BMO AI Chief Sees Shift to Small Models as Frontier LLM Costs Mount

前沿LLM成本攀升 BMO AI主管预见行业转向小模型 企业AI优势将源于嵌入式工作流程执行与特定领域的专用小模型，而非在未能带来可衡量回报的情况下挤压银行利润率的高成本前沿大语言模型。 Kristin Milchanowski阐述了Bank of Montreal如何将人工智能与量子计算转化为对利润率的可衡量影响，而非将AI视作同质化的技术竞赛或单纯的降本手段。她强调在部署智能体自动化之前应先优化人工工作流程，建立严格的模型风险治理以加速落地采用，并将工具直接集成到企业核心基础设施中。此外，Milchanowski还详细介绍了BMO在信贷决策方面的实际应用，预测全行业将从高成本的前沿LLM转向小语言模型（Small Language Models），并概述了该行在长期量子研究及后量子加密就绪方面的规划。 BMO首席AI与量子官Kristin Milchanowski指出，AI已从技术竞争演变为利润与执行周期；随着基础模型日趋同质化，企业的核心壁垒在于将智能深度嵌入工作流并建立客户信任。 (2:28) 企业在推进智能体自动化之前必须先优化既有人工流程；BMO通过AI整合11个传统系统并提炼数百份信贷与法律文件，使信贷决策周期缩短16%，交易胜率提升1.5倍。 (6:03) 企业级AI应避免单纯追求削减成本的“效率陷阱”，而应以驱动营收增长、提升客户体验与加快决策速度为目标，并遵循明确经济收益、共情设计及预先建立信任三大原则。 (12:51) 在银行环境中部署自主AI智能体的最大挑战并非生成能力，而是如何规模化解决智能体身份标识（Identity）、自主协同编排（Orchestration）以及人工介入的问责边界（Accountability）。 (19:42) 受前沿大模型高昂代币成本与ROI压力的驱动，Milchanowski预测未来12至18个月内银行业AI将大规模转向更具性价比、领域专精的小型语言模型（SLM）。 (31:49)

Artificial intelligence in banking has shifted from an arms race over frontier models into an execution cycle where durable value depends on workflow redesign, governance, and revenue expansion rather than commoditized technology, according to Bank of Montreal Chief AI and Quantum Officer Kristin Milchanowski. Execution and Workflow Discipline Over Model Parity (2:28) Because major financial institutions license the same foundational models from a small group of frontier labs, access to raw models ceases to be a competitive edge and can trigger margin compression. Milchanowski argues that performance instead hinges on operational execution, customer relationships, and embedding tools seamlessly into daily workflows. Deployments must be treated with scientific discipline rather than 'investor theater,' measuring whether AI disappears into the background to deliver bottom-line results. In BMO's credit decisioning workflows—such as the AIM and ARM tools—the bank mandates eliminating operational friction before applying automation. By having AI query up to 11 siloed legacy systems and synthesize 200 to 800 risk and legal documents, BMO reduced end-to-end credit decision cycle times by 16% and achieved a 1.5x improvement in client engagement and deal win rates compared to non-AI-assisted interactions. Tracking baseline human metrics prior to deployment ensures that reported productivity gains reflect verifiable financial lift rather than speculative assumptions. Revenue-Led Scaling and Eliminating Broken Processes (12:51) Milchanowski warns that hyperfocusing on efficiency turns enterprise AI into an outdated cost-cutting exercise reminiscent of Six Sigma programs that historically failed to create lasting value. Instead, prioritizing revenue growth, customer experience, and user empathy fosters adoption that naturally drives operational efficiency and improves decision velocity. To build institutional trust, BMO deployed Microsoft Copilot to all 55,000 employees alongside an enterprise-wide 'AI for All' training program before rolling out specialized tools. Automating an already flawed workflow represents the single largest execution failure in enterprise AI, producing faster operational waste rather than value. To enforce process discipline, BMO requires leaders to eliminate at least one operational bottleneck or redundant process each month, tracking accountability through quarterly reviews and annual leadership cohorts. Agent Architecture, Governance, and Core Infrastructure (19:42) Transitioning from basic chatbots to autonomous enterprise agents introduces three core engineering hurdles: identity, orchestration, and accountability. Every agent must possess a persistent identity documenting its creator, purpose, and data boundaries; dynamic orchestration to coordinate with other agents without manual routing; and strict failure protocols determining when human intervention must be triggered. Rather than slowing innovation, upfront Model Risk Management serves as an architectural accelerator by establishing edge cases and operational guardrails before deployment. To prevent architectural fragmentation, BMO strictly ingests AI capabilities directly through its primary cloud infrastructure stack—AWS, Microsoft, and IBM—or approved Model Context Protocol environments. Organizationally, primary profit-and-loss divisions retain budgetary and application ownership, while Milchanowski's central team engineers reusable horizontal capabilities and establishes responsible AI taxonomies that resolve compliance ambiguities. Quantum Benchmarks and the Transition to Small Models (27:56) In quantum computing, Milchanowski differentiates current 'quantum utility'—verifiably solving real business problems on quantum hardware, achieved only recently—from 'quantum advantage,' where quantum machines definitively outperform classical architectures. BMO uses an IBM quantum partnership to model seismic data, leveraging earthquake prediction across fault-line client regions as an algorithmic sandbox for the bank's capital markets business, which processes over one billion transactions daily. While she anticipates quantum advantage demonstrations in laboratories within 12 months, commercial production will take years and emerge in molecular sciences before financial services. Over the next 12 to 18 months, Milchanowski forecasts an industry-wide pivot away from expensive frontier large language models toward open-source Small Language Models (SLMs) and targeted question models. High token consumption costs currently outpace realized business value for many enterprises, making right-sized, domain-specific models a more economically sustainable path to return on investment. While financial institutions face external compliance mandates—such as two-year NIST timelines for post-quantum encryption—the primary barrier to enterprise AI remains internal: organizations must overcome legacy spreadsheet workflows and fragmented databases to achieve real-time decision velocity without eroding regulatory trust.

