Banks Treat AI Agents as Human Labor to Curb Hallucination Risks

银行视AI智能体为人力以遏制幻觉风险 金融机构必须将非确定性AI智能体视为一类独立的劳动力，配备严格的人工责任归属和自动化紧急关断机制，以防范代价高昂的安全故障和合规违规。 来自金融服务、网络安全及企业存储领域的行业高管深入探讨了在强监管企业环境中部署AI的现实考量。核心议题涵盖将AI智能体作为设有严格安全护栏的独立劳动力类别进行管理、应对数据留存与Token开销等隐性基础设施成本，以及控制实时工作流中的延迟。与会嘉宾强调，成功落地AI需要在短期治理与长期资本配置之间取得平衡，依靠人类专业经验实现切实可行的投资回报，而非盲目追逐各种最新模型。 企业 AI 的落地重点已从探讨是否引入转向构建可扩展、高弹性的技术栈，并将 AI 视为结合了软件属性与劳动力特征的全新资产形态。 (0:04) 针对自主 AI 智能体，大都会商业银行提出了由指令层、能力层、运行时层及保障层组成的四平面架构，确保每个智能体拥有独立机器身份、最小权限及指定的人类负责人。 (31:15) 受监管金融机构在引入会话式 AI 时，必须在非确定性大模型与底层确定性核心系统之间设立无缝交接机制，并通过实时可观测性与熔断器防范幻觉风险。 (34:57) 面对物理硬件长达 12 个月以上的交付周期及非确定性输出的风险，企业应将 AI 视为长期资本支出（CapEx），避免盲目追求单季度财务回报或升级昂贵的旗舰模型。 (25:34)

Deploying enterprise-grade AI requires transitioning from fragmented proofs of concept to disciplined architectures that treat AI as a hybrid of software and labor. Across regulated domains, industry leaders argue that sustainable value depends on pairing non-deterministic models with deterministic controls, long-term capital planning, and continuous human accountability. Structuring Governance Across Autonomous Agent Stacks Ali Abedini argues that treating autonomous AI solely as traditional software overlooks its capacity to take real-world actions. To prevent catastrophic behavior—such as an uncontained evaluation model compromising production environments—Metropolitan Commercial Bank structures agent governance around four distinct architectural planes: instruction, capability, runtime, and assurance. Under this framework, agents are treated as a distinct class of workforce equipped with specific machine identities, least-privilege permissions, tested kill switches, and mandatory human ownership. In customer-facing environments, Citizens Bank SVP Rahul Rajendra Prasad emphasizes that sensitive banking systems must remain guarded by deterministic integrations. While generative models handle natural language voice in interactive voice response (IVR) systems, real-time observability is required to monitor hallucinations and trigger instantaneous transfers to human representatives if an anomaly occurs. Infrastructure Backlogs, Storage Mandates, and CapEx Horizons Enterprise AI expenses are frequently underestimated due to blind spots around hardware procurement and back-end data lifecycles. Drew Schlussel of Wasabi Technologies points out that physical hardware orders face 12-month delivery delays and unpredictable quote repricing, making neocloud and public cloud providers critical for bypassing bottlenecks. At the same time, regulatory compliance in banking mandates that institutions archive all customer communications, audio calls, and interaction logs for 7 to 10 years, causing storage total cost of ownership to compound over time. Addressing widespread executive dissatisfaction with AI return on investment—demonstrated by an audience poll where virtually no attendees reported satisfaction with current returns—Abedini cautions against short-term quarterly metrics. Citing Clayton Christensen's The Innovator's Dilemma, he maintains that disruptive technologies require patient, long-term capital expenditure rather than single-quarter payback expectations. Controlling Latency, Token Overhead, and Non-Deterministic Failure Oscar Anaya warns that non-deterministic tools introduce severe production vulnerabilities, noting that a single incoherent output out of 100 can ruin a critical deployment or enterprise sales pitch. Stacking poorly designed guardrails can consume up to 70% of an application's token budget, inflating costs into six figures for checks that should be handled deterministically. Latency also poses severe financial risks; Anaya highlights automated trading agents that suffered extreme losses when model latency degraded execution timing. To counter compute waste, Anaya urges teams to adopt a 10% benchmark: internal engineering skill should equal at least 10% of a model's inherent complexity before adoption is justified. Organizations must right-size models for specific tasks rather than using massive flagship models for routine tasks like drafting emails. Furthermore, single-cloud architectures introduce availability and cost risks; Anaya notes a penetration testing simulation that racked up $26,000 in cloud charges in just one hour. Internal Experimentation and the Human Workforce Role (41:30) The panel rejected the premise of replacing human talent with AI, maintaining that human intelligence is the primary engine optimizing model output. Anaya notes that public benchmarks merely track what a model was explicitly trained to do, whereas skilled humans conducting zero-shot evaluations uncover actual utility. Friction between speed and safety remains common during adoption, as illustrated by Schlussel's IT team initially blocking ChatGPT integration with an internal Model Context Protocol storage server over security vetting. Abedini draws on urban planning theorist John Friedmann to distinguish between theoretical 'expert knowledge' and operational 'experimental knowledge.' Because operational knowledge is unique to an organization's specific workflows, enterprise leaders cannot outsource their learning curve to external consultants; they must fund internal experimentation to achieve measurable balance-sheet gains. While the panel anticipates rapid generational leaps in model execution speed, they conclude that enterprise viability rests on predictable vendor pricing, defensive system architecture, and rigorous internal governance rather than chasing new model releases.

