Banks Face Regulatory Barriers Over AI Explainability, Executives Warn

高管警告：AI可解释性监管要求令银行业面临落地阻碍 金融机构在部署人工智能模型时陷入困境，对模型可解释性和董事会责任的严格监管要求导致运营成本过高。 在AI4行业会议上，来自Commercial Bank International、BillGO和Global Payments的高管指出，金融机构在落地人工智能时面临严峻的合规摩擦。应用AI的主要阻碍并非模型调用或订阅费用，而是满足监管对模型可解释性与可审计性的严苛要求，以及董事会所承担的最终法律责任，这些因素往往导致运营与改造成本居高不下。尽管各机构已在边缘防欺诈、小语言模型POS终端、智能支付及合成数据信贷评分等场景展开部署，但因高管层普遍担忧法律风险与责任归属，许多项目仍倾向于规避风险或依赖人工流程。为此，高管们主张通过制定企业级AI章程、将合规前置嵌入产品研发流程，以及采用确定性AI框架来逐步化解这一矛盾。 监管对模型可解释性与可审计性的强制要求，叠加上董事会承担的个人法律责任，使得银行部署AI的合规及治理成本极其高昂，导致管理层因避险而暂缓落地或重回人工流程。 各金融机构正尝试多项前沿落地，包括Global Payments在搭载12 TOPS算力NPU的POS机上运行本地小语言模型，BillGO构建零人工干预的智能支付网络，以及CBI利用合成数据开展AI信贷评分。 为解决黑盒效应和追责难题，高管们建议推行跨部门企业级AI章程，将合规要求作为主动特性嵌入产品开发周期，并向可追溯、确定性的AI架构演进。

Fintech and banking leaders report that embedding artificial intelligence into financial infrastructure unlocks operational gains from edge hardware to credit underwriting, but broad implementation remains constrained by board-level legal liability, governance overhead, and asymmetrical trust standards. Edge Hardware, Agentic Workflows, and Synthetic Underwriting Financial institutions are shifting AI deployment from experimental software to practical edge and workflow automation. Arjun Wadwalkar details how Global Payments—which processes approximately $3.7 trillion across more than 6 million merchants—is deploying handheld Genius POS devices equipped with 12 TOPS neural processing units. These devices run local small language models directly on the edge, enabling merchants to adjust menus and configurations via natural language while screening transactions for fraud without cloud latency. In back-office and customer workflows, Satya Panda outlines BillGO's development of an agentic payment network engineered to reach a zero-touch operational state, eliminating residual manual reconciliation tasks for merchants. At Commercial Bank International (CBI), Giovanni Everduin reports that an 18-month deployment of AI credit scoring incorporated synthetic data to successfully underwrite credit for historically underserved demographics, specifically women lacking traditional credit files. Escalating Fraud and Algorithmic Trust Asymmetries (8:58) Deploying frontier models in financial operations has introduced a symmetrical escalation in security risks. Wadwalkar notes that while modern models excel at evaluating device, merchant, and transaction records concurrently to reduce false positives, Everduin warns that defensive gains are offset by a rapid rise in AI-generated deepfakes, hyper-realistic voice clones, and automated spear-phishing campaigns targeting bank personnel, who remain the primary attack vector. Simultaneously, a behavioral trust asymmetry complicates deployment. Everduin observes that while CBI's internal benchmarks show machine learning models consistently outperform human decision-makers, organizations demand flawless auditability from software while tolerating systemic bias and un-auditable errors from human staff. Despite current friction, Everduin forecasts that public acceptance of AI access will normalize within six to twelve months, mirroring historical comfort with email platforms scanning user data. Regulatory Liability, Model Explainability, and Enterprise Charters (12:37) The primary barrier to enterprise banking adoption is organizational and legal liability rather than model access fees. Everduin emphasizes that regulatory mandates for strict explainability make deployments costly, noting that an internal audit revealed third-party subcontractors using untracked shadow AI tools. Because legal accountability falls on the board of directors, risk committees frequently delay deployments or default to manual processes due to personal liability fears. To bridge this accountability gap, Wadwalkar points to the legal precedent set when an airline was held liable for financial losses caused by an inaccurate chatbot policy, arguing that black-box models without designated human owners are untenable in production. Wadwalkar advocates establishing enterprise AI charters that align risk, engineering, and compliance executives around fixed operational guardrails. Panda adds that overcoming board-level resistance requires shifting toward deterministic, fully auditable frameworks that guarantee predictable outputs. Radical Transparency and the Draft Shock Dilemma (40:04) Adopting AI tools is reshaping internal corporate dynamics. Panda describes BillGO's collective intelligence model, which shifts internal corporate knowledge from being 95% confidential to 95% accessible by default across emails, chat channels, and notes. This centralized access enables models to draft routine updates automatically; Panda reports that drafting executive updates, which previously required one to two hours of employee labor, now takes one minute of generation and five minutes of review. However, widespread generation introduces unexpected workflow bottlenecks. Everduin highlights "draft shock," where team members leverage generative tools to instantly produce voluminous, highly polished 500-page reports. This capability shifts executive workloads from task execution to exhausting evaluation, forcing senior leadership to spend significant time critically auditing aesthetically professional documentation that may lack substantive business value. While localized AI applications deliver measurable speed and hardware independence, enterprise-wide scaling in finance remains bottlenecked by unresolved governance questions: how boards can manage liability for non-deterministic systems, and whether organizations can absorb the audit overhead of machine-generated volume without eroding productivity.

来自金融与支付行业的专家指出，尽管人工智能在端侧商户工具、信贷普惠和自动化协同中展现出显著效率，但在金融风控对抗、合规可解释性要求以及董事会法律责任等多重约束下，规模化落地的核心瓶颈已从模型研发转向组织治理、责任归属与工作流重构。 端侧应用与信贷普惠：微型模型落地实践 金融机构正从边缘硬件与模型微调切入推动AI务实落地。Global Payments产品管理总监Arjun Wadwalkar介绍，该公司年处理交易规模约3.7万亿美元，覆盖100多个国家逾600万商户；其Genius POS平台推出了搭载12 TOPS神经网络处理单元（NPU）的手持终端，在本地端侧运行小型语言模型（SLM），使小微商户无需登录云端即可通过自然语言修改菜单和配置，并在本地实现实时欺诈监测。BillGO首席技术官Satya Panda表示，该公司正构建针对小微企业的智能体（Agentic）支付网络，核心理念是在交易流程中消除全部人工操作，实现零接触业务闭环。 在传统银行业务方面，迪拜商业银行（CBI）首席战略与创新官Giovanni Everduin披露，该行在过去18个月中应用了AI信用评分技术，并结合合成数据解决缺乏传统征信记录群体（尤其是女性）的信贷评估问题，在严守传统承保标准的前提下扩大了信贷供给范围。 双向博弈：AI防御提效与网络黑产升级 (8:58) AI在金融安全领域呈现出鲜明的双刃剑特征。Wadwalkar指出，支付清算的本质是风险管理业务，前沿AI模型能够并行评估交易、设备与商户的历史数据，表现明显优于传统静态规则系统，其理想状态是在后台隐形运行以降低正常交易的误报率，同时将高风险威胁无缝路由至人机协同审核流程。Panda同样主张金融机构必须采用主动预防型工具，替代损害客户体验的传统阻断式规则。 然而，Everduin警示防御端与黑客黑产的技术演进正在同步升级。针对银行员工与客户的AI网络钓鱼、诈骗邮件、伪造语音及深度伪造（Deepfake）视频正呈现“疯狂”增长态势，而一线员工依然是防御体系中最脆弱的攻击切入点。Panda补充指出，业内近期出现的安全事件（如Anthropic发布新工具后又收紧权限）表明，威胁行为者利用前沿商用模型攻击企业系统的响应速度正在急剧加快。 合规摩擦与归责真空：算法黑盒的法律约束 (12:37) 严格的监管要求与法律责任认定正在成为金融机构采纳AI的关键摩擦点。Panda认为监管合规不能仅作为产品周期尾声的审核环节，必须作为原生功能嵌入产品研发全流程。Everduin直言合规科技带来了高昂的运营和财务摩擦，监管机构对模型完全可解释性与可审计性的强制要求，往往导致银行部署AI的综合成本居高不下；企业应用AI的真实开销主要源于组织重组、流程再造、合规治理和潜在监管罚款，而非模型API的调用费用。 这种法律风险放大了金融管理层的避险心态。Everduin披露，CBI在风险排查中发现第三方供应商和分包商已在私下使用未经登记的AI工具，由于最终法律责任由银行董事会承担，面对无法量化的个人法律责任，高管层更倾向于暂停AI项目或退回人工流程；在内部风险委员会中，极少有业务负责人愿意为AI潜在失误承担责任。Wadwalkar引述加拿大航空（Air Canada）因客服Chatbot提供错误退票信息被法院判决承担全额赔偿的法律判例，强调AI模型决不能沦为不可解释的黑盒，每一项自动化决策都必须对应具体的人类责任人。 信任不对称与组织重塑：极端透明与草稿冲击 (36:17) 人机信任机制在金融场景中表现出显著的不对称性。Everduin指出，在银行内部的概念验证（POC）测试中，机器学习模型在决策准确率上几乎总是远超人类，但组织出于对机器出错的极度担忧而设置了苛刻的治理框架；相比之下，缺乏审计路径且充斥主观偏见的人类决策却长期被体系容忍。他预测公众对AI的信任度将在半年至一年内达到类似日常使用Gmail开放数据权限的普遍接纳水平。 在企业协作形态上，Panda在BillGO推行了以“极端透明”为基础的集体智能架构，将Slack沟通、会议记录、邮件及知识库等95%的内部数据默认向AI模型开放，使员工编写管理周报的时间从原本的1至2小时缩减为1分钟生成加5分钟人工复核，Wadwalkar亦将日常办公工具接入定制Claude智能体以自动处理常规沟通。不过，Everduin对这种全透明模式表达了审慎忧虑，并提出大模型普及引发了“草稿冲击”（Draft Shock）——员工能瞬间生成数百页格式精美的冗长报告，导致高管的核心精力严重受阻于对缺乏实质内容的华丽文档进行耗时甄别。 金融行业在AI应用上面临技术效能突出与组织治理滞后的深层矛盾。尽管端侧模型交互、信贷普惠和内部自动化已展现出清晰的投入产出比，但如何建立确定性与可审计的数据谱系、消除董事会与业务线在法律追责上的避险阻力，并应对“草稿冲击”引发的管理审核瓶颈，仍是机构在迈向高度自主化金融网络前尚未解决的核心课题。

## Background 本内容由 Ai4 频道于 2026 年 8 月 22 日发布在 YouTube 平台上。该圆桌论坛由《The Deep View》主编 Jason Hiner 主持，汇聚了来自迪拜商业银行（CBI）、BillGO 和 Global Payments 等机构的高管，探讨如何在满足强监管与风控要求的前提下，将 AI 落地于金融科技、反欺诈与支付流程之中。 ## TL;DR 论坛深入探讨了金融与支付领域从概念验证走向规模化落地的核心挑战与实践经验。嘉宾们分享了端侧 SLM 硬件终端、智能代理（Agentic）支付网络以及合成数据信用评分等具体应用，并剖析了深度伪造与新型黑产对传统防御体系带来的严峻威胁。针对合规难题，专家们强调必须将治理原生融入产品生命周期，建立企业级 AI 宪章以打破数据孤岛与责任归属困境。此外，讨论还提出了分阶段提升系统权限的“信任自主权阶梯”，并反思了 AI 泛滥导致的“精致草稿垃圾”对管理层决策效率的反噬。 ## Atmosphere & style 现场对话务实而犀利，充满了传统金融机构与激进科技先锋之间的理念碰撞与幽默调侃。来自传统银行的嘉宾以坦率自嘲的口吻揭示了董事会的合规焦虑与数据孤岛现实，与 Fintech 创业者推崇的“极端透明”形成了鲜明对照。整场讨论既有对前沿代理商业（Agentic Commerce）的技术兴奋，又透露出对 AI 幻觉和法律追责的理性敬畏，真实还原了金融精英在创新与风控钢丝上博弈的现场感。 ## Key takeaways - [03:40](https://www.youtube.com/watch?v=CULFxCllIJw&t=220s) 端侧硬件、智能代理与合成数据正在推动金融落地：Global Payments 探索搭载 12 TOPS NPU 的端侧 AI 收银终端，BillGO 打造全自动代理结算，银行则利用合成数据消除信贷历史偏见并推进普惠金融。 - [08:58](https://www.youtube.com/watch?v=CULFxCllIJw&t=538s) AI 在升级金融反欺诈能力的同时也带来了新型黑产威胁，微调大模型虽能实现后台静默拦截，但黑产与诈骗团伙同样利用前沿模型赋能犯罪，攻防双方进入长期拉锯战。 - [11:22](https://www.youtube.com/watch?v=CULFxCllIJw&t=682s) 音视频深度伪造（Deepfake）与高保真社工钓鱼呈爆发态势，在人机协同风控体系中，终端的“人类”（员工与客户）依然是黑客利用尖端 AI 武器突破的最大攻击向量。 - [13:06](https://www.youtube.com/watch?v=CULFxCllIJw&t=786s) 合规与监管不能充当系统上线前的“事后阶段门禁”，而必须在系统架构之初作为产品原生特性（Feature）融入全生命周期。 - [14:32](https://www.youtube.com/watch?v=CULFxCllIJw&t=872s) 董事会因承担个人连带法律责任而产生合规焦虑，加上第三方外包供应链中的 AI 盲区及高昂治理成本，促使传统银行在激进落地上面临巨大阻力并倾向观望。 - [16:30](https://www.youtube.com/watch?v=CULFxCllIJw&t=990s) 金融核心系统必须拒绝不可预测的黑盒模型，依靠具备完整数据血缘（Data Lineage）的端到端确定性模型，以明确决策的具体责任归属主体。 - [18:34](https://www.youtube.com/watch?v=CULFxCllIJw&t=1114s) 银行内部创新的“游说”策略在于绕开抗拒的业务部门，优先为合规主管与首席风控官开发提效工具，将其转化为制定行内护栏的 AI 坚定布道者。 - [21:42](https://www.youtube.com/watch?v=CULFxCllIJw&t=1302s) 企业级 AI 宪章（AI Charter）由跨职能团队共同评审确立，作为大型组织内部不可逾越的硬性护栏与统一创新边界。 - [23:31](https://www.youtube.com/watch?v=CULFxCllIJw&t=1411s) 金融与商户服务正在从通用的“一对多”模式彻底转向“一对一”超个性化交付，商户开始要求平台直接输出动态定价、智能排班等可指导经营的高阶洞察。 - [26:29](https://www.youtube.com/watch?v=CULFxCllIJw&t=1589s) 纯粹的支付通道费率模式已被严重商品化，金融科技企业必须在支付之上构建垂直整合的 SaaS 软件，以全链路数据洞察构建高附加值护城河。 - [29:40](https://www.youtube.com/watch?v=CULFxCllIJw&t=1780s) 代理商业（Agentic Commerce）需要打破传统卡基网络，探索专为自主 Agent 设计的原生支付清算轨道，并确立严格的身份授权与资金流向溯源机制。 - [34:04](https://www.youtube.com/watch?v=CULFxCllIJw&t=2044s) 由于用户对个人财务高度敏感且银行底层数据孤岛割裂，AI 财务系统必须遵循“只读-建议-协助-完全执行”的信任与自主权阶梯（Trust Autonomy Ladder）逐步赋能。 - [36:53](https://www.youtube.com/watch?v=CULFxCllIJw&t=2213s) 加拿大航空等司法判例表明法院认定企业须对大语言模型的幻觉输出承担连带法律赔偿责任，绝对准确性是金融与支付部署 AI 的核心基石。 - [40:04](https://www.youtube.com/watch?v=CULFxCllIJw&t=2404s) 组织协作正从信息孤岛转向“极端透明”的集体智能中枢；同时企业必须构建统一的知识图谱底座，以避免 AI 泛滥导致的“精装垃圾文档”造成管理瓶颈。