## Background 本内容由频道 Ai4 于 2026 年 8 月 22 日发布在 YouTube 平台上。页面简介指出，AI 竞争已从单纯的技术周期转向利润与执行力周期，讨论聚焦于银行如何超越模型本身，以高度的纪律、信任与可衡量的业务成效将智能化规模化落地为真正的商业价值。 ## TL;DR 在这场深度对话中，蒙特利尔银行（BMO）首席人工智能与量子官 Kristin Milchanowski 深入剖析了银行业 AI 从“实验演示”迈向“利润创造”的实战路径。她强调，当顶尖大模型逐渐走向通用化，机构的真正护城河在于无缝嵌入深层业务流的端到端执行力与风险治理，而非单纯的技术竞赛。BMO 摒弃以单纯降本增效为核心的打法，转而以决策速度与收入增长为导向，在信贷审批自动化、多 Agent 协同以及量子计算的前瞻探索中，严格确立“先建立信任再规模化”的治理框架与人机协同机制。 ## Atmosphere & style 整场对话展现出一种务实、严谨且极具专业纵深感的高管访谈氛围。主持人循序渐进地抛出关于 ROI、技术泡沫与执行瓶颈的尖锐追问，而拥有量化交易与模型风控背景的嘉宾则以科学家式的理性思维与生动的赛车、六西格玛类比沉着应对，言谈间充满冷静的工程纪律与战略笃定。现场没有浮夸的技术狂热，而是充盈着对大型金融机构现实复杂性、流程痛点及治理本质的深刻洞察，带给观众一种兼具宏观战略高度与微观落地手感的沉浸式思辨体验。 ## Key takeaways - [02:29](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=149s) AI 已从单纯的技术军备竞赛进入“利润压缩与执行力周期”：通用大模型正迅速商品化，银行真正的护城河并非模型参数，而是如 F1 赛车团队般的端到端协同能力与客户信任。 - [06:04](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=364s) 推行“先精简优化、再智能自动化”的落地法则：避免用前沿技术重复低效流程，在剔除冗余后再利用智能体打通 11 个跨系统数据孤岛，成功将信贷决策周期缩短 16%。 - [09:56](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=596s) 警惕单纯追求降本增效的“六西格玛”陷阱：过度关注效率红利极易被行业竞争迅速抹平，AI 投资必须聚焦于驱动营收增长、提高赢单胜率与提升企业整体决策速度。 - [14:55](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=895s) 规模化 AI 的三项核心法则：坚持经济学效益先行、带着共情进行创新、在规模化推广前先建立员工与客户对技术的信任，通过全员赋能培育技术采纳文化。 - [19:42](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=1182s) 银行级自主智能体（Agent）的三大架构支柱：必须具备可追溯的身份元数据、跨节点自主编排能力，以及内生于系统架构的风控边界与人工降级接管机制。 - [23:23](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=1403s) 坚决拒绝零散外挂组件（No Bolt-ons）：技术外购本质上依然需要深度工程自研，AI 能力必须原生嵌入银行核心云基础设施与专有数据流之中。 - [26:14](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=1574s) 负责任 AI 治理框架能以清晰度消除流程模糊性：通过预设标准化分类法与风控升级路径，终结跨部门无休止的推敲争论，大幅提升项目的决策落地速度。 - [27:56](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=1676s) 量子计算布局重在底层算力函数的跨界复用：针对地震预测研发的复杂时间序列量子算法，可直接迁移并应用于资本市场每日超 10 亿笔交易的高频数据分析。 - [31:54](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=1914s) 未来 12 至 18 个月小语言模型（SLMs）将成为主流：面对大语言模型高昂的 Token 算力成本与难以衡量的 ROI，小巧、快速且专注精准提问的模型将更具性价比。 - [33:38](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=2018s) 后量子加密（PQC）面临合规倒计时：金融机构必须在未来两年内严格对齐 NIST 等监管标准，积极布局向后量子加密架构的全面迁移。 - [33:38](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=2018s) 中央团队与业务线的高效协同机制：中央 AI 团队专注攻坚平台通用底层“大石”并制定全行治理规则，各业务线掌握预算主导权并针对客户需求灵活定制。 - [36:21](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=2181s) 以嵌入式智能体终结“Excel 汇总孤岛”：将 Agent 直接植入核心财务与 CRM 系统，使原本耗费多团队数周时间的手工数据收集转化为分钟级实时财务预测。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=4s) 角色背景与银行 AI 的治理与构建并重 来自路透社旗下金融评论机构 Reuters Breakingviews 的美国专栏作家塞巴斯蒂安·佩莱赫罗（Sebastian Pellejero）开场介绍了本场对话的嘉宾——满地可银行（BMO）首席人工智能与量子官克里斯汀·米尔恰诺夫斯基（Kristin Milchanowski），共同探讨银行业如何将 AI 与量子计算转化为真正的商业价值。米尔恰诺夫斯基首先分享了自己的职业与学术背景：作为一名科学家，她早年从交易员起步，曾是一名量化分析师（Quant），并在量子数学领域取得博士学位，长期深耕于模型风险管理。她感慨道，如今金融服务业正处于 AI 与量子技术相互交融的历史性交汇点。 在拥有约 6 万名员工、全球市值排名第 75 位的上市公司 BMO 中，米尔恰诺夫斯基掌管着整个企业的 AI 战略，其核心工作主要由两大支柱构成：治理与构建。在治理层面，面对银行业已应用 38 年的传统 AI 以及新兴的自主智能体（Agentic Capabilities），她负责制定政策、流程与智能体治理框架，并牵头由 15 家金融服务机构同行组成的每月闭门交流组，共同对齐标准；在构建层面，她的团队致力于打造可扩展的企业级核心底层能力，使各业务线的 P&L（损益表）团队能够在此基础上根据自身产品需求进行定制与灵活调用。 - **机构与组织规模**：BMO（Bank of Montreal）拥有约 60,000 名员工，为全球第 75 大公开上市公司。 - **同行协作机制**：牵头由约 15 家大型金融机构 AI 负责人组成的月度智库，交流 Agent 治理规范与实操经验。 - **双轨职责**：兼顾企业级 AI/Agent 风险治理制度制定与端到端核心技术能力的工程化构建。 --- ### [02:29](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=149s) 告别技术竞赛：利润压缩周期与 F1 赛车式执行力 针对行业描述中“AI 不再是单纯的技术周期，而是一个利润与执行力周期”的论断，米尔恰诺夫斯基给出了深刻剖析。她指出，当全行业所有机构都能获取完全相同的顶尖基础模型和平台时，军备竞赛式的技术较量便不复存在，随之而来的是残酷的“利润压缩”（Margin Compression）。在此背景下，银行若想从竞争中脱颖而出，决胜关键不在于模型本身，而在于如何通过技术去强化人际关系与客户信任，让技术无形且无缝地融入业务体验之中。 谈及如何甄别真实的业务价值落地与浮于表面的“投资人表演”（Investor Theater），米尔恰诺夫斯基指出，科学家的底层思维是“单次实验成功绝不等于可以规模化复现”。BMO 首席财务官（CFO）最看重的指标正是 AI 是否与核心业务流紧密挂钩。如果一项技术必须刻意向用户强调其存在，说明它并未真正深入嵌入流程；理想的状态是技术完全隐形，消除一切卡顿与割裂感。 为了具象化这种差异化优势，米尔恰诺夫斯基引用了一级方程式赛车（F1）作为生动类比：即便各支车队起点处配置了性能极其接近的赛车，最终决定胜负的依旧是赛车手的操控、维修团队（Pit Crew）的协同配合、整个团队的运营纪律以及在高压环境下的执行表现。在银行业，AI 亦是如此——真正的护城河来自于围绕模型落地的端到端执行力。 ```mermaid flowchart TD accTitle: AI Competitive Differentiation Framework accDescription: Flowchart contrasting the commoditized model layer with proprietary organizational execution. subgraph TechLayer["通用技术层 (无实质差异)"] A["顶尖基础大模型"] B["标准化模型服务平台"] end subgraph ExecutionLayer["执行与运营层 (核心护城河)"] C["深层业务流无缝嵌入"] D["严谨的模型治理与风控"] E["高协同性的人机交互"] end subgraph ValueLayer["最终商业价值"] F["增强客户信任与人际连接"] G["抗衡行业利润压缩"] end TechLayer --> ExecutionLayer ExecutionLayer --> ValueLayer ``` --- ### [06:04](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=364s) 落地信贷决策与跨系统整合：优化流程与人机协同 在具体的端到端 P&L 落地案例上，米尔恰诺夫斯基详细介绍了 BMO 在信贷决策产品（如 AIM 与 ARM 系统）中的生成式 AI 应用。传统信贷审批链条是一个“荒谬且耗时数小时甚至数天”的混乱流程。在引入 AI 之前，她对业务部门提出了硬性要求：首先自行优化业务流，剔除陈旧、低效和无意义的环节——"我们不需要用先进技术去重复愚蠢的事情"。在人类协作流程精简到最优之后，团队再利用智能体（Agent）自动化串联关键节点，最终成功将信贷决策周期缩短了 16%。 米尔恰诺夫斯基进一步总结了当前整个金融服务行业普遍聚焦的两大主战场： 1. **内容再呈现与智能检索（Content Representment）**：深入调阅 200 至 800 份厚重的文档，精准提取所需的关键风险条款或跨文档共性信息； 2. **跨异构遗留系统的数据抓取与打通**：大型银行往往背负着长达上百年演进中累积的离散系统（Disparate Systems），信贷经理或反金融犯罪调查员为了撰写一份备忘录（Memo），往往需要反复切换登录多达 11 个不同的系统。如今，AI 与自主智能体能够代为穿梭于各个孤岛系统之间提取数据。 在此过程中，米尔恰诺夫斯基特别强调必须严格保持“人机协同”（Human-in-the-loop）。衡量 AI 价值的前提是必须精确测量原先人类单独操作时的基准耗时，再对比 AI 辅助后的耗时差值；很多企业盲目上线系统并臆断其产生了效益，却从未在初始阶段对人类作业基准进行过严肃度量。 ```mermaid flowchart LR accTitle: Credit Decisioning Workflow Optimization accDescription: Step-by-step optimization flow from legacy process cleanup to agentic integration and baseline measurement. subgraph Step1["阶段一：业务精简"] A["传统冗长信贷流程<br/>(耗时数小时至数天)"] --> B["剔除冗余低效环节<br/>(纯人工作业优化)"] end subgraph Step2["阶段二：技术赋能"] B --> C["智能体 Agent 串联流程"] C --> D["跨 11 个异构系统抓取数据"] end subgraph Step3["阶段三：价值产出"] D --> E["人机协同审核 (HITL)"] E --> F["信贷决策时间缩短 16%"] end ``` --- ### [09:56](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=596s) 董事会的 ROI 追问：从效率指标到胜率与客户参与度 面对动辄数十亿美元的 AI 资本开支，董事会不可避免会追问投资回报率（ROI）。米尔恰诺夫斯基坦言，目前几乎没有任何一家银行在财务损益表（P&L）中设立独立的“AI 收入/成本”会计科目，因此评估必须落实在具体的业务产出指标上。 在向董事会汇报时，BMO 重点呈现了多维度的实际成效： - **客户互动与赢单率提升**：其 AIM 系列产品赋能下，业务团队获得了相比以往高出 1.5 倍（+150%）的客户互动胜率与业务转化表现； - **运营效率与周期压缩**：大幅缩短前中后台复杂业务的运转周期。 米尔恰诺夫斯基强调，证明 AI 价值的核心在于事前就必须明确定义所追求的商业价值指标（Business Value Outcome）；若缺乏明确目标，最终就无法评判落地成败。 在回应佩莱赫罗关于“效率红利往往会被行业竞争抹平或重新投入消耗，而非直接沉淀为净利润”的质疑时，米尔恰诺夫斯基直言，她加入 BMO 的使命绝不是为了单纯搞一轮“降本增效”（Efficiency Play）。她回顾道，那种追求极致效率的做法就像当年盛行的“精益六西格玛”（Lean Six Sigma）时代，在座的许多人都亲身经历过，但正如历史所见，通用电气（GE）当年押注这套模式的最终结局并不理想，单纯的效率游戏无法构成持久的商业壁垒。 ### [12:51](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=771s) 转变思维：从盲目追求效率转向收入增长与决策速度 Kristin Milchanowski 回顾了过去企业推行精益管理与六西格玛（Six Sigma）的教训。虽然少数企业（如丰田汽车通过其看板系统 Kanban）取得了巨大成功，但大量公司在单纯追求效率的打法中收效甚微。她明确指出，过度聚焦于“效率提升”往往会导致企业错失“收入增长”的大好机会。对于任何组织而言，营收增长都至关重要。 在蒙特利尔银行（BMO），Kristin 推行了一种全新的思维转变： - **以共情与信任为基石**：带着共情与信任去构建 AI 系统，信任会直接推动员工与客户的采纳率（Adoption），而广泛的采纳最终会转化为实质收入。 - **收入驱动效率**：当团队全力聚焦于业务增长、客户体验与收入时，流程中的效率提升会作为“副产品”自然而然地发生；反之，若一开始仅仅要求“提高效率”，就会彻底丢掉业务增长的大局。 - **决策速度（Decision Velocity）**：这是企业接下来的核心竞争壁垒。将 AI 深度嵌入基础设施中，让企业预测、供应链以及各项运营决策以更快的速度流转，从而全面驱动收入。 ```mermaid flowchart LR accTitle: 收入驱动型 AI 价值链路 subgraph Mindset["核心驱动力"] A["带着共情与信任构建"] --> B["提高用户采纳率"] end subgraph Outcome["业务成效"] B --> C["驱动业务与收入增长"] C --> D["自然带动运营效率提升"] end ``` --- ### [14:55](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=895s) 叫停项目与规模化 AI 的三项核心法则 当被问及是否曾主动“扼杀”过某些 AI 投资项目时，Kristin 解释道，作为一名科学家以及科学家团队的领导者，他们对所有事情都持有“假设驱动”的态度。当事实依据和外部环境发生变化时，创新项目必须随之调整。在 BMO，他们曾因数据成熟度不足、控制措施或治理机制尚不完备而调整过项目的上线时间、缩减范围甚至直接叫停。在银行业中，“负责任地推进往往比抢先一步更为重要”，叫停或暂停项目不是反创新，而是为了在技术迭代过快时保护银行的核心资产——客户信任。 针对企业规模化落地 AI，Kristin 提出了三项关键原则： 1. **从经济学效益出发（Economics First）**：在写下第一行代码之前，必须彻底厘清这项技术能带来何种业务价值，以及价值将以何种形式兑现。 2. **带着共情去创新（Innovation with Empathy）**：没有共情的创新，就像没有信任的效率。构建产品时要考虑终端同事和客户是否真正愿意、渴望在日常工作中使用它。 3. **规模化之前先建立信任（Trust Before Scaling）**：虽然技术人员天然渴望快速铺开产品，但 BMO 采取了循序渐进的策略。他们为全行 55,000 名员工全面上线了 Microsoft Copilot 并推出了 "AI for All" 全员培训计划，以培养员工的好奇心肌肉（Curiosity Muscle）。