的数据连接与流转编排。 - **产品化评估模型**：围绕“数据可用性”、“业务回报预估”、“技术实现难度”与“风险层级”四维矩阵筛选落地上线项目。 --- ### [35:04](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=2104s) 智能体合规框架与“了解你的智能体”（KYA）机制 随着大量自主 AI 智能体代表企业和个人用户发起交易、保管资产并进行高频价值流转，合规与监管治理框架的演进迫在眉睫。银行业界达成的普遍共识是，无论监管层还是银行自身，都必须先建立起清晰的“合规与负责任（Ethical and Responsible）”AI 使用政策，管理层必须在团队开发或引入工具之前完成制度定调。虽然银行处于强监管行业，但 AI 本身正被用于自动化生成合规报告与审计流水；鉴于当前银行业 AI 监管整体仍处于萌芽阶段，金融机构多采取与监管方协同探索、共同制定行业标准的稳妥路线。 针对传统 KYC（了解你的客户）和 KYB（了解你的企业）在自主智能体时代的延展，业界提出了 KYA（Know Your Agent，了解你的智能体）框架构想。当智能体开始独立代表实体做出信用评估与金融交易决策时，验证智能体背后的技术工具链、底层模型以及授权逻辑，与当前企业合规体系中审查供应商使用的技术工具逻辑如出一辙，KYA 正在成为金融机构风控体系的自然延伸。 - **制度先行原则**：管理层必须在 AI 工具技术部署之前，确立负责任与伦理治理政策底线。 - **监管协同与自动化合规**：AI 工具在反洗钱、审计和合规报表自动化方面展现潜力，银行业正与监管机构共同确立早期规范。 - **KYA（Know Your Agent）机制**：随着智能体承接交易与信贷决策，针对智能体身份、所用技术架构与权限归属的穿透式合规认证成为必然。 --- ### [38:19](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=2299s) 告别同质化：用 AI 赋能复杂决策与深层客户关系 对于希望推动生成式 AI 与智能体从试验走向落地的银行从业者，核心策略在于明确竞争定位与聚焦真实痛点。从 2007-2008 年金融危机以来的相当长时期内，公开上市银行的估值长期承压，市盈率往往徘徊在 10 倍左右，而公用事业公司（Utility Firms）的市盈率却高达 20 倍。部分分析师曾鼓吹银行应像公用事业公司一样被估值，但这意味着银行沦为了提供标准化商品的管道服务商。如果金融产品完全同质化为大宗商品，银行将只能在价格战中恶性循环。 银行建立护城河的关键在于差异化，尤其是商业贷款、企业融资以及依赖深层人际互动与专业信任的“关系型借贷”（Relationship-based Lending）。AI 的真正价值不是取代银行家，而是充当“决策增强器”，帮助从业人员更好地评估不确定性、辅助复杂的授信与风控决策，并深化与客户的商业纽带。对于刚开始 AI 探索的团队而言，务实的起步方式是紧盯业务一线员工当前面临的最大痛点与低效瓶颈，优先通过技术消除流程摩擦、提升运转效率。 - **警惕公用事业化陷阱**：避免产品沦为同质化的大宗商品以致陷入恶性价格竞争。 - **聚焦差异化核心资产**：将资源投向需要复杂评估的商业信贷与关系型借贷场景。 - **AI 定位为能力增强**：运用 AI 辅助评估复杂不确定性、提升决策质量，而非简单替代人工。 - **以一线痛点为切入点**：优先改造流程冗长、员工挫败感强烈的低效环节，快速验证落地价值。 ### [41:14](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=2474s) 从低垂果实到差异化创新：AI 实践演进与圆桌总结 在探讨如何推动全员拥抱技术变革时，嘉宾进一步阐明了快速启动与持续积累实战经验的重要性。通过在早期切入那些直观且易于上手的“低垂果实”（low-hanging fruit），团队能够在反复的实践与迭代（repetitions）中积累实战信心。这种由浅入深的落地过程，能够切实赋能并启发整个组织内的员工，促使大家像 Daniel Chi 教授所强调的那样去深入思考更深层次的战略议题：“我们该如何实现差异化？在哪些业务节点实现差异化？又该如何通过这项技术达成这一目标？”正是这种从简到繁的积累，为整个组织构建起统一的 AI 思维方式奠定了坚实基础。 在圆桌会议的尾声，主持人向参与讨论的嘉宾 Daniel Chi 教授与来自万事达卡（Mastercard）的 Federico Cohen Freue 表达了由衷的感谢，并向现场的所有参会者致意。他由衷感慨，在场的每一位从业者都在共同推动银行业与金融科技前沿边界的拓展（pushing the frontier forward），并祝愿大家在接下来的 Ai4 大会日程中收获满满。 - **快速实践的飞轮效应**：从简单且显而易见的场景切入积累实战经验，是推动全组织建立“差异化竞争意识”和掌握高阶 AI 赋能方式的前提。 - **推动行业前沿**：呼吁银行业与金融机构从业者继续勇于尝试、探索真实落地，共同拓宽生成式 AI 与 AI Agent 的能力边界。 ## Notable quotes > [10:25](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=625s) 我们是否正在将人类完全踢出业务流程？事实恰恰相反，目前的实践是在强化人类的能力，大幅提升生产力。 — Jonathan King > [16:33](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=993s) 商业贷款历来是银行业最核心的“现金牛”与盈利支柱，而商业贷款的核心基石是“关系型银行”（Relationship Banking）。 — Daniel Chi > [16:33](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=993s) 把钱借给不需要钱的人；而在客户真正缺钱时，银行反而不愿贷款以规避下行风险。 — Daniel Chi > [20:18](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=1218s) AI 信用分析依赖客观数据、能更精准地量化实际风险，使银行可以按照资产真实的风险水平定价，而无需额外附加高额溢价来对冲信息不对称或借款人的道德风险。 — Daniel Chi > [23:17](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=1397s) 万事达卡的底线是让“Agent 驱动的商业（Agentic Commerce）”达到与当下人们线下刷卡或手机轻触支付（Tap to Pay）相同的信任度与安全基准。 — Federico Cohen Freue > [31:49](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=1909s) 数据战略是整个体系的基石，但这并不意味着所有底层数据必须在起步时就做到完美无缺，而是要建立起全局性的持续治理流程。 — Federico Cohen Freue > [35:04](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=2104s) 管理层必须在团队开发或引入工具之前完成制度定调。 — Federico Cohen Freue > [38:19](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=2299s) 银行建立护城河的关键在于差异化，尤其是商业贷款、企业融资以及依赖深层人际互动与专业信任的“关系型借贷”。 — Daniel Chi > [38:19](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=2299s) AI 的真正价值不是取代银行家，而是充当“决策增强器”，帮助从业人员更好地评估不确定性、辅助复杂的授信与风控决策，并深化与客户的商业纽带。 — Daniel Chi > [41:14](https://www.youtube.com/watch?v=S4BEP_0Tu44&t=2474s) 通过在早期切入那些直观且易于上手的“低垂果实”，团队能够在反复的实践与迭代中积累实战信心。 — Federico Cohen Freue ## Mentioned Ai4, Daniel Chi, Federico Cohen Freue, Jonathan King, Mastercard, Coinbase Ventures, University of Nevada-Las Vegas (UNLV), Lee Business School, ChatGPT, Coinbase GPT (CBGPT), Agent Pay, GENIUS Act, Federal Reserve, Bank for International Settlements (BIS), Wells Fargo, J.P. Morgan, Amazon, Walmart, Google Workspace ## Video description Federico Cohen Freue, Senior Vice President, AI & Data Operations, Mastercard Jonathan King, Principal Investor, Coinbase Ventures Daniel Chi, Professor and Chair, Department of Finance, University of Nevada-Las Vegas Abstracts Banks are rapidly moving from experimentation to real deployment of GenAI and AI agents—but success requires more than just the latest model. In this panel, banking and technology leaders will explore how financial institutions are leveraging GenAI to streamline operations, enhance customer service, improve risk analysis, and boost productivity across the enterprise. We’ll discuss the infrastructure, data strategy, and governance models needed to support safe and scalable rollouts, as well as key lessons from early use cases in front-office and back-office functions. Whether you’re building a virtual banker or automating internal workflows, this session will provide a clear look at what it takes to make GenAI real in banking. [Source](https://www.youtube.com/watch?v=S4BEP_0Tu44)

[10:25] Jonathan King: 我们是否正在将人类完全踢出业务流程？事实恰恰相反，目前的实践是在强化人类的能力，大幅提升生产力。 [16:33] Daniel Chi: 商业贷款历来是银行业最核心的“现金牛”与盈利支柱，而商业贷款的核心基石是“关系型银行”（Relationship Banking）。 [16:33] Daniel Chi: 把钱借给不需要钱的人；而在客户真正缺钱时，银行反而不愿贷款以规避下行风险。 [20:18] Daniel Chi: AI 信用分析依赖客观数据、能更精准地量化实际风险，使银行可以按照资产真实的风险水平定价，而无需额外附加高额溢价来对冲信息不对称或借款人的道德风险。 [23:17] Federico Cohen Freue: 万事达卡的底线是让“Agent 驱动的商业（Agentic Commerce）”达到与当下人们线下刷卡或手机轻触支付（Tap to Pay）相同的信任度与安全基准。 [31:49] Federico Cohen Freue: 数据战略是整个体系的基石，但这并不意味着所有底层数据必须在起步时就做到完美无缺，而是要建立起全局性的持续治理流程。 [35:04] Federico Cohen Freue: 管理层必须在团队开发或引入工具之前完成制度定调。 [38:19] Daniel Chi: 银行建立护城河的关键在于差异化，尤其是商业贷款、企业融资以及依赖深层人际互动与专业信任的“关系型借贷”。 [38:19] Daniel Chi: AI 的真正价值不是取代银行家，而是充当“决策增强器”，帮助从业人员更好地评估不确定性、辅助复杂的授信与风控决策，并深化与客户的商业纽带。 [41:14] Federico Cohen Freue: 通过在早期切入那些直观且易于上手的“低垂果实”，团队能够在反复的实践与迭代中积累实战信心。

[4:20] Federico Cohen Freue: repeatability, low risk, and messy unstructured data. [12:28] Federico Cohen Freue: Product Onboarding Assistant [16:33] Jonathan King: Coinbase GPT (CBGPT) [23:17] Federico Cohen Freue: Agent Pay [35:04] Jonathan King: Know Your Agent [38:19] Daniel Chi: downward spiral [41:14] Jonathan King: repetitions [41:14] Jonathan King: pushing the frontier forward