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=CULFxCllIJw&t=4s) 开场介绍：将 AI 从概念落地到真实商业 论坛主持人、《The Deep View》主编 Jason Hiner 欢迎现场观众来到 Ai4 峰会首日。他指出，当前行业中充斥着关于前沿大模型与技术快速迭代的讨论，但企业从业者真正关心的是实际落地：面对陈旧未优化的 IT 基础设施以及对变革心存顾虑的团队，如何将 AI 转化为切实的业务成效。随后，三位来自金融与支付领域的嘉宾依次进行了自我介绍。 来自阿联酋迪拜商业银行（Commercial Bank International）的首席战略与创新官 Giovanni Everduin（大家习惯称他为 Gio）率先发言。他同时也是该行创新实验室与企业风投（CVC）的创始人。在正式发言前，他打趣地对着摄像机强调免责声明：“法律团队特别交代过，我今天分享的内容纯属个人观点，不代表银行或国家的立场。” 紧接着，B2B 支付平台 BillGO 首席技术官 Satya Panda 进行了介绍，说明公司正利用 AI 加速安全智能的支付流程，并聚焦于解决 AI 时代的风控与反欺诈瓶颈。最后，Global Payments 产品管理总监 Arjun Wadwalkar 介绍了公司背景：作为一家《财富》500 强收单机构（近期完成对 Worldpay 的整合），Global Payments 业务覆盖 100 多个国家，服务超过 600 万家商户，年交易处理量高达约 3.7 万亿美元。Arjun 专注于公司 Genius 智能 POS 平台的端侧 AI（Edge AI）体验与硬件战略，致力于为中小型企业（SMB）打造 AI 原生商业层。 ```mermaid flowchart TD accTitle: 嘉宾背景与业务定位分布 subgraph CBI["迪拜商业银行 (CBI)"] G["Giovanni Everduin<br/>首席战略与创新官"] G1["传统银行数字化转型"] G2["信贷普惠与决策治理"] end subgraph BILLGO["BillGO"] S["Satya Panda<br/>首席技术官 (CTO)"] S1["B2B 智能代理支付网络"] S2["小微企业全自动结算"] end subgraph GP["Global Payments"] A["Arjun Wadwalkar<br/>产品管理总监"] A1["Genius 智能 POS 平台"] A2["12 TOPS 端侧 SLM 硬件"] end ``` ### [03:40](https://www.youtube.com/watch?v=CULFxCllIJw&t=220s) 落地实践：端侧硬件、智能代理与信贷普惠 在 Jason 提出关于“目前正在推进或已部署且最具业务成效的 AI 项目”时，Arjun 首先分享了 Global Payments 针对 SMB 商户的痛点所构建的方案。小微商户日常背负着极重的手动运营包袱，例如频繁调整菜单、设置促销活动和在后台管理页面中反复切换。Global Payments 正在试点 AI 辅助配置菜单与工作流，并通过数据反馈快速迭代。更具突破性的是，他们计划在年内至明年初推出一款 **AI 原生手持收银与运维终端**。该设备内置 **12 TOPS（每秒万亿次运算）** 的神经处理单元（NPU），可直接在端侧本地运行小型语言模型（SLM）。商户只需语音对设备说“把商品价格调整为某某金额”，无需登录云端控制台即可完成修改；该终端还同时具备端侧实时反欺诈能力。 Satya Panda 介绍了 BillGO 的实践。小微企业主的核心诉求是专注主业，而不是将时间消耗在收款与对账上。BillGO 正在打造一套“智能代理（Agentic）支付网络”，其核心理念是：“把钱打给小微商户并不意味着支付流程的终结；只有当小微商户完全不需要做任何后续操作时，支付才算真正完成。” 为了让不具备技术背景的商户轻松上手，他们将 AI 交互设计得像使用 Google 搜索一样简单直接，全面实现后台自动化结算。 Giovanni Everduin 则从受监管商业银行的视角分享了两大核心项目： - **AI 驱动的普惠信用评分（Financial Inclusion）**：银行在过去一年半中引入了 AI 信用评估，不仅大幅缩短了贷款审批与承销周期，更重要的是利用合成数据（Synthetic Data）打破了传统信贷偏见。分析历史数据时发现，女性等特定少数群体的信贷样本极度匮乏，导致传统风控模型对其不利。通过引入合成数据建模，银行能够为缺乏传统征信记录的群体开辟全新的信贷资金池。 - **高管决策支持与影子 AI 治理**：针对管理层普遍使用 ChatGPT 或 Claude 辅助撰写邮件和决策的现象，Gio 指出由于无法直接复制粘贴内部信息，高管们往往会用手机拍照、传到个人工具处理后再经 WhatsApp 转发至 Teams。这种做法不仅流程繁琐，更存在泄露战略上下文与敏感信息的巨大合规隐患。银行正着手构建一套严密的安全治理架构，让员工在满足强监管、低风险偏好要求的前提下合规使用大模型能力。 ### [08:58](https://www.youtube.com/watch?v=CULFxCllIJw&t=538s) 双刃剑：AI 驱动的反欺诈升级与新型黑产威胁 Jason 引导讨论进入风控与欺诈防范这一金融核心议题。Arjun 强调，金融与支付本质上是“风险管理行业”。与传统基于静态规则的模型相比，当前经过微调的前沿大模型（Frontier Models）在识别复杂行为模式方面表现极为出色，能够无缝整合交易历史、设备指纹及商户信誉等多维参数。这种升级不仅大幅降低了误报率（False Positives），还实现了静默式的后台拦截与必要时的人机协同（Human-in-the-Loop）升级审核机制，做到了“让真正合法合规的用户在交易过程中感受不到风控拦截的存在”。 Satya 对此表示赞同，并指出大型合规企业过去的思维定式往往是通过死板的规则“限制所有人进行操作”，而在 AI 时代，企业应当主动拥抱前沿工具，以积极进取的态度部署前瞻性风控，而非一味封闭受限。 然而，Giovanni 随即提出了尖锐的逆向观点：前沿模型是一把双刃剑，防守方技术升级的同时，黑产与诈骗团伙也在利用同等强大的模型赋能网络犯罪。如今生成的钓鱼邮件与欺诈话术在专业度与拟真度上均呈现出指数级上升，给传统防御体系带来了前所未有的严峻挑战。 ### [11:22](https://www.youtube.com/watch?v=CULFxCllIJw&t=682s) 人类仍是最大的攻击面：AI 深度伪造与欺诈攻防战 谈及当前金融安全所面临的前沿威胁，嘉宾们直言，基于语音通话和视频的深度伪造（Deepfake）仿冒攻击数量正呈现出“彻底疯长、击穿天花板”的爆发态势。AI 的强大之处恰恰在于其“润物细无声”，但在欺诈场景中，这一优势却变成了极具杀伤力的武器——绝大多数普通人根本无法分辨自己面对的是真人还是高度拟真的生成式 AI 工具。在金融风控体系中，业内经常探讨“人机协同”（Human-in-the-loop），然而现实中最脆弱的环节依然是处于终端的“人”。无论是被钓鱼的普通储户，还是银行内部经办人员，人类本身已经成为黑客利用先进 AI 武器进行突破的最大攻击向量（Attack Vector）。 ```mermaid flowchart TD accTitle: 欺诈者利用尖端 AI 与金融机构防护的动态博弈 subgraph Attackers["攻击方：威胁行为体"] A["黑客与欺诈团伙"] --> B["尖端前沿模型与伪造工具"] B --> C["高保真音视频冒充 / 钓鱼"] end subgraph Vulnerability["攻击切入点"] C --> D["终端人质 / 员工 / 客户<br/>（人类是最大脆弱面）"] end subgraph Defense["防守方：金融与监管机构"] D --> E["银行风控与合规系统"] E --> F["安全防护与介入机制"] F -.->|"动态博弈 / 相互制衡"| B end ``` 这种前沿技术的滥用正在引发全行业的警惕。嘉宾特别提及此前 Anthropic 的 Mythos 漏洞等风波，模型厂商之所以被迫撤回相关能力，正是因为欺诈分子和威胁行为体直接把最前沿的基础模型调转枪口对准了防御方。在接下来的相当长一段时间内，金融机构与黑客之间注定是一场残酷的拉锯战（Tug of war）。尽管技术演进的终局可能是防御技术与进攻手段在相互制衡中建立新的动态平衡与安全护栏，但在可以预见的当下，这对任何一家强监管企业与商业银行而言，都是一场实打实的“噩梦情境”。 - **攻击爆发点**：AI 生成的合成语音和换脸视频真假难辨，欺诈者针对终端人员实施的精准社工冒充成为银行的首要风险。 - **武器化反噬**：攻击者利用各家大模型厂商最前沿的模型反制防御系统，迫使 AI 开发者因安全漏洞不得不回撤模型。 --- ### [13:06](https://www.youtube.com/watch?v=CULFxCllIJw&t=786s) 合规不可作为事后补丁：将监管治理融入产品全生命周期 面对极其严苛的监管环境，初创金融科技公司与大型银行面临着截然不同的管理重构挑战。长期以来，传统大厂与银行的惯常做法是：产品团队自行闭门造车设计系统，临上线前风控官（Risk Officer）突然介入并抛出一整套繁复的合规红线，随即引发全团队的鸡飞狗跳和混乱补救。Satya Panda 强调，在 AI 时代，合规（Compliance）与监管绝对不能再充当流程末尾的“阶段门禁”（Stage Gate），而必须原生植入整个产品生命周期本身。“合规不应被视为产品设计完成后的事后补救，而必须作为一项核心功能（Feature），而非限制创新的枷锁。” 要真正实现这一点，企业首先必须在战略上统一业务目标，将业务、技术、合规与法务等所有利益相关方召集在同一个房间里，共同制定组织的“AI 章程”（AI Charter）与“AI 政策”。唯有全员在顶层目标上形成共识，才能将合规治理拆解为细分的工程化维度：引入安全护栏、建立敏捷的工单升级处理机制，并在关键决策节点严格设置人工审核环节。在金融领域，触犯合规底线是不可承受之重，稍有不慎便会直接危及整个企业的生死存亡。 - **从“门禁”转向“原生特性”**：合规审查必须前移至系统架构之初，成为产品特性的一部分，杜绝“先上线再擦屁股”的开发模式。 - **组织协同机制**：通过设立统一的 AI Charter，拉通业务、风控与工程团队，避免跨部门各自为战带来的合规脱节。 --- ### [14:32](https://www.youtube.com/watch?v=CULFxCllIJw&t=872s) 董事会的焦虑与“等待策略”：金融机构落地 AI 的隐形成本与合规两难 来自传统银行实操一线的体验却充满了现实的骨感与无奈。Giovanni Everduin 坦言，当内部还在摸索如何使用 AI 时，监管机构的要求已经排山倒海般压来——每一条判定都必须具备完全的“可解释性”（Explainable）和“可审计性”（Auditable）。在严苛的合规重压下，很多时候银行甚至觉得“干脆不用 AI 反而更省事，宁愿多雇些人去做传统的人工操作”。因为引入 AI 的真实成本远不止采购 API Token 或软件订阅费，企业组织流程重塑、工作流再造、搭建治理架构的沉没成本，以及潜在的高昂监管罚单，构成了难以承受的隐形成本。 ```mermaid flowchart TD accTitle: 银行 AI 落地的组织压力与推迟决策循环 A["监管严格要求：可审计与可解释"] --> B["外包与三方供应链存在 AI 盲区"] B --> C["最终合规责任全链条归咎于银行"] C --> D["董事会面临个人连带法律责任"] D --> E["管理层极度焦虑，风险厌恶"] E --> F["ROI 收益模糊 vs 失败案例屡见报端"] F --> G["最终决策：选择观望两三年，推迟激进落地"] ``` 更棘手的问题潜藏在企业供应链的黑盒之中。当银行试图盘点全行现有的“AI 敞口与资产清单”时，惊恐地发现团队对实际暴露面几乎一无所知——因为银行采用的各类供应商及其二级外包团队，私下全都在使用形形色色的第三方 AI 工具。而监管机构的逻辑极其简单直白：任何交付给终端客户的链条中出了差错，最终唯一的责任人就是银行本身。 > “当我们承担最终责任时，意味着我的董事会要承担连带责任。而一旦董事会要负责，他们就会变得极其焦虑脆弱——因为没有任何董事愿意承担个人层面的法律追责。” 董事会的恐慌迅速向下传导为巨大的决策阻力。在没有人能给出确定性 ROI、满屏都是同行因激进上马 AI 而成本失控被迫回滚的负面新闻面前，传统的银行管理层最自然的避险选择便是：“让我们再观望个两三年，看看别人怎么做。”过度的监管与未知的风险，直接将受监管金融机构推向了极度谨慎的保守深渊。 --- ### [16:30](https://www.youtube.com/watch?v=CULFxCllIJw&t=990s) 拒绝黑盒：责任归属难题与端到端确定性模型 AI 决策的“责任归属”（Accountability）是横亘在金融创新面前的另一道高墙。Arjun Wadwalkar 明确指出，部署在金融核心系统的 AI 模型绝对不能是一个不可预测的“黑盒”；AI Agent 或算法做出的每一笔决策、每一次拦截或放行，背后都必须有一个具体的人作为责任主体（Accountable Owner）。只有明确责任人，才可能实现可信赖、负责任的 AI（Responsible AI）。 然而，这一理想在现实开会时遭遇了残酷的人性拷问。Giovanni 生动还原了银行风险委员会上的戏剧性一幕：当所有人都在为 AI 的颠覆性前景叫好时，一旦有人抛出“如果这个系统出事捅了娄子，谁愿意站出来担责？”整个会议室往往鸦雀无声，最终只有创新团队负责人苦笑着举手——“因为大家都知道，我的本职工作就是‘打碎鸡蛋来摊煎饼’（Break eggs to make omelets）的创新背锅人。”普通业务部门对 AI 内部机理本就缺乏认知，自然没人愿意为一项不可控的新技术押上职业生涯。 为了打破这种“问责即背锅”的文化僵局，技术底层必须走向工程化蜕变。未来的风控体系必须依赖“端到端确定性模型”（Deterministic Models）与完善的“数据血缘”（Data Lineage）。系统不能允许相同的输入产生截然不同的随机输出，必须保障全链路的端到端可追溯性。唯有当底层工具能够提供无可辩驳的数据审计链条，管理层和董事会才能获得推进行业变革的安全感；否则，所谓的“责任制”不过是出事时用来拉人当替罪羊的说辞。 - **责任人的博弈**：“成功有无数个父亲，失败却成了孤儿”，业务部门因不了解模型机理而普遍拒绝为算法故障担责。 - **技术解法**：摆脱大模型的概率随机性，打造输入输出完全可溯源、具备数据血缘分析的确定性金融风控体系。 --- ### [18:34](https://www.youtube.com/watch?v=CULFxCllIJw&t=1114s) 创新的“游说”策略：让合规与风控主管成为 AI 布道者 面对全行上下普遍的避险心态，创新推动者必须转变打法。在银行这种天然低容忍度的土壤里推 AI，不能靠在台上空喊口号，而必须展开扎实的内部“教育”与“政治游说”。Giovanni 指出，如今行业大会充斥着过度炒作与耸人听闻的负面叙事，社交媒体（如 X / Twitter）上的焦虑情绪更让决策层把目光死死锚定在 AI 失控的反面案例上。出人意料的是，在银行内部最难攻克的反而不是把守红线的合规人员，而是本应从降本增效中获益最多的业务部门。 ```mermaid flowchart LR accTitle: 创新团队突破合规防线的内部合作路径 subgraph Step1["传统路径（阻力重重）"] A["创新团队"] -- "强行推销" --> B["业务部门（疑虑抗拒）"] end subgraph Step2["迂回游说路径（顺水推舟）"] C["创新团队"] -- "优先赋能解决痛点" --> D["合规主管 & 首席风控官 (CRO)"] D -- "牵头制定实操护栏" --> E["联合背书 & 制度出台"] E -- "主动布道推行" --> F["全行各业务线顺利落地"] end ``` 基于这种生态，聪明的创新领导者选择采取一种“迂回进攻”的政治策略：绕开业务端，优先联合首席风控官（Chief Risk Officer, CRO）与合规主管（Head of Compliance）。由于规章制度本就由风控团队起草，创新团队先去攻克那些“看起来不怎么光鲜性感、但能大幅减轻合规风控工作负担”的内部场景。一旦合规部门切身体会到了 AI 工具带来的便利与可控性，他们便会放下敌意，转而以规则制定者的身份在银行内部替创新团队自发“布道”，主动制定务实可行的护栏规则。 这种打法被主持人 Jason 打趣地称为典型的“美式游说”（Lobbying）。与此同时，Arjun 也呼应了这种跨部门拉通的重要性：他们通过在企业内部建立共同的“AI 章程”，让业务、技术与合规三方平起平坐地坐在一张桌子前，赋予每个关键人对 AI 路线图的参与感与所有权。唯有当所有利益攸关方都觉得这个方案有自己的心血和利益在其中，金融领域的 AI 试验才能真正走出无休止的怀疑，迈出规模化落地的第一步。 - **优先击穿合规痛点**：优先选择为合规与风控部门开发提效工具，把防守方转化为坚定的盟友和行内布道者。 - **利益绑定消除壁垒**：通过制度化的 AI Charter 机制共享所有权，让原本相互制约的监管职能与创新职能站在同一阵线。 ### [21:42](https://www.youtube.com/watch?v=CULFxCllIJw&t=1302s) 企业级 AI 宪章与跨组织协同规范 随着每周都有新的 AI 模型和技术创新涌现，许多企业正面临信息过载的困境。来自高管层（C-suite）的统一指引往往是要求团队提高效率、加速采用 AI 并提升客户体验，但要在规模化企业中有效编排并落地这一战略，建立一套全员高度认同的“最低共识基线”至关重要。主持人 Jason Hiner 半开玩笑地询问企业制定的“AI 宪章”（AI Charter）到底是一份怎样的文件，由谁来起草，并调侃道：“希望你不会说是 Claude 写的。” 嘉宾解释称，AI 宪章是所有利益相关方共同参与审查与制定的集体产物，一旦确立便如同“刻在石头上”的硬性护栏（hard guardrails）。其核心目的在于明确组织运营的基本底线与边界框架。当团队在日常创新或尝试突破常规时，一旦涉及潜在风险，所有人都可以随时对照该基线进行复核，明确创新的安全边界所在。在大型组织中（例如拥有 3 万到 3.5 万名员工的企业），要让全体员工在 AI 这类颠覆性技术上步调一致极其困难，因此已有不少大型企业在过去 6 到 8 个月中推行了全公司范围的 AI 准则，甚至将其实践公开发布为开源规范供行业参考。 - **AI 宪章的核心定位**：作为组织内部的创新安全底线，为跨部门业务与技术探索提供明确的操作边界（Operating Box）。 - **制定与推行机制**：由跨职能利益相关方共同评审敲定，避免盲目冒进带来的合规与风控隐患。 - **大企业落地挑战**：对于数万名员工规模的机构，统一认知是 AI 规模化赋能的前提。 --- ### [23:31](https://www.youtube.com/watch?v=CULFxCllIJw&t=1411s) 客户期望的重塑：从“一对多”到“一对一”超个性化 谈及 AI 在金融界面与个性化体验上的变革时，主持人和嘉宾探讨了年轻一代客户对银行业务提出的新诉求——他们希望金融服务更加无缝、移动化且“不挡道”。来自传统银行视角的 Giovanni Everduin 给出了极为坦率的回答，幽默地劝告大家：“如果你作为客户有这种期待，千万别来找我们银行，去找这些初创科技公司吧。因为我们只会坐等他们把产品做出来，等技术相对成熟了，我们要么直接抄过来，要么把他们买下来再落地。”他直言，年轻客户若指望传统银行能迅速响应这些前沿期望，往往会大失所望，而这背后受制于银行严苛的合规与风险架构。 随后，初创生态的嘉宾指出，金融服务正在从传统的“一对多”（开发一套方案套用给数百万客户）演进为真正由 AI 驱动的“一对一”超个性化服务。在技术工具完备的背景下，深入理解单个客户并提供定制化服务已不再是让用户发出“哇”惊叹的加分项，而是变成了基本预期。以服务商户（Merchants）的技术平台为例，随着一些技术前沿的商户开始自主尝试基于 Claude Code 等工具编写代码或搭建自己的 Agent，他们反过来要求底层平台输出智能化洞察，例如“在当前环境下动态价格调整方案该如何制定”或“人员排班应如何基于预测进行优化”。如果企业无法跟上步伐将 AI 深度融入日常业务，长远来看将难以生存。 ```mermaid flowchart LR accTitle: 客户服务模式演进 subgraph Traditional["传统金融服务架构"] A["单一标准化产品方案"] --> B["覆盖数百万普通客户 (1-to-Many)"] end subgraph AI_Era["AI 驱动的新型架构"] C["全链路数据与智能 Agent"] --> D["动态定价 / 运营调度建议"] D --> E["超个性化独立服务 (1-to-One)"] end ``` - **服务模式转型**：从通用型“一对多”（One-to-Many）彻底转向基于实时洞察的“一对一”（One-to-One）超个性化交付。 - **商户端智能化诉求**：商户不再满足于单纯的交易系统，而是要求获得动态调价、智能排班优化等可直接指导运营的智能指标。 - **行业分化态势**：初创科技公司负责前沿试错与体验定义，传统大型银行则倾向于在模式成熟后通过技术采购或资本并购进行整合。 --- ### [26:29](https://www.youtube.com/watch?v=CULFxCllIJw&t=1589s) 支付领域的变革：数据编排、SaaS 垂直整合与大宗商品化困境 支付作为资金流转的基础设施，始终是金融科技中增长最为迅猛的板块。嘉宾剖析了支付产业链的割裂现状：在传统卡基支付（Card Payments）中，没有任何单一机构能够独立完成端到端的资金划拨，链路中分散着 Visa/Mastercard 等卡组织、发卡行（Issuer）、收单行（Acquirer）以及收单受理端等多个主体。在这一生态下，谁拥有全链路网络中最完整的数据可见度，谁就能基于分析与智能算法提供前所未有的高价值服务。 多位嘉宾达成共识：“纯粹的支付通道本身已经被严重商品化（Commoditized）。”单纯依靠通道费率的模式难以为继，企业要想突围，必须在支付能力之上构建垂直整合的 SaaS 解决方案，统揽整个商户的经营生态。支付只是其中一个触点，通过处理商户数据与消费者交易数据所提炼出的商业洞察，才是构建高附加值服务的护城河。Giovanni 补充道，尽管支付行业资金规模庞大，但在过去十年中，除区块链和 USDC 等加密稳定币催生了全新商业模式外，传统信用卡支付的底层机制并没有发生根本性改变。 | 维度 | 传统支付通道模式 | AI 与数据驱动的垂直 SaaS 模式 | |:--- |:--- |:--- | | **盈利逻辑** | 依赖交易流水通道费（严重同质化、费率微薄） | 依赖增值服务订阅与全链路经营软件赋能 | | **数据利用** | 仅用于基础交易清结算与事后风控记账 | 汇聚网络全链路数据，实时提供经营预测与个性化洞察 | | **客户黏性** | 极易被低费率竞品替代，迁移成本低 | 深度嵌入商户日常排班、库存、营销与定价，生态锁定强 | --- ### [29:40](https://www.youtube.com/watch?v=CULFxCllIJw&t=1780s) 代理商业（Agentic Commerce）与下一代支付原生网络 讨论进一步延伸至“代理商业”（Agentic Commerce）与自动化支付协议的前沿领域。