在推进具体用例试点时，通过让高管亲自参与沟通、让员工在日常中熟悉并建立信任，从而在后续全面推广时收获了极高的采纳率。 - **最大的执行失败模式**：“如果你将一个糟糕的流程自动化，你得到的只会是一堆更庞大的垃圾（If you automate a bad process, you're just going to get more garbage）。” - **每月消除瓶颈机制**：业务部门不能偷懒不做功课。Kristin 要求业务领导者每个月必须主动消除一个流程瓶颈或砍掉一个冗余流程，并通过季度高管会议与年度考核机制进行严格的跨部门问责。 --- ### [19:42](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=1182s) 银行级 AI Agent 落地：身份、编排与风控内生化 谈及当前备受瞩目的智能体（AI Agents），Kristin 强调，在银行部署具有自主行动能力的 Agent 与普通的 Chatbot 完全不可同日而语。当前最大的挑战并非 Agent 能否生成有价值的输出（这一能力已被证实），而是大规模落地时的三大难题： ```mermaid flowchart TD accTitle: 生产级 Agent 核心落地架构 subgraph AgentGovernance["Agent 治理与运维体系"] ID["1. 身份 (Identity)<br/>创建者 / 预定用途 / 适用场景 / 训练数据 / 访问权限"] ORCH["2. 编排 (Orchestration)<br/>自主发现协同 Agent / 动态跨节点通信调度"] RESP["3. 问责与风控 (Accountability)<br/>故障降级路径 / 明确人机权责分工节点"] end ID --> ORCH ORCH --> RESP ``` - **身份（Identity）**：每个创建出来的 Agent 必须携带完整的身份标识与元数据，包括：由谁构建、为什么目的而构建、在何处使用、使用了哪些数据构建、被允许访问哪些内部系统。 - **编排（Orchestration）**：Agent 之间需要具备自主协同的能力，无需人类每次手动介入指令（例如无需人工指定“A 去找 B”），Agent 能够自主判断在特定情境下该调用哪一个专业 Agent 协同解决问题。 - **风控背景的赋能**：Kristin 曾在摩根大通（JPMorgan Chase）从事模型风险管理（Model Risk Management）工作。她借用赛车类比：“在赛车中，油门有时比刹车更能帮你摆脱困境；而在模型风控中，它教会你在项目一开始就提出最尖锐的问题。” - **把治理内生于架构**：设计初期必须追问“它的使用边界是什么？”以及“系统一旦失败会发生什么？”在银行业，一旦技术失效，必须精准知道如何将流程切回人工操作，以及人在整个流程中接管的具体节点。如果在后期才补救，治理就会成为瓶颈；而如果在早期就设计清楚，治理就成了系统架构的一部分。 --- ### [23:23](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=1403s) 自研、外购与生态伙伴策略：依托核心云底座 在“自研（Build）、外购（Buy）还是合作（Partner）”的选择上，Kristin 结合自己曾担任大型咨询公司合伙人的经历指出，并没有放之四海而皆准的模式。虽然在 BMO 的某些决定在外界看来或许有争议，但战略执行非常明确： - **锚定三大核心科技供应商**：BMO 的底层基础设施完全运行在 AWS、微软（Microsoft）和 IBM 之上。 - **不采购杂乱的外挂组件（No Bolt-ons）**：为了实现极高的决策速度（Decision Velocity），AI 必须原生嵌入底层基础设施。BMO 拒绝盲目采购零散的独立工具，只有当三大底层云平台整合或认证了特定初创公司的应用、MCP（模型上下文协议/工具）或能力时，BMO 才会引入使用。 - **“外购本质上依然是自研”**：从合作伙伴处购买解决方案绝不意味着即插即用，银行依然必须投入大量工程力量，根据自身的专有数据、业务流程和合规要求进行深度定制。因此在 Kristin 看来，“买”的背后依然需要扎实的“造”。 ### [26:14](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=1574s) 建立负责任 AI 治理框架：消除模糊性以加速规模化落地 谈及由智能体（Agent）赋能的业务流程与问责机制时，Kristin Milchanowski 强调科学家团队必须高度依赖风险控制。在银行业，AI 产出的结果只有在“可信”（Trusted）的前提下才有实际业务意义，因此治理机制在技术与业务之间扮演着至关重要的桥梁角色。BMO 建立了一套标准化的“负责任 AI 框架”（Responsible AI Framework），其中包含详尽的分类法体系（Taxonomy）。在这套框架下，各个层级明确知晓风控阈值、控制措施、升级路径以及各项职责的具体归属人。 这种清晰度赋予了团队更快的执行速度，因为决策无需在每个环节被反复重新推敲和争论。在传统银行体系中，真正拖慢创新步伐的往往是模糊性（Ambiguity）。如果没有完备的治理机制，团队就会陷入无休止的“这个试点到底该不该做”的重复决策循环中——即便两个方案之间往往只有两句话的微小差异。正如 Kristin 所总结的，在 AI 竞争中最终胜出的组织，不一定是最快上线模型的机构，而是那些能够以负责任的方式规模化扩展 AI、并在此过程中始终维系客户与监管信任的机构。 ```mermaid flowchart TD accTitle: 负责任AI治理架构与决策闭环 accDescription: 展示标准分类法、控制阈值与明确问责如何消除模糊性，实现AI规模化与可信交付 subgraph Governance["负责任 AI 框架 (Responsible AI Framework)"] A["统一分类法体系\nTaxonomy"] --> B["预设风控阈值与控制措施\nThresholds & Controls"] B --> C["明确升级路径与问责人\nEscalation Paths & Accountability"] end subgraph Execution["业务赋能与决策加速"] C --> D["消除重复推敲与流程模糊性\nEliminate Relitigation"] D --> E["高决策速度落地试点\nDecision Velocity"] end subgraph Outcome["长期竞争优势"] E --> F["可信的业务结果\nTrusted Business Value"] F --> G["负责任的规模化扩展\nResponsible Scaling"] end ``` --- ### [27:56](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=1676s) 跨界布局量子计算：从地震预测算法到资本市场海量交易分析 针对主持人提及 BMO 提交了一项用于预测地震的量子算法专利这一“看似疯狂”的举动，Kristin 阐述了背后的战略逻辑。BMO 与 IBM 达成了战略合作，并获得了使用其量子计算机的权限。Kristin 带领着一支专业的量子研究团队，正在推进多个能为银行创造长远业务价值的项目。尽管这些项目为银行财务报表带来实质贡献可能还需要数年时间，但团队的所有科研探索都是以底层通用函数为导向展开的。 之所以选择地震预测作为突破口，一方面是因为 BMO 超过 60% 的客户居住在地质断层带（Fault zones），关乎客户的生活福祉，与银行“在商业与生活中勇于向善”（Boldly Grow the Good in business and life）的使命契合；另一方面，地震预测是一项对海量数据进行时间序列预测（Time forecasting）的极致计算挑战。对于在资本市场每日处理超过 10 亿笔交易的金融机构而言，处理地震数据所研发的量子计算函数，完全可以复用到资本市场的高频数据分析与复杂场景预测中。 谈到量子计算的发展现状时，Kristin 区分了两个关键技术节点： - **量子实用性（Quantum Utility）**：当前科学界已经达成这一阶段。这意味着研究人员可以在量子计算机中运行经典算法并获得真实有效的结果，用于解决现实商业问题，而这在三年前是无法做到的。 - **量子优势（Quantum Advantage）**：指量子计算环境在效率、速度和智能化上全面超越经典计算环境。Kristin 预测科学界有望在未来 12 个月内实现这一里程碑，但这仅代表实验室验证，距离机构级生产环境投产仍有很长的路要走，且首批投产的行业预计会是分子科学领域的化工、制药和生物科技，而非银行。 --- ### [31:54](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=1914s) 未来 12 至 18 个月行业预测：大语言模型的成本困境与小语言模型的兴起 当被问及未来 12 到 18 个月银行业 AI 领域会出现哪些出人意料的趋势时，Kristin 大胆预测小语言模型（SLMs, Small Language Models）将成为行业主流。目前行业普遍面临大语言模型（LLMs）成本高昂的困境，各大机构在 Token 计费上投入巨资甚至持续亏损，而长期的商业价值转化却充满不确定性。 Kristin 认为，行业正在认识到“提出更好的问题”才是关键所在，专注于精准提问的模型（Question models）将更加受到青睐。随着前沿基础模型公司开源其建模技术，越来越多机构能够构建出小巧、精炼且成本效益极高的小语言模型。这些小模型不仅运行速度更快、计算开销更低，还能让业务端更快衡量并兑现切实的商业投资回报。 --- ### [33:38](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=2018s) 现场问答：后量子加密（PQC）的合规挑战与业务线协同运营模式 在观众互动环节，针对白宫近期签署的量子加密行政命令及企业的应对准备，Kristin 指出，尽管宏观政策导向纷繁复杂，但技术团队必须聚焦于美国国家标准与技术研究院（NIST）等监管机构设立的硬性合规标准。企业面临在未来两年内必须完成向后量子加密（Post-Quantum Encryption, PQC）迁移的明确合规时限，所有金融机构都必须对此予以高度重视并积极布局。 ```mermaid flowchart LR accTitle: 中央团队与业务线的协作运营架构 accDescription: 展示中央AI团队提供基础设施与治理，业务线掌握预算与定制落地的协作模式 subgraph CentralTeam["中央 AI / 量子团队 (Central Team)"] direction TB C1["攻坚核心底层能力\n'Big Rocks' Core Capabilities"] C2["制定全行 AI 治理与规则\nGovernance & Controls"] end subgraph BusinessLines["业务线 (Lines of Business)"] direction TB L1["掌握预算与业务决策权\n'P&Ls Run the World'"] L2["根据客户需求定制场景\nTailored Builds"] end C1 -->|"提供平台基础设施"| L2 C2 -->|"提供合规治理指引"| L1 ``` 随后，一位来自银行业的与会者提问了业务线（Lines of Business, LOB）与中央团队之间的运营模式： - **业务线主导资金与场景**：业务线深谙客户需求并掌握预算分配权（“损益部门掌控一切”），部分业务线甚至配备专属的数据科学家团队进行自主开发。 - **中央团队攻坚核心平台能力与治理**：Kristin 的团队专注于解决全行通用的底层“核心大石”（Big rocks），并输出统一的治理标准。中央团队与业务线建立互信伙伴关系，业务线在遵循既定合规治理规则的前提下，自主裁量并落地具体业务场景。 --- ### [36:21](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=2181s) 现场问答：以嵌入式 AI 智能体打破“Excel 汇总孤岛”，实现高决策速度 关于如何利用智能体 AI（Agentic AI）提升决策速度（Decision Velocity），Kristin 举了一个银行业极为普遍的痛点案例：每当 CEO 要求进行月度或季度财务预测时，各业务线和部门通常需要抽调 8 到 10 名员工，花费数周时间在庞杂的 Excel 表格中收集、填写并层层上报数据，最后汇总成一份全行预测呈现给管理层。这种割裂的人工流程完全无法提供应有的决策响应速度。 ```mermaid flowchart TD accTitle: 传统手工预测与嵌入式AI智能体预测对比 accDescription: 对比传统跨部门Excel层层汇总流程与嵌入式AI智能体直连核心系统的实时预测流程 subgraph Traditional["传统预测流程：耗时数周 / 跨团队割裂"] T1["CEO 提出预测需求"] --> T2["各业务线手工统计"] T2 --> T3["8-10人跨部门协作与多份 Excel 汇总"] T3 --> T4["人工层层上报与反复校对"] T4 --> T5["耗时数周交付 CEO"] end subgraph Agentic["Agentic AI 嵌入式流程：实时驱动"] A1["CEO / 管理层实时触发指令"] --> A2["AI 智能体直接连接核心财务系统与 CRM"] A2 --> A3["自主提取多维数据并动态建模"] A3 --> A4["分钟级生成精准预测输出"] end ``` Kristin 指出，真正的决策提速不在于单点工具的使用，而在于将 AI 智能体直接嵌入到存储核心财务数据和业务流转的底层系统（如全行内部财务系统或 Salesforce 等 CRM 系统）中。智能体能够自主提取实时数据并自动生成预测洞察，省去跨部门人工倒腾数据的繁琐环节。银行业必须在技术架构与业务领导力上建立高度纪律性，将前沿智能技术深度植入核心基础设施，才能真正释放“嵌入式决策速度”的业务价值。 ### [39:00](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=2340s) 结语与致谢 在对话的尾声，主持人对嘉宾 Kristin Milchanowski 的精彩分享表达了诚挚感谢，并向现场观众推介她的著作，热情呼吁大家“去读读她的书（Please check out her books）”。同时，他也欢迎大家持续关注路透热点透视（Reuters Breakingviews），以获取关于全球经济与前沿动态的“敏锐洞察（incisive takes on what's happening in the world）”。 整场讨论在热烈的感谢与掌声中圆满落下帷幕，双方致意并向现场组织者表达了谢意。 - **嘉宾推荐**：呼吁关注 BMO 首席 AI 与量子计算官 Kristin Milchanowski 的相关著作。 - **媒体关注**：推荐关注路透热点透视（Reuters Breakingviews）对全球市场与科技趋势的深度评论与分析。 ## Notable quotes > [02:29](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=149s) 单次实验成功绝不等于可以规模化复现 — Kristin Milchanowski > [06:04](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=364s) 我们不需要用先进技术去重复愚蠢的事情 — Kristin Milchanowski > [14:55](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=895s) 负责任地推进往往比抢先一步更为重要 — Kristin Milchanowski > [14:55](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=895s) 如果你将一个糟糕的流程自动化，你得到的只会是一堆更庞大的垃圾（If you automate a bad process, you're just going to get more garbage） — Kristin Milchanowski > [19:42](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=1182s) 在赛车中，油门有时比刹车更能帮你摆脱困境；而在模型风控中，它教会你在项目一开始就提出最尖锐的问题。 — Kristin Milchanowski > [23:23](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=1403s) 外购本质上依然是自研 — Kristin Milchanowski > [27:56](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=1676s) 在商业与生活中勇于向善（Boldly Grow the Good in business and life） — Kristin Milchanowski > [31:54](https://www.youtube.com/watch?v=HxIPxdWtw3U&t=1914s) 提出更好的问题才是关键所在 — Kristin Milchanowski ## Mentioned Kristin Milchanowski, Sebastian Pellejero, Reuters Breakingviews, 满地可银行 (Bank of Montreal / BMO), Microsoft, Microsoft Copilot, AWS, IBM, JPMorgan Chase, 通用电气 (GE), 丰田汽车 (Toyota), NIST (美国国家标准与技术研究院), Salesforce, AIM, ARM ## Video description Sebastian Pellejero - U.S. Columnist, Reuters Breakingviews Kristin Milchanowski - Chief AI and Quantum Officer, BMO AI is no longer a technology cycle — it is a margin and execution cycle. Competitive advantage will not come from access to models, but from the ability to operationalize intelligence at scale with discipline, trust, and measurable outcomes. • Shift from experimentation to value creation • Execution discipline model superiority • AI tied to revenue, efficiency, and risk outcomes • Trust and governance as prerequisites for scale [Source](https://www.youtube.com/watch?v=HxIPxdWtw3U)

[2:29] Kristin Milchanowski: 单次实验成功绝不等于可以规模化复现 [6:04] Kristin Milchanowski: 我们不需要用先进技术去重复愚蠢的事情 [14:55] Kristin Milchanowski: 负责任地推进往往比抢先一步更为重要 [14:55] Kristin Milchanowski: 如果你将一个糟糕的流程自动化，你得到的只会是一堆更庞大的垃圾（If you automate a bad process, you're just going to get more garbage） [19:42] Kristin Milchanowski: 在赛车中，油门有时比刹车更能帮你摆脱困境；而在模型风控中，它教会你在项目一开始就提出最尖锐的问题。 [23:23] Kristin Milchanowski: 外购本质上依然是自研 [27:56] Kristin Milchanowski: 在商业与生活中勇于向善（Boldly Grow the Good in business and life） [31:54] Kristin Milchanowski: 提出更好的问题才是关键所在

[2:28] Kristin Milchanowski: AI is no longer a technology cycle — it is a margin and execution cycle. [2:28] Kristin Milchanowski: successful experimentation is definitely not the same thing as repeating the outcome [16:06] Kristin Milchanowski: Innovation without empathy is efficiency without trust. [18:36] Kristin Milchanowski: If you automate a bad process, you're just going to get more garbage. [21:42] Kristin Milchanowski: the accelerator can get you out of more trouble than the brake. [26:14] Kristin Milchanowski: results only matter if they can be trusted results [27:56] Kristin Milchanowski: boldly grow the good in business and in life. [33:38] Kristin Milchanowski: the P&Ls run the world.

[0:04] I'm Sebastian Bel Harrow. I'm a US columnist  of Reuters Breaking Views, uh the financial [0:08] commentary arm of Reuters. Uh joining me  today, Kristen Mulchinowski. Um you know, [0:12] Kristen's the AI chief AI and quantum officer at  [ __ ] Um we're here to talk about how banks are [0:19] turning AI and quantum into value. Kristen, maybe  you can start off by sort of telling everyone what [0:24] your role entails. Yeah, I um so I think first  it's helpful just to say I'm a scientist and so [0:31] my background is in uh model risk. Um I started on  the trading floor uh with my my financial career. [0:40] Uh so I was a quant um my PhD is is in quantum  mathematics and so this is sort of like this [0:48] moment in time when all of these things are coming  together in financial services both the AI and [0:53] the quantum side of the house. Uh so it's really  really exciting. The uh thing that we do at Beimo [1:00] so I run uh AI for for the enterprise and a lot of  my job right now is governance as all of you can [1:08] imagine. Um setting up all of the policies and the  procedures and how do we want to handle agents, [1:14] how do we want to handle agentic capabilities um  in addition to the AI that we've been using all [1:21] of us for 38 years. All of that's still happening  and being built as well. So big part of the job is [1:27] designing that governance. I have um I lead a peer  group. There's about 15 of us across different [1:35] financial services uh companies and we meet to  talk about this this stuff uh once a month. Um [1:44] so we're trying to also lead uh amongst each other  to so that we can kind of combine notes. But a big [1:51] part of my job is also building. So we build a lot  of this capability for our firm. Um we're trying [1:57] to do it in a way that scales. Uh we're trying to  build the core capability so that then in each of [2:03] our P&Ls they can take that core capability  and tailor it for their own product needs. [2:09] That's just a little bit you know. So I've got a  big governance side to my job and then a big uh [2:14] building side of the job. and we're about 60,000  people and we are um the 75th largest globally [2:23] publicly traded company in the world. So you know  our recess description says that AI is you know no [2:29] longer a technology cycle. It's a margin and an  execution cycle. I think that's a strong claim. [2:34] Convince the skeptic in the audience like where  specifically has AI shown up in in BMO's numbers? [2:40] Well, I think when I talk about it being a margin  cycle, what I mean by that is when we all get our [2:46] hands on the same exact capability, it's no longer  a technological race, what it is is a margin [2:55] compression. So, how can you make more money uh  than your competition if you're all running the [3:01] same model and using the same model platforms?  And so to do that in a margin compression world, [3:08] we really believe um the technology is going to  help, but it's really the relationships and the [3:14] trust. And so designing the technology that helps  enable more trusted relationships is is our path [3:21] for that. You know, every bank, every company  says it's moving from experimentation to to value [3:27] creation. like what's your test for, you know,  whether that's real at an institution versus, [3:33] you know, investor theater in some sense. You  know, one of the things that that scientists [3:38] learn really early on is that successful um  experimentation is definitely not the same thing [3:44] as as repeating the outcome. So just because you  can do it once doesn't mean that you can repeat [3:50] it at scale. and testing whether AI is attached  to a workflow that really matters is one of the [3:57] metrics our CFO cares about the most. Um, and  so if you have um, you know, to tell people that [4:05] the AI exists, it probably isn't embedded deeply  enough. Um because what I want is that um I want [4:16] the technology to sort of disappear and be more  seamless ex client experience and execution for [4:22] both of our teammates and um and our clients.  So that experience is what we're after. And if [4:28] they feel like they're dealing with some sort of  clunky thing, then that that isn't going to end up [4:33] scaling either. Um the things that I look for that  that matter are whether that use case is tied to a [4:42] meaningful outcome. I think a lot of times people  get so focused on the use cases without really [4:48] understanding what that end business value output  is. Um that that component is really important [4:55] when you're moving from experimentation to to  value creation. you you said that models won't [5:03] be the differentiator if everyone's buying the  same front tier models from the same three or [5:07] four labs like what exactly is proprietary about  a bank's AI capability? The the thing that is is [5:15] the differentiator is um you know it's a little  bit like racing. So you've if any of you follow [5:24] Formula 1, I'm a big I'm a big uh yeah we got  some in here. um you you've got two teams that [5:31] can start remarkably similar cars. Uh but it the  outcome is really going to depend on the driver, [5:39] the pit crew, the operating discipline of that  team and how well they're able to perform under [5:45] pressure. And AI is going to be the same way. the  model that matters. It's it's the execution of [5:52] it that