Banorte, CI&T Say AI Halves Loan Times as Regulatory Lag Holds Back US Banks

Banorte与CI&T称AI使贷款耗时减半，监管滞后制约美国银行业 金融机构正借助AI加速贷款处理与联络中心自动化，但美国监管的不确定性及Token成本上涨恐将阻碍更广泛的企业级应用。 行业专家探讨了AI在银行业中从概念炒作向切实运营价值的务实转型。专家小组审视了诸多现实应用，例如在客户开户环节实现90%的文件验证自动化、加快按揭贷款处理速度，以及通过对话式聊天机器人承接数百万次客户互动。他们还探讨了企业面临的紧迫挑战，包括Token成本优化、监管合规，以及保持人工介入（human-in-the-loop）安全机制以防止模型幻觉并维护客户信任。 AI 正在推动高品质金融辅导服务的普及，让普通客户也能获得以往仅限 VIP 的个性化建议，并为银行摆脱遗留系统、重塑核心业务流程提供了创新空间。 (0:04) 墨西哥北方金融集团（Banorte）依托合成数据生成与严格的评估框架（evals）测试各类复杂场景以防范模型幻觉，并坚持将模型 IP 与敏感数据保留在私有数据中心。 (2:12) 中小型金融机构聚焦于房贷审批等高摩擦环节，借助 AI 将放款周期从 45 天大幅压缩至 20 天，同时通过模型适配优化与上下文管理等 FinOps 手段抑制激增的 Token 成本。 (10:23) 在落地应用方面，Banorte 利用 AI 实现了 90% 提交证件的自动化审核，其智能对话系统每月分流 200 万至 300 万条客户咨询（相当于节省 500 名人工坐席），并开展超个性化利率与内容推荐。 (16:32) 尽管 AI 能够实现贷款全流程直通处理，但在合规与可追溯性要求下，大部分金融机构在最终放款与审批决策上仍严格保留人工审核环节（human-in-the-loop）。 (21:00)

Financial institutions are leveraging artificial intelligence to automate intensive workflows and scale personalized services, but realizing sustainable value requires balancing deployment speed against strict model governance, token costs, and regulatory uncertainty. Front-to-Back Operational Deployment and Personalization (16:32) Olga Ceja and Young Pham observe that artificial intelligence is shifting banking from static legacy processes to dynamic, democratized services. Personalized financial coaching and customized credit terms, historically reserved for VIP clients, can now be scaled to mass-market retail customers. In front-to-back operations, banks are targeting high-friction bottlenecks: Pham notes that AI-assisted documentation in mortgage lending can cut closing timeframes from 45 days down to 20 days, while client onboarding cycles across small business and commercial lending are shrinking from weeks to a single day. At Grupo Financiero Banorte, Mexico's second-largest bank, Ceja details tangible production deployments. The institution's conversational chatbot resolves between 2 million and 3 million customer inquiries monthly, deflecting work that would otherwise require an estimated 500 contact center agents. Furthermore, the bank combines biometrics and AI document extraction to automate 90% of identity verification during digital account opening, leaving only 10% of edge cases for manual staff inspection. Governance, Synthetic Evaluation, and Human Safeguards (2:12) Because customer trust is a bank's foundational asset, AI adoption relies on validation systems that diverge sharply from standard software testing. Ceja explains that Banorte utilizes comprehensive evaluation frameworks to control communication style and eliminate hallucinations. To accelerate this without exposing customer data, engineering teams generate synthetic data to simulate complex edge-case scenarios and varied user personas. Additionally, Banorte mandates internal intellectual property ownership over models and isolates sensitive records within private data centers. Pham points out that banking is primed for model management due to decades of experience running automated fraud detection and credit risk models. However, compliance standards continue to restrict complete autonomy. While straight-through processing and loan pre-approvals are increasingly executed by AI, full loan underwriting and ultimate funding steps still require a human-in-the-loop to guarantee legal auditability and end-to-end decision traceability. Token Economics, FinOps, and Institutional Disparity (10:23) Accelerating enterprise adoption introduces sharp cost pressures linked to model token consumption. Pham recounts that his firm achieved 99% internal employee adoption but saw token expenses quadruple over a single year due to deficient forecasting. To prevent cost overruns without imposing counterproductive hard caps, Ceja outlines Banorte's FinOps practices: right-sizing models to task complexity rather than using top-tier models for basic queries, substituting expensive proprietary APIs with open-source models, gatekeeping whether a task requires an agent or a simple retrieval-augmented generation (RAG) tool, and truncating context windows to avoid resending conversational history. These technical and economic demands are exacerbating a split between large institutions and community lenders. Pham warns that major banks possess the capital to engineer proprietary walled gardens and own their data architectures, whereas smaller credit unions and community banks lack these resources. Consequently, smaller players must avoid broad enterprise overhauls and confine their AI investments to isolated, high-impact tasks to remain viable. Navigating Regulatory Volatility Across Jurisdictions The contraction of experimentation cycles from years down to days has outpaced public policy in several regions. In Mexico, while specific AI legislation remains nascent, existing statutes strictly govern cloud deployments. Ceja cautions against heavy-handed over-regulation that could impede consumer benefits, arguing that financial institutions already have a proven track record of auditing traditional machine learning algorithms for credit decisions. Pham contrasts the regulatory landscapes of the European Union and the United States. While European institutions operate under the structured framework of the EU AI Act, US regulatory guidance shifts unpredictably every two to four years across political cycles, as evidenced by stalled Consumer Financial Protection Bureau initiatives. Rather than waiting for comprehensive federal legislation, Pham advocates that American banks establish voluntary, industry-wide standards and interoperability protocols, mirroring the historical governance model of interbank payment rails. The panelists conclude that an unvetted deployment resulting in large-scale data loss or systemic error represents the primary existential threat to financial AI, meaning long-term adoption depends on self-imposed industry standards and verifiable traceability rather than raw technological speed.

银行业高管指出，人工智能正在推动金融专业知识的普及并重构银行核心业务流程，但机构必须在加速产品上市与遵守合规风控、算力成本管理及防范信任危机之间取得平衡。 金融服务普及与内部工作流重构 (0:04) 行业专家认为人工智能正在打破传统银行业的壁垒。墨西哥第二大银行 Grupo Financiero Banorte 的 Olga Ceja 指出，以往只有贵宾客户才能享受到专属客户经理与高级财务导师服务，而生成式人工智能能够将个性化财务顾问推向普通大众，帮助所有客户做出更明智的财务决策。 拥有30年银行业经验的 Young Pham（曾供职于 JPMorgan Chase）表示，银行在过去长期使用同一种模式处理遗留系统更换、数字用户体验优化以及基础反欺诈等老旧问题。人工智能赋予了机构重新审视内部运营与产品准入的创新自由，使银行能够从根本上重新设计核心业务流程。 评估体系、数据安全与规模分化 (2:12) 客户信任是银行的核心资产，金融机构在部署技术前必须进行广泛实验与严谨的模型评估体系建设。Ceja 解释称，与传统软件测试不同，人工智能评估侧重于防止模型幻觉并控制沟通语气。银行通过生成合成数据来模拟复杂测试场景与多元客户画像，大幅缩减了前期准备时间。此外，银行必须完全拥有模型知识产权，并确保敏感数据绝不离开本地数据中心。 Pham 强调，银行业早已具备成熟的机器学习治理与欺诈监控机制，但行业正面临严重的两极分化。大型机构有实力构建封闭花园并完全掌控模型与数据，而美国数以千计的中小型社区银行及信用合作社则缺乏底层基础设施，难以跟上技术步伐。同时，合规审计与投产准备流程滞后，成为大行将技术实验快速推向市场的核心制约。 聚焦高摩擦场景与算力成本管控 (10:23) 面对资源劣势，中小金融机构放弃了全面的企业级改造，转而将人工智能精准聚焦于高摩擦的独立任务。Pham 举例说明，在房屋抵押贷款处理与信贷审批中应用人工智能整理文档，能将贷款发放周期从传统的45天大幅压缩至20天，为人力有限的机构带来了显著效率提升。 然而，Token 消耗失控与算力成本成为落地过程中的普遍阻碍。Pham 透露其所在机构因预测偏差，单年 Token 消耗量激增三倍。为此，Ceja 提出了一套成本优化策略：合理调整模型规格，避免对基础问答调用昂贵的大模型；积极采用开源模型以减少对专有 API 的依赖；设立需求准入门槛，评估任务是否必须使用智能体或仅需简单的 RAG 对话检索；以及严格优化上下文窗口，杜绝重复调用冗长历史记录。 落地实效、人机协同与全球监管差异 (16:32) 在实际业务方面，Banorte 已将人工智能深度植入前台运营：数字开户时采用生物识别与文档抽取技术，自动验证了90%的官方文件，仅余10%需人工介入；智能客服每月承接200万至300万条咨询，相当于节省了500名坐席人力；银行还能利用人工智能实现个性化利率定价与定制营销内容生成。尽管如此，Pham 指出在全流程信贷审批与放款环节，为满足可追溯性与监管合规要求，机构仍必须设置人工介入节点（Human-in-the-loop）。 针对监管环境，Ceja 认为只要机构能保障决策的端到端可审计与可解释，监管并非不可克服，但需警惕过度监管扼杀创新。Pham 则比较了跨区域态势：欧洲银行业已依据《欧盟人工智能法案》建立起清晰的治理框架；而美国由于政治周期导致政策频繁变动，银行界应效仿支付清算体系主动建立行业标准与互操作性规范，而不是被动等待联邦法案。 专家一致警告，银行业最大的潜在灾难在于大型金融机构的不审慎部署导致数千万客户数据泄露或受损，这将彻底摧毁公众对银行体系的信任。