当前行业普遍的做法是在既有的信用卡网络与收单结算链路之上搭建代理协议，但多位专家尖锐地指出，真正的颠覆性机会在于打造“原生面向智能体（Native Agentic）的全新支付轨道”，而非生搬硬套数十年前为人类刷卡设计的传统基础设施。 此外，行业正在迎来底层协同的重要转折点。自 2025 年底以来，包括 Visa、FIS 以及 Google 等行业巨头开始联合推进 AI 时代的新型支付生态合作。类似于全球通用的卡组织清算规则，银行、Fintech 公司与卡组织之间正在建立统一的协议、平台与治理政策。当 AI Agent 具备自主发起与完成交易的能力时，底层风控与审计面临着前所未有的严苛拷问：“这笔交易到底是谁授权的？由谁发起？资产最终归属是谁？资金流向何处？”整个生态必须在底层建立更加坚固的身份认证、授权追踪与端到端控制链条。 - **新型支付轨道需求**：打破传统卡组织收单清算闭环，探索专为自主 Agent 交易设计的原生清算网络。 - **巨头生态结盟**：科技巨头（如 Google）与金融基建巨头（如 Visa、FIS）联手制定 AI 支付治理标准与互操作框架。 - **Agentic 交易核心治理命题**： - **授权机制**：如何界定人类对智能体的委派范围与权限阈值。 - **发起与溯源**：在全自主交易链路中实现端到端的责任追溯。 - **资产归属与流向**：确保自主支付过程中的资金安全与法律合规。 --- ### [31:40](https://www.youtube.com/watch?v=CULFxCllIJw&t=1900s) 生成式 AI 与数字银行助手：心理防线与消费者接受度 当生成式 AI 为大众接入银行与复杂金融系统提供全新交互界面的同时，用户的实际心理接受度却展现出意料之外的冷淡。主持人 Jason 向现场观众发起互动调查：“在座各位有多少人希望由 AI 助手来协助你做出明智的财务决策，或者对你现有的财务决策提出反馈意见？请举手。” 令人意外的是，在满场全是非洲与全球 AI 领域从业者及专家的会议厅里，全场最终仅有 3 个人举手。Jason 不禁打趣道：“真是个挑剔的群体，要知道这可是一个专门讨论 AI 的行业峰会。” 针对这一悬殊的反差，嘉宾从心理学与用户行为角度给出了深度剖析。传统银行很早就在客户服务中使用基于规则的确定性（Deterministic）系统，但当技术迈向生成式与主动建议时，阻力不仅来自于技术本身，更来自于人类对财务健康的情感敏感度。嘉宾坦言自己也没有举手，因为“个人的财务健康在心理层面上是极具情绪化（Emotional）且高度敏感的隐私话题”。在大多数情况下，人们在心理上其实早已清楚自己财务状况的症结所在，面对冷冰冰的算法审视自己的财务收支与决策缺陷，往往存在天然的心理防御与信任壁垒。 ### [32:56](https://www.youtube.com/watch?v=CULFxCllIJw&t=1976s) 银行业部署 AI 投顾的法律责任与数据孤岛困境 Giovanni Everduin 指出，让 AI 介入个人财务决策存在两大核心阻碍。第一是用户心理层面的排斥——绝大多数人并不希望有人对自己的花钱方式指手画脚，更不用说是一个能够洞察自己所有资金流向的“感知实体（sentient being）”。第二则是银行业面临的极其现实的法律责任（Legal Liability）问题。银行很难直接上线一个提供财务建议的 AI 工具（尽管他本人非常讨厌狭义的聊天机器人概念），因为一旦 AI 给出错误建议，责任认定与合规追责将变得极其棘手。 此外，Giovanni 破除了外界对银行业数据能力的幻想。许多客户严重高估了银行数据的基础设施建设水平，误以为银行内部的数据井井有条，能够轻而易举地分析出用户的财务画像并给出指导。而现实情况是，银行的底层数据高度碎片化，割裂地散落在各种遗留系统之中；Giovanni 戏称，所谓的“系统”可能只是一张存放在某位员工桌面电脑里的 Excel 表格，这使得银行在现阶段根本无法支撑起高阶 AI 智能所需的连贯数据基础。 --- ### [34:04](https://www.youtube.com/watch?v=CULFxCllIJw&t=2044s) 信任与自主权阶梯（Trust Autonomy Ladder） 针对用户对 AI 接触财务数据的不适感，Satya Panda 提出了任何企业和金融科技机构在构建 AI 系统时都必须遵循的“信任自主权阶梯（Trust Autonomy Ladder）”。目前多数银行的聊天机器人功能极为有限，仅仅停留在修改密码或查找基础信息等简单应答阶段。Satya 强调，企业绝不可能指望用户在第一天就把全部银行信息和投资支配权拱手相让，AI 系统必须像人类建立信任一样，分阶段逐步提升自主权： ```mermaid flowchart LR accTitle: AI Trust Autonomy Ladder subgraph Level1["阶段一"] A["只读权限 Read<br/>访问基础信息"] end subgraph Level2["阶段二"] B["智能建议 Recommend<br/>提供优化/整合建议"] end subgraph Level3["阶段三"] C["步骤协助 Assist<br/>经用户确认后代办"] end subgraph Level4["阶段四"] D["完全执行 Execute<br/>获得长期委托授权"] end A --> B B --> C C --> D ``` 在互动讨论中，Giovanni 抛出了一个尖锐的观点：当用户在类似 Claude 等工具中频繁点击“允许（Allow）”或“危险操作中继续”时，这究竟代表着真正的“信任”，还是单纯出于“人类的懒惰与追求便利”？他坦言自己往往并不具备验证系统是否安全的工具和背景知识，只是不想反复处理弹窗确认，于是选择“点击确定并祈祷一切顺利”。Satya 对此回应称，尽管用户习惯各异，但开发者有责任提供兜底保障，包括透明的审计追踪、每日/每周执行动作汇总邮件，以及可供用户随时调大或收紧权限的“控制旋钮（dials）”。 - **只读阶段（Read）**：AI 仅具备数据读取权限，不进行任何主动干预。 - **建议阶段（Recommend）**：AI 在分析数据后提出建议（例如“建议合并账单或优化理财”），决策权完全在用户手中。 - **协助阶段（Assist）**：AI 规划具体操作步骤并询问“是否需要我代为执行”，由用户逐次确认。 - **执行阶段（Execute）**：在长期建立信任后，用户授予 AI 明确的完全执行权限。 --- ### [36:53](https://www.youtube.com/watch?v=CULFxCllIJw&t=2213s) 人机信任不对称与加拿大航空 Chatbot 诉讼判例 Giovanni 深入剖析了人机交互中存在的“信任不对称（Asymmetry in Trust）”现象。人们对风险的承受度主要由熟悉感和心理舒适度决定：在银行开展的概念验证（POC）基准测试中，算法模型在决策表现上几乎每一次都以压倒性优势战胜人类；然而，管理层却会对模型施加极其严苛的合规束缚与可解释性审计，仅仅因为担心“机器可能会犯错”。相比之下，人类在信贷审批等决策中几十年如一日地充斥着严重的偏见与失误，但由于无法看透人类大脑的思考过程，人们反而对人类决策的固有缺陷习以为常。Giovanni 认为，这种对 AI 施加的超额审视源于技术的新颖性，预计在半年到一年内，大众对 AI 的接纳程度就会像如今信任 Gmail 托管个人全部生活一样自然。 Arjun Wadwalkar 随即以真实的法律案例佐证了 AI 输出失误对企业的严重后果： | 案例要素 | 详细情况（加拿大航空 Air Canada 诉讼案） | |:--- |:--- | | **事件起因** | 客户通过加拿大航空基于 LLM 的智能客服咨询机票退改签政策 | | **AI 故障表现** | 聊天机器人发生严重幻觉（Hallucination），向客户提供了虚假的“全额退款”承诺 | | **后果与纠纷** | 客户误信 AI 导致错过航班且无法按政策退款，随后将加航告上法庭 | | **加航辩护主张** | 辩称该承诺系聊天机器人的独立输出，航司不应承担连带法律责任 | | **法庭最终判决** | 法院判定客户胜诉，责令加航全额退款并赔偿客户由此产生的一切额外损失 | Arjun 强调，虽然现代用户在大模型普及后对交互体验的期望大幅提高，但也深知系统存在幻觉风险。因此，金融与支付科技提供商必须将输出的**绝对准确性（Accuracy）**放在首位，即便初期只能提供最基础的信息，唯有确凿的准确率才能为逐步扩展复杂功能奠定信任基石。 --- ### [40:04](https://www.youtube.com/watch?v=CULFxCllIJw&t=2404s) 颠覆传统协作：从信息孤岛到“极端透明”与集体智能 主持人 Jason Hiner 将话题引向工作组织方式的根本变革，邀请 Satya Panda 分享其从 OpenAI 首席执行官 Sam Altman 处受到启发的“令人不适的透明度（Uncomfortable Transparency）”理念。Satya 指出，历史上无论是企业还是个人，都倾向于将信息据为己有，甚至在组织内部将信息壁垒当作权力的“排他性俱乐部”乃至政治武器。传统企业中约有 95% 的信息处于保密或孤岛状态，而 AI 时代正在将这一比例彻底倒置——95% 的内部信息都将走向公开透明。 Satya 阐述了 BillGO 正在推行的激进变革：全公司员工互相开放所有 Confluence 页面、Slack 频道消息、会议纪要、内部邮件甚至主管的日历日程（仅当事人明确标记为 Private 时才保密）。AI 工具接入这些海量上下文后，帮助员工摆脱起草邮件等琐碎事务，聚焦于高阶批判性思考。大部分企业仅将 AI 作为边缘外挂，而他们正将模式彻底颠覆（Flip the Script）——将 AI 打造为组织正中央的“集体智能（Collective Intelligence）”中枢，全员围绕 AI 协同运转以达成指数级的生产力释放。 面对这种近乎全裸的透明工作模式，Giovanni 幽默地表示：“如果是我个人，绝对不会想这样工作（笑）。”他打趣道，光是想到政府层面的数据监管就已经让人神经紧绷，更别提让公司拥有这种程度的全方位监视了。但他由衷钦佩敢于率先试水的先锋企业（Trailblazers），正是这些探索者承担了试错风险，去探明水下究竟是一片坦途还是“布满鲨鱼（Shark-infested）”。最后，Giovanni 话锋一转，提出了对 AI 泛滥的深层担忧——当所有人都依赖 AI 快速生成大量看似完整却毫无灵魂的冗长内容时，组织正面临严重的“草稿垃圾（Draft Slop）”膨胀问题。 ### [43:19](https://www.youtube.com/watch?v=CULFxCllIJw&t=2599s) AI 生成长篇文档带来的“精致垃圾”与管理瓶颈 在讨论生成式 AI 深入团队日常所带来的连锁反应时，讲者指出当下职场出现了一种令人啼笑皆非的极端现象：每当团队遇到任何问题，团队成员提交上来的往往是一份长达 500 页、目录与索引一应俱全的“惊人”详尽文档。如果员工足够敏锐，他们会在提交前刻意抹掉典型的 Claude 或 ChatGPT 语言风格；但如果不够老练，那些由大语言模型自动生成的八股文结构就会原封不动地暴露出来。 这种看似极具生产力的高效输出，反而在管理层引发了严重的效率瓶颈。少数高管或决策者不得不耗费极其庞大的时间与精力，对这些文档进行严格的批判性评估。讲者无奈地直言，自己如今的大量时间都被迫消耗在“评估那些包装得无比精美，但内容实质却粗制滥造的垃圾文档（evaluating crappy documents that are incredibly wellcrafted）”上面，原本旨在提效的技术反而给组织带来了新的负担。 ### [43:50](https://www.youtube.com/watch?v=CULFxCllIJw&t=2630s) 构建统一智能底座与个人 AI Agent 的提效闭环 针对这一挑战，Arjun Wadwalkar 提出了应对策略，他非常赞同建立“通用最低智能层（Common Minimum Intelligence Layer）”平台并围绕其构建专用 AI 的思路。他强调，如果没有为 AI Agent 提供扎实的企业上下文与知识图谱（Knowledge Graph），Agent 就无法具备执行任务的先决知识，其表现也绝不可能达到预期。因此，搭建这个企业级的统一智能底座是发挥 AI 真实价值、让每一分订阅和研发投入物有所值的核心前提。 Arjun 分享了自己在个人层面的落地实践：他将自己的日常邮箱、Slack 等工作通讯账号全面接入 Claude，构建了一个深受信赖且经过严密安全配置的专属个人 Agent，如今他几乎所有个人层面的信息处理都交由该 Agent 执行。目前，他正积极推动将这一模式拓展至正式的企业工作场景中，期望实现由 Agent 自动生成每日高管邮件摘要、并直接辅助起草回复。 随后，Arjun 以职场中人人都感同身受的周报撰写为例，直观展示了 AI 带来的真实投资回报率（ROI）：在引入自动化工具前，为了给上级领导梳理本周成果与重要进展，撰写一份合格的高管汇报往往需要花费 1 到 2 个小时；而借助深度整合上下文的 AI 工具，整个初稿生成只需不到 1 分钟，自己仅需花费约 5 分钟进行人工审阅与把关。这直接为高管挽回了近两小时的高价值时间，使管理层得以从机械的文字整理中彻底解脱。 ```mermaid flowchart TD accTitle: Weekly Update Workflow Comparison subgraph Manual["传统汇报流程（耗时 1-2 小时）"] M1["翻找分散记录<br/>（邮件/Slack/看板）"] --> M2["人工拼凑整理数据"] M2 --> M3["逐字撰写高管汇报草稿"] end subgraph AIAssisted["AI 智能底座流程（耗时约 6 分钟）"] A1["接入统一智能层与通讯流<br/>（Claude Agent 自动抓取）"] --> A2["AI 自动结构化汇总<br/>（生成耗时 1 分钟）"] A2 --> A3["人工快速核对审查<br/>（审阅耗时 5 分钟）"] end ``` - **知识图谱的重要性**：AI Agent 必须依赖底层统一的知识层作为信息底座，否则缺乏上下文将导致输出偏离预期甚至失控。 - **高管周报时间成本对比**：传统纯手工撰写 executive update 需要 1 至 2 小时；经由定制化 AI 工具后，1 分钟生成草稿，5 分钟完成审阅，净节省约 2 小时的高管生产力。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=CULFxCllIJw&t=4s) 法律团队特别交代过，我今天分享的内容纯属个人观点，不代表银行或国家的立场。 — Giovanni Everduin > [03:40](https://www.youtube.com/watch?v=CULFxCllIJw&t=220s) 把钱打给小微商户并不意味着支付流程的终结；只有当小微商户完全不需要做任何后续操作时，支付才算真正完成。 — Satya Panda > [13:06](https://www.youtube.com/watch?v=CULFxCllIJw&t=786s) 合规不应被视为产品设计完成后的事后补救，而必须作为一项核心功能（Feature），而非限制创新的枷锁。 — Satya Panda > [14:32](https://www.youtube.com/watch?v=CULFxCllIJw&t=872s) 当我们承担最终责任时，意味着我的董事会要承担连带责任。而一旦董事会要负责，他们就会变得极其焦虑脆弱——因为没有任何董事愿意承担个人层面的法律追责。 — Giovanni Everduin > [16:30](https://www.youtube.com/watch?v=CULFxCllIJw&t=990s) 因为大家都知道，我的本职工作就是‘打碎鸡蛋来摊煎饼’（Break eggs to make omelets）的创新背锅人。 — Giovanni Everduin > [23:31](https://www.youtube.com/watch?v=CULFxCllIJw&t=1411s) 如果你作为客户有这种期待，千万别来找我们银行，去找这些初创科技公司吧。因为我们只会坐等他们把产品做出来，等技术相对成熟了，我们要么直接抄过来，要么把他们买下来再落地。 — Giovanni Everduin > [26:29](https://www.youtube.com/watch?v=CULFxCllIJw&t=1589s) 纯粹的支付通道本身已经被严重商品化（Commoditized）。 — Arjun Wadwalkar / Giovanni Everduin / Satya Panda > [29:40](https://www.youtube.com/watch?v=CULFxCllIJw&t=1780s) 这笔交易到底是谁授权的？由谁发起？资产最终归属是谁？资金流向何处？ — Arjun Wadwalkar / Giovanni Everduin / Satya Panda > [31:40](https://www.youtube.com/watch?v=CULFxCllIJw&t=1900s) 个人的财务健康在心理层面上是极具情绪化（Emotional）且高度敏感的隐私话题 — Arjun Wadwalkar / Giovanni Everduin / Satya Panda > [40:04](https://www.youtube.com/watch?v=CULFxCllIJw&t=2404s) 如果是我个人，绝对不会想这样工作（笑）。 — Giovanni Everduin > [43:19](https://www.youtube.com/watch?v=CULFxCllIJw&t=2599s) 评估那些包装得无比精美，但内容实质却粗制滥造的垃圾文档（evaluating crappy documents that are incredibly wellcrafted） — Giovanni Everduin ## Mentioned Ai4, Air Canada, Anthropic, Arjun Wadwalkar, BillGO, ChatGPT, Claude, Claude Code, Commercial Bank International (CBI), Confluence, FIS, Genius POS, Giovanni Everduin, Global Payments Inc., Google, Jason Hiner, Mastercard, Mythos, OpenAI, Sam Altman, Satya Panda, Sheela Ursal, Slack, The Deep View, Teams, USDC, Visa, WhatsApp, Worldpay, X (Twitter) ## Video description Giovanni Everduin – Chief Strategy & Innovation Officer, Commercial Bank International Arjun Wadwalkar – Director, Product Management, Global Payments Inc. Sheela Ursal – CEO, Fintech Next Satya Panda – CTO, BillGO Jason Hiner – Editor-in-Chief, The Deep View Fraud, compliance, and user expectations are converging to make AI a core enabler of modern financial software. This panel brings together fintech leaders who’ve deployed AI to improve fraud detection, risk scoring, and payment flows, while delivering better experiences to end users. Panelists will share lessons on balancing innovation with regulatory guardrails, scaling AI across high-volume financial systems, and measuring ROI in terms of both efficiency and customer trust. Learn what the next generation of AI-driven fintech and payments looks like—and how enterprises can capture its value. [Source](https://www.youtube.com/watch?v=CULFxCllIJw)

[0:04] Giovanni Everduin: 法律团队特别交代过，我今天分享的内容纯属个人观点，不代表银行或国家的立场。 [3:40] Satya Panda: 把钱打给小微商户并不意味着支付流程的终结；只有当小微商户完全不需要做任何后续操作时，支付才算真正完成。 [13:06] Satya Panda: 合规不应被视为产品设计完成后的事后补救，而必须作为一项核心功能（Feature），而非限制创新的枷锁。 [14:32] Giovanni Everduin: 当我们承担最终责任时，意味着我的董事会要承担连带责任。而一旦董事会要负责，他们就会变得极其焦虑脆弱——因为没有任何董事愿意承担个人层面的法律追责。 [16:30] Giovanni Everduin: 因为大家都知道，我的本职工作就是‘打碎鸡蛋来摊煎饼’（Break eggs to make omelets）的创新背锅人。 [23:31] Giovanni Everduin: 如果你作为客户有这种期待，千万别来找我们银行，去找这些初创科技公司吧。因为我们只会坐等他们把产品做出来，等技术相对成熟了，我们要么直接抄过来，要么把他们买下来再落地。 [26:29] Arjun Wadwalkar / Giovanni Everduin / Satya Panda: 纯粹的支付通道本身已经被严重商品化（Commoditized）。 [29:40] Arjun Wadwalkar / Giovanni Everduin / Satya Panda: 这笔交易到底是谁授权的？由谁发起？资产最终归属是谁？资金流向何处？ [31:40] Arjun Wadwalkar / Giovanni Everduin / Satya Panda: 个人的财务健康在心理层面上是极具情绪化（Emotional）且高度敏感的隐私话题 [40:04] Giovanni Everduin: 如果是我个人，绝对不会想这样工作（笑）。 [43:19] Giovanni Everduin: 评估那些包装得无比精美，但内容实质却粗制滥造的垃圾文档（evaluating crappy documents that are incredibly wellcrafted）