is going to create that separation um and  that differentiating component of it. So I kind of [5:59] use a lot of car racing analogies, but maybe you  should walk me through, you know, one deployment [6:05] where you can draw a straight line from from the  AI to the P&L outcome, you know, whether it's like [6:11] revenue or cost or loss avoidance like Yeah. Um  so so my financial friends in the audience we have [6:17] AIM and ARM in our world and these are things  where we're trying to bring credit decisioning [6:23] um to to our teams um better smarter faster  and we have developed u with generative AI [6:30] um these products that are able to um to reduce  our um our time by 16% um that helps these bankers [6:42] make these credit decisions that much faster. Uh  and and that's probably a pretty familiar use case [6:49] with most all of you. It's in banking we're  a little bit limited right now on what kind [6:54] of use cases we can implement uh at scale. But  these few that are really really good uh we're [7:01] repeating them in a lot of different places across  the bank. And that is is the best example that I [7:07] can give of how it's carrying through end to end.  So we we took a workflow that we knew was bonkers. [7:16] So it's a workflow that that took several hours um  sometimes even days of how information comes back. [7:23] We looked at how do you optimize that workflow?  So first my ask of the business partners was hey [7:29] optimize your own business. Get get some of the  junk out of your process. get some of the the [7:35] stagnant parts out that you don't that you know  you already don't need because we don't need to [7:39] use technology to repeat silly things. And then  when that process is optimized as best as it can [7:46] be from a human connectivity standpoint, we  then try to help enable that chain of flow [7:52] with the technology u with our agent creation  and things like that. And then that's how that [7:58] um that workflow is then reduced in time. the the  content representment is pretty much where all of [8:04] financial services is right now. Going into 200  800 different documents, finding the answer of [8:12] what you need or finding the risk, finding  the clause like a in a legal document. Um, [8:18] finding those commonalities in the different  documents and bringing those back in. That's [8:23] that's where most of us in financial services  is playing. The other thing that is really [8:29] um nice and and end to end for us is um most  all banks um so this isn't unique to just ours [8:39] uh you all are living this experience where we  have these disparit systems technology stack has [8:44] been accumulated over a hundred years. Um, and in  that that means that our bankers or our financial [8:53] crime teams have to access sometimes 11 different  systems to do their one job to write the one memo [9:01] or to write the one um, you know, yeah, one one  memo. And so we can use AI and we can use the [9:09] agents now to help with even just getting the  information from those systems. But I guess one [9:16] of the big things you're hearing from me even in  this description is that we still very much have [9:20] that human in the loop, but we're able to track  now because we understood what their human process [9:26] was. We optimized that. So we knew how much time  it was taking the human to do the job and now we [9:33] know how much it's taking them with the complement  of AI, which is kind of what you need to be able [9:40] to measure if something's working or adding value  or not. A lot of times we just implement the tech [9:46] and then we assume that it's adding value, but you  never really tracked how long or how much it took [9:52] the the human to do the job in the first place.  Yeah. And you know, banks are spending billions [9:57] of dollars on AI. You know, claims about the value  rooted from these initiatives are vague. You know, [10:03] when you go to your board, when boards ask like  where's the return? like what's the honest state [10:08] of measurement and like what is what is being  implemented that is still sort of difficult or [10:13] harder to to measure still yeah the um I think  the the P&L outcome right that's really really [10:20] hard with with this equation is um is that none  of us I I'm I'm think it's safe to say none of [10:30] us but I should probably say better like almost  none of us perhaps um have AI as a standard and [10:37] loan P&L line item. um yet people are asking the  question how is AI doing X right in these board [10:46] meetings and other so when you want to measure  a specific business outcome across use cases [10:52] um you you we're going to see value show up in  different ways uh we've we've publicly discussed [10:59] examples where AI is cont uh contributing to  stronger client engagement so we talk about that [11:05] with the board um improved win rates Um, so our  one of our AIM products actually wins 1.5 better [11:16] client engagement than than without it. So, um,  we're able to bring that measurement back to to [11:23] our boards. Um, and the that operational  efficiency that I was mentioning earlier, [11:29] uh, it's just reducing those cycle times um, in  various parts of our business. The question is [11:37] is like AI does create value. We all we all know  that it does, but it's really about can we scale [11:45] that value responsibly and consistently across the  enterprise. And to prove that you can or can't, [11:52] we really have to start with that business value  measurement that you're trying to look for. If [11:58] you don't know what business value outcome you're  trying to achieve, then at the end it's hard to [12:04] um say that you successfully did it or not.  You know, like efficiency gains have a way of [12:10] getting reinvested or being competed away rather  than just dropping to the bottom line. Like why [12:15] shoulders believe that AI expands margins rather  than just sort of resetting the cost of of doing [12:22] business? I'll answer your question. I I I want to  start though with uh this the same topic but just [12:30] to share one of the things that I shared with our  leadership team is I'm not here uh I didn't come [12:37] to be to be for an efficiency play like that  that's that lean six sigma days um a lot of us [12:44] in this room did that and and we lived through  that um GE didn't actually turn out so well um [12:51] after all of that right uh So, and there was a few  companies who did great. Toyota did great um with [12:58] their combine system, but a lot of companies  just didn't do great with six sigma and lean [13:04] principles. And that's really the efficiency play.  