## Background 本内容由 Ai4 于 2026 年 8 月 22 日发布在 YouTube 平台。根据页面简介，本次圆桌论坛汇聚了来自 Axios、CI&T 和墨西哥北方银行（Grupo Financiero Banorte）的行业专家，探讨银行业如何在前台到后台的全栈业务中落地 AI，涵盖反欺诈、客户服务与主动风险管理等高价值实战场景。 ## TL;DR 圆桌探讨了银行业从早期概念验证向企业级规模化落地演进的真实路径。墨西哥北方银行与 CI&T 的高管深入剖析了如何通过合成数据评估套件（Evals）、模型精准路由及精细化 Token 成本控制，在保障数据隐私和合规的前提下化解幻觉风险。讨论进一步展示了 AI 在小微企业入驻自动化、端到端房贷提速、超个性化实时交互以及全员虚拟助手赋能中的落地成果，并强调“合规左移”与保持人机协同（Human-in-the-loop）是守护银行核心资产——“信任”的关键护栏。 ## Atmosphere & style 整场对话在专业务实与幽默风趣之间保持了绝佳平衡，主持人 Nathan 频繁抛出大富翁卡片、查天气乱象等机智抓人的梗，有效活跃了严谨的合规议题。两位嘉宾配合默契，既有顶层战略架构的宏观拆解，又有来自一线银行落地与数千名员工转型的真实账单困境和技术取舍，毫无空泛的行业公关辞令。现场氛围从开场对‘更笨银行业’的调侃，渐次深入到对信任危机、监管雷区与人机关系的严肃思考，并在结尾‘逃离沙盒’的欢声笑语中圆满落幕，兼具干货深度与现场感染力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=z27M2kzQQEk&t=4s) AI 正在重塑银行业务模式，将以往专属高净值客户的私人财务顾问与教练服务普惠至普通大众，并推动底层业务流程的根本性再造。 - [01:52](https://www.youtube.com/watch?v=z27M2kzQQEk&t=112s) 信任是金融机构的核心资产，银行通过构建合成数据压力测试、模型评估套件（Evals）以及阶梯式门禁，在人机协同与自动化之间把控风险。 - [06:11](https://www.youtube.com/watch?v=z27M2kzQQEk&t=371s) 虽然沙箱隔离与“围墙花园”保障了安全与知识产权确权，但也进一步拉大了大型银行与中小金融机构之间的基础设施成熟度差距。 - [10:23](https://www.youtube.com/watch?v=z27M2kzQQEk&t=623s) 中小金融机构应聚焦单点高价值流程（如将抵押贷款放款周期减半），并通过模型精准路由（Model Routing）与轻量化模型控制暴涨的 Token 算力账单。 - [13:42](https://www.youtube.com/watch?v=z27M2kzQQEk&t=822s) 企业在推进 AI 落地时必须审视 ROI，避免盲目上马复杂 Agent，通过轻量级 RAG 架构、开源模型以及精细化上下文管理实现降本增效。 - [16:32](https://www.youtube.com/watch?v=z27M2kzQQEk&t=992s) AI 正在核心业务工作流中规模化落地，例如 Banorte 实现 90% 的开户文件由 AI 自动审核，并借助 AI 驱动千人千面的超个性化信贷营销。 - [21:00](https://www.youtube.com/watch?v=z27M2kzQQEk&t=1260s) 信贷最终放款决策仍需保留“人机协同（Human-in-the-loop）”机制，以确保决策链条满足严格的合规护栏与端到端可审计性要求。 - [22:50](https://www.youtube.com/watch?v=z27M2kzQQEk&t=1370s) 全球 AI 监管呈现分化格局，金融体系面临的最大潜在威胁并非技术故障本身，而是不当使用 AI 导致客户数据泄露所引发的信任危机。 - [27:12](https://www.youtube.com/watch?v=z27M2kzQQEk&t=1632s) 实体网点正加速从处理存取款的“事务型柜台”向“综合财务咨询中心”演进，人机协作模式让员工有更多精力为客户提供高价值专业指导。 - [28:56](https://www.youtube.com/watch?v=z27M2kzQQEk&t=1736s) AI 的核心价值在于知识赋能而非单纯人力替代，例如辅助数千名房贷专员迅速掌握跨线金融产品销售能力，实现低成本业务扩张。 - [30:34](https://www.youtube.com/watch?v=z27M2kzQQEk&t=1834s) 百年信誉与合规资质是银行不可替代的护城河，结合 AI 激活沉睡的客户数据，使得银行相比外部独立 AI 工具能提供更可靠的财务建议。 - [34:24](https://www.youtube.com/watch?v=z27M2kzQQEk&t=2064s) 企业级 AI 正在跨越 PoC 鸿沟迈向规模化量产，大型银行已实现呼叫中心与网点全员普及专属虚拟助手，大幅缩短业务升级与响应时间。 - [37:05](https://www.youtube.com/watch?v=z27M2kzQQEk&t=2225s) 超个性化服务高度依赖“实时感知”，银行系统能在客户就医支付或购票等紧急时刻动态调额，并通过最契合的渠道进行精准触达。 - [39:17](https://www.youtube.com/watch?v=z27M2kzQQEk&t=2357s) 推行“合规左移（Early Compliance Engagement）”方法论，在设计初期邀请法务合规团队共同制定数据边界与分流机制，可从源头化解监管阻力。 - [41:24](https://www.youtube.com/watch?v=z27M2kzQQEk&t=2484s) 在小微企业（SMB）入驻与审查场景中，AI 重点承担多源非结构化报表的自动化解析与跨文档对账比对，有效防范数据矛盾与潜在欺诈。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=z27M2kzQQEk&t=4s) 银行业重塑：从“更笨”到普惠化与流程再造 圆桌对话在轻松幽默的氛围中拉开帷幕。主持人 Nathan Bomey 开玩笑地抛出话题：既然本次论坛的主题是“更智能的银行业（Smarter Banking）”，那是否意味着银行业过去一直处于某种“更笨的银行业（Dumber Banking）”时代，如今正借助人工智能迈向更宏大的新纪元？墨西哥北方银行（Grupo Financiero Banorte）人工智能与创新执行总监 Olga Ceja 给出了非常肯定的回答。她指出，AI 在银行业的应用核心在于“产品与服务的普惠化”，并赋能客户做出更明智的财务决策。在传统模式下，只有高净值 VIP 客户才能配备专属客户经理；而有了 AI，银行能够为每一位普通客户提供随手可得的专属财务教练与关系顾问，这正是银行业变得更加智能的体现。 CI&T 首席战略官 Young Pham 结合自己在摩根大通（JPMorgan Chase）等机构长达 30 年的从业经验补充道，过去数十年间，银行业总是在用相同的方式解决老问题——无非是更换核心系统、改造前端数字化体验、升级风控与反欺诈工具。但生成式 AI 的到来带来了前所未有的创造力空间，让从业者能够重新思考银行内部的底层业务流程，甚至从根本上重塑银行自身的运营模式与产品交付方式。 --- ### [01:52](https://www.youtube.com/watch?v=z27M2kzQQEk&t=112s) 信任基石：合规底线、评估套件与合成数据 针对“金融涉及老百姓的切身资产，理应是 AI 最晚落地的领域”这一传统观点，Olga Ceja 深入剖析了金融机构在拥抱创新时的严谨逻辑。她强调，虽然墨西哥目前尚未出台专门针对 AI 的法规，但银行在云端产品使用方面面临着极其严格的监管约束。作为全球监管最严苛的行业之一，也是墨西哥第二大银行，Banorte 销售的核心资产就是“信任”，因此绝不能承担失去客户信任的代价。为了确保 AI 输出的绝对可靠，Banorte 建立了一套与传统软件测试完全不同的评估机制（Evals）： ```mermaid flowchart TD accTitle: 金融级 AI 评估与上线流程 subgraph Prep["1. 测试准备与数据生成"] A["定义测试用例与边界"] --> B["生成 AI 合成数据与多元画像"] end subgraph Eval["2. 安全与性能评估套件 (Evals)"] B --> C["幻觉检测与准确性验证"] C --> D["沟通语调与合规性审查"] end subgraph Deploy["3. 部署与资产管控"] D --> E["模型资产自研与 IP 确权"] E --> F["本地数据中心封闭运行"] end ``` 这种评估并不等同于传统的 QA 性能测试，其核心是消除幻觉（Hallucinations）并规范 AI 与客户交互时的应答方式。传统测试很难覆盖海量真实场景，而借助 AI 生成的合成数据（Synthetic Data），团队能够在短时间内构建出各类复杂场景与多元用户画像（Personas）进行极端压力测试。同时，银行必须对模型知识产权（IP）拥有自主掌控权，并确保核心敏感数据绝不流出本地数据中心。 Young Pham 对此持高度乐观态度。他指出，银行其实比大多数行业更适合落地 AI，因为传统机器学习早已深度应用于反欺诈检测模型中，金融机构与生俱来就拥有极其完善的治理与合规文化。面对模型的准确性与幻觉隐患，银行的关键在于建立清晰的阶段性管控门禁（Stage Gate），明确在何时需要保持“人类介入监督（Human in the loop）”，何时才能开放全流程自主自动化（Full Autonomy）。 - **合规核心要点**：严格遵守云监管标准，将数据限制在本地数据中心内，坚持自主掌控模型知识产权。 - **评测创新突破**：利用 AI 自动生成多元合成数据与用户画像，摆脱传统测试耗时冗长的困境。 - **治理与落地原则**：设置阶梯式门禁，根据风险等级在“人机协同”与“端到端全自动”之间精准切换。 --- ### [06:11](https://www.youtube.com/watch?v=z27M2kzQQEk&t=371s) 围墙花园：沙箱隔离与大中小银行的成熟度差距 听到 Olga 谈及幻觉，Nathan 打趣道，自己经常幻想银行系统发生幻觉、误以为自己的账户里有巨额存款，就像玩大富翁（Monopoly）时抽到“银行出错，向你支付 10 美元”的机会卡一样。随后，主持人引申到近期 OpenAI 模型与 Hugging Face 相关的安全事件，抛出关于“沙箱（Sandbox）”安全隔离的核心疑问：银行业究竟如何保证所有实验都被牢牢锁在安全边界之内？ Olga 澄清说，沙箱机制对银行而言早已是家常便饭。安全团队在数年前就搭建了高安全级别的沙箱环境，任何新开源模型或刚面世的技术工具，在进入正式技术评估前都必须在沙箱中进行严格隔离测试。墨西哥监管机构此前也针对金融科技（FinTech）推出了监管沙箱机制。Young Pham 紧接着强调，打造“围墙花园（Walled Garden）”以实现知识产权和学习模型的确权是行业共识。然而，这也折射出残酷的市场现实：全美有数千家金融机构，资金雄厚的大型银行能够轻松构建成熟的技术基座与沙箱体系，而广大中小银行在基础设施成熟度上面临着巨大的资源鸿沟。 --- ### [08:50](https://www.youtube.com/watch?v=z27M2kzQQEk&t=530s) 研发提速与投产摩擦：敏捷实验 vs 监管流程 Olga 进一步指出了当前银行在 AI 落地中面临的真正痛点：实验速度与投产流程之间的严重脱节。在过去，一项新技术的实验与概念验证（PoC）往往需要耗费数年才能走向生产环境；而在生成式 AI 时代，团队完成一项技术验证甚至只需要几天或几周。这种极速的研发周期彻底颠覆了以往的产品上市时间（Time to Market）。 然而，当技术团队在极短时间内完成测试并取得优异评估结果后，后续流程却不得不撞上高度受监管行业的传统审批高墙。合规审查、风险评估、监管报备等上线流程依然繁琐漫长，导致前端爆发的实验速度与后端沉重的上线流程之间产生了显著的摩擦与张力。 --- ### [10:23](https://www.youtube.com/watch?v=z27M2kzQQEk&t=623s) 破局中小机构与算力账单：场景切入与 Token 成本困境 针对中小金融机构如何在 AI 浪潮中生存的疑问，Young Pham 指出，小型银行和信用社无需追求全面的系统级重构，而是应当聚焦于痛点极高、产出明确的单一垂直任务。以住房抵押贷款流程为例，许多小型机构过去极度依赖人力，审批放款周期长达 45 天；通过引入 AI 辅助进行贷款文件处理与承保分析，能够将放款周期大幅压缩至 20 天，在不推翻现有体系的前提下实现产能跃升。 | 机构维度 | 核心策略 / 面临痛点 | 具体表现与落地举措 | |:--- |:--- |:--- | | **大型银行** | 全栈自研、合规沙箱与全流程赋能 | 掌控模型知识产权，构建合成数据评估闭环，推进个性化财务顾问普惠化 | | **中小金融机构** | 聚焦单点高价值流程（AI-assisted） | 专注抵押贷款处理、智能核保，将放款周期从 45 天缩短至 20 天 | | **全行业共性困扰** | FinOps 与算力 Token 成本暴涨 | 99% 员工采用 AI 导致 Token 消耗年增 4 倍；需转向“按场景选配模型”精细化运营 | 但谈及中小银行面临的 Token 消耗与算力成本时，Young Pham 坦诚地苦笑道：“这个问题我也没有标准答案（This one I don't have an answer for it）。”随着 AI 商业化公司追求盈利，算力与 Token 费用居高不下，有时人力替代节约下的成本几乎被高昂的 Token 账单全额吞噬。一些客户为了控制未知风险，甚至对 Token 配额设置了硬性上限（Hard Cap），但这严重压制了生产力。以 CI&T 自身为例，虽然实现了全公司 99% 的员工在日常工作中使用 AI，但过去一年内 Token 消耗量直接翻了四倍，其预算预测难度极大。 面对这一困局，Olga Ceja 分享了 Banorte 的务实解法：推行模型分级与精准路由策略（Model Routing）。并非所有业务场景都需要调用参数量最大、最智能、成本最高的大模型；针对简单的问答（Q&A）或基础查询，合理分流至轻量化模型，是金融机构控制 FinOps 成本、实现健康商业闭环的关键手段。 ### [13:42](https://www.youtube.com/watch?v=z27M2kzQQEk&t=822s) 成本控制、技术务实与 AI 落地最佳实践 企业在推进 AI 落地时必须保持清醒的技术选型与成本意识。针对特定用例选择特定模型至关重要，而随着开源模型性能的快速提升以及门槛的降低，头部大模型厂商面临着降价以提供更便宜算力接入的巨大压力。企业在立项时首先要学会设置用量上限并做好成本管控。在项目优先级排序中，核心考量应当是“使用 AI 究竟能为组织带来什么实际效益”以及“这是否真的必须用 AI 解决”。当前许多部门往往一窝蜂地举手要求上马 AI Agent，但深入评估后会发现，一个轻量级的检索增强生成（RAG）聊天机器人就完全足以胜任工作。 在技术实现层面，工程团队必须掌握优化上下文管理的最佳实践。如果在每一轮多轮对话中都无脑注入完整上下文，Token 消耗量将会迅速爆炸并带来高昂账单。媒体也广泛关注过这方面的落地乱象：Axios 报道过某位 CTO 发现公司员工居然大量使用昂贵的 AI 工具来“查天气”，主持人 Nathan Bomey 也打趣说自己前几天刚用 AI 来解析荷马史诗《奥德赛》的主题——这些显然都不是企业买单的初衷。技术存在并不等于一定能普及，正如“翻汉堡机器人”早在多年前就已发明，但因成本过高至今未被快餐巨头普遍采纳，银行引入 AI 也必须经过严格的投资回报率审视。 ```mermaid flowchart TD accTitle: AI 落地选型与成本控制路径 subgraph S1["需求评估与选型"] A["业务部门提出 AI 需求"] --> B{"是否真正需要 AI / Agent?"} B -- "是，复杂自主决策" --> C["专项 AI Agent / 定制模型"] B -- "否，基础问答/检索" --> D["轻量级 RAG 架构 / 开源小模型"] end subgraph S2["工程与成本控制"] C --> E["精细化上下文管理"] D --> E E --> F["设定 Token 上限与消费配额"] F --> G["ROI 与业务价值闭环"] end ``` --- ### [16:32](https://www.youtube.com/watch?v=z27M2kzQQEk&t=992s) 业务工作流落地：Banorte 的自动化与超个性化实践 AI 在银行业的应用重心已经从前一年的软件开发生命周期（SDLC 加速、代码重构、合成数据生成、数据管道现代化）全面转向核心业务工作流。特别是在包含大量重复任务的业务环节中，AI 的效率提升尤为显著。以贷款审批、中小企业及商业客户入职开户为例，以往耗时两到四周的繁琐流程，如今在 AI 赋能下一至两天即可完成，极大加速了客户价值变现周期。 墨西哥北方银行（Grupo Financiero Banorte）在数字开户、客户服务以及个性化营销方面取得了突破性进展： - **数字开户与生物识别验证**：通过生物识别技术确保“人证合一”，并利用 AI 自动化验证证件真伪及提取信息。目前 Banorte 高达 90% 的客户提交文件均由 AI 自动审核完成，仅有 10% 流转至人工复核，支持客户在数分钟内完全线上办理开户、房贷、信用卡及各类贷款。 - **海量客服分流**：Banorte 智能客服聊天机器人每月处理 200 万至 300 万条消息互动。如果缺少这一 AI 系统，银行需要额外雇佣约 500 名坐席人员才能承接相同体量的咨询请求。 - **AI 驱动的超个性化（Hyper-personalization）**：打破传统千篇一律的通用信贷利率推送，AI 能够针对每位客户生成完全定制化的产品方案（包括差异化利率），甚至连推送营销时展示的文案与视觉图像都是由 AI 针对单个客户实时动态生成的。 --- ### [21:00](https://www.youtube.com/watch?v=z27M2kzQQEk&t=1260s) 信贷全自动审批的技术可行性与合规边界 关于“AI 是否能完全脱离人工干预独立作出贷款决策”，业内目前的技术能力与实际部署之间存在一道明确的合规防线。虽然在预审批（Pre-approval）和直通式处理（Straight-Through Processing, STP）阶段 AI 已经大幅提速，但大部分金融机构仍未实现全流程脱手放款。这并非技术无法做到，而是受到了严格的“合规护栏”（Compliance Guardrail）限制。 全自动信贷审批必须满足可审计性（Auditability）和全链路可追溯性（Traceability）。为了保障公平与合规，当前大多数银行在最终贷款批准与资金发放（Funding）环节依然保留了“人机协同”（Human-in-the-loop）机制，让人工决策者作为关键把关人。 --- ### [22:50](https://www.youtube.com/watch?v=z27M2kzQQEk&t=1370s) 全球监管格局与信任危机推演 面对银行业强监管的属性，不同司法管辖区的 AI 监管成熟度呈现出显著差异： - **欧盟领先**：欧盟出台的《人工智能法案》（EU AI Act）与治理框架走在全球前列，为欧洲银行合规应用 AI 提供了清晰的标准体系。 - **墨西哥推进**：监管部门与政府正在展开立法对话；银行业过去在信贷审批中长期运行机器学习模型并接受审计，只要 AI 的每一步决策保持透明与可追溯，监管对接并不会构成无法逾越的阻碍，但业界普遍希望监管节奏不要拖慢创新步伐。 - **美国滞后与自律探索**：由于政治周期波动，监管政策推进迟缓（如原本备受看好的 CFPB 开放银行规则被搁置）。在缺乏自上而下统一立法的现状下，美国银行业可能需要借鉴以往支付体系的建设经验，自发形成行业互通的统一标准与实践准则。 当被问及“最担心的噩梦场景”时，嘉宾指出：最可怕的不是技术故障本身，而是某家大型金融机构因不当、不严谨地使用 AI，导致数千万客户的敏感数据泄露或权益受损，从而引发公众对整个银行体系信任的断崖式崩塌。在银行网点柜员持续减少、线上化普及的背景下，人类员工在 AI 时代应当如何重新定位自身角色与客户关系，成为了行业下一阶段亟待解答的核心议题。 ### [27:12](https://www.youtube.com/watch?v=z27M2kzQQEk&t=1632s) 实体网点的职能演进：从事务性柜台到专业财务指导 在金融领域，很多产品具有极高的专业门槛。Grupo Financiero Banorte 的人工智能与创新执行总监 Olga Ceja 指出，银行网点正在经历深刻的定位转变。过去柜员和网点员工被繁重的日常操作（如转账、取现）所占据，而未来随着基础业务的数字化，员工将有更多时间专注于复杂的金融产品。当客户面临办理房贷、配置股票或投资理财等重大决策时，他们依然渴望与真人进行沟通和确认。这种“数字化与人类智慧结合的混合模式”（hybrid approach between digital and human）正在重塑网点，使其从一个单纯的资金出入场所，演变为提供高价值财务指导的专业咨询空间。 ```mermaid flowchart TD accTitle: 银行网点职能与AI赋能演进流程 subgraph Routine["传统模式：事务型网点"] A1["日常操作/柜台排队"] --> A2["手工转账与现金存取"] end subgraph Transition["数字化与AI赋能转型"] B1["自动化/数字化迁移"] --> B2["释放员工时间与精力"] B2 --> B3["AI实时辅助与知识库赋能"] end subgraph Future["未来模式：全方位咨询空间"] C1["房屋贷款决策验证"] C2["定制化股票与投资咨询"] C3["全生命周期财务规划"] end Routine --> Transition Transition --> Future ``` - 客户在办理大额房贷或复杂投资时，仍然高度依赖与真人的面对面验证与互动。 - 实体网点的角色正加速从“交易处理中心”向“综合财务咨询中心”迁移。 --- ### [28:56](https://www.youtube.com/watch?v=z27M2kzQQEk&t=1736s) 赋能而非替代：AI助力房贷顾问跨界提供全方位金融服务 CI&T 首席战略官 Young Pham 强调，AI 在金融服务中的核心价值在于知识共享与赋能，而非单纯的人力裁减。他分享了一个典型案例：全美排名前十的一家大型抵押贷款机构旗下拥有 3,000 名贷款专员（Loan Officers）。该机构收购了一家小型银行后，面临着如何拓展存量业务、吸收存款并推广银行全线产品的挑战。他们没有耗费巨资去额外雇佣上千名传统银行家或柜员，而是借助 AI 工具对这 3,000 名原本只精通房贷的业务员进行全方位赋能。 借助 AI 的实时支持，这些贷款顾问能够轻松理解并向客户推介储蓄产品、大额存单（CDs）、货币市场基金以及子女大学教育金规划。这展现了与“AI 取代人类”截然相反的路径——“赋能人类去建立更深厚的财务关系”（giving humans the capability to have a deeper financial relationship）。从成本和扩张效率来看，这种通过 AI 实现现有团队能力跃升的策略，成为了一种极具进攻性的增长手段。 - 贷款专员通过 AI 辅助迅速掌握存款、CD 与货币市场工具的销售与咨询能力。 - 避免了新招募 1,000 名银行客户经理的庞大开销，实现了高性价比的跨业务线扩张。 --- ### [30:34](https://www.youtube.com/watch?v=z27M2kzQQEk&t=1834s) 财务咨询中的信任基石与合规边界 主持人 Nathan Bomey 提到，尽管很多人已开始尝试使用 AI 来制定财务计划或获取股票投资建议，但自动化投资建议的可靠性依然存疑。针对客户未来是倾向于脱离银行生态自行获取建议，还是继续依赖银行系统这一问题，Young Pham 坚信银行在信任层面具有不可替代的护城河。他以自己陪同小学生参观银行金库的经历做比喻：孩子们对坚固金库的惊叹，折射出银行业历经数百年建立的“绝对信任”。这种信任关系如同病人看医生——人们或许会在网上自行搜索，但在做出关键决策时，依然需要可信赖的专业顾问进行把关。 针对 Nathan 提出“若千分之一的几率下 AI 给出了极度糟糕的投资建议该怎么办”的担忧，Young Pham 风趣地指出“FINRA（美国金融业监管局）只允许银行内特定具备资质的人犯错，而不是允许所有人都在那里瞎建议”。Olga Ceja 补充道，银行拥有海量客户数据，过去的痛点在于无法充分处理和利用这些数据回馈客户，而 AI 正在补齐这块拼图。尽管年轻一代更习惯通过数字化手段接触金融，但随着财富的积累，他们最终仍会向全方位的真人专业服务靠拢。 - 银行的百年信誉是外部独立工具难以企及的核心资产。 - 监管合规框架（如 FINRA 规定）严格约束了机器与人员提供金融建议的资质与边界。 - 银行利用 AI 能够激活过往沉睡的海量数据，让客户经理在沟通中掌握更精准的洞察。 --- ### [34:24](https://www.youtube.com/watch?v=z27M2kzQQEk&t=2064s) 跨越 PoC 鸿沟：大型企业从概念验证迈向规模化量产 针对 Uber CTO 近期引发热议的言论——即“目前尚未看到 AI 投资带来显著的面向消费者的核心功能”，现场嘉宾进行了深入探讨。Young Pham 认为，行业目前正处于从概念验证（PoC）向企业级规模化落地（PoC to Scale）的关键转折点。过去各机构主要处于试水和探索阶段，缺乏规模化成效并不代表技术本身不行，而是源于企业过往的探索方式较为分散。如今，企业对商业模式和长期 ROI 的理解更加深刻，大量投资正集中用于将试点项目转化为支柱业务的生产力工具。 Olga Ceja 指出，今年正是大型企业停止空谈实验、全力推进实际落地的关键年份。以 Banorte 为例，AI 已经在全员生产力上带来了实打实的飞跃：呼叫中心（Contact Center）的每一位座席代表以及网点的客户经理如今都配备了专属的虚拟助手。员工不再需要挂起电话到处寻找主管协助，虚拟助手能够实时提供解答，极大提升了业务流转效率，并直接驱动了营收与运营表现的改善。 - 大型机构的 AI 策略正在从边缘试水全面转向对“枢纽业务环节”（lynchpin parts）的规模化改造。 - 呼叫中心座席与网点客户经理已实现 100% 普及虚拟助手支持，大幅削减了人工升级查询的时间损耗。 --- ### [37:05](https://www.youtube.com/watch?v=z27M2kzQQEk&t=2225s) 问答：实时数据流与超个性化（Hyper-personalization）业务落地 在现场提问环节，有观众询问如何将海量数据转化为真正驱动业务或改善客户体验的具体用例。Olga Ceja 以 Banorte 的“超个性化实时交互”为例做了解答。她指出，仅拥有数据远远不够，关键在于“实时性”（real time）的结合与响应。 当系统监测到客户正在购买机票，或者在医院急需支付大额医疗费用而面临资金不足时，AI 能够根据其实时上下文与预授信模型，即刻触发并调高该客户的信用额度。此外，超个性化还体现在内容表达与沟通渠道的精准匹配上：系统会根据客户的偏好画像（例如科技型用户或关注教育的用户），以最契合的语调生成定制化文案，并通过客户最常用的渠道进行触达（例如在墨西哥市场最为主流的 WhatsApp，或是针对特定群体的邮件与短信），从而在正确的时间、正确的渠道推送最贴切的方案。 - 实时场景感知：针对医院急诊支付或机票预订等紧急场景，实时触发预授信额度提升。 - 多维度触达适配：结合用户兴趣画像定制文案，并在 WhatsApp、SMS、Email 等不同渠道间智能分流。 --- ### [39:17](https://www.youtube.com/watch?v=z27M2kzQQEk&t=2357s) 问答：让法务与合规前置，破解自动化落地的监管阻力 现场观众进一步提问：当银行在房贷审批或信用分析等关键流程中准备推出自动化模型时，该如何说服内部与外部的监管机构？Young Pham 给出了极为清晰的落地方法论：**合规左移（Early Compliance Engagement）**。 在传统的软件交付模式中，法务与合规团队往往在方案构建完成后才介入进行事后审查，这极易引发合规阻力甚至导致项目流产。而 CI&T 倡导并验证成功的模式，是将法务与合规团队从一开始就拉入业务流程的共创设计中。在梳理房贷评估的工作流时，团队会与合规专家共同探讨：“我们能否抽取这些特定数据项用于自动化资产评估？如果不被允许，我们该如何内置人工分流与分级处置机制（triage）？”这种将合规专家转化为设计参与者的方式，彻底打破了以往“合规只会说不”的安全壁垒，极大加速了自动化方案的顺畅合规上线。 - 将合规角色从“构建后审查者”重塑为“设计初期共创者”。 - 在自动化流程中预先设计合法的数据抽取边界与异常分流机制，从源头消解内外部监管疑虑。 ### [40:29](https://www.youtube.com/watch?v=z27M2kzQQEk&t=2429s) 监管合规边界与内部工作流的审计可追踪性 谈及金融机构在 AI 落地过程中与监管机构的互动，嘉宾明确指出，目前大多数生成式 AI 和自动化工具的应用仍牢牢圈定在“内部工作流”之中。对于金融机构而言，与监管沟通最核心的前提在于确保整个流程具备端到端的“可审计性与可追溯性”（audibility and traceability）。银行和技术团队并没有试图发明全新的信贷产品，也没有去触碰那些需要外部监管机构进行根本性审批的高风险业务底线。 正因如此，目前落地的很多技术痛点在机构内部已经得到了较好的解决与控制。然而，这其中存在一条不可逾越的红线：如果系统一旦越界，开始面向那些本不具备资质或不该接受建议的客户提供未经许可的自动化“财务与投资建议”（financial advice），那么面对监管机构的性质就将发生彻底改变，引出更为严峻复杂的合规质询。 - 关键考量：合规的核心在于 AI 决策链条中每一步的“可审计性”与“可追踪性”，确保系统有据可查。 - 业务边界：当前聚焦于优化内部作业流程，严格避开直接提供外部自动化投资/理财建议或重塑底层核心信贷产品的监管雷区。 --- ### [41:24](https://www.youtube.com/watch?v=z27M2kzQQEk&t=2484s) 小微企业（SMB）入驻审核与欺诈防范的技术拆解 在最后的观众提问环节，一位参会者针对此前讨论中提到的小微企业（Small Business）入驻场景发问，询问这究竟主要侧重于贷款发起（loan origination）、承保（underwriting）、财务报表拆解（financial spreading），还是贷后流程？ Young Pham 对此进行了详细拆解，指出技术方案大量呼应了 Olga 此前分享的思路，尤其重度依赖于智能文档审核与对账验证。针对小微企业，AI 系统正在自动化处理包括公司章程、小企业个人信用材料以及各类损益表在内的多源非结构化凭据。AI 的核心任务不仅在于核验“这份文件是否真实有效”，更在于跨文档比对申报数据之间的矛盾与细微差异，以便在早期排查潜在的欺诈行为。相比于针对个人客户仅需调用 Experian 进行标准信用评分查询，小微企业的身份验证要繁琐得多；AI 与外部企业数据接口深度打通后，免去了人工肉眼核对不一致项的枯燥步骤，极大压缩了手动、重复性劳动的耗时。 ```mermaid flowchart TD accTitle: 小微企业 AI 自动化入驻与审查流程 subgraph S1["材料提交与摄取"] A["小微企业多源文档<br/>(公司章程 / 财务损益表 / 申报材料)"] end subgraph S2["AI 自动化验证核查"] B["文档真实性核验<br/>(鉴伪 / 防篡改)"] C["跨文档数据对账<br/>(识别矛盾 / 欺诈筛查)"] D["自动化企业身份核验<br/>(直连外部数据征信接口)"] end subgraph S3["审批与落地"] E["加速入驻与承保<br/>(告别低效人工初审)"] end A --> B A --> C A --> D B --> E C --> E D --> E ``` - 核心审查环节：AI 重点承担小微企业的公司章程、个人证明及财务报表的自动化解析与真伪鉴别。 - 差异化对账：利用模型横向比对多份提交材料间的细微矛盾，防止客户虚报数据或进行欺诈。 - 自动化质跃：相比个人简单的 Experian 征信，小企业审查通过 AI 串联多方系统，直接取代了传统耗时费力的人工作业。 --- ### [42:59](https://www.youtube.com/watch?v=z27M2kzQQEk&t=2579s) 尾声趣谈：从监管“沙盒”到童年沙滩 活动临近尾声，主持人 Nathan Bomey 幽默地借用媒体惯例做结。他提到 Axios 旗下所有通讯刊物通常都以“一件大事”（one big thing）开篇，因此在活动结束时，他们最喜欢以“一件趣事”（one fun thing）来收尾。借着当下金融科技圈最炙手可热的流行词“沙盒”（sandbox），Nathan 笑着抛出了一个应景的双关提问：两位嘉宾小时候到底在真正的沙坑（sandbox）里玩过吗？ 台下顿时响起会心的笑声。Young Pham 笑着回忆道，自己小时候确实没怎么去过社区那种沙坑，因为天气实在太热、太潮湿了；不过，他们当时在海滩上度过了漫长的童年时光，所以在他眼里，广阔的“海滩就是我们的大沙盒”。Nathan 随即打趣总结道：“这叫成功逃离了沙盒（Escape the sandbox）！”在一片轻松欢快的气氛中，主持人号召全场为嘉宾的深度分享热烈鼓掌，圆满结束了本次关于银行业 AI 实战价值的研讨。 ## Notable quotes > [10:23](https://www.youtube.com/watch?v=z27M2kzQQEk&t=623s) This one I don't have an answer for it — Young Pham > [27:12](https://www.youtube.com/watch?v=z27M2kzQQEk&t=1632s) hybrid approach between digital and human — Olga Ceja > [28:56](https://www.youtube.com/watch?v=z27M2kzQQEk&t=1736s) giving humans the capability to have a deeper financial relationship — Young Pham > [30:34](https://www.youtube.com/watch?v=z27M2kzQQEk&t=1834s) FINRA 只允许银行内特定具备资质的人犯错，而不是允许所有人都在那里瞎建议 — Young Pham > [39:17](https://www.youtube.com/watch?v=z27M2kzQQEk&t=2357s) 合规只会说不 — Young Pham > [42:59](https://www.youtube.com/watch?v=z27M2kzQQEk&t=2579s) 海滩就是我们的大沙盒 — Young Pham > [42:59](https://www.youtube.com/watch?v=z27M2kzQQEk&t=2579s) Escape the sandbox — Nathan Bomey > [06:11](https://www.youtube.com/watch?v=z27M2kzQQEk&t=371s) 银行出错，向你支付 10 美元 — Nathan Bomey ## Mentioned Young Pham, Nathan Bomey, Olga Ceja, Grupo Financiero Banorte, CI&T, Axios, JPMorgan Chase, OpenAI, Hugging Face, CFPB, FINRA, Uber, Experian, WhatsApp, EU AI Act ## Video description Young Pham - Chief Strategy Officer, CI&T Nathan Bomey - Reporter, Axios Olga Ceja - Executive Director, Artificial Intelligence and Innovation, Grupo Financiero Banorte Banks are no longer asking if they should use AI, but where it brings the most value. This panel dives into front-to-back use cases across the banking stack. Learn how leading institutions are deploying AI to catch fraud faster, serve customers better, and manage risk more proactively. Walk away with tactical insights on what’s working, what’s hype, and what’s next [Source](https://www.youtube.com/watch?v=z27M2kzQQEk)

[10:23] Young Pham: This one I don't have an answer for it [27:12] Olga Ceja: hybrid approach between digital and human [28:56] Young Pham: giving humans the capability to have a deeper financial relationship [30:34] Young Pham: FINRA 只允许银行内特定具备资质的人犯错，而不是允许所有人都在那里瞎建议 [39:17] Young Pham: 合规只会说不 [42:59] Young Pham: 海滩就是我们的大沙盒 [42:59] Nathan Bomey: Escape the sandbox [6:11] Nathan Bomey: 银行出错，向你支付 10 美元