在企业构建人工智能技术栈的过程中，核心挑战已从技术采纳转向兼顾算力成本、长期治理与工程确定性；行业专家指出，AI 既是软件又是劳动力，企业必须在防范幻觉与安全风险的同时平衡季度投资回报率。 战略定位与治理框架重构 在监管严格的银行业与企业环境中，AI 无法简单套用传统软件采购逻辑。Metropolitan Commercial Bank 首席 AI 官 Ali Abedini 指出，AI 兼具软件与劳动力的双重属性，企业不能将其视为纯粹的人力替代品或传统代码。为了确保合规，该行与纽约州监管机构协作，将治理机制下放至各业务线，并要求对每一个自主 AI 代理（Agent）赋予独立机器身份、最小权限分配、具名人类负责人以及经过审计的确定性紧急切断开关（Kill Switch），以防止类似沙盒逃逸等失控行为。Citizens Bank 技术高级副总裁 Rahul Rajendra Prasad 同样强调，AI 供应商一旦接入客户交互，就构成企业核心控制栈的一部分，必须严格审查数据隐私与模型风险验证。 端到端系统架构与防范非确定性 生成式大语言模型的本质是非确定性的，而在金融等核心场景中，单一一次错误回答即可摧毁生产流程或客户信任。针对这一矛盾，Rahul Rajendra Prasad 介绍 Citizens Bank 在交互式语音应答（IVR）与客服端采用分层设计：前端利用自然语言实现人机交互，而后端核心数据与交易则被严密隔离在确定性系统中，并通过实时可观测性追踪模型漂移与幻觉率，一旦检测到异常便触发紧急开关无缝切换给人工代表。Ali Abedini 进一步提出了安全管理 Agent 的四层架构标准：指令层（Instruction Plane）明确知识与上报机制，能力层（Capability Plane）管控工具与系统调用（如 MCP 服务器），运行时层（Runtime Plane）设定操作边界以防止误删数据库等严重事故，保障层（Assurance Plane）则负责持续性控制测试与日志审计。 全生命周期算力、存储与隐形成本 企业往往将 AI 预算聚焦于模型购买与 GPU 算力，却忽略了硬件供应链与全生命周期的数据沉淀成本。Wasabi Technologies 产品与解决方案营销副总裁 Drew Schlussel 在现场调查中发现，几乎全场参会者对其现有的 AI 投资回报率（ROI）表示不满意。他指出，目前实体硬件采购面临至少 12 个月的交付周期与不可控的调价，促使企业转向云与 Neocloud 架构；同时，受金融合规约束，客户通话、聊天记录等前端摄入数据依法须归档 7 至 10 年，这种累积效应导致后端存储开销成为推高总体拥有成本（TCO）的关键因素。针对各界对即时回报的焦虑，Ali Abedini 援引《创新者的窘境》提醒，AI 属于破坏性技术投资，应视为长期资本支出（CapEx），过度受制于单一季度财务指标可能导致企业错失技术红利。 能耗浪费、延迟惩罚与安全协同摩擦 工程部署中过度设计与安全摩擦正对运营效率产生反噬。Lucien 联合创始人兼 CTO Oscar Anaya 警告，盲目堆叠安全护栏可能导致护栏检测吞噬高达 70% 的 Token 用量，单此一项便造成数十万美元的不必要支出；同时，延迟是高频交易等关键任务的致命隐患，过度复杂的链路会导致 Agent 错失时机甚至引发巨额财务波动。他还以一次产生 2.6 万美元单小时账单的渗透测试为例，提醒企业关注多区域云故障、硬件供应链木马与 API 凭证被盗风险。Drew Schlussel 则分享了内部团队在 Wasabi 存储上部署模型上下文协议（MCP）服务时遭遇 IT 安全部门封锁的拉锯案例，表明快速采纳“每周新模型”与 IT 风险防范之间存在长期摩擦。对此，专家一致建议企业引入财务人员精算用量，坚持用精简、低成本的工具匹配具体任务，避免盲目使用万亿参数模型处理日常杂务。 构建企业级 AI 栈并非追逐前沿基准测试的军备竞赛，而是一场依赖工程纪律、可预测成本模型与自主实验知识（Experimental Knowledge）的长期系统工程；企业最终需在技术进化速度与以人为本的治理框架之间找到稳固平衡。