And so, I didn't really buy into hyperfocused on [13:11] efficiency means you're going to miss out on  revenue. Uh revenue growth is really vital for [13:18] all of our organizations. And if you focus your AI  initiatives and your team and how you think about [13:26] uh building with empathy and building with  trust, trust drives adoption. Adoption then [13:32] can drive revenue, right? Like it's all it's all  associated with that. If you focus on revenue, [13:39] you naturally end up becoming more efficient in  places. And that's that's the like the mindset [13:47] shift that I was bringing to our team at Beimo was  let's focus on revenue, let's focus on growth and [13:54] client experience and you'll n naturally become  more efficient. But if you just ask us to make [14:01] things more efficient, then you're going to miss  out on revenue plays the other way is, you know, [14:08] obviously better. So when when you're asking like  this kind of question, it's it's tough because a [14:15] lot of people most people are focused on those in  um efficiency plays and um and they're not seeing [14:24] that outcome. So you know, I I really think that  decision velocity is what's coming next for all of [14:31] us. How do you embed AI into your infrastructure  so that all the decisions that you're making, [14:37] your forecasts for for your firm, your supply  chain components, all of those um parts when you [14:46] have faster decision velocity that helps you drive  the revenue and um yeah. Is there an AI investment [14:55] that you killed? Like what is what what has that  taught you about like what does and doesn't work? [15:01] Um, you know, I I really think that some of the  you know, the way that uh we build at AI at at [15:12] Beimo, again, I'm I'm a scientist and I lead a  team of scientists and we have um hypotheses, [15:18] right, uh for everything that we do. And so when  evidence changes, um innovation programs need to [15:26] operate the same way. Um so we've had situations  where we've adjusted the timing, we've adjusted [15:31] the scope or the deployment approaches because  that the governance or the controls or the data [15:36] readiness needed to mature more. Um so it's not an  anti-inovation uh approach. It's it's just that in [15:46] a bank it's how you protect trust and uh moving  responsibly is often more important than moving [15:55] first. So there were things that we've killed or  just paused because it the technology was moving [16:01] a little bit too fast and we needed to catch  the rest of the program up. You know your book [16:06] has you know tens of principles on scaling AI. Um  you know if there were three that you wanted the [16:14] audience to really take hold like which three and  which one do you think you know enterprises sort [16:20] of violate most most often? Um, I am a big fan of  starting with the economics, really understanding [16:27] the business value that you're um, and how  that's going to show up before you even start [16:33] building. I'm a big big fan of that. Um, I really  value empathy. Innovation without empathy is uh, [16:42] efficiency without trust. Um, and when you think  through what does that mean? What does it mean [16:48] to build with empathy? It means um I'm going to  build a product that Iramm wants to use at the end [16:54] of the day or that Sunny wants to use at the end  of the day. Like these are things that I want my [16:59] clients and my teammates to be eager to use. And  so I'm thinking about all of that and the design [17:06] principles upfront. Um so that that's one and um  building trust before scaling. It's really really [17:16] um as a scientist or a technologist, we really  want our product to be out there and at large [17:22] and everywhere and we want this massive adoption  all the time, right? And one of the things at [17:28] Beimo that that is is really true is that all of  our people like we rolled out co-pilot to every [17:36] single one of our employees, 55,000 employees,  and we also rolled out an AI for all training. We [17:43] didn't want anyone to be left out if they we want  to build this curiosity muscle. Um we we want them [17:50] to be leaning into the subject. But one of the  but when it comes to the use cases and the pilots, [17:58] we really want people to trust and part of that  is do they see it in their everyday life? Are [18:05] they getting comfortable with it? Do they know how  to be curious with it before we start bringing in [18:10] other use cases? And building that trust, having  them see my face, having them be part of some of [18:16] our learnings, like all of that really helps. So  then when you do have a pilot that's successful [18:24] and you can start rolling it out to the masses,  you have a better adoption rate, which is what [18:29] technologists are are all after. What's like the  like single biggest like execution sort of failure [18:36] mode you've seen in your career in this space?  Like where where do you see people sort of falling [18:40] into hang-ups? Um if you automate a bad process,  you're just going to get more garbage. Um and no [18:53] one really likes to do their homework. Um, but  really making sure that your business partners, [18:59] if they want the best out of this technology, they  need to do their homework. So, I ask my leaders [19:05] every time I'm in a room with them, I ask them  to get rid of one um bottleneck a month or one [19:12] inefficient process that they don't need anymore  a month. And I asked them to show up that way. [19:18] And we we have a mechanism for this. Like we we  actually hold each other accountable across the [19:23] firm. Um we we've got a a quarterly leader  meeting. We have an annual cohort uh um that [19:30] we get together. And this is one of the metrics  that we talk about is like have you been removing [19:35] bottlenecks every month. So that I think that's my  you didn't say pet peeve, but that's my pet peeve. [19:42] Um maybe we can move to agents. Everyone's excited  about them. They're everywhere. Um but like an [19:48] agent that acts autonomously inside of a bank is  a bit of a different animal than a chatbot. Um, [19:53] what's been sort of the hardest unsolved  problem in deploying agents in production [19:58] at a bank? The hardest problem is not whether or  not agents can generate useful outputs. They can [20:06] uh we all uh experience that and know that  now to be true. The hard problem is identity, [20:13] orchestration and accountability at scale. And  what I mean for that, so what I call identity, uh, [20:22] just to level set definitions cuz we all come from  a little bit different of a place. Um, identity [20:27] is when I create an agent or my teammate Sebastian  creates an agent or Scott creates an agent, I need [20:35] it to live with the identity of who built it, what  it was built for, where it's going to be used, [20:43] what data it uh was built with, what systems  it's allowed to access. It need it needs this [20:50] identity to be able to live with that agent that  was created. And then orchestration, just again [20:59] to make sure all of our definitions are the same,  orchestration um that um is a kind of an unsolved [21:06] or becoming solved problem is I need that agent to  be able to connect with other agents more readily [21:15] without me having to say, "Hey, Sebastian's agent,  please go talk to Scott's agent." like I wanted to [21:22] be able to know that you should reach out to Scott  this time for this particular question and then [21:28] next time you can come to me like if the three  of us are representing agents which is really bad [21:33] because you're not supposed to size or whatever  that fancy word is Mr. reporter who writes very um [21:42] you know you came up through model risk management  at JP Morgan you know literally policing models [21:47] um how does that shape how you build now does it  make you slower or faster in in implementing the [21:52] stuff you know I um it it makes me faster in the  ways that matter um I again using a racing analogy [22:01] I I learned in racing that uh the accelerator can  get you out of more trouble than the brake and Um, [22:09] model risk teaches you to ask the really hard  questions at the beginning. And it's it's like [22:16] what's that intended use? What are the limits?  