[0:04] Olga Ceja: democratize access to high-quality financial products and aid customers in making smarter decisions [0:04] Olga Ceja: a personalized financial coach can be put into the hands of every single customer [0:04] Young Pham: grant institutions the creative freedom to rethink internal operations and broaden product access [2:12] Olga Ceja: relies on extensive experimentation and rigorous evaluation frameworks ("evals") to ensure the accuracy and tone of AI-generated responses [2:12] Olga Ceja: preventing hallucinations and controlling the style of communication [2:12] Olga Ceja: synthetic data, allowing testing teams to simulate numerous complex scenarios and diverse customer personas [2:12] Olga Ceja: maintain complete ownership of their models' intellectual property and ensure sensitive data never leaves their internal data centers [2:12] Young Pham: determine where to maintain a human-in-the-loop versus when to make a workflow fully autonomous [2:12] Nathan Bomey: bank error in your favor [10:23] Young Pham: reduce loan closing timeframes from 45 days down to 20 days [10:23] Young Pham: token consumption quadrupled over a single year due to inaccurate forecasting [10:23] Olga Ceja: Matching specific use cases to appropriate models rather than using the largest, most expensive model for basic Q&A [16:32] Olga Ceja: AI automatically validates 90% of submitted official documents, leaving only 10% requiring manual human inspection [16:32] Olga Ceja: handles 2 million to 3 million customer messages monthly. Without this system, the bank would require an additional 500 contact center agents [26:21] Young Pham: warns against an impractical deployment by a major financial institution that exposes tens of millions of customers, permanently eroding public trust

[0:04] We're really excited to have you here. This is  going to be a fun conversation. Now, the title [0:09] of the panel says something about smarter banking,  which does suggest there might be something such [0:14] as dumber banking perhaps. Uh are we in are  we moving from the dumber banking era into [0:20] something that's going to be much more grand and  powered by AI? I absolutely uh think that there's [0:28] um the the use of of AI in banking absolutely will  democratize the access to better products for our [0:37] clients and uh help them make smarter decisions.  So for example, right now uh you can have uh an [0:48] specific executive or relationship manager uh  for every single client just VIP clients and uh [0:56] with AI now you can have a a relationship person  a coach a financial coach uh at your hands for [1:05] every single person. So it's going to be smarter.  My guess it's yes, of course. Yeah, I I would echo [1:14] what Olga said and I would add on top of that.  Um I've been banking uh services uh banking side [1:22] JP Morgan Chase for 30 years in the industry. The  excitement is the reinvention of all the workflows [1:30] that quite honestly we've been fixing the same  way. It's always the same thing. replacement of [1:35] core, looking at the digital experiences, looking  at risk, fraud tools, all that. But now we get to [1:42] have so much more creativity to start thinking  about what we can do, how we can even change [1:48] the way the banks themselves operate to provide  access to those products uh that that you talked [1:52] about. yoga. I feel like in the early days of  the AI conversation, a lot of the focus was on, [1:59] okay, the last place you'll ever see AI is in  finance because, you know, we have to be careful [2:05] with people's money. So, with that in mind, how  does that guide the decision-m that you make? [2:12] Um, of course, for example, at least in in Mexico,  there's still no specific regulation for AI, [2:19] but we do have a specific reg regulation  for products that we use in in the cloud. [2:26] uh we are one of the most regulated uh industries  worldwide and we need to be very very careful as [2:35] you said because it's people's money and if  something we sell at least from from Banort [2:43] that it's that second largest bank in Mexico it's  trust so you can't lose trust so how do we deal [2:50] with that uh we have a lot of experimentation uh  First we uh run a complete set of evals that make [3:03] us certain of the answer that the AI it's going  to be giving to our customers. So um it's it's [3:13] uh of course the evals and I I I would say that  there are many kind of evals right and it's not [3:23] also like a software development um system that  you go into Q&A uh and and into into testing and [3:33] performance. It's more about uh no hallucinations.  it's more about the way that the AI will give the [3:43] answer to to your customer. So I think um the the  the way that we select it's about running these [3:52] evils and making sure that we have many many uh  test scenarios. Now it's also uh the the the the [4:04] good purpose of this is that you can create a lot  of synthetic data that in the past giving a lot of [4:12] scenarios was completely difficult take a lot of  time and now with the use of AI you can run a many [4:21] tests with different scenarios different personas  because you have the way to create these different [4:28] personas with synthetic data through AI. So, um I  think that is something that we we create and also [4:37] that we need to own the the intellectual property  of the models and we need to make sure that the uh [4:46] information doesn't go out from our inh from from  our data centers. So that is something that we as [4:54] financial institutions we need to be very careful.  Yeah. Um, I actually think banks are more primed [5:03] to be able to implement AI. Traditional machine  learning and AI have existed in fraud models, [5:10] fraud detection uh, within banks and good strong  governance and compliance has always been there, [5:16] right? This is not a new way of operating  relative to maintaining that trust that that [5:22] you talked about. It's adding on another  dimension. Um so a lot of the work we do [5:29] um the it's it's reinventing sort of how we look  at some of the ideas um so that if there are it to [5:38] me AI is is constantly developing so if there are  uh accuracy problems hallucinations understanding [5:48] how to stage gate that and really under as  technology is deployed really understanding [5:54] where to put it where do I still insert human  in the loop as part of this governance model. [5:58] When do I make it fully autonomous uh in terms of  automating a workflow? Um but I I I'm I'm actually [6:06] very bullish. I'm going to be the most optimistic  person here and really be the cheerleader for this [6:11] crowd. Well, Olga, when you said something about  hallucinations, I started dreaming about the bank [6:16] hallucinating that I have more money than I  [laughter] do. But um you know, it's kind of [6:19] like when you play Monopoly, you know, you get  that community chess card and it says bank error [6:23] in your favor. collect $10. I think could be more  maybe. But um okay, so this week uh we actually I [6:29] say last week I believe it was we saw in the news  open AI talking about this breach essentially that [6:35] happened with one of its models and hugging face  was the uh victim so to speak which I think all [6:41] for this sake of this conversation it sort of  tells us okay what are the dangers and the the [6:46] big word of the month is sandbox like keeping  everything in the sandbox. How do banks ensure [6:53] that it all stays in that sandbox? Well, at least  um this is as as as you were saying, this is not [7:01] new for the banks. We have been experimenting with  new technologies uh for for a couple of years back [7:08] and the creation of sandbox is something regular  for for for the banks and uh very important for [7:15] our security teams that they they consider these  sandboxes to be secure and that you can experiment [7:24] with open source models with tools that just  came out yesterday before you run all the the [7:31] the the technical evaluations and everything but  the creation of sandbox we have been doing it [7:37] for for a couple of years back actually also from  from regulation um the the regulator put a sandbox [7:46] uh at least in in in Mexico for some fintech  erh things so I I guess it's it's now for us [7:54] it's it's normal the kind of of sandboxes that  we use for experimentation yeah And and I would [8:01] echo that deming. Uh you said something uh uh  a little earlier about owned IP and really uh [8:09] owning not just the data but the the the learning  models that are associated. I think that's a big [8:14] topic of conversation uh having those sandboxes  sort of uh wall gardening what you need to do so [8:22] that you can have areas for experimentation.  