[0:04] Jason Hiner: Hiner emphasizes that the goal of the panel is to focus on driving real, measurable business results despite these operational constraints. [0:04] Giovanni "Gio" Everduin: disclaimer that his statements reflect his personal views, not those of his bank or the UAE government. [8:58] Arjun Wadwalkar: financial processing is fundamentally a risk-management business. [12:37] Satya Panda: regulatory compliance can no longer function as a final check at the end of a product cycle. Instead, compliance must be embedded directly into the product development lifecycle as an active feature. [12:37] Giovanni "Gio" Everduin: the true cost of AI is not model API tokens or subscription fees, but rather organizational restructuring, workflow redesign, compliance governance, and potential regulatory fines. [18:18] Giovanni "Gio" Everduin: convincing "everyone—inside the bank, outside the bank." [36:17] Giovanni "Gio" Everduin: human decision-making—such as traditional credit decisioning—has historically been fraught with bias and error without comparable scrutiny. [36:17] Giovanni "Gio" Everduin: human reasoning cannot be audited or inspected internally, yet society remains comfortable with human flaws while demanding perfection from AI. [36:17] Arjun Wadwalkar: while user expectations have surged after interacting with advanced consumer models like ChatGPT and Claude, technology providers must prioritize output accuracy to establish a reliable trust model. [40:04] Satya Panda: uncomfortable transparency [42:25] Giovanni "Gio" Everduin: draft shock [42:25] Giovanni "Gio" Everduin: jittery

[0:04] All right, thank you for coming. Uh, welcome to  AI4. Uh, hope everybody's first day is off to a [0:10] great start. We've got a wonderful panel here  to talk about finance, fintech, payments, uh, [0:18] and how AI fits in. And the great thing is uh  you know in talking about the panelists before [0:24] uh with the panelists before we came on um  we're going to get to the stuff that we hope [0:28] you really care about which is like you hear lots  of things about AI and all of the the models and [0:35] all of the advances and how fast is everything  moving and we know that what you really want [0:40] to know is like okay all that's great uh how  can I actually put this to use when I go home [0:45] and I have infrastructure that's old that's not  optimized and have people whose mindset, you know, [0:52] really is not ready to sort of change in advance.  How do I actually go and take these things and get [0:58] real results, you know, with them? So, that's  what we're going to try to do. Um, my name is [1:03] Jason Heiner. I'm the editor-inchief of the Deep  View, which is a publication solely focused on AI, [1:10] uh, especially aimed at, uh, professionals  and people who are out there trying to put [1:15] AI to work in the real world as well. I'm going  to let our guests introduce themselves because [1:20] this is who you really want to hear from. So,  we'll we'll start and go down the line. Uh good [1:25] afternoon everyone. My name is Giovanni Everdan,  but people call me Gio. I am the uh chief strategy [1:30] and innovation officer for a listed bank out of  Dubai where I also founded our innovation lab [1:35] and our corporate VC. Uh and my legal team really  wanted me to tell you because I like to keep it [1:40] real that whatever I'm going to tell you today is  my personal opinion. Does it reflect the bank or [1:44] the country in whatever capacity? I'm looking at  the camera right now. Hello everyone. Uh my name [1:51] is Satya Panda. I'm the chief technology officer  at Bilo. Uh we are a uh bit B2B payments company. [1:59] We leverage AI to accelerate payments uh in a  smarter secure manner. and uh I'm going to be [2:05] talking about how in the age of AI we can overcome  some of the risk and uh uh fraud constraints that [2:12] we have. Hey, thanks guys. Hey, hello everyone.  My name is Arjun Welker. I am director of product [2:19] management at global payments. So, Global  Payments is a Fortune 500 uh payment acquirer, [2:25] probably one of the largest after our recent  world pay acquisition. We process somewhere [2:30] around $3.7 trillion of uh annual transaction  volume. We are in 100 plus um countries and and [2:37] we service 6 million plus merchants. And uh I  specifically lead payment strategy and edge on [2:45] the edge AI experiences and and edge AI hardware  in store for our Genius point of sale platform. [2:51] It's a vertically integrated SAS platform uh that  we offer to enterprise and SMB merchants. And [2:58] I'll be talking about AI native commerce uh layer  that we are building uh and uh AI native you know [3:06] workflows for SMB merchants uh so they can focus  on their businesses while we help them deliver [3:13] technology that can makes make their lives easier.  Thanks for having us. Oh yeah, thanks Gio. So, [3:22] um, like Gio mentioned, I I do want to call out  looking at the camera, all views that I express [3:27] here today are my own and do not represent the  views of my employer. This is a difference between [3:32] listed companies and private companies. [laughter]  We're allowed to say whatever we want. Um, no. So, [3:40] let's um let's get straight to it and talk about  what's the project that you're working on inside [3:47] your company right now or you're taking AI and  using it um either have already deployed it or are [3:53] working on um currently getting ready to deploy  it that you're most proud of and that you know is [3:59] seeing sort of real business impact. Let's start  the opposite direction and and go this way. Yeah. [4:04] Yeah. So, we develop technology for SMB merchants,  right? And uh when you look at these merchants, [4:11] they have a huge [snorts] operational overhead,  you know, managing menus, managing promotions, [4:17] managing their uh customer experiences and  navigating through these administrative screens. [4:24] Um and what we are trying to do is we are trying  to embed AI to improve that experiences and and [4:29] make it more efficient. Uh so we are developing AI  assisted menu configuration changes, AI assisted [4:37] um workflow changes uh and I know they are  like small iterative changes but we are testing [4:43] those to first see the feedback and and see how  impactful it is for those SMB merchants. The other [4:49] thing that we are doing is that we are bringing to  market this AI first handheld that has like neural [4:57] processing unit on board 12 ter operations per  second that can run small language models on the [5:03] edge. Uh that allows these merchants to interact  with that handheld payment and operational device [5:10] to run their uh full stores like they can ask that  handheld to hey update the product price from this [5:17] to this. they don't have to log into any like  cloud portal to do it. Similarly, that handheld [5:23] has on the edge AI fraud detection capabilities um  that that we plan to introduce. So, these are some [5:30] innovative things that we are working on and plan  to introduce uh later in this year and early part [5:35] of part of next year. Uh at Bilgo we uh primarily  serve small business customers and uh for small [5:43] business customers uh we want them to focus on  their business and not necessarily spend their [5:49] time receiving or processing payments and uh just  like uh Ajun out here we are building a agentic [5:55] payment network that accelerates payments for  the small businesses and our our moto is that the [6:03] payment is not done just because we are handing  the small businesses payment. The payment is [6:09] actually done when the small business doesn't have  anything else to do. And that's where our agentic [6:13] payment platform is providing the necessary  tools to the small business that allows them to [6:20] um leverage AI. Uh they are not necessarily very  AI proficient. So we are making it as use as easy [6:26] as using a Google so that they can focus on their  business while we take care of their payments. [6:33] Yeah. So we're a we're a bank. So there's two  things. one that is not necessarily very sexy but [6:38] very useful and one that is very important I think  for what the country and the country is the UAE [6:45] Dubai where I live and so the thing that's really  important for the country is financial inclusion [6:49] and what we're doing is what we've done over  the last year and a half we've implemented uh AI [6:54] powered credit scoring now at the simplest level  that for a bank it just means we can do credit [7:00] scoring faster which means we can underwrite loans  faster and quicker the turnaround time is faster [7:04] and we can reach people better. But the thing  that's really important from a financial inclusion [7:08] perspective is when we started looking at this  stuff and looking at our data and benchmarking it, [7:12] there's a lot of underrepresented um demographics,  right? A lot of minority groups. In particular, [7:18] for us, for instance, women, not enough women who  have taken loans. And as a result, the results [7:22] are very skewed uh against their favor. And so  now using synthetic data and our credit models, [7:28] we can actually open up credit pools for more and  more people that may or not have the traditional [7:33] credit history that we would need to be able to  underwrite them. So that's been really good. I [7:37] think the thing that I'm probably most excited  about though is trying to use AI for executive [7:43] decision-m. Now like most of you probably  everyone in the company uses chat GBT, Claude, [7:49] whatever it may be, right? And so how do you  kind of how do you bring that into a governance [7:55] and structure it in a way where you can actually  use that as a team because the problem we have [8:01] right now I was telling the guys earlier is that  everyone in our leadership team uses chat GBT or [8:05] claw to write their emails and I know this  because I know how you talk and I see your [8:09] emails and there's a massive gap right but what  you're doing when you write an email through claw [8:14] or chatg or whatever is most people take a photo  of the email because they can't really share it [8:18] and [snorts] then they get the result and then  they copy paste past it through WhatsApp into [8:22] Teams and it's like a whole thing. But here's  the issue with that. You're still sharing a lot [8:25] of sensitive confidential information. Even if  it doesn't have client information or customer [8:30] names, [snorts] it still has contextual strategic  information about the company. And so we're trying [8:35] to figure out how can we give people access to  these tools because they clearly want it and [8:39] it's clearly driving better decision-m. But how do  we do it in a way that works for a bank, a highly [8:44] regulated bank with a very low risk appetite  and with a very jumpy regulator waiting to slap [8:49] us if anything goes wrong? All right, Arjun, you  mentioned fraud protection. Um fraud prevention, [8:58] uh risk management, just a huge part of what you  all do all the time, what everybody, you know, [9:03] here um has to think about and care about. How is  AI um helping um and and maybe in one ways how is [9:12] it also you know creating new challenges too?  