[0:04] Good morning everyone. [0:06] Morning. All right. Um, good to see you [0:09] in my hometown. I'm just my office just [0:12] not even a mile away that direction. Um, [0:15] I I see the strip every day from my [0:17] window. So, um, I'm Daniel Chi. I'm [0:20] professor and chair of the finance [0:22] department in the Lee business school at [0:25] UNLV, University of Nevada, Las Vegas. [0:28] Um so I've been a professor for 20 [0:33] years. This year is my uh exact 20 year [0:36] anniversary as a professor. Um so I've [0:39] been teaching and research on various [0:42] finance topics. And most recently um I'm [0:45] uh leading the effort to build out a [0:48] comprehensive finance curriculum on [0:51] financial technology and artificial [0:53] intelligence. Um do I have a minute to [0:57] tell you about uh my department? Okay. [1:01] Um just to to my horn. Um so our [1:04] department um the finance department has [1:08] seen a tremendous growth in the last [1:09] three years. Our enrollment has [1:11] increased from about a thousand students [1:14] to um this semester is 2,700 students. [1:18] So uh over 150% increase. uh we have a [1:23] commercial banking program and we have a [1:25] commercial banking advisory board. I [1:28] think that's one reason I'm sitting [1:29] here. Um I communicate uh regularly with [1:33] my banking advisory board uh members. Uh [1:36] and AI is absolutely a essential topic [1:40] that they are talking about. [1:42] >> Awesome. Thanks. Awesome. Um so I I lead [1:45] a team at Mastercard called AI and data [1:47] operations. um that that really means a [1:50] couple of things, but it's ultimately [1:53] responsibility over how Mastercard is is [1:56] using AI and advancements in data to [2:01] push our strategic initiatives forward. [2:03] So, what that means in practice is I I [2:05] focus on our enterprise and data [2:07] strategy. So, where are we going? What [2:09] does it mean for us? What are the [2:10] strategic bets we're willing to place? [2:12] And then actually execute on some of [2:13] those bets. um how we use AI internally. [2:18] So how are we deploying it for ourselves [2:20] and to to the benefit of our employees [2:22] and making sure that our employees have [2:24] access to it as well as uh building our [2:26] data infrastructure uh that powers it [2:29] because you can't really scale it [2:31] without having like a data [2:32] infrastructure at scale. And then [2:34] probably my favorite part of of the role [2:36] is the people side of things. So how do [2:39] we train all our employees to to use it? [2:41] What do they need to know? Uh how do we [2:43] communicate about it? what do we you [2:45] know what messages do we want to make [2:46] sure are being heard in the ecosystem [2:47] which you know with with respect to [2:49] today not only do I have an internal [2:51] lens but it's actually required me and [2:53] my team to have a really really good [2:55] understanding of how is it being used [2:57] across the banking sector not just in [2:59] payments but across the life cycle [3:01] because we engage with the with [3:03] institutions in the ecosystem across all [3:05] of these fronts and so understanding [3:07] what needs they are what's working [3:08] what's not working has been actually [3:09] pretty important for our understanding [3:11] and to support them as best as we [3:13] Awesome. [3:15] Thank you guys for the for the intro. [3:16] So, you know, as I think about like the [3:18] broader AI sector, you know, in the [3:20] popularity that it's gained today. I'm [3:22] like, well, you know, technically AI has [3:24] been around since the 1950s, right? If [3:27] you think about the original Turing test [3:28] and like a lot of focus around machine [3:30] learning, but it wasn't until 2022 when [3:33] we saw the emergence of chat GPT and [3:35] chat GBT actually became the fastest [3:37] growing consumer app, right? I think I [3:39] saw a stat that they grew to 100 million [3:41] users in less than 6 months or so, [3:44] right? So, fast forward to now, right? [3:47] We have all of these LLMs. We're now [3:50] starting to see the rise of AI agents or [3:52] like these autonomous software entities [3:54] that are purpose-built to accomplish a [3:56] said goal or task. And so, we're quickly [3:58] moving from kind of theory to reality. [4:02] And that's where I want to focus on the [4:03] the real practical you know um [4:06] implications for AI within the banking [4:08] and financial services sector. So [4:10] Frederrico I want to start with you in [4:11] terms of like what are you seeing right [4:13] in terms of Gen AI or AI agents and how [4:16] they're driving real value today in the [4:18] banking and financial services sector. [4:20] >> It's a it's a great question. Uh I would [4:22] I would summarize it in in two patterns. [4:24] One that is customerf facing and one [4:25] that's internally facing. So customerf [4:27] facing the the general pattern that [4:29] we're seeing with all the more advanced [4:31] agentic applications that are resonating [4:34] is really about a couple of things. It's [4:36] about repeatability, lowrisk and like [4:40] messy unstructured data. And so which is [4:42] why you know when you when you look at [4:44] the media you hear things like customer [4:46] support. Why is customer support being [4:48] one of the the the initiatives where [4:51] this technology and this these methods [4:53] seem to be resonating is because well [4:54] you do it hundreds thousands hundreds of [4:56] thousands of times a day typically with [4:58] the same corpus of information typically [5:00] with some little bit of of context and [5:03] it's easy to have a human in the loop [5:05] right you don't have to fully automate [5:07] it to get 80 90% of the overall value [5:10] and so uh on a consumerf facing basis or [5:13] on a customerf facing basis we're seeing [5:16] like reducing a lot of the friction in [5:17] customer support. That means like simple [5:20] questions and things to ask fully [5:22] automating them because again they're [5:23] low risk. They don't necessarily require [5:25] an action but also one of the ones that [5:27] doesn't get talked about a lot but is is [5:29] getting quite a lot of traction is [5:31] getting the right information to your [5:34] bankers and to your customer service [5:35] representatives. So think about a wealth [5:37] manager or a commercial banker. One of [5:39] the classic big problems in commercial [5:42] banking has been how do I get my banker [5:44] the right information? what products to [5:46] cross-ell, what context they need to [5:48] have to the client and all of a sudden [5:49] now that is being much more actively [5:52] embedded and you are freeing and [5:54] empowering the banker to have much more [5:56] meaningful conversations and take out a [5:59] lot of the arduous work that they used [6:00] to be there before. So customerf facing [6:03] low-risk information is is one. On an [6:07] internal facing basis, one that I I [6:10] think is starting to get a lot of [6:11] traction as well that we're seeing, [6:13] we're starting to to look into it as [6:15] well is there's a number of, you know, [6:18] very important back office processes [6:21] that up until a year or two ago before [6:23] we had multimodal models [6:25] were very manual. Think about things [6:27] like invoice reconciliation. Think about [6:30] doing KYC or KYB. Typically, that [6:33] requires gathering information from [6:35] somebody, but also it's not always text. [6:37] It means you have an invoice. You have [6:38] to look at incorporation documents. You [6:40] have to actually search databases for [6:42] sanction screening. That is starting to [6:44] become quite interesting because all of [6:46] a sudden, now you can actually have an [6:48] agent do your database screening. You [6:50] could have an agent do, you know, [6:51] looking at a number of PDFs that you get [6:54] sent. And the barrier to entry to that [6:57] is much lower now because you don't have [6:59] to prepare a system to understand what [7:01] the document's going to look. You can [7:03] just say I will give you a document, [7:05] figure out where these five pieces of [7:06] information are, give it to another [7:08] agent. And so again, you're freeing the [7:12] underwriter, the you know, the [7:14] compliance officer to go away from [7:16] having to do all that orchestration work [7:18] to actually work on the really hard ones [7:22] where like where you don't have [7:23] confidence on what to do where you [7:25] actually need a much more critical eye [7:26] to do further research. And so it's I I [7:29] think you know it's fundamentally [7:31] changing the type of activities that [7:33] these people do honestly to focus more [7:35] on their specialty rather than the [7:37] orchestrating or manual efforts that had [7:39] to be done just because of the systems [7:40] we had. [7:42] >> Thank you so much for that. And [7:43] professor Chi similar question right [7:45] like where are you seeing real value? [7:46] What are you hearing from either [7:47] students or some of your university [7:48] partners? [7:49] >> Yes. Yes. So I'm I'm going to that's [7:52] great perspective uh from a practitioner [7:55] u point of view. I'm going to um have a [7:59] slightly more academic view on AI and [8:01] banking. You know, if we think about [8:02] banking, what banking is um now nowadays [8:06] banks are more diverse. Uh commercial [8:08] banking plus asset management plus other [8:11] things. But traditionally, uh we needed [8:14] banks to do what we call credit [8:16] intermediation. [8:18] They take deposits from someone and they [8:21] make loans to someone else. Right? This [8:23] credit intermediation is sort of the [8:26] core of banking or nowadays we call [8:28] commercial banking. So we think about [8:30] two tasks that commercial banking needs [8:34] to do. One is to process to aggregate [8:37] and process information [8:39] and second is to make credit decisions. [8:43] Right? So now we think about how can AI [8:47] help one or both of that task those [8:49] tasks. Well, in the past we had [8:51] automation, [8:53] we had computers, right? Since many [8:56] decades ago, we had automation, [8:58] computers, computing powers that really [9:01] helped us pro aggregate and process the [9:04] information more efficiently. [9:07] And now the second task which is to make [9:10] credit decisions that is really human [9:12] intelligence, right? So up to now we [9:16] have not been able to delegate the [9:18] credit decisions. um other than except [9:21] for those very simple you know consumer [9:23] standardized consumer loans right we [9:25] have not been able to delegate those uh [9:27] credit decisions to computers because [9:30] they don't they did not have what we [9:32] call intelligence a human intelligence [9:34] because if you think about credit [9:36] decisions you know if if you study [9:38] banking you know the five C's of credit [9:41] of making a credit or making a loan the [9:44] five C's are character [9:48] capital, capacity, [9:51] collateral and conditions and the the [9:54] first one is character. How? So that is [9:56] very human intelligent intelligence [10:01] based. How how can you assess a [10:03] borrower's character with a computer [10:06] right you it's difficult but nowadays I [10:09] think for for uh gen AI that can [10:13] actually aggregate comprehend and create [10:16] something that is beyond simple data [10:19] points that can aid the second task of [10:23] of commercial banking which is making [10:25] credit decisions. I think that is where [10:27] the real value is to banking. uh because [10:30] after all when you think about what is [10:32] the difference between banking and the [10:34] other sectors say Amazon or or Walmart [10:39] right it's very different Amazon and [10:41] Walmart they provide a platform for [10:45] merchandise [10:46] banking provide a platform for for [10:49] service which is and credit which is [10:52] which is intangible and and very very um [10:55] dependent on human intelligence so I [10:58] think that's the real value is of AI to [11:00] bank. [11:01] >> Awesome. Yeah, I definitely resonate [11:03] with with both of your points, right? I [11:04] think like, you know, just kind of [11:06] reemphasize this notion of one, we're [11:08] kind of like entering this like phase of [11:10] like next generation automation powered [11:13] by AI where we're taking like what [11:15] traditionally used to be these very [11:16] complex workflows and we're introducing [11:18] these agents and these more intelligent [11:20] thinking machines into the mix. And a [11:23] key point it was uh you know I was doing [11:25] a round table yesterday about you know [11:26] just the rise of agentic commerce and a [11:29] question that always came up is like are [11:30] we completely removing humans out of the [11:32] loop and I think what we're hearing is [11:34] that we're actually augmenting what us [11:36] as human users can do right and we're [11:38] kind of making things more productive [11:40] and one of we see it very similarly at [11:43] at Coinbase right I think you know we [11:45] wrote a thesis last year about the [11:46] intersection of crypto and AI and one of [11:49] the things that we identified was that [11:50] you know we actually see generative AI [11:52] and natural language interfaces becoming [11:54] the tri the primary transaction modality [11:57] right to whether you're engaging with [11:58] finance you're buying stable coins etc [12:01] and so I really think it's going to [12:03] transform you know the way like humans [12:05] and machines interact within the digital [12:07] economy so with that framing in mind my [12:10] next question just really you know we [12:11] kind of talked about okay these are some [12:13] of the benefits this is where we can see [12:14] value now let's talk about the [12:17] applications of that value so ferico [12:19] maybe start by sharing like what's like [12:21] one example use case that you've seen go [12:23] all the way from pilot to production and [12:25] you know what made it work. [12:28] >> That's a good