I do think it's much easier if you're a large [8:29] financial institution versus uh some of the  smaller firms that may be represented here [8:34] right out of the thousands of banks in in the US.  I think at least relative to that maturity to have [8:43] uh that that technology infrastructure is probably  the the disparity that exists in in the market [8:50] today. Well, that's a really interesting go ahead.  Let me tell you the the the challenge that I see [8:55] that it's different from the past experimentation  with other technologies. So experimentation would [9:02] have take years until you you see something in  production. So I guess the challenge for big banks [9:10] and regulated industry is now that experimentation  just takes sometimes weeks or days and absolutely [9:20] the time to market it's changing completely with  this technology. So I think that is one of the [9:26] challenges because if you want to put something  into production uh very quickly and that you have [9:34] tested already the balance goes good the problem  is all the processes that we need to take as [9:42] u as a regulated industry to put something into  production for our clients. So I think that's [9:47] the difference that experimentation is now a with  a completely different speed that in the past. I [9:55] think it's fascinating because I don't think  the average person necessarily thinks about [9:59] experimentation is something that banks do. Um  but young you you mentioned something I think [10:04] would be interesting for the next portion of our  conversation. The sort of the tension between the [10:08] big banks and the small banks on this issue.  Um, in the US, we know we have a, you know, [10:12] pretty vibrant community of smaller financial  institutions. Not as much the case the rest of the [10:18] world, which is interesting. When it comes to AI,  how can the smaller institutions keep up? Yeah, [10:23] I I think certainly um the there are different  ways to look at the idea and we talk about human [10:33] in the loop versus full automation or what's  AI assisted. there's a significant amount of [10:38] task that they can isolate on and go through,  right? Uh a lot of credit unions we work with, [10:44] uh mortgage loan processing is a huge issue.  So, being able to even isolate I AI use cases [10:50] on that end, uh and eliminate some of that, uh it  really expedites your loan closing time frame. So [10:58] in those scenarios, it's not necessarily a full  transformation of the bank, but certainly I can [11:04] still have productivity and really start to apply  it um in areas that are are that we can really [11:11] improve upon, right? Doing underwriting and loan  close from 45 to 20 days. Uh very unimaginable [11:20] for a smaller institution that was really relying  on human uh capital as part of that. But when I [11:28] when I think about small institutions and AI, the  first thing that comes to to mind for me at least [11:34] is cost concern. Certainly a big conversation  in the last several months in AI is, you know, [11:39] token usage and can a smaller how can a smaller  bank keep up if AI costs so much? You want to go? [11:50] This one I don't have an answer for it. [laughter]  it we took FinO ops and the idea of token usage I [11:57] think is very it's top of mind for everyone  um there's different models at institutions [12:04] some some hard cap it right we work with clients  that hard cap the availability of their tokens [12:11] um which I don't think it's a productive model but  it it guards around uh unforeseen risk right now [12:18] um I think token costs are going to continue to go  because ultimately these AI companies have to make [12:24] money. Um, and just even thinking about any cost  savings I'm making on some of the human capital [12:33] replacement, it's almost negligible sometimes with  the token cost. So, I wish I had a better answer [12:39] for everyone on that. We struggle with it as a  company. We were very um uh we went hard on AI and [12:45] then our our token consumption as a technology  delivery company uh quadrupled uh in the last [12:54] year. Um so the good news 99% of our people use  AI in what they do. The bad news is we we've not [13:03] accurately forecasted how to maintain that cost.  Um so we'll go I'm hoping you have a better answer [13:08] than I do. I don't know if it's a better answer  but I guess uh everyone is struggling with that [13:14] and uh there are a lot of techniques. Um so  for example choosing the right model for every [13:22] specific use case you do not need to use um uh  the big the biggest or the or the most intelligent [13:31] model for asking simple questions Q&As's um so so  that is one one technique that we we we do choose [13:42] a specific model for a specific use case um also I  think this is changing with now the performance of [13:50] open-source models that are more accessible  and I guess that it's going to be pushing [13:57] uh the big companies to reduce also some some some  cost access uh models that are that are cheaper. [14:06] Uh but I think it's something that everyone needs  starting to to to learn how how do we do it? um [14:16] capping the the the consumption and um also one  one important thing is uh for example in in the [14:25] in the prioritization of the of the initiatives  that we want to develop what is going to be [14:32] really the benefits of using something with AI  is it something that really needs AI because now [14:42] everyone needs some some use case in in their in  their in their area and they raise a hand and hey [14:49] I need a use case um with an agent well maybe it's  not even an agent it's just a simple retrieval [14:57] augmented generation chatbot that can perform  the the the task so it's knowing the benefits [15:04] that it's going to have for the organization it's  knowing the techniques also in in the more in the [15:11] technical side there are techniques on how to  build better. How do you manage the context? [15:18] So if you are always uh having inputting  the the context every single conversation, [15:25] well of course the token consumption will will go  very big. So you also need to have best practices [15:32] when you are getting into the development and  use of the of the AI. Yeah, I agree. You know [15:38] the best practices we've reported on that a  lot at Axio. In fact, one of my colleagues had [15:43] a story that got a lot of attention because a CTO  told her that a lot of their employees were using [15:48] AI to check the weather, [laughter] which wasn't  exactly what they en envisioned as the usage for, [15:55] you know, for the tool. Um, but, you know, I  I used AI the other day to explain the themes [16:01] of the Odyssey to me. You know, [laughter] I  don't think that's what they wanted either. But, [16:05] um, but no, I I do think uh I do think costs  are a big part of this conversation. And I've [16:09] always said that just because technology exists  doesn't mean it'll be adopted. And you know, [16:15] I a good example would be burger flipping robots,  you know, uh have existed for a long time, [16:22] but most fast food companies haven't adopted them  because they're too expensive. So, I do think [16:26] it's