Anyone want to take that one? So I mean I look [9:18] at uh finance and and payments like majorly like  it's a riskmanagement business and with all these [9:26] frontier models tuned for fraud prevention. What  is happening here is that um they are extremely [9:32] efficient in recognizing patterns like what's the  transaction history, what's the device history, [9:39] what is the merchant history, you know all these  different parameters go into a fraud prevention [9:44] model that uh traditional rulebased static models  were not that efficient. So these frontier models [9:53] are are really enhancing that experience and the  best fraud prevention and protection experiences [10:00] that the customer and the end user doesn't  feel. It reduces the false positives and [10:05] it's it makes sure that uh the fraud detection is  happening behind the scenes and there is human in [10:12] the loop workflows identified for escalating uh  fraud that is identified to make sure that that [10:19] it's effectively uh processed. I'll probably add  to that that fraud and risk management what in a [10:27] traditional sense has evolved significantly  and in the age of AI I think the many big [10:34] regulated companies they think about hey how do  I stop everybody from doing things rather than I [10:40] think what Arjun mentioned I want to echo that is  that we have the tools available to us that can [10:47] u provide fraud detection fraud prevention in a  proactive manner that we never had in the past. [10:52] There are many tools I've used in the past like  the rule-based ones. So absolutely I think for [10:57] companies to really be very competitive and be  at the forefront of this they should encourage [11:02] innovation in this space rather than restrict. I  let me flip the argument then because I completely [11:09] agree with you guys. But you know what the biggest  problem for us is right now? All of those things [11:12] you're saying, all these frontier models,  they're also enabling fraud and fraudsters, [11:17] right? And so the level of sophistication  in fishing and fraudulent scamming emails [11:22] and calls and videos has been absolutely insane,  through the roof for us. And to your point, right, [11:28] it's that the techn great technology technology  where people don't even know what it's there. But [11:32] the flip side of it is the same thing. I was like  a lot of people don't understand how good people [11:37] have become at some of these impersonations  right through AI tools and our biggest risk [11:42] factor is still people right we talk about humans  in the loop humans at the end of the loop that are [11:46] being targeted with people with AI tools that's  still the biggest you know attack factor for us [11:52] as a bank absolutely if I can add to that that is  totally a situation that we have to handle maybe a [11:58] month ago somebody some of you may know you would  have followed anthropics mitos and you know fable [12:04] and all of that fiasco, right? They had to take  it back exactly for that reason because fraudsters [12:09] and the threat actors are actually using these  very frontier models against us. And that is where [12:14] the there'll be a period of time where there's  always be a tug of war. But then eventually I uh [12:19] I think we'll get to a balance point where uh one  side or the other side are going to be keeping us [12:25] in check and there'll be safeguards coming into  place. But absolutely for the foreseeable future, [12:30] this is a nightmare situation for most banks  and regulated companies. You all operate in [12:37] very regulated environments. Although it's a  little different in the like the size of the [12:42] uh you know the operations and and uh yeah SA  you're um working in a little bit more of a [12:49] startup environment. So that said, um what is the  impact of AI when it comes to you know regulation, [12:58] governance, all of that? How has that changed  the game? What's that look like? I think the [13:06] compliance and the regulation it shouldn't be like  a stage gate at the end. And the way the modern [13:11] systems have to be designed is they are part of  the product life cycle and part of the product. [13:16] It cannot be an afterthought that I I've designed  the product now how do I make this compliant and [13:20] risk-f free. Uh and that is where some of  the newer technologies that we have with AI [13:26] are going to be helping us immensely right in uh  traditionally I have worked in big tech companies [13:31] and and banks as well and it has always been okay  product goes and designs their own systems and [13:38] then our risk officer comes like hey these are the  regulations and guidelines you need to follow and [13:44] then there is a scramble. So I feel we have an  amazing opportunity to really embed this process [13:49] into the product development and it's a feature  rather than a limiting constraint. Um, I think [13:56] it starts with having a common business objective  and bringing all the stakeholders in the same room [14:03] and having an AI charter and an AI policy as an  organization that everybody can align on and then [14:11] uh it goes into a multitude of dimensions, right?  like adding safeguards, making sure there are uh [14:18] proper workflows in in place where escalations can  happen, there are humans in the loop because one [14:24] thing that we cannot do is is violate compliance,  right? That puts the whole business at risk for [14:32] us. Honestly, we're we were still scrambling  figuring out how we're going to use this AI stuff [14:36] and then the regulator comes and goes like now  it needs to be explainable and auditable and all [14:40] of this stuff. Honestly, sometimes it's easier to  just not use it frankly and just, you know, invest [14:48] a little bit more in people doing more manual work  because the cost and when I say cost is not just [14:53] the cost of buying tokens or a subscription model.  It's the cost of changing the organization and [14:57] building the workflows and doing the governance  and everything else and the potential fines from [15:02] the regulator. It makes it really really hard  for us to do this. I I think when we started [15:08] our journey um we looked at what is our actual  inventory what's our current exposure as an [15:13] organization to AI and guess what we have no clue  at all because a lot of our vendors right have [15:19] subcontractors and all of them use AI tools and  from the regulator we're ultimately responsible [15:24] if anywhere in the chain of anything that goes to  a client eventually we're accountable and when I [15:29] say we are accountable my board is accountable and  when my board is accountable they get really jumpy [15:34] They don't like to be personally reliable,  right? And so when they get really jumpy, [15:38] they put a lot of pressure on us. And so what  happens? What do you think? Right? So either [15:42] we invest a ton of money and a ton of time in  something where frankly no one can really tell [15:46] me what the roe is going to be and how much better  things are going to be. And then you read all of [15:50] these stories about companies that went full AI  and had to roll it all back and the cost went [15:54] crazy, right? All of that stuff. If you think a  leadership team really wants to go down that path [15:58] and take a risk, if they can just choose to like  let's just wait it out, wait a couple more years, [16:03] see what everyone else is doing. So that's where  we are right now, which is we are governed by [16:07] regulations. And it's great. Don't get me wrong,  I'm looking at the camera. Regulations are great, [16:12] but at the same time, too much regulations in an  emerging tech environment can be very restrictive [16:20] and it will for most companies that are heavily  regulated, they will rather c they will rather [16:25] be cautious and are on the side of caution than  the other way around. Yeah. And I'd like to add [16:30] to that a little bit like I think accountability  is is very important. Your AI model cannot be a [16:36] black box. every decision that u your AI model  or AI agent makes there has to be an accountable [16:43] owner behind it. I just wanted to add to that and  and that would help you be compliant and make sure [16:49] you are deploying responsible AI. But and so but  this is an interesting thing because we literally [16:54] had our risk committee last week two weeks ago and  this is the question who's going to be accountable [17:00] everyone agrees that this is great right when  when everything goes success as many fathers but [17:04] ultimately who wants to be accountable for when  this goes bad and guess how many people want to [17:08] raise their hand. It's ultimately me because my  job anyway is to kind of, you know, I usually say [17:13] my job is to break eggs to make omelets. I'm  the innovation strategy guy in the bank. But [17:18] at the same time, nobody else wants to take that  risk because it's still a very significant risk, [17:22] right? So, how do you drive a cultural change when  there's an accountability for something that most [17:27] people don't really understand what it is or  how it works? I think I think those are some [17:32] of the current constraints we have. But as how do  we move look at the future, right? uh in terms of [17:37] looking at the future I feel uh we we have to  build to he used something like auditability [17:43] or data lineage and end to end deterministic  models we cannot have AI blackbox to his words [17:50] do different things of the for the same set of  input so there are uh models evolving processes [17:56] evolving in the risk space where we there'll be  end toend auditability uh deterministic models and [18:02] how we can educate and bring the workforce and the  leadership into the game because To Joan's point, [18:07] most of the big tech and bank environments, the  leadership and the board, they're not necessarily [18:13] well versed with this and they are riskaverse to  his point. But as we evolve one year, two year, [18:18] five years from now, the path is to how  do we provide that confidence with this uh [18:22] auditability and data lineage and accountability  is obviously big part of it. But without this [18:28] uh entire framework, accountability is just a  name, you know, to throw under the bus. Gio, [18:34] I want to double click on what you said. You're  you're in this bank has this sort of low tolerance [18:40] for for risk. Any bank, any bank, every every  bank. Um you're the innovation guy. Uh you have to [18:49] sort of be the one sort of pushing on AI and sort  of helping drive change in the organization. What [18:55] does that look like for you? Who who do you have  to convince? Where have they been more open to [18:59] like doing some experiments? where is there sort  of more resistance and you expect that like we're [19:04] just not going to be ready for that for a while.  