question. So [12:31] I'll I'll share I'll share a couple of [12:32] observations. One one use case that [12:36] we've seen [12:38] we've seen go from you know kind of [12:39] basic to scale. Uh we've talked about it [12:42] externally in in the past we call our [12:44] product on boarding assistant. So think [12:46] about our business when we are [12:48] onboarding banks onto the network that's [12:50] a highly technical process right and it [12:53] it it can involves a number of things [12:55] just you know simple plugging into the [12:57] network all the way to using enablement [12:59] of other tools you may be interacting [13:00] with different rails you may be using [13:02] fiat currency stable coins so it's it's [13:05] quite complicated depending on the [13:07] pattern you have to do and it's an [13:09] involved process that requires [13:10] partnership with a number of teams at [13:12] our with our customers so we said how [13:14] can How do we make that a better [13:16] experience? It's complicated. It's [13:18] technical in nature. And what we said [13:20] is, you know, much like what you what [13:22] you just said, we don't want to fully [13:23] automate this process. We actually think [13:25] there's a very important human component [13:27] in there, which is working with the [13:28] client, understanding what they need, [13:31] solving problems, anticipating what [13:33] could come. And we just we just the way [13:35] this works is you have to work amongst [13:37] teams. And so you can't today is easily [13:40] interject, you know, an LLM even even [13:42] listening to calls and really do that. [13:44] So we said where where's the real value [13:46] for for artificial intelligence there [13:48] and we said the real value is twofold. [13:52] It's one actually going through and [13:54] identifying all of the corpus of [13:56] knowledge the technical documentation we [13:58] have product onboarding material product [14:00] management notes etc and identify like [14:03] what does the customer need or what does [14:05] our person need to provide to the [14:07] customer to move them through the next [14:09] step. But two, it's an ever evolving [14:11] process, right? So as you work with [14:13] different clients, you'll come up with [14:14] new best practices, you refine it. And [14:16] so we wanted to make sure that as part [14:18] of that, we fed that that emerging [14:20] intelligence back into the system. And [14:22] so we've developed our own multi- aent [14:24] system that effectively does that. So [14:26] when when we have um when when we're now [14:30] we're supporting our customers in [14:32] onboarding onto new products, we use [14:34] that tool behind the scenes to empower [14:36] our people. And what we found is like [14:38] over time because we are involving [14:42] uh new intelligence into the system as [14:44] part of just where the human is in the [14:45] loop the system gets better over time as [14:48] well and learns and modifies based on [14:50] how our products are evolving. So it [14:52] hasn't tasked us with you know [14:54] consistently replacing the corpus of [14:57] knowledge or updating everything but [14:58] rather we can do it in a much more [14:59] iterative way. And so how did we get to [15:02] scale? We actually started with one [15:03] product. We say we're going to get it [15:05] right. we're going to listen to a very [15:07] kind of convict uh you know a group of [15:10] um agents on our end that have a lot of [15:12] high conviction for how to do this that [15:13] they're excited about doing this and we [15:15] spent quite a bit of time testing it [15:17] before we started expanding on a product [15:19] byproduct basis uh and it's working [15:21] right so now what we found is that on [15:23] average you know these lengthy [15:24] integration processes we're we're [15:26] eliminating around 20 28 hours on [15:29] average from the technical [15:30] implementation [15:32] um and that's good for a number of [15:33] reasons but the primary one is we're [15:35] just helping our clients get to where [15:37] they need to go faster, right? And it's [15:39] a better experience because they get [15:41] more direct answers. They get more [15:43] defined guidance from us. And I think [15:45] it's it's those things that work. What I [15:48] will highlight under the scenes is is [15:49] this undertone of we tested it with one [15:51] product, we got it right, we listen. [15:53] taking a product management approach to [15:56] doing this I I I think is quite [15:58] important because [16:00] um you know I think what I what I'm [16:03] personally seeing in the space a lot is [16:04] we talk a lot about well have you either [16:06] experimented or have you done it at [16:07] scale and and it's not a binary decision [16:10] you have to go like with any modern [16:12] software product you have to go through [16:14] steps get an MVP out prove it right [16:16] listen to your c to your users and start [16:19] refining from there that gets you to the [16:21] outcome faster too right because you put [16:22] something in people's hands and that [16:24] maybe solves half the half the the task [16:26] you're trying to solve, but it gets you [16:27] solving half the task before you, you [16:29] know, um much quicker than you'd solve [16:31] the the full amount. [16:33] >> I like the approach there in terms of, [16:35] you know, we think about it very [16:36] similarly at Coinbase, like taking a [16:38] crawl, walk, run approach, um to AI. Um [16:41] for us, we released an internal tool [16:43] called Coinbase GPT. Um and we started [16:46] with a very small use case, which just [16:48] how do we kind of automate very routine [16:50] customer support use cases, right? [16:52] especially as we onboard new customer [16:54] support team members, things around [16:56] login issues and how do you kind of find [16:58] um the relevant, you know, kind of [17:00] resolution workflow in a sense. We've [17:02] since now kind of expanded our CBGP to [17:05] all employees internally at the company. [17:07] So now CBGPT is being used to help our [17:11] product managers write better product [17:13] specs, right? We're using it to query [17:15] like the thousands of documents that the [17:18] company has created across our G Suite. [17:20] So you know certainly on that journey of [17:22] like you know crawl walk run start small [17:24] start with a use case prove out the [17:26] effectiveness and then scale from there. [17:28] Um professor G similarly like in terms [17:31] of like what are some of the use cases [17:32] you're seeing from perhaps like your [17:33] commercial banking partners that you [17:35] know has actually is live in production [17:37] today that you're pretty excited about. [17:40] So um um just out of curiosity if if