an interesting conversation for banks. Now,  let's pivot to some of the real promise of AI when [16:32] it comes to banking. What are some of the most  interesting uh I guess examples of innovation that [16:38] you're seeing when it comes to you know AIdriven  advancements in banking? Um I I touched on it uh a [16:48] little bit before. I think last year was a a big  concentration relative to how AI really sped up [16:55] the software development life cycle. So as banks  you could do more. You could rebuild applications [17:00] you could modernize. you can create synthetic  data, you could modernize your data pipelines. [17:06] I'm we're we're seeing a fundamental shift  and we're actively participating in a shift [17:11] where now it's being applied into those business  workflows. Uh I talked about the lending workflow. [17:18] Um but what we're seeing as a common pattern is  uh particularly areas where a lot of the the task [17:25] uh redundant task and those patterns exist where  automation can go in there to really make it more [17:31] efficient. Uh other than uh credit underwriting  uh customer onboarding particularly for [17:38] um uh small business uh commercial clients things  like that that exist there. Um those those to me [17:47] have been the the really fun use cases to go on.  Um because what AI has done is really develop not [17:55] just those capabilities to do it faster but  you're you're doing value realization with [18:00] those customers because you're getting them  uh access to you as a financial institution [18:05] sooner. Right? So if you're onboarded it doesn't  take you two to four weeks anymore. It's one day, [18:11] two days a small business and it's a laborious  task. Same thing for mortgage, same thing for uh [18:18] insurance. Uh a lot of those capabilities there.  Yeah, I would agree. Uh for us, one of the most [18:25] important use cases it's customer boarding. So you  need to validate that the person that it's opening [18:31] the account, it's really the person that it's  telling you it is. So we use a lot of biometric [18:40] um to validate this. So um not al biometric  technology but also uh the customer needs to [18:49] load some documents many documents. So we need to  verify if the uh document it's official uh extract [18:58] information and I can tell you that 90% of the  documents are now validated automatically through [19:05] AI and only 10% needs to go to a human validation.  So that it's a completely different experience for [19:13] our customers that can open an account completely  digital in just minutes couple of minutes or apply [19:21] to a mortgage a credit card a loan u everything  that it's digital that is something that it's [19:28] uh very very successful and talking about  customer support I guess it's one of the other [19:35] u use cases that are very very good for us so for  example our our chatbot if it's getting I guess [19:45] two two to three million messages in monthly  basis and if we do not have that chatbot we [19:54] do need additionally uh additional persons  in the contact center getting through them [20:01] uh approx like five uh 500 persons getting into  the the the the queries that the chat at what [20:10] it's already answering. So those are things that  are getting completely different. Um there's also [20:17] one important strategy for us in Banorta that we  call hyperpersonalization. So hyperpersonalization [20:24] it's just not getting a generic offer for for you.  And what it's a generic offer well the credit card [20:33] with the with the interest rate. So now through AI  we can make these specific offers per customer and [20:43] interest rate different for you and an interest  rate different for you. Even the message and [20:48] the image that you receive are created with AI  and they are hyperpersonalized for every single [20:54] customer. It's pretty fascinating. And so and I  think Young mentioned earlier the role that AI can [21:00] play in risk modeling um fraud detection that sort  of thing. How close though are we to getting to a [21:06] place where maybe AI could start to make lending  decisions without a human ever being involved? [21:15] I am I going to ask chat GPT? Well, you know,  preapproval is part of the lending process. That's [21:22] that's in there. Uh that's happening a lot faster.  Um I ultimately it's it's a compliance issue I [21:29] think that that exists as part of that. So uh the  comfort level with the financial institution now [21:36] there's got to be audibility and traceability  talked about that even on offers when we build [21:41] those personalization um that exists. So um having  that built in I think eases the comfort level. Um [21:50] most most of the clients we work with are are  not there yet. Uh technically it's capable. [21:56] It it is just more of a compliance guardrail  that that is is sort of limiting uh that that [22:04] final step I think. What do you mean by  compliance guardrail? Well uh uh making [22:10] a a a full loan decision without some aspect  of of approval. So within the loan origination [22:18] process certainly there you know having um uh  straight through processing is is always existed [22:25] um running pre-approvals with better information  that you're getting um exist but then uh [22:34] uh full loan approval and funding uh at at some  levels depending on the institution. Um there's uh [22:43] um part of that ex uh has a human as part of that  uh decision criteria. Certainly banks obviously [22:50] operate one of the most regulated aspects of  society. When it comes to regulation though, [22:56] you mentioned earlier in Mexico effectively  nothing when it comes to AI certainly not [23:00] much in the US as well. That may change, but how  do you think the regulatory landscape needs to [23:06] evolve to help banks take advantage of this? Help  banks to help banks take advantage of the benefits [23:13] of AI. Do they do they need to have some sort of  regulatory framework to be able to succeed here or [23:19] is there too much gray area right now that will  prevent them from really being able to do much? [23:24] I think definitely we need some kind of  regulation. Um as every single new new technology [23:32] um it needs to be used with ethics responsibility  on how are you using it. Uh I guess it's going to [23:41] be very very soon. Uh there are some conversations  at least in in in Mexico within the the regulators [23:47] to put some and the government to put uh this kind  of regulation. um just uh what I hope is that that [23:57] doesn't um slower our our innovation processes  and the use of AI in benefit of our customers. [24:05] As as you were saying, we already used AI for many  many years back in order to give a a an approval [24:14] of a loan of any single product. there are a lot  of machine learning um models already there that [24:21] are being audited by the regulators. So I think  as long as you can trace every single decision [24:29] that the AI is doing and and it's transparent  for for for this I don't think we will have [24:36] any issue with that. Any thoughts on that the  regulatory landscape? Uh it it's interesting. Um [24:44] uh we work globally. Um EU banks that we work  with are way ahead of where we are. Uh the EU AI [24:53] uh uh governance framework that they've put  out it has started to help um create regulatory [25:00] frameworks for those institutions. Um the US it  every two to four years I I I don't know who to [25:09] bet on. You know, I I was bullish on um uh open  banking and the CFPB ruling that was supposed to [25:16] happen two years ago that now is sort of gone by  the wayside there. I do think um what banks can [25:24] do is certainly the the traceability that that  we've described on there. Um but I do think that [25:34] uh for banks to really do it to ensure trust  which I think is the important issue is that [25:41] uh uh within banks themselves particularly in  the US we can set common standards and practices. [25:48] We've done it on payments. We've done it on these  things. So there's operability on the banks. Um, [25:53] but I I I think that's probably going to happen  more organically than uh more of a top- down uh [26:01] regulators sort of coming in at least in the  US until sadly something catastrophic happens. [26:07] Sure. Well, I do imagine that that will  happen at some point and then it'll be up [26:11] to the industry to communicate and to navigate  through that. What's your worst nightmare? Um [26:21] um just a impractical use of it um um by uh one of  the large financial institution that exposes tens [26:31] of millions of customers and there's a significant  lack of trust um in banking and and the system. [26:39] Yeah. [clears throat] Or losing data. Sure. Sure.  