What is what does the texture of that look like? [19:08] You being the guy who has to, you know, convince  people like let's try some of this AI stuff. Yeah, [19:13] I have to convince everyone like inside the bank,  outside the bank, everyone has to be convinced. [19:19] Um but more so before they can be convinced,  they have to be educated. Frankly speaking, [19:23] a lot of like and partially it's some of these  conferences that are part of the problem. there's [19:27] so much hype and stuff and narratives and you  know if you're on Twitter like I am or X like [19:33] you just get bombarded with opinions and views  and it's very hard for someone to stay on top [19:38] of what's really happening right a lot of these  you know let's call them problems that happen [19:44] um like with uh with mythos or anything else  they get oversized exposure and so people kind [19:50] of anchor in on these negatives so that's the  hard part surprisingly in at least in our bank [19:56] the business has been the hardest even though I  think they would benefit the most. The two leaders [20:01] that I have been able to do a lot of work with are  the head of compliance and our chief risk officer. [20:07] And it's also strategically for me because if  I can do something with risk and compliance to [20:11] make their lives easier, which I can, they're  going to help me then roll that out and because [20:16] they're anyway writing most of the governance,  right? And so my strategy was let me start there, [20:20] which is not necessarily the sexy bits, but it's  the part where I anchor in the governance and the [20:25] comfort and then let them go out and evangelize  on my behalf instead of me pushing it because [20:29] obviously I'm also I'm biased. My incentives  literally my my KPIs are to drive innovation, [20:36] right? Like so the more I do the better for me.  So for them is the opposite. So if I can convince [20:41] the people that have to build the guard rails of  the value of it, then I think they can build very [20:46] practical guardrails. So that's a great takeaway.  So if you can partner with some of the people who [20:51] are writing the the guidelines, deliver something  for them, then potentially they can help you sort [20:57] of drive some I think in the US you call that  lobbying. [laughter] I love it. All right, [21:02] Arjun, I want to double click on something  you mentioned which was an AI charter. Um, [21:07] what does that look like in your company? Has that  enabled you to do some of the similar things? Has [21:11] it enabled you to sort of move things forward in  ways, you know, that that have been helpful? Yeah, [21:16] 100%. And I I think uh what that does is it brings  all the stakeholders on a level playing field. Uh [21:22] you bring them in a room, they all have input to  it. So they all feel that they own a part of it, [21:29] right? And that's how you you bridge those gaps  and and bring in people to believe in the vision [21:35] and and innovate in in the AI space because  every month or even month is is is long like [21:42] every week there's new models new innovation and  everybody's overwhelmed with the information and [21:48] the seauite guidance that all organizations are  receiving is to get more efficient adopt more AI [21:54] make your customer experiences better but to be  able to orchestrate that at scale we have to make [22:01] sure that there is a common minimum uh alignment  guideline that that everybody believes in. So is [22:08] AI charter a document and like who who writes  it and hope you're not going to say claude. Um [22:14] yeah it's it's a collective thing right like  all stakeholders review it and and then that's [22:18] set in stone to make sure these are some hard  guard rails that that everybody has to follow. [22:24] So at any given point uh when anybody goes off  track or tries to do something out of the box [22:29] tries to innovate that that might look at look as  risk they can go back and review okay what is our [22:35] baseline and and what is the box in which we need  to operate. How long have you had a AI charter [22:42] you know at your company? So u we have had it now  for like more than like six to eight months. Okay [22:49] very good. Um would you recommend that any or  large organizations consider making one? It's [22:55] been that helpful. I have actually read like a  lot of companies are coming up with uh something [23:00] of this concept of an AI charter or or an AI  guideline that uh organizationwide they have to [23:06] adopt. It's it's not new like a bunch of companies  are are coming up with their policies to make sure [23:13] especially large ones right it's very hard to get  30,000 people or 35,000 people on on the same page [23:19] with respect to a transformative technology like  AI so uh I believe a lot of them are adopting it [23:25] and and some of them have also published it and  made it open source very good all right let's talk [23:31] about customer expectations because one thing that  AI can do is it can drive um it can help drive a [23:37] lot of this like personalization, right? So,  more and more customers that come to or finance [23:43] um banks um their expectations are different,  right? Uh and especially the younger they get [23:51] like they just want it to get out of the way. They  want it to be mobile, all of these things. How is [23:56] um AI impacting personalization and interface and  all of those things from a customer standpoint? [24:03] I'm just going to say don't come to us as a bank,  come to these guys because we're just going to [24:07] wait what these guys ship and once it's a little  bit more mature, we're going to copy it or buy [24:11] them and then we're going to implement it. I think  as a customer today, especially younger customers, [24:17] whatever expectations you may have, you're going  to be vastly disappointed in your bank's ability [24:21] and willingness to meet those expectations.  And there's very valid reasons for that. But [24:26] I think these guys can probably talk a little  bit better about how you really do it. I think [24:30] it's a fascinating topic, right? So traditional  fintech banks and any of any of these environments [24:34] right has been like a many to one. You build a  solution sorry one to many. You build a solution [24:40] that applies to millions of customers and in  where we are evolving right now it is probably [24:45] uh one to one. is getting into one to one  especially in a smaller startup environment, [24:50] right? We have the tools and the abilities to  really understand every customer and provide [24:56] the services in a much more personalized manner  where it is onetoone and the customers are going [25:02] it's not going to be like you know oh wow  they're going to be expecting this more and [25:05] more and any company or any organization that  doesn't steer in the direction to provide this [25:10] hyperpersonalized services are going to lose the  race. I mean in our case we we service merchants [25:17] right and and what they expect is that um since  we have uh all the data for them and and we run [25:24] all their technology they do expect uh some amount  of intelligence coming out of these platforms now [25:30] um after they have all experimented with cloud  code like some high higher tech merchants have [25:35] implemented their own agents they expect us to  provide those personalized recommendations like [25:42] um what would a dynamic price change look like  in their environment or or what would a schedule [25:48] change look like? Like they they look for those  intelligent indicators that can help improve uh [25:54] their operations and that's something that's now  becoming more and more like a basic expectation [26:01] when you go out and and try to sell tech to to  these merchants. uh and and I believe as you [26:07] know AI gets more intertwined in our daily  lives even more than what it it already has [26:13] um these expectations will continue to grow and  organizations that don't keep up um and embrace uh [26:22] AI and and embrace this tech like it'll be very  hard for them to survive in in the long term. [26:29] All right. We promised in the uh the name of this  that we'd talk about payments, you know, too. And [26:35] uh Gio, I think you were saying like payments  is the one area that keeps growing, you know, [26:39] in in this space. So, let's talk a little bit  about payments and transactions. Uh it's a very [26:45] dynamic area. People still need to move money. You  know, how about when it comes to um to this where [26:51] is AI, machine learning, sort of these advances?  