What happens when it fails? Like when's the last [22:22] time your team asked you in this design process,  what happens if this fails? That identity piece is [22:31] so important because if something is failing and  you have to go back to just the human workflow, [22:39] especially in banking, you got to know where to  stick stick the person back into the process and [22:45] which part of the process was being handled by the  technology versus the human. Um so those limits [22:52] um are are really really important and if you  ask those questions late governance becomes a [22:58] bottleneck but if you ask them early governance  is is part of the architecture. Um, so I guess my [23:05] background kind of made me focused on building  systems that can res withstand the scrutiny [23:14] um responsibly with with uh time um in building  systems. Should banks build? Should they buy? [23:23] Should they partner? Like where has [ __ ] drawn  that line? And maybe where have you changed your [23:28] mind in in those different buckets? Um so so I I  the answer to this really depends. Uh I be before [23:39] with Beimo I was uh a partner of a big consulting  firm and so I've seen a lot of different um [23:47] institutions. Sometimes building is absolutely the  best way to go because that's the culture of the [23:53] organization and they do it really really well. um  sometimes building is is the best option and the [24:01] way to go. So my answer I just wanted to caveat  that with because we're doing it a certain way at [24:08] Bimo doesn't mean that it makes sense for everyone  to do it the same way. But when it came to this [24:14] particular technology, one of the things that was  not still isn't a popular decision. Um, but I'm [24:23] I'm adamant that it is our decision and the the  teams are following it, which is nice of them. Um, [24:32] but we only leverage our big tech providers  with this newer technology, meaning uh we run [24:40] an infrastructure stack on AWS, uh, Microsoft and  IBM. That's that's our infrastructure stack. Uh, [24:49] that's not a pitch, that's not a sale. just is is  where we are. And so if if I'm going to leverage [24:56] this new technology, it's going to be through  those platforms. Now, if those platform providers [25:03] have worked with a startup or have integrated um  an app or an MCP or or something through them, [25:11] then we will use it. Um but I'm not I'm not going  to go out and just buy everything. And the other [25:20] reason is because I'm trying to integrate it into  our infrastructure. I don't want bolt-ons because [25:26] I want that decision velocity to happen. And the  only way you can get there is if you embed the AI [25:32] into your infrastructure stack. Um, so for us,  we buy a lot, but it also it's it's this is not [25:42] unfamiliar to anyone in this room. when you buy  something, you still have to tailor it to um your [25:47] institution, your data, your workflows, everything  else. So, um buying is still very much building, [25:53] but um but we do buy a lot through our partners.  Um maybe we talk about like trust and governance [26:01] and risk as you've mentioned. um like when an  AI agent makes a bad call that in the process [26:08] should like cost customers money or lead to  mistake like who's accountable how are you [26:14] guys structuring accountability for you know uh  these processes these agent enabled processes [26:23] we depend and we being like the um  my team of scientists we depend on [26:30] uh the controls because um results only matter  if they can be trusted results. Uh so governance [26:38] serves that bridge for for AI. Um and a concrete  example of that is our responsible AI framework. [26:48] It's got a taxonomy. Uh in that taxonomy, they  know exactly what the thresholds are. They know [26:54] exactly what the controls are, escalation pass,  who's accountable for which piece, and they can [27:00] move faster because the decision um doesn't have  to get relitigated at every stage. Um in a bank, [27:08] it's the ambiguity that's going to slow you  down. So, if you don't have all of that in place, [27:14] um you're you're going to be stuck in all of  those redecisions every time, like should we [27:21] be doing this pilot? should we be doing that? And  the there's only two sentences that are different [27:26] between the two things, but um you know, if you  get that standardized process in place, you can [27:31] help reiterate. Um I don't really think that the  organizations that win with AI are necessarily [27:39] going to be the fastest ones. Um I think it's  going to be the ones that can scale AI responsibly [27:48] um while maintaining that trust. I I think  that's that's going to be the end. Yeah. Spent [27:56] a lot of time on AI, but you also are in charge of  quantum uh which is exciting. Everyone's excited [28:01] about. Um you know, [ __ ] filed a provisional  patent on a quantum algorithm to help forecast [28:06] earthquakes. Um which is, you know, to me a  crazy thing for the bank to do. Um what's the [28:11] strategic logic and when does like quantum produce  anything that a CFO would recognize as as value? [28:17] Yeah. So, um we are we've partnered with IBM um  and we have access to their quantum computer. Uh [28:27] I've got a quantum research team and we're we  are working on uh several projects uh that add [28:35] business value to to Beimo. Those are going  to be several years down the road. We're not [28:41] not claiming that the business value comes  tomorrow by any stretch of the imagination. [28:46] Um but in our research everything that we  do is um based on a function. So um I'm [28:55] very passionate about earthquakes. Uh and over  uh 60% of our clients live in a fault tolerant [29:03] uh location. So it's very relevant to our client  base. Uh and then our our mission statement is to [29:11] boldly grow the good in business and in life. And  so if we can contribute to a scientific community [29:17] um to provide prediction for earthquakes uh well  in advance then you know that that's kind of like [29:24] the how or or the why is not so crazy. Um, but  with the with the earthquake prediction part, [29:31] you're dealing with a ton of data. And if your  bank or your financial institution deals with [29:39] um with capital markets, you're also dealing with  over a billion transactions a day. Um and so if I [29:46] can learn a way to ingest and time forecast um  all the data needed for earthquake prediction [29:55] then we can start doing the same u use the same  function for capital markets. And where would you [30:02] sort of describe where like quantum is today? you  know is it real but early or is like some of the [30:07] industries quantum positioning marketing some  combination of the two like what has been your [30:13] take on all the developments in recent months?  Yeah. Um so we believe we've reached uh as a a [30:20] scientific community quantum utility uh quantum  utility and quantum advantage. The difference [30:25] between those two phrases quantum utility is that  you can run a classical algorithm and you can get [30:32] inside of quantum and you can get a real answer.  Um so you can solve a real business problem [30:39] utilizing a quantum computer. Uh and and society  can do that now. Uh so that's a big step forward. [30:46] We weren't able to do that three years ago.  