you [17:44] want to uh show me uh who are from the [17:48] banking industry if you can please all [17:51] right okay we have a lot of bankers [17:52] great okay now we can talk about uh [17:55] banking um so the use case I've spoken [17:58] with to prepare for this panel I spoke [18:01] with several of my adi advisory board [18:03] members um their use case have not gone [18:06] beyond simple automation [18:10] um the um they have probably deployed [18:12] tools that are now labeled AI but are [18:16] very similar to what we had before the [18:18] AI boom right so again you think about [18:22] uh banking uh credit intermediation [18:26] uh we have a a term called relationship [18:28] banking right relationship it's very [18:31] important um less important for probably [18:35] consumer products but very important for [18:38] commercial loans and that is the the [18:41] cash cow that is the the workhorse right [18:44] of of banking revenue and profit is [18:47] commercial loans um and there's a thing [18:49] called relationship banking so we do not [18:51] want to replace the relationship with a [18:54] computer but we want to enhance the [18:57] relationship and in that sense I think [19:00] we are still a long a ways away from [19:03] actually a use case in in that direction [19:05] but let me again my bankers are So they [19:08] understand this and let's let me um um [19:12] give you some recent evidence uh on this [19:14] research evidence on this front. So uh [19:18] um bank of international settlement BIS [19:21] recently uh published a working paper on [19:25] AI and relationship banking relationship [19:28] lending you know if you are in [19:30] commercial lending you understand that [19:32] you want to lend money to people who [19:34] don't need it right when they do need it [19:36] you don't want to lend them right so [19:39] that because banks you know want to [19:42] protect the downside risk so what [19:44] usually happens in the last hundreds of [19:47] years in commercial banking is if you [19:51] are a banker and you have a customer, [19:54] you want to charge them more [19:58] than you than you should in good times [20:01] so that you can lend to them in bad [20:04] times, right? So when the risk is low, [20:07] you want to overcharge the risk to [20:09] compensate for the loan you are going to [20:11] give them when the risk is high. That is [20:14] when they exactly need the loan. When [20:16] you do not want to lend them the loan is [20:18] when exactly when exactly when they need [20:20] it. So this is this is a counter [20:22] cyclical of credit supply and interest [20:26] rate. Right? So banks bankers justify [20:29] that by okay we provide the downside [20:31] protection when time is tough. Therefore [20:34] we charge you more um when time is good. [20:38] And everyone that is sophisticated [20:40] enough understand this and they accept [20:42] this. But this new research paper shows [20:44] that when using a sample of Italian [20:47] banks, banks in Italy, right, the the [20:51] those banks, a sample of banks, they [20:53] they implemented AI tools on credit [20:56] analysis before extending the loans and [20:59] they find that the AI tools actually [21:03] um attenuates that counter cyclicality [21:06] meaning in English I'm going to explain [21:08] is that in good times the bank those [21:11] banks charge lower rates than they [21:14] otherwise would and in bad times they [21:16] still lend extend the credit to those uh [21:20] firms at the normal rate they they used [21:23] to charge. So so basically this is good [21:25] for banks because banks now become more [21:27] competitive because their AI analysis [21:32] rely more on objective data uh because [21:36] their AI analysis can better analyze [21:39] risks. Therefore, they can price their [21:42] loans at actually the risk level that [21:46] the loans deserve, not at the risk level [21:49] plus plus a premium to mitigate what we [21:53] call borrow hazard or information as [21:54] symmetry problem. So, actually the AI [21:57] tools help uh break down the information [22:00] symmetry or moral hazard between the [22:02] borrowers and the lenders that make the [22:04] banks more competitive. that also makes [22:06] the borrowers better off because now [22:09] they enjoy a lower interest rate. So [22:11] those tools I think know the banks will [22:14] bankers will really enjoy and the [22:16] developers here can really think about [22:19] right how to [22:21] know help the bankers know become more [22:23] competitive in that sense. [22:25] >> I I I agree. I'd add one thing. One one [22:27] of the themes we talk a lot internally [22:29] at Mastercard when it comes especially [22:31] to small B businesses is one of the [22:33] common problems that every small [22:35] business in the world has is they want [22:37] access to capital and they need it fast [22:38] and they need it fluidly and information [22:40] for them is not always guaranteed. And [22:42] so one of the one of the pieces that I [22:45] agree like AI is really bringing is the [22:47] ability to make better credit decisions [22:50] with more unstructured data with more [22:53] non-traditional data is making you know [22:57] is really bringing some of the elements [22:58] that you get with typical wholesale [23:01] commercial relationship banking to much [23:03] longer tail and and what it translates [23:05] to is you know entrepreneurs and small [23:08] businesses can get exactly to to profess [23:10] point access to capital faster and [23:14] typically more fluidly, especially in [23:16] times of need. [23:17] >> Awesome. Really impressive to see so [23:19] many bankers in a room and thank you [23:21] that they asked the crypto guy to [23:22] moderate this panel. So, thank you all [23:24] so much. Um, you know, so speaking of [23:27] that, like you know, one of the use [23:28] cases that we've been thinking about, [23:30] especially with everything we've seen in [23:32] like the regulatory um space is stable [23:35] coins, right? So we think of stable [23:36] coins as programmable money, digital [23:39] dollars always on 24/7, you know, [23:41] crossborder near instant settlement. And [23:44] so this idea of like how do you combine, [23:46] you know, stable coins with agents that [23:48] perhaps may have like their own [23:50] self-custody wallets and can transact [23:52] and exchange value on behalf of like the [23:54] users, on behalf of enterprises, etc. [23:56] So, you know, very curious on what are [23:58] you all's thoughts as it relates to like [24:00] this intersection of stable coins, [24:02] tokenization, and AI agents as it [24:05] relates to how that might transform the [24:07] banking sector. And Feder Rico, I want [24:08] to start with you because, you know, I [24:10] saw that you