Um how are these things making that better? [26:57] And what is what does that look like? How much  is that a part of you know your thinking when [27:00] you think about innovation right now? Maybe I can  talk about I deal with payments all day long. So [27:06] uh traditionally in a payment sense if you  understand the payments market there are [27:10] different types of players in movement money.  There is no one single company that moves money [27:15] from end to end. Uh in terms of cards for example  there's a visa masterard, there is a issuer, there [27:21] is an acquirer, there is a permanent acceptance.  There's various pieces where we are heading is [27:26] where the the company or the organization that  has the most visibility with respect to data [27:33] across the network across the end toend ecosystem  is going to be able to provide the services for [27:38] the entire network provide the capabilities that  is going to be very very valuable and tying back [27:44] to the previous personalization without the  relevant data doing that personalization is [27:49] going to be challenging. So uh data is going to  be the key and based on the data the insights [27:54] and the analytics and the intelligence is going  to drive lot more capabilities and services that [27:59] that are unimaginable today. 100% uh data is  gold and and whoever owns the maximum amount [28:07] of that payment orchestration data would can  like is is best position to win that race. And [28:14] um I mean at the end of the day payments are  commoditized. So um you have to survive by [28:22] building value added services on top of payments  like a vertically integrated SAS offering that can [28:28] power the whole merchant ecosystem and payments  is just a part of it and you can get a lot of [28:34] those indicators from the data that the merchant  is uh processing like the merchant data you get [28:41] the customer data you get to provide and offer  those value added services that are meaningful. [28:49] Do you want do you want to answer this one? No,  [laughter] you you kind of did on the last one, [28:54] didn't you? I payments is commoditized. I think  that's a great statement and I'm honestly I'm [28:59] still waiting for something really exciting to  happen in the payment space. When it comes to AI, [29:03] I haven't really seen it. I think for me  personally, I think blockchain capabilities [29:08] are far more interesting when it comes to  payments. Um but yeah, I I don't know why I never [29:13] understood but payments is such a fast growing  industry. Um clearly there's a lot of money to [29:18] be made but I think you guys will agree it hasn't  fundamentally changed the way payment schemes work [29:22] right in the last 10 years or so until the  what you mentioned about blockchain and and [29:27] USDC and and all that like that has significantly  changed and created new business models but yes [29:34] uh credit card payments uh has not fundamentally  changed in years right and what is happening now [29:40] is that there are value added services being  created on top of it like aentic commerce and [29:46] agentic payment protocols are coming up that are  leveraging those traditional rails and what I feel [29:52] is prime to be disrupted is a new rail that can  cater to the agentic commerce uh ecosystem instead [30:02] of leveraging those traditional rail like nobody  has thought through uh an a native agentic payment [30:09] rail. everybody is uh creating protocols that  follow the same card network acquiring processing [30:17] flow. Uh and I think that's the next uh era where  where innovation and and new business models uh [30:24] would come into play. One thing I'll add totally  agree with that. One thing I'll add and the future [30:28] of payments with respect to innovation right I  think uh November December 2025 is when all of the [30:35] big tech companies like Visa FISA Google they've  all partnered to start providing uh partnering in [30:42] the ecosystem the new world of payments ecosystem  in the age of AI and there have been several [30:47] collaborations that dem democratizes it's not  like you know any company will do whatever they [30:52] want and that there's a realization that just like  a common scheme team where you can use a card in [31:00] any location right so everybody's cooperating  a similar type of cooperation is starting with [31:04] between fintech companies banks and then schemes  to provide the common rails common platform common [31:11] policies and it is going to it's been just 6  months but it's going to continuously evolve and [31:16] at the ground level to either of the other points  uh things may not change but at the surface how [31:21] we operate how do we uh track end to end when  agentic agents are going to do transactions. [31:28] There has to be questions around who author who  authorized it, who initiated it, who owns this, [31:34] where is going going all of these have to be hands  have to be much more strengthened. All right. A [31:40] similar uh topic to the one we talked a a little  bit ago is generative um AI is opening these new [31:48] possibilities for consumers in uh accessing  banking and financial systems. Right? So, I'm [31:56] going to ask the audience a question, you know,  first, which is digital banking assistance. How [32:02] many um people in the room would like to have AI  help you um you know, make really good financial [32:11] decisions, maybe give you feedback on some of your  financial decisions. Raise your hand. All right, [32:16] there's three of you. Thank you. All right,  appreciate that. Um so, tough crowd and this [32:20] is an AI conference. So, um what are you all  seeing? are you seeing your customers you you all [32:26] were early adopters banks were uh early adopters  in c using customer service for um you know this [32:34] so you you've used some of these systems those are  deterministic systems but still you know what what [32:39] are you all seeing in terms of customers wanting  to use the tools so I I'm going to make a guess [32:44] why not many people raised their hand including  myself I think you know financials your financial [32:51] health is very emotional and I think My sense  psychologically is that most people know what [32:56] they should be doing or shouldn't be doing. They  don't necessarily want someone to remind them of [33:00] that. Definitely not a sentient being that has  access to all of the things that you're doing with [33:04] your money. And so that's number one, and that's  a serious thing. I don't necessarily think people [33:07] want that. The second thing is the reason why even  if you wanted it, you're probably not going to get [33:12] it is legal liability. It's very simple. It's very  hard for us as banks to have an AI chatbot over I [33:19] hate chatbots but an AI tool out there that gives  you advice because now to your point right how do [33:24] we ensure that that advice is right who's being  held accountable if the advice is not right and [33:29] on top of that I think again most customers  vastly over overestimate how well our data as [33:36] banks is organized and structured right so it  may look like we have all of this information [33:40] of you but we can easily tell you what you should  be doing the truth is all of that data is highly [33:45] fragmented across multiple systems. And by system,  I'm being very generous. It could be an Excel [33:50] sheet that sits on someone's desk, right? And so  it's very difficult for us to even get to that [33:55] level of intelligence. Yeah. I I think um there's  a little bit of uncomfortableness in uh any [34:04] anybody uh to have their information available to  a agent or a you know a chatbot or whatever have [34:11] you. What is going to be important? uh many of  the banks you would have used their chat bots they [34:16] just provide basic information you know how do I  set my password where can I find this information [34:21] but that that's barely what they do most banks  right where I think it is heading is uh in [34:26] terms of any any organization whether they're bank  fintech any organ or for that matter have to have [34:32] something something I call as a autonomy ladder  right on day one you cannot expect consumers to [34:38] just hand away their banking information and all  of their financial information and let you do go [34:43] you know go wild with whatever investing or  whatever the trust ladder that I would say is [34:48] that for any AI system that anybody builds for  any purpose they have to have a trust autonomy [34:55] ladder so they any any tool have to first get  the ability to read just read and that's that's [35:02] people will be able to give that information once  you uh your tool is able to read AI agent is able [35:07] to read then they can recommend things okay you  know what I just saw all of this information maybe [35:13] you should consolidate this or maybe you should  do this. It's just recommendation. That is the [35:16] second level of trust autonomy ladder. Then once  you earn a little bit more trust hey this is right [35:22] that's where you go to the third level where you  say you know what I can do these steps for you. [35:27] Would you like me to do it? And then the customer  says hey you're right go ahead and do it. And the [35:32] final is where execution where the customer gives  you permission. This is the autonomy ladder that [35:37] any system has to follow and without which any  anything anybody builds is is set to fail. And I [35:43] mean you you see this every day. I'm sure some  of you use cloud right. How often does it say [35:48] when you update something it says uh allow me to  do this dangerously allow all the time. You know [35:54] after doing it four times you get irritated and  say what? Okay just go do it. I trust you enough. [36:00] That that's an example of a trust autonomy ladder.  