Um we being the scientific community. Then [30:54] uh quantum advantage is something we all hope for  and quantum advantage is when you become better, [30:59] smarter, faster than in a quantum environment than  doing it in a classical environment and that's the [31:07] game changer moment really um for everyone and  I think that will happen in the calendar year [31:16] um or even in the next 12 months um but that  doesn't mean it's going to be in production. [31:23] in a theater near you soon. Um, there's a lot more  that has to happen between that being achieved [31:30] once in a lab and being production ready for  institutions. And it's probably not going to be [31:36] first in banking. It'll probably be your molecular  science places. So, your chemical companies, [31:43] your pharmaceutical companies, biotech. Does  that help? Yeah, a little bit. Yeah, [laughter] a [31:49] little. [gasps] Um let's do one prediction for  the audience and then we'll open up to questions. [31:54] Like what will be true about AI and banking in 12  to 18 months that you know almost nobody in this [32:00] room sort of believes today? I think in the next  12 to 18 months people will realize that small [32:07] language models are the way to go. Um that that  we need to learn how to ask better questions. Um, [32:19] and so question models, large or small, will be  more popular in the 12 to 18 month time frame. But [32:26] goodness, this space is so hard to have a crystal  ball. Um, there's always something new happening, [32:32] but that's where I would go. I think small  language models are probably going to um be [32:36] very popular. We'll probably be on stage next year  talking about small language model success. Um, [32:42] large language models cost way too much money. Now  everybody's trying to make money with tokens. Um [32:48] a lot of us are losing money with tokens. Um and  and the value is five years in questionable. Um [32:57] so I I really do think we're going to shift to  small language models because we'll be able to [33:01] say ah you know we can run these a lot smarter and  cheaper and and and start measuring that business [33:08] value faster. Why do you think that hasn't gotten  in traction yet? just because of the speed or [33:14] because of the pace of development with the large  models. Well, the frontier model companies came [33:19] out with the big shabbam. I mean, they came  out with the large language models first. Um, [33:24] and now that a lot of them are open source um  modeling techniques, others have then created [33:30] the small language models. Um, so I think that's  it's just a the way they were developed. Um, [33:38] let's open up to questions. Raise your hands. We  have a microphone. Um, I guess we'll start. He's [33:43] walking that way, so we'll start there. It's  a deep breath. [laughter] Hey, thank you for [33:49] uh this uh session. I have a question related to  the recent executive order the president signed [33:54] regarding the quantum encryption mandate. Uh how  the enterprises are getting ready for this and [34:01] what's your take on that? This happened last week  or week before. Yeah. Yeah, I'm I'm aware of the [34:07] news, but I um I I don't pretend to understand um  what the why or the how is. Um so unfortunately, [34:15] as a scientist, I'm not able to really um I'm not  I'm not sure for me. I mean, there are very real [34:24] uh deadlines. Um so NIST um and other there's  other regul regulations, not what you just brought [34:30] up, but there's other regulations that exist  and all of us need to follow that. I think all [34:34] of us I'm I'm generalizing I have no idea what  your jobs are. Um but we have to follow those [34:41] standards in the next two years. Um so those are  the deadlines I'm focused on because they're very [34:47] clear regulated needs for postquantum encryption.  Um and I don't know how many people are taking [34:53] that seriously but we we all need to be. Kristen,  can you talk about what the So I work at a bank [34:59] so I can relate to a lot of what you just said.  Can you talk about the operating model between [35:04] the lines of business and the central team in  driving AI use cases forward, creating backlogs, [35:12] accountability, build? Yeah, our line of business  is key. Uh, and they they run the world. Um, [35:18] so our lines of business, they they choose where  they want their money spent. They know their [35:23] clients the best. Um, sometimes they even have  the own teammates who have the scientists and the [35:29] capability to build it themselves. Um so our lines  of businesses play a really vital role. Um my my [35:37] team handles big rocks. Um so core capabilities  that the whole firm needs. Um and and then the [35:44] lines of business can tailor those things once you  know once they're provided. Um and then my team is [35:50] doing the governance. And even with that with the  governance I kind of have built relationships with [35:57] the lines of business and then with that I then  say like please follow these rules and then the [36:05] lines of business can choose whether or not they  want to follow the rules. Um but they they all do [36:10] um at our firm that's that's part of our culture  but that's kind of the the difference. The P&Ls [36:15] run the world. Thanks Kristen. This was a great  session. Thank you. Yeah. Um, you talked about [36:21] decision velocity. Can you uh talk a little bit  more about where you see things going? Especially [36:30] at a bank, your decisions depend on how quickly  you can come up with insights. Um, and that's what [36:37] accelerates uh uh decisions. um in AI in agentic  AI uh what do you see happening in that space and [36:48] the evolution that's happening and where things  are going you talk a little bit about that sure um [36:54] so anyone's CEO asks for a forecast pro probably  at least monthly or quarterly right um and that [37:02] forecast process has to go out to every single  line of business each of the individual teams [37:07] the COS all come up with their forecast. It's  usually in an Excel file. It takes several weeks [37:15] uh or a week um of maybe 8 to 10 people. Like  everyone lived this experience. Yeah. Okay. Um [37:24] and so I'm just giving you an example, Sunny,  of like that's like when you want to forecast, [37:29] you have to go out to every single person who  reports to you. You have to go then their team has [37:33] to go to work on it and then you raise it back up.  you roll it up into some overarching forecast and [37:40] you present it back to the CEO. Um, and and that  is not embedded decision velocity. That's not. So, [37:49] how do we take the AI agents and embed it into  our um into our systems that have all of that [37:57] financial data and it can be extracting that  and formulating the forecasts without having [38:04] to compile all these. So that's that's how I want  that decision to be sped up uh in a very real way. [38:11] Um then you could you could use a different  example. You can use your Salesforce. Again, [38:16] I'm not um I'm not being a sales agent for for  anything, but it just happens to be one of the [38:21] systems we use. Uh for Salesforce, if you want  something, you have to go to Salesforce and get [38:27] it in each business line or, you know, sometimes  even within each department. Like what? Why do [38:35] I have to do that, right? Like I shouldn't have  to. And so the decision velocity could be sped up [38:42] um a whole lot more if if we embed um the those  technologies into the core infrastructure and come [38:49] up with a disciplined approach technologically  and business leadership wise to to harness that [38:56] velocity. Thanks everyone for coming. I  hope this was useful. Uh you know Chris [39:00] is an author. Please check out her books and you  know please come to Reuters Breaking Views for [39:05] incisive takes on what's happening in the world.  So thanks again. Yeah, thanks Ashley. [applause]