all announced agent pay. [24:12] Um, that kind of, you know, touches on [24:13] those aspects. So, why don't you kind of [24:15] give the audience a sense of what that [24:16] is and you know, what what we can expect [24:18] to see in the future. [24:19] >> Yeah. No, happy to do it. Um, it's it's [24:22] a topic of a lot of interest for us. So, [24:25] and I think that intersection is getting [24:26] very interesting. So, for for context [24:29] for everybody, we uh in in late April, [24:31] we announced something called agent pay, [24:33] which is effectively making the [24:36] capabilities in our network [24:39] uh flexible enough to now allow AI [24:43] agents to make transactions on behalf of [24:46] a human. Now, why did we get there? and [24:50] and what does it really mean is is where [24:52] I'll start and I'll link it to to what [24:54] it means from a stable coin perspective. [24:58] One of the things that we're seeing [24:59] especially on the consumer side is that [25:00] a growing percentage of internet traffic [25:03] when it comes to product discovery is [25:05] starting to happen in AI agents. Right? [25:08] So whereas before I might go to a [25:10] particular website or I might look up [25:11] best shoes, now I'll go to the agent of [25:14] my choice and I'll ask I want to go [25:17] running you know or I need new new pair [25:19] of shoes for running. Here are the [25:21] details of what how long I run and my [25:24] gate and you know the type of shoa that [25:26] I want etc. And it can take all that [25:29] information and now produce [25:30] recommendations for me. that that that [25:32] pattern is only growing and it's it's [25:35] already non-trivial when you think about [25:38] in terms of internet traffic and when [25:40] you think about that experience there's [25:42] one thing that is not closing the loop [25:44] today which is the payment piece so [25:47] today you go through that pick your the [25:49] LLM that you use per on on your personal [25:51] life try it and what happens today with [25:54] the exception of one today it doesn't [25:56] make the payment on your behalf but [25:59] what's the challenge there even if they [26:01] do it today. Typically what's happening [26:03] is it's a screen scraping and they're [26:04] putting in your details and that is not [26:07] particularly secure at scale. And so [26:10] what we said is we need to enable the [26:12] ecosystem. This use case is organically [26:14] coming up. We believe it's only going to [26:16] continue to grow as these tools mature [26:18] and now agents are effectively available [26:21] across not just the consumer side but [26:23] the business side. And so why do we need [26:26] to enable that? Well, when you tap your [26:29] card, when you tap your phone, when you [26:31] make a transaction online, regardless of [26:33] what um what currency you're using, most [26:36] people in the world don't have a concern [26:39] in terms of the safety, the [26:41] trustworthiness, and the safeguarding of [26:43] that transaction being routed and [26:44] actually being executed. The baseline [26:47] for us is that agent commerce has to hit [26:50] that same bar. You have to have to be [26:52] able to trust as much as possible, which [26:54] means you have to understand how it's [26:56] being used. You have to understand the [26:57] safeguards and guard rules that we're [26:58] putting in at the agent level and at the [27:01] network level to make sure that that [27:03] stays. And that it's equally as secure, [27:05] if not more, than what you benefit from [27:07] today. That's the bar. And so what we [27:11] created with Agent Pay is is an [27:12] evolution of some of our capabilities in [27:14] the network to enable that. Um, [27:18] but it's also about choice, right? We [27:22] are not out there saying every [27:23] transaction in the world is going to be [27:24] a gentic. Like just for the record, I [27:26] intend to continue going to restaurants [27:28] and just tapping my card when they bring [27:30] the terminal. Uh, that's not going away, [27:32] right? And so what we see is it's about [27:34] choice. And another piece that's very [27:36] important for us in terms of choice is [27:38] is the method of payment, right? So one [27:40] of the things that we're already doing [27:41] is supporting stable coin in our [27:43] network. So that you know you can have [27:46] your card linked to to a wallet with [27:47] stable coins and and have those those [27:49] capabilities embedded because our goal [27:51] is to power commerce whether it's [27:52] through agents through other methods and [27:54] taking advantage of stable coins [27:56] especially now that you know we're [27:57] starting to get a regulatory framework [27:58] in in a number of countries for it. And [28:00] so for us the goal ultimately is is [28:04] providing that choice that safety and [28:06] that trustworthiness in any way you you [28:09] execute commerce. That's kind of the way [28:12] we see it and why, you know, for us [28:14] things like stable coin and agent-driven [28:16] commerce are quite interesting vectors [28:19] that, you know, we believe we're only [28:20] kind of at the beginning of right now. [28:22] >> Professor Chu, what's your academic take [28:24] on like this intersection of like stable [28:26] coins and AI agents? [28:29] >> Um, the the Genius Act on stable coin is [28:32] it's so new. I think we are all trying [28:34] to figure out the implications. uh but [28:36] we know what it says uh in the genius [28:39] act which is uh mostly banks essentially [28:43] banks can issue the stable coins back by [28:45] one by one one to one by US dollars [28:49] so um [28:52] there are various implications one by [28:54] one okay um I won't take too long maybe [28:56] three minutes if you think about this in [28:58] a few years potentially we can have [29:01] thousands of bank issued stable coins [29:06] right? U Mastercard probably have to [29:08] handle you know all these different [29:10] thousands of coins that essentially have [29:13] the same value which is a dollar to a [29:15] dollar. [29:16] Um you probably don't remember but about [29:19] 100 years ago until 100 years ago we did [29:22] have a system like that. It's gold based [29:25] currencies. At that time every bank [29:28] could issue their paper currency. [29:31] Right. Right now you take out your [29:32] dollar bill you see Federal Reserve note [29:36] only until 100 years ago it was Wells [29:39] Fargo note JP Morgan note you know US [29:43] bank note and all the notes and those [29:45] notes were allowed to be issued because [29:48] they were backed theoretically by gold [29:52] right and then and then we know what [29:56] happened later is is didn't work at that [29:59] time so now we um had we now we have [30:03] only one currency which is US dollar and [30:05] you should buy Federal Reserve. So, [30:08] stable coins the use case um we have [30:11] credit cards, we have PayPal, we have [30:14] other weo other ways to pay. Um stable [30:18] coins is an option. Maybe some people [30:20] prefer that option but but if you think [30:22] about theoretically it's not that new at [30:25] least not entirely new right earlier [30:28] currencies were backed by gold. Now it's [30:31] backed by US dollar which is not backed [30:33] by gold anymore. So how stable that is [30:36] is as stable as US dollar is relative to [30:39] other currencies and your purchasing [30:40] power. Okay. The next point and I will [30:43] be done after this that I think banks [30:45] really bankers really need to think [30:47] about is [30:50] uh there's a clause in the stable in the [30:53] genius act which is very advantageous to [30:56] bankers which is if you issue stable [30:59] coins to your customers to your [31:01] depositors you don't have to pay [31:03] interest rate you don't have to pay [31:05] interest now checking account pays like [31:08] some interest and savings account pays a [31:11] little bit more interest but if you are [31:13] a banker you can issue stable coins and [31:15] you don't have to pay interest [31:17] that seems a good deal I think that's [31:19] why bankers now suddenly are embracing [31:22] uh that's one reason not the only reason [31:24] are embracing the stable coins because [31:26] now they see asset right that they can [31:31] sell get the money put into the Federal [31:34] Reserve system now now right now is four [31:37] four and a quarter% annual rate from the [31:39] Federal Reserve and they pay zero on [31:41] that deposit basically. So yeah, if yeah [31:46] that's good business model. [31:49] >> Thank you professor G. Um so you know [31:51] we've talked about you know kind of like [31:53] practical use cases and I think for the [31:55] practitioners in the room we love to [31:57] learn you know kind of maybe lean a [31:59] little bit more into you know actual [32:01] implementation details right and talk [32:03] about like some of those challenges. So [32:05] for Rico, especially from your lens, [32:06] right, leading like the AI and data [32:08] operations um um strategy at Mastercard, [32:11] how do you think about, you know, [32:12] infrastructure and data strategy as it [32:15] relates to scaling up like a AI program [32:17] specifically in this sector? [32:19] >> It it's pretty critical. Um you know, I [32:23] I would say we we take a couple of of [32:26] pillars to it. The first one is data [32:29] strategy is actually quite fundamental. [32:31] That doesn't mean all your data has to [32:33] be perfect, but it means you you [32:35] actually do need a strategy behind it. [32:36] So we um one of my peers who's actually [32:39] also attending the conference uh one of [32:41] his teams is our data strategy team and [32:43] we have people working with every part [32:46] of the organization on data strategy. So [32:49] what data do you have? How should you [32:51] store it? How do you use it responsibly [32:53] and in a compliant way? And how do you [32:55] set it up in a way that's going to allow [32:57] you to do everything? So it's not a [32:58] point in time work but rather it's it's [33:00] continuous continuous work. So that that [33:02] piece is important and and one of the [33:05] things that comes out very very valuably [33:07] from that effort is what is your most [33:09] valuable data like what do you [33:11] prioritize using first and that actually [33:13] requires expertise in using data and [33:16] data management which is not a [33:17] discipline that everybody has. So that's [33:19] kind of the first piece. The second is [33:22] based on that work we make sure that [33:23] from an infrastructure standpoint [33:26] you know we are in a place where we can [33:28] enable the organization. So over the [33:30] last year my team spent quite a lot of [33:32] work thinking about what is the pattern [33:35] to start getting all of the critical [33:36] masses of data in the company in a place [33:41] so that then you can build artificial [33:42] intelligence on top of it. So that [33:43] that's been a journey right uh because [33:45] it's not just building a data lake it's [33:47] it's actually much more about [33:48] orchestration right just having the [33:50] right connection in place and then very [33:53] tactically um you know like I was saying [33:56] so my team is responsible for internal [33:58] use of AI what that means in practice is [34:00] I have a product management team we take [34:02] a product lens to it and each one of [34:03] those product managers goes very deep [34:05] with different parts of the organization [34:07] and says what are your use cases and [34:09] they evaluate the visibility based on [34:11] what data do you have available how much [34:13] value is it going to get? How difficult [34:14] is it to create this? You know, is it [34:16] high risk, low risk? I could take [34:19] everything I just said, take AI out of [34:21] it and say this is how you think about [34:24] technology deployment or, you know, [34:26] empowering of an organization using [34:28] technology, right? AI brings in very [34:31] interesting standpoints because it's [34:32] nondeterministic and you actually need, [34:34] you know, expertise in developing [34:36] agentic systems and and and and whatnot. [34:38] That's that's really emerging. But the [34:40] rest of the discipline we've actually [34:42] found [34:44] grounding in what we know works has been [34:46] quite successful. So if I if I were to [34:48] say something for for the group as you [34:51] were thinking about what's the right way [34:53] to start on the journey or to to [34:55] accelerate taking a product management [34:57] lens to it and focusing what is feasible [35:00] today that there's plenty to do with [35:02] that already. [35:04] Professor Ch, I want to kind of like [35:06] shift the question a little bit for you [35:08] and you kind of think about, you know, [35:10] if we if we imagine a world where we [35:12] have like these in number of agents that [35:14] are transacting on behalf of users, on [35:17] behalf of enterprises and they are [35:19] custodying value, exchanging that value [35:21] at a very rapid scale. That must imply [35:25] that there we probably need some type of [35:27] governance framework for these agents, [35:29] right? like there's still compliance um [35:32] like kind of like requirements that we [35:34] have to adhere to. How do you think [35:35] about that like you know like how do we [35:37] how should we adapt like governance and [35:38] compliance as it relates to like AI [35:40] within this sector? [35:42] >> Absolutely. Great