Um all right. So let's talk a little bit about the [26:45] human's role um you know in terms of back office  but in terms of the human relationship in banking [26:50] itself with the customer. We know that banks have  been evolving for years now certainly declining [26:55] number of tellers um and I'm sure many of you are  able to do most of your banking without actually [27:00] speaking to a human but the human does come into  play a role in various stages of the banking [27:05] process. what role is there going to be left for  the human when AI really continues to advance [27:12] uh in finance? So I think there's very uh there  are products that are very specialized. So I [27:20] guess that um my my idea is that these kind of  persons can have more time really to specialized [27:29] in these kind of products so that they can give  and have more time to give a better guidance to [27:36] to the customers. So whenever you for example make  a decision uh to get a mortgage there are specific [27:46] decisions that you want to validate with a human.  So uh this kind of guidance and that the the the [27:54] tellers and the persons at the branches have  more time to to really guide our our customers [28:00] with this specialized information. I think that  is one of the the thing instead of making routine [28:06] tasks at the at the at the branch. I guess that  it's something that we look at that it's u it's [28:14] a hybrid approach between digital and human. There  are still a lot of people that uh goes to the goes [28:21] to the branches but for this kind of of of of  things for the mortgage or getting guidance for [28:29] investments for stocks. So maybe right now our our  branches are not um prepared for for that and I [28:40] think that will be a shift that instead of having  just making transfers or getting money in the [28:47] branch that it's a space for getting a financial  guidance. I agree the the financial guidance piece [28:56] um you have more access to information knowledge  sharing so that relationship uh there's greater [29:03] expertise uh uh that someone that is in contact  with a customer can have that can talk about a [29:10] more holistic financial picture. Um I'll use a  story of a of a client of ours who's a top 10 [29:17] u mortgage provider in the country. They've got  3,000 loan officers throughout the US. Um they [29:23] acquired a bank, small bank, but now um they have  the ability to actually start to be competitive [29:32] uh in the banking, grow deposits and go through  that with the use of AI. It's not hiring bankers [29:39] and tellers. It's giving that capability for their  3,000 loan officers to actually understand bank [29:47] products, savings products, CDs, money markets.  And so, one of the ways they've used it isn't a [29:53] replacement of human capital, but how do I make  people who are fantastic at the mortgage business [29:59] be able to talk about savings products, sa  helping their kids uh for college? And so [30:05] it to me it's a very interesting inverse of what  people always talk about relative to AI replacing [30:13] humans but giving humans the capability to have a  deeper financial relationship. Um and from a cost [30:20] perspective obviously um um not having to go out  and build another thousand bankers in field is has [30:28] been a really highly aggressive growth strategy  for them to be able to do that versus just a cost [30:34] uh replacement strategy. Yeah. Um, you mentioned  financial advice or investment advice. And it is [30:41] interesting because I do think a lot of people  are already using AI in some capacity to help [30:46] develop a financial plan or to help potentially  give them stock advice. And I don't know if you [30:52] should take that stock advice or not, but um,  but it's always hard to predict. I mean, I was on [30:56] the SpaceX earnings call yesterday and everything  sounded good, but if you look at the stock today, [31:00] it doesn't look so good. So, I don't know. I  mean financial advice when it comes to investment [31:04] advice in banks, how much do you think that is is  going to be done outside of the bank's ecosystem [31:09] in in terms of the customer is doing it on their  own versus the bank is going to help provide it [31:13] to them. Banks are still the trusted resources. I  do think that people and I do it too. You're going [31:22] to do research. You're going to look at things.  But um um I I was on a field trip with some kids [31:31] um that we went to a bank, you know, for my  field trip to a bank. That's what that's what [31:36] the kids school elementary wanted to do. Um what  an aquarium. What was interesting was every kid [31:43] there said, "Wow, can we see the safe? Can we  see the safe? This is the most trusted place." [31:51] That does sound fun. Everywhere. Right. And  so as as as you look at these five, six years, [31:56] even even older, that that trust is been around  for hundreds of years. So there it to me, [32:05] it's no different than talking to a doctor. I'm  still going to look at other stuff, but when [32:10] it comes down to real financial preparation and  planning, I talk to my banker. I talk to my loan [32:16] officer. I talk to my uh wealth manager. Um, I may  bring up ideas and they may shoot me down and go, [32:24] "Space X is a terrible investment. Don't do it."  Uh, but uh um I think what's interesting with AI [32:30] is giving them the ability to help uh communicate  more and understand and have more power [32:38] uh to really drive the relationship. Yeah. And  the bankers will have more data that before with [32:45] a without AI it was impossible just to process  all all this kind of data. So banks we have uh [32:53] plenty of data. The problem was that we were not  using that data in the benefit of of our customers [32:59] but data is there. So I think with with AI I  completely agree that uh you still find someone [33:08] uh at the bank that you that you trust. Even  with the use of AI though and investment advice, [33:14] it does mildly concern me only because what if  you know one in a thousand cases the bank gives [33:21] me really bad advice with the AI just you know for  some reason says you know go put it all in red not [33:28] to you know with apologies to this institution.  [sighs] Well, FINRA only allows certain people [33:33] to be wrong at the bank as opposed to everyone uh  at the bank. But uh I I I do think also there is [33:45] uh we always look at a younger generation of  people that that always come up particularly [33:50] that that look at access to financial products,  right? And and to your point, Nathan, there may be [33:57] a slightly different way that they want to engage.  um and AI may certainly help them as part of that. [34:03] But we also see that where as they accumulate  wealth, it's digital bank to uh something that has [34:12] more of a full service relationship. So um really  haven't seen that model change uh tremendously. So [34:18] I do think that you know that opportunity does  still exist at the banks. Uh I have one more [34:24] question and I think we're going to have just a  few minutes for questions from you. So, if anyone [34:29] has a question or two, think of those. Uh, and  we'll be around with the microphone in a moment. [34:32] Uh, tomorrow at one of the keynote sessions, I'll  be interviewing the CTO of Uber, and he made some [34:39] waves, uh, a few weeks ago by saying that at this  point, they haven't really seen any significant [34:45] consumerf facing features come out of their AI  investments. Um, and I've heard some pretty strong [34:50] takes on both sides of that. Some people who think  he was right on and some people who think he was [34:54] way off. What do you think? I I I I think um I ran  a session yesterday at a round table going from P [35:06] to scale. Um I think uh how AI has been adopted.  