And so here's the question. Is it trust or is it [36:04] just laziness and comfort? Because I think for  me a lot of it is just laziness. It's I don't [36:09] necessarily trust it because I don't have the  tools or the context to be able to know if I [36:13] can trust it. I just know I trust myself not  want to have to deal with this all the time. [36:17] So I'd rather just click yes and pray and hope  for the best. I mean it's individual preference, [36:22] right? So it's your money. You can do anything you  want with it. But at at the same time, I think we [36:27] do have some responsibility to take care of. uh  and there are tools like any tool you build if you [36:32] can give the transparency and let's say hey there  is an auditability and you publish a end of day [36:36] end of week email say hey these are the actions  I've taken these are the results of that action [36:41] and and there are controls that you can put like  if something outcomes are not happening as per the [36:46] uh what happens you can put like dials where  customers can dial in dial out a ton of things can [36:53] I can I say something about trust because there's  one thing I find really fascinating which is there [36:57] seems to be some asymmetry in trust when it comes  to people versus machines, right? [snorts] And I [37:02] think a lot of it has to do with comfort because  I think your risk appetite is primarily driven [37:05] by the comfort you have with a certain risk and  the exposure of it. And so we talk about AI and [37:11] auditability and explainability and trust and all  of this stuff, right? Every time we do a PC in our [37:16] bank, which is like a proof of concept, and we  actually benchmark human decision-m against the [37:20] machines or, you know, a model, the model almost  always wins spectacularly, right? But at the same [37:26] time, we put all of these constraints around it  because god forbid the model might be wrong. And [37:30] that's the right thing to do because the model  might be wrong. But what about all these people [37:33] that have been wrong for the last decades? Like  we as humanity are terrible at doing some of these [37:39] things. Credit decisioning, highly biased. But  here's the problem. It's a person. You can't look [37:44] inside their mind to know what kind of bias they  have. There's no auditability on their reasoning [37:48] model, right? But we're perfectly comfortable with  that. But when it's a machine, we need to have [37:52] all of these governance things. And again, I get  it. We should, but I think we should also be very [37:57] cognizant of the fact that it's because it's still  a new technology. Give it 6 months, a year, maybe [38:03] less, and all of us will be as comfortable with  AI as we currently are with Gmail where we give it [38:09] access to pretty much everything in our lives. And  I I just wanted to add to that like the liability [38:15] piece Dio touched on initially. Um, I wanted to  give you an example like there was a a lawsuit [38:22] between Air Canada and and one of their customers  who was chatting with their chatbot, right? An LLM [38:29] based chatbot about their refund policy. Uh, and  the chatbot hallucinated and provided the wrong [38:36] wrong refund information and the the client  missed their flight thinking they'll get full [38:41] refund. That never happened. And they actually  took Air Canada to court, right? and Air Canada [38:47] lawyers were like, "Oh, it's the chatbot." And the  court decided in favor of uh the customer and uh [38:56] he won the lawsuit against Air Canada and they had  to uh give him like give him all the money back [39:01] and along with any additional you know expenses  they incurred. So all that comes to say is that [39:09] yes customers after you know chatting merchants  and customers after chatting with claude chat [39:15] GPT their expectations have increased a lot  but at the same time uh they understand that [39:22] these uh systems hallucinate and where we all  as technology providers come into the picture [39:28] is that we have to 100% make sure that whatever  we are exposing has that trust model and we are [39:38] increasingly focusing on accuracy of the output  even if it's a very basic model and providing [39:43] very basic information. Accuracy is what will  win trust and slowly allow you to do a lot more [39:51] through these chat bots than you can initially.  All right, last question. I want to zoom out a [39:58] little bit from from just sort of fintech, but  I I was thinking about something that uh that [40:04] sat you we we talked about before we started,  which is that you talked about what I would [40:09] call sort of radical transparency. you called it  uncomfortable transparency that you had heard in [40:14] uh another conference you were at that Sam Alman  was talking about and you really flipped my [40:20] understanding of this is not something I expected  to hear from somebody who worked at a fintech you [40:25] know but talk a little bit about that and then and  then yeah for uh for the rest of us maybe think [40:31] about you know what are what are the ways that all  of this that's happening in AI is changing sort of [40:36] the ways we think about both both our work and our  our companies. Yeah, absolutely. So uh yeah this [40:41] was at a conference where Sam was talking about  you know where we are heading. Uh historically [40:46] all of us whether companies, humans, we like to  keep information to ourselves and in many cases [40:52] information in organizations information is used  as uh a exclusive club and sometimes as a weapon [40:58] too and where with leveraging AI and uh is coming  is uh it's going to be uncomfortable transparency. [41:06] U maybe traditionally 95% of the information was  like confidential or kept in silos. where we are [41:12] heading is probably 95% is going to be publicly  available where we have to be uncomfortable uh [41:18] with the transparency like right now each one of  my teammates and in my company we have access to [41:25] everybody's information such as any confluence  pages they have put in any slack messages [41:31] uh any meeting notes uh emails these tools have  access to the emails and this is really helping [41:37] them deliver exponential outcomes do their  job much much faster so they can focus focus [41:41] on critical thinking rather than drafting how to  write an email. So all of us have to be that way [41:47] like everybody in my company can see my calendar  like if I have to make something confidential [41:52] private it's on me or my executive assistant to  mark that as private. So in this way what we call [41:59] in our company is collective intelligence.  We have developed to flip the script. [42:04] Most companies today have kind of workforce and  their AI is kind of bolted in pockets everywhere [42:10] and we are flipping the script where we call  we have built AI as a collective intelligence [42:14] at the center and our workforce works around it  and this is how we are going to be working and [42:20] delivering exponential outcomes rather than being  constrained. That's a that's a radical change in [42:25] the way that you operate. So Arjun and Gio, how is  AI changing the way you operate, the way you work? [42:33] um you know personally not like this [laughter]  definitely not like this oh god I already feel [42:40] very jittery with the government having that level  of oversight uh let alone my my company but I I [42:45] sort of understand it and I sort of appreciate  it too I think it takes a lot of courage to go [42:49] down this path I also think you need trailblazers  people are willing to take these risks right to [42:53] make it more comfortable for everyone else and to  kind of see that the water is not shark invested [42:58] or maybe it is shark invested and you want to stay  the hell out for me personally honestly AI I I'll [43:02] talk about something that's a negative because  I think we talked a lot about the positives, [43:05] but one of the biggest problems I have with AI  and people using AI now is that we have something [43:11] called draft schop. And what it really means is  your ability to produce comprehensive drafts is [43:19] insane these days. Every time we have a problem,  someone in my team will now come with like the [43:24] most amazing documentation, 500 pages, index,  whatever, whatever. Right? If they're good, they [43:29] would have removed all of the typical claude or  chatgptiles. If they're not, all of that typical [43:34] structuring is still there. The problem with that  is it creates a bottleneck because now there's [43:38] a handful of people, executives or whatever, who  have to evaluate that critically all the time and [43:44] that's taking a lot of my time now is evaluating  crappy documents that are incredibly wellcrafted. [43:50] Very good, Arjun. Yeah, I I think um I agree  with that whole common minimum intelligence [43:55] layer platform and building a AI around it because  uh if you don't provide information to your own [44:03] agent, it does not have the knowledge graph to  perform and and it won't beat your expectations. [44:09] So that is absolutely necessary and in my personal  world that's what I've done uh connect connected [44:14] everything you know all my email slack account  everything to claude and and I I now literally [44:20] operate everything through through my cloud agent  and and I trust it because I have set it up in a [44:25] way that I know it's it's secure and and I can do  that uh and we are trying to I [laughter] want to [44:31] push that and bring it into into my workplace  where every day I can get professional email [44:36] summaries and and send replies eyes through the  cloud agent itself. But it's a journey. But again, [44:41] I I would say that that establishing that  intelligence layer is of utmost importance because [44:46] that is what will give you the the most accurate  responses and and the best value for your money [44:52] and your subscription. Maybe I'll give you an  example of something that all of us do and it's a [44:57] very simple example. Everybody can relate. You're  all professionals see out here at the end of the [45:01] week or whatever your schedule is. uh how much  time do you spend sending an update to your boss [45:06] and what are the accomplishments of the week? I'm  sure all of us do in a timely manner. U any any [45:11] guesses how much time you spend today without AI.  Let's hear it. Let's hear it. I mean before we use [45:19] build tools for that we used to spend like between  one to two hours because these are executive [45:23] updates. Today I have a tool that does this within  one minute and I probably spend about five minutes [45:28] to review things. It just gave me my two hours  back. Two hours of my time is much more valuable.