question. So I talked [35:43] to the bankers and uh unanimously they [35:47] say we absolutely first need a ethical [35:51] and a responsible policy on AI use not [35:54] only for the industry for the regulators [35:56] but for each bank. the banks I mean the [36:00] the management has to set the policy [36:03] before their team can develop or adopt [36:05] AI tools right so that that governance [36:08] uh policy is absolutely critical um and [36:12] I think I heard I heard the terms [36:14] ethical and responsible you know quite [36:16] often in that sense and also if you [36:19] think we think about banks are highly [36:20] regulated highly regulated so I did ask [36:23] them you know [36:26] uh what are the implications for [36:28] compliance and they say actually [36:32] AI may actually help with compliance in [36:34] terms of generating the compliance [36:36] reports may uh doing auditing and all [36:39] that. Uh but the um regulation on AI in [36:44] banking is still in the infant stage. So [36:47] we don't know the what the regulations [36:49] will be and therefore the banks need to [36:51] sort of wait and work with the [36:52] regulators and sort of set that industry [36:55] standards um for the regulations so they [36:59] can do the compliance but they there [37:01] there does not seem to be a a serious [37:04] concern right now as far as to [37:06] compliance [37:09] um u because again we are so new in the [37:12] process but it's certainly something [37:14] that we can we should seriously think [37:17] Um, Feder Rico, want to follow up on [37:19] that with you. Something you said [37:20] earlier about, you know, kind of agents [37:22] perhaps sitting and like helping to [37:24] automate like compliance workflows etc. [37:26] where, you know, there's linky processes [37:28] for KYC, KYB. But in a world where, you [37:30] know, these agents are transacting, is [37:32] there a know your agent type of [37:34] framework, right? A KYA type of [37:36] framework that, you know, we need to [37:38] kind of consider agent identity, etc. [37:40] >> Yeah, I I I think there is. It's [37:42] something that, you know, we're actively [37:44] thinking about is as you as you start [37:47] having agents represent businesses or [37:50] consumers now making you know financial [37:53] transactions, creditworthiness [37:55] decisions. [37:57] We you know I I do think that there is a [37:59] space for for know your agent framework [38:01] in the same way that you know if you [38:03] think about your own compliance programs [38:05] today you may be asking your suppliers [38:07] or vendors what tooling or technologies [38:10] they're using already. So it it's not [38:13] that far of of a of a jump uh if if you [38:17] will. [38:19] >> Awesome. Thank you. And I and I know [38:20] that we're coming, you know, uh to to [38:23] close. And so wanted to just kind of end [38:24] on the last question is just, you know, [38:26] what is the advice that you would give [38:27] all of the bankers in in the room that, [38:30] you know, want to get started, they want [38:31] to start operationalizing Gen AI and AI [38:33] agents for their respective, you know, [38:35] companies, etc. What advice would you [38:37] get them give them? Professor Chi, we'll [38:38] start with you. [38:39] >> Um, thank you. Um great question. I I [38:42] think you know I've been thinking about [38:44] AI as a educator. We are developing AI [38:46] curriculum in finance and bankers are [38:48] thinking about how to stay competitive [38:51] and we become more competitive. I think [38:53] banking um you know the bankers here u I [38:56] don't know whether you remember [38:58] um a few years ago you know or [39:02] from the time of the uh 078 crisis until [39:06] a few years ago uh bankers were at least [39:09] publicly listed banks were quite [39:12] frustrated with their u stock valuation [39:16] um P 10 times yet utility firms uh P 20 [39:20] times and people thought were there were [39:23] there were analysts actually talking [39:24] about banks should be uh valued like [39:27] utility firms uh you know rather than 10 [39:30] times P they should be 20 times P and I [39:32] would say no you don't want to be uh a [39:35] utility firm utility firm is a commodity [39:39] provider you you want to stay away from [39:42] commoditize your products right if you [39:45] just become another bank uh just make [39:49] your products commodities there's really [39:51] no competitive advantage other than [39:53] pricing right so that is a a downward [39:56] spiral there's there's no end in sight [39:58] so [40:00] I would advise my bankers on my advisory [40:04] board rather than try to commoditize [40:06] your products um probably some products [40:09] you want to commoditize and compete on [40:11] price uh but there are uh products you [40:14] want to differentiate differentiate and [40:17] those are the commercial loans the [40:18] business loans the relationship based [40:20] lending mean that that you want to uh [40:23] differentiate your bank from other banks [40:25] and that is really where AI comes in is [40:29] to facilitate those relationship to [40:31] enhance [40:33] uh complex decision making to help you [40:37] better assess uncertainty right that [40:41] that's where your value is is to assess [40:43] uncertainty and makes complex decisions [40:45] you want the tools not to replace you [40:48] but to help you enhance your decisions [40:51] Rico. [40:52] >> No, I think that's well put. I I I had [40:54] one thing which is if you're at the [40:56] start of the journey, focus on existing [40:59] problems like where your people have [41:02] frustration, where you know the process [41:05] is slow or where there's benefit to to [41:08] accelerate [41:11] that. Um that is where I would start. [41:14] And why? Because as you get moving and [41:17] you get those repetitions, [41:19] that's what's really going to empower [41:20] the rest of your people to also start [41:23] thinking about exactly what Professor G [41:25] said. How do you differentiate? Where do [41:27] you differentiate? How do you use the [41:28] technology that way? It's it's about [41:31] getting that practice in with the simple [41:33] with the lowhanging fruit that's very [41:35] apparent and approachable that then will [41:38] set you up to bring the entire [41:39] organization to think in the way that [41:41] the Dr. Chi is uh is me, Professor Chi [41:43] is mentioning. Awesome. Well, Professor [41:46] Chief Fder Rico, I want to thank you [41:47] both so much for joining me on this [41:49] panel and I want to thank everyone for [41:50] attending. Right. I think the persons in [41:52] this room, we're pushing the frontier [41:54] forward. So, thank you all so much. Hope [41:56] you enjoy the rest of your AI4.