It has been very much at an experimentation stage [35:15] and really thinking about how to to get those  ideas out there. I do think there's a fundamental [35:21] shift right now to scale those commercially within  an enterprise and particularly for a bank. So, [35:29] um I don't think it was a lack of of capability or  that the technology wasn't there. It was just how [35:38] organizations uh started to dip their toe into it.  But we're we're seeing way more investments around [35:46] it. uh way more understanding of the business case  and the long-term impact that it has to the the [35:52] business case. So I I would probably argue that  true maybe there's not large scale significant [36:00] impact but it's it's there a lot of the work  we're doing is scaling those pilots um to really [36:09] transform these these lynchpin parts of the bank.  Yeah. I I think this year it's the the year that [36:16] uh enterprise large enterprises are getting really  into implementing it and stop just experimenting. [36:24] I guess uh at least we have seen very very  good results in productivity. Um, for example, [36:30] for the contact center, every single employee in  the contact center has uh a virtual assistant. [36:36] Instead of taking the the phone and trying to find  a supervisor getting the the the query solved, [36:44] uh, also our executives at the branches, people  at the branches have also these virtual assistant [36:51] helping them. So of course we have seen very very  good use cases and I I think that this year will [36:59] be shifting these these revenues and and and  these results. Certainly sounds like we're on [37:05] the doorstep. So with that we will like to maybe  take a couple questions if we can fit them in. I'm [37:10] sure someone has something pressing. Only good  questions allow. [laughter] Well my mom always [37:16] said you you always ask the question it doesn't  matter what you get. [laughter] So you were [37:21] talking about earlier about using the data and  can you speak to a use case that you've actually [37:29] utilized some of that data to either improve  the business or the consumer side preferably [37:34] the business because there's always a ton  of data but what's the right use case that [37:38] you've seen actually come to fruition. Yes. So  um specifically this case of hyperpersonalization [37:47] so you know where is your your customer making  transactions. So for example when you're buying um [37:58] uh a flight okay or maybe you're at the hospital  and you're getting out of money. So if you have [38:06] that data in real time, you can make the decision  to allow uh to increase the the the the limit of [38:16] that customer that it's already pre-approved. So  it's not just only using the data but using it at [38:23] real time. So that is one of the of the points and  hyperpersonalization for the messages for getting [38:31] of a specific your interest. How do you want me  to talk to you? What are your interests? Maybe you [38:37] are very uh tech person. Some persons are maybe  more uh education. So it's it's just that getting [38:46] all these data so I can give you the best message  also for you, not only the the offer and and the [38:53] the the right time for you and the right channel.  So some people it's just at least in Mexico the [39:00] the the most uh used channel it's WhatsApp. So  okay I know that this person it's through WhatsApp [39:08] but other persons are using still email or SMS. So  it's the finding the right time the right channel [39:17] with the right message for the customers. How how  do you deal with the regulators especially when [39:23] you're about to roll out an automated solution  say uh in in your mortgage process or like credit [39:29] analysis process I mean you developed something  and you you are like on the verge of rolling it [39:34] out so how do you convince the regulators both  internal and external well in internally uh one of [39:44] the the models to actually do a more successful ai  deploy deployment is really bringing in compliance [39:52] early. So it it is a creative process versus uh  a review of a build solution. So we're seeing way [40:00] more uh understanding to where the legal teams,  the compliance teams sort of participate when we [40:08] we start to look at that workflow. Can I pull  these data values to do automated appraisal? [40:14] um and and if not, how do we think about that  triage that's built in? So, at least internally, [40:23] that has actually unlocked the capability a lot  more versus the safeguard of we can't do it. We [40:29] don't know if it's going to be accurate. I I don't  you know, we So, at least on that end, we've, [40:36] you know, that that's a lot more solved that we're  seeing with the financial institutions there. Um [40:43] externally um I think most of the work that we're  doing are really within the internal workflows and [40:52] uh the the biggest part of just working with the  regulators I think we talked about it ensuring [40:58] that there's audibility and traceability  throughout um but we're not uh inventing [41:05] any new credit products. we're not doing  some of these things that are fundamentally [41:11] uh necessary for the external regulators to come  in. Uh now if we start giving financial advice [41:18] for certain people that uh shouldn't then I then I  think it's a bit bigger question there. I think we [41:24] have a time for one more and then I'll finish this  up. Thanks very much. Um my question is I believe [41:29] there is a reference made earlier to use case uh  related to uh small business on boarding. uh would [41:37] would love if you could unpack that a little bit  more. Are you talking about more so on the side [41:41] of the loan origination, underwriting, financial  spreading, uh post origination, all of the above? [41:47] Uh appreciate it. Yeah, you were talking about  the small business. Uh um a lot of the things [41:55] that Olga talked about particularly when you look  at document review um corporate charter personal [42:03] C charter for small business sort of going through  uh verification of um income statements and things [42:10] like that. So that's a lot of the things that  we are running AI on to to do the same thing. [42:16] You know is this an authentic document? um are  there discrepancies in what they're providing [42:23] to reconcile that so that they may be conducting  fraud? Um identity checkpoints that become more [42:29] automated uh as opposed to a customer who  we you know you can run an Experian check uh [42:35] uh versus a business and and linking direct  to some of those services there. So all those [42:42] tasks are really just accelerated by building that  in. So you're not building in the manual review. [42:48] You're not trying to look for the inconsistencies.  Um uh particularly for small business. And so [42:54] we're we're really just expediting a lot of more  of those manual repetitive tasks there. Well, [42:59] at Axios, we like to end all of our events with  one fun thing since all of our newsletters start [43:04] with one big thing. Uh and since sandbox is the  word of the month, did you play in a sandbox as [43:11] a kid? I did. You did? I did. Or was it too  sandy for you young? [laughter] I've never [43:18] been and I really like it. And it was too hot  to play in the sandbox. It was too humid. But [43:25] uh we spent a lot of time uh at the beach. So  we we consider that our big sandbox. Escape [43:31] the sandbox. All right. Well, how about a hand  for our panelists? Thank you so much. Thank you.