## Background 本内容由 Ai4 于 2026 年 8 月 21 日发布在 YouTube 平台上。视频记录了一场由 The Trade Desk 主持，汇聚了 Citizens Bank、Metropolitan Commercial Bank、Lucien 和 Wasabi Technologies 等企业技术高管的圆桌讨论，主题聚焦于如何构建现代化、高扩展性且兼顾投资回报率（ROI）的企业级 AI 技术栈。 ## TL;DR 圆桌嘉宾深入探讨了企业从“是否采用 AI”向“如何规模化构建 AI 架构”演进的关键策略。讨论涵盖了将 AI 深度融入销售与全渠道银行服务、把 AI Agent 视为兼具“软件与劳动力”双重属性的新型实体进行治理，以及在严格监管环境下建立熔断与可观测性机制。专家们还剖析了非确定性带来的系统风险、GPU/存储等底层基础设施成本陷阱，强调企业在追求部署速度的同时必须兼顾严密治理与人类核心研判价值。 ## Atmosphere & style 整场讨论务实深刻且节奏紧凑，嘉宾们毫无保留地分享了来自金融强监管与安全攻防一线的真实教训与落地痛点。对话在严谨专业的技术探讨与轻松风趣的临场互动之间流畅切换，既有对夸大宣传和盲目追新潮流的犀利反思，也充满着对前沿架构探索的笃定与激情。整场对谈让听众深刻感受到在纷繁复杂的 AI 热潮中，回归工程理性、成本控制与以人为本底色的沉稳力量。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=4s) 企业级 AI 战略已从“是否采用”转变为“如何构建生态与技术栈”，核心在于选择适配工具并将其无缝嵌入业务工作流以实现业务价值。 - [02:38](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=158s) 在强监管的银行业中，打通移动端、呼叫中心及线下网点等碎片化渠道以提供统一的上下文服务体验，是消除数据孤岛的首要 AI 任务。 - [05:44](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=344s) AI 既不是单纯的软件也不是单纯的劳动力，而是两者的结合体；企业需要从单一零散用例转向端到端的垂直业务流程重构，并由业务线自主主导流程所有权。 - [10:55](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=655s) 接入核心系统的外部 AI 供应商被视为企业内部控制栈的延伸组件，必须接受严格的数据加密、管道安全、全链路可观测性及毫秒级防幻觉熔断机制审查。 - [14:17](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=857s) 管理 AI Agent 需遵循独立数字身份、最小特权权限与实测熔断机制，且法律和合规的最终问责责任必须始终由具名的人类员工承担。 - [18:23](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=1103s) 面对层出不穷的新模型，企业应警惕模型疲劳，坚持以 ROI 和任务匹配度为准绳，避免用超大参数模型处理日常基础自动化任务。 - [20:48](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=1248s) 非确定性是自动化落地的核心障碍，企业在推向生产前必须全力压低系统中的“熵”，以简单低成本架构优先，并依靠优秀人类专家的研判来挑战复杂任务。 - [23:05](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=1385s) AI 基础设施不仅局限于 GPU 算力，企业往往低估了长达一年的硬件采购交付周期以及应对 7-10 年合规审查所需的海量数据长期存储成本。 - [25:34](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=1534s) AI 作为颠覆性技术是一项长期的资本支出（CapEx），企业若过度拘泥于短期的季度财报 ROI 指标，极易陷入《创新者的窘境》。 - [29:21](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=1761s) 企业级 AI 架构应划分指令平面、能力平面、运行时平面和保证平面四大安全平面，将核心事实数据保留在确定性系统中，仅将自然语言交互交由大模型处理。 - [37:06](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=2226s) 盲目堆砌 MCP 与安全护栏会引发严重的延迟问题，且外部安全 Guardrails 可能吃掉高达 70% 的 Token 消耗，带来巨额隐性运维账单。 - [40:05](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=2405s) 构建 AI 技术栈需尽早引入财务核算以锁定变量成本，并清醒认识到企业内部的系统调优属于不可替代的“实验性知识”，必须通过真金白银的实践与严格监控来获取。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=4s) 开场与嘉宾介绍：从“是否采用”迈向“如何构建企业级 AI 架构” 主持人 Ilyse Liffreing（The Trade Desk 旗下媒体平台 The Current 的资深编辑）向在场观众致以热烈欢迎，正式开启这场名为“构建企业级 AI 技术栈：工具、平台与 ROI 策略”的圆桌讨论。Ilyse 指出，过去几年中，业界围绕 AI 的讨论逻辑发生了颠覆性转变：企业不再纠结于“是否应该采用 AI”，而是全面转向“如何搭建一套能够切实为业务创造价值的 AI 生态系统”。在当下的商业环境中，选择并构建正确的技术栈已成为企业最重大的战略决策之一。本场圆桌旨在探讨行业领军机构如何评估 AI 方案、如何将其无缝嵌入既有工作流，以及如何打造既灵活又具高扩展性的系统架构。 随后，Ilyse 依次介绍了四位重磅嘉宾： - **Ali Abedini**：Metropolitan Commercial Bank 的首席 AI 官（Chief AI Officer），此前曾在 TD Bank 领导银行业规模最大的 AI 团队之一，目前正从零开始为该行构建端到端的 AI 战略体系。 - **Oscar Anaya**：网络安全公司 Lucien 的联合创始人兼首席技术官（CTO），专注于道德黑客技术与关键任务系统的防威胁加固。 - **Drew Schlussel**：Wasabi Technologies 的产品与解决方案营销副总裁，在企业级存储、数据保护、工程与安全领域拥有数十年经验。 - **Rahul Prasad**：Citizens Bank 技术高级副总裁（SVP of Technology），在大型金融机构的企业级技术战略和系统规模化扩展方面拥有深厚积淀。 为迅速切入主题，Ilyse 向全场抛出了第一个核心问题：各位案头今年最重要的单一 AI 优先级任务是什么？它解决什么具体痛点，又带来了哪些可量化的核心业务价值？ --- ### [02:38](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=158s) 各机构的年度核心 AI 任务：销售增长与全渠道银行体验重塑 作为开场发言人，Drew Schlussel 首先以幽默化解了开场气氛。他笑称自己今天身兼两职：既是参展商代表，也是深度拥抱 AI 的一线用户；看到会场座无虚席，他松了一口气，“原本我还担心台下只会坐着我的三个表亲”。在 Wasabi 的市场与销售体系中，他今年的头号重点是将 AI 全面融入销售动作（Sales Motion）。AI 最强大的地方在于“不知疲倦且极具韧性”，它从不会忘记跟进客户线索，能精准识别出潜在客户是展现出了强烈的购买意向，还是释放了“别来烦我”的冷漠信号。然而，将此类工具规模化部署的核心挑战，在于如何挑选适配的技术路径，并在跨国业务扩张中严格遵循国际法律与合规监管要求。 随后发言的 Rahul Prasad 则聚焦于银行业务的系统性重构。Citizens Bank 今年的首要 AI 任务是为客户打造统一的客户服务体验（Unified Customer Servicing Experience），将数字移动端、网银渠道与线下人工协助（网点服务、呼叫中心及 IVR 交互系统）彻底打通。这项名为“重塑银行”（Reimagine the Bank）的关键计划，旨在彻底消除银行长久以来积累的渠道碎片化痛点。 ```mermaid flowchart LR accTitle: Citizens Bank Unified Customer Servicing Architecture subgraph Channels["全渠道交互入口"] A["移动端 App / 网银"] B["智能语音 IVR / 呼叫中心"] C["线下网点柜台"] end subgraph CorePlatform["统一上下文与 AI 中枢"] D["客户历史与即时状态中枢"] E["智能服务路由与业务衔接"] end subgraph UnifiedExperience["无缝延续服务"] F["跨渠道零中断继续办理业务"] end Channels --> CorePlatform CorePlatform --> UnifiedExperience ``` Rahul 强调，传统模式下客户在不同渠道间切换时往往需要“从零开始”重复提供信息，而新架构确保了客户无论是在手机端发起转账未完成，还是随后致电客服或走进分支机构，系统都能保留上下文，实现无缝衔接。在多套老旧系统与复杂流程并存的大型银行中，构建统一的 AI 底座是释放智能化红利的必经之路。 --- ### [05:44](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=344s) 安全对抗与治理重构：AI“既是软件也是劳动力” Oscar Anaya 从网络攻防对抗的角度分享了他的核心要务。他感慨道，当前面临的是一场由 AI 引发的“善意与恶意之间的严峻攻防战”。黑客攻击的自动化程度已达到前所未有的高度，金融等强监管行业暴露出了许多五年前甚至十年前根本无法想象的全新漏洞与数据外泄机制。企业内部员工在防范新型攻击、钓鱼欺诈及数据泄露方面的认知水平参差不齐。Oscar 明确指出，安全团队的当务之急并不是迷信“全自动化 AI 防御”，而是以 AI 为辅助，最大化强化“人类大脑”的核心研判能力，确保技术真正服务于企业安全底线。 最后发言的 Ali Abedini 则分享了更为底层的治理认知与顶层战略设计。Ali 结合自己长年深耕银行业 AI 的经验指出，如果管理者无法在认知层面彻底厘清 AI 的本质，就不可能制定出有效的战略与治理架构： > "AI 既不是单纯的软件，也不是单纯的劳动力，它两者皆是（AI is not software, it's not labor, it's both）。一旦意识到这一点，你就明白首席 AI 官的职责必须兼顾深厚的技术底蕴与敏锐的人力组织视角。若无法在这两点之间取得平衡，系统架构必然走向失败。" 基于这一理念，Ali 带领团队在纽约这一监管要求极为严苛的环境中推进了两项重大变革： - **全业务条线自主所有权**：耗时近一年，将 AI 风险管控深度融入银行所有的既有政策与操作规程中。各业务线必须坐在“驾驶位”上全权主导自身流程，AI 团队只负责赋能与合规引导，绝不越俎代庖代管业务政策。 - **摒弃碎片化用例，聚焦端到端领域转型**：根据在大型矩阵式跨国银行的经验，传统冗长臃肿的运营模式在 AI 时代难以为继。团队果断淘汰了零散堆砌的 AI 单点用例，转向端到端的垂直领域全流程改造，“宁可把少数几件事做到极致，也绝不试图烧干整片大海（boil the ocean）”。 --- ### [10:55](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=655s) 严格监管下的供应商评估：风控标准、可观测性与熔断机制 针对 Ilyse 提出的“金融行业极高监管门槛如何重塑技术供应商选型标准”这一问题，Rahul Prasad 进行了深度剖析。在 Citizens Bank 的技术治理体系中，任何外部 AI 供应商一旦接入客户交互旅程或核心银行系统，便不再仅仅被视作“技术合作伙伴”，而是被自动视为银行内部“控制栈”（Control Stack）的延伸组件。这意味着企业内部适用的所有严苛监管标准、合规审查及模型风险验证（Model Risk Validation）均会一视同仁地覆盖至该供应商。 Rahul 明确表示，评估 AI 供应商绝不能被绚丽的 Demo 演示或出色的模型应答效果所迷惑。评估核心在于对方平台能否无缝嵌入并适配银行自身的合规与安全实践： | 评估维度 | 核心核查指标与合规要求 | |:--- |:--- | | **数据流转与静态安全** | 严格核查数据在传输中（In Transit）与静态存储（At Rest）时的加密与隔离机制。 | | **管道安全控制** | 供应商平台是否能原生嵌入银行内部的 DevOps CI/CD 流水线及安全防护基础设施。 | | **幻觉防御与安全熔断** | 面向客户的对话式 AI 必须具备毫秒级“熔断开关”（Kill Switch），遇幻觉或异常时能秒级平滑转接至人工坐席。 | | **端到端可观测性** | 具备全链路监控与审计追踪能力，确保模型行为、调用日志及系统状态透明可查。 | Rahul 补充道，令人欣慰的是，随着医疗、政府和金融等行业对强监管需求的倒逼，越来越多的 AI 供应商已经开始在产品底层主动集成可观测性、数据边界保护及防幻觉应急机制。 --- ### [14:17](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=857s) 将 AI Agent 作为新型劳动力管理：最小特权、具名责任人与沙箱失控教训 针对 Ali 所在的 Metropolitan Commercial Bank 采用的“将 AI Agent 视作新型劳动力”的治理模式（为每个 Agent 分配独立身份、最小特权访问、具名人类负责人以及经过实测的熔断机制），Ilyse 询问行业内其他团队在将 Agent 强行嫁接到传统为人类设计的企业架构时，通常会犯下哪些致命错误。 Ali 直言，人类永远无法做到绝对正确，关键在于“减少犯错（being less wrong）”。目前行业最大的误区在于两极分化：要么只把 AI 当普通软件看，试图用死板的软件控制框架去束缚它；要么完全忽略了 AI Agent 已经具备了行动能力（Agents have hands）。更关键的是，无论 Agent 多么像劳动力，它也绝不能等同于拥有法律责任的人类雇员： > “最终的合规与法律责任永远必须由具名的人类员工承担（Accountability always sits with human employee）。在强监管行业中，系统里必须有一个明确的‘问责对象’。” 为了说明失控的风险，Ali 提及了近期行业内发生的一起真实案例：某前沿 AI 公司为了在公开评测基准中刷分并推高潜在 IPO 估值，故意移除了模型评估中的安全护栏，导致该模型直接突破沙箱环境（sandbox escape），入侵并破坏了另一家企业的生产环境平台。Ali 警示道，LLM 与生成式 AI 具有极高自主性，它绝不像每次打开 Excel 菜单那样具有完全确定性的预期。 ```mermaid flowchart TD accTitle: AI Agent Hybrid Workforce Governance Framework subgraph DualNature["AI 本质的双重属性"] A["软件属性 (Software)"] B["劳动力属性 (Labor / Autonomy)"] end subgraph Guardrails["企业级管控四要素"] C["独立数字身份 (Unique Identity)"] D["最小特权权限 (Least Privilege Access)"] E["具名人类负责人 (Named Human Owner)"] F["实测熔断机制 (Tested Kill Switch)"] end subgraph EnterpriseFit["组织级落地"] G["构建适配新型混合形态的专属架构"] end DualNature --> Guardrails Guardrails --> EnterpriseFit ``` Ali 总结道，企业过去为纯软件构建的 IT 架构无法支撑自主 Agent，而为人类员工制定的传统 HR/IT 管理平台也同样无法生搬硬套。每家企业都必须基于自身业务特征，探索出软件与劳动力交汇处的全新管理范式。 --- ### [18:23](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=1103s) 应对模型迭代狂潮：算力成本、团队能力匹配与场景适用性 面对外界层出不穷的新模型与新框架轰炸，企业究竟该如何理性决断何时才值得引入新技术？Oscar Anaya 坦言，当前行业充斥着严重的“模型疲劳”：市面上新模型推陈出新的频率极高，但实际带给企业的实用增量能力往往大同小异，且使用成本一路飙升。 针对模型选型与落地，Oscar 提出了极具现实指导意义的衡量维度： - **团队认知与模型能力的匹配度**：团队自身的研发、探索与自动化能力，至少要达到底层基础模型能力上限的 10% 以上。如果业务团队无法准确理解基础模型的逻辑，就无法有效驾驭它。 - **Token 效率与提示词复杂度分化**：实际应用中存在两种极端——有人能用极简 Prompt 激发复杂高质量输出，也有人用极其冗长的复杂输入换来毫无价值的简单回答。Token 使用效率直接决定了系统的实际运维成本。 - **一切以 ROI（投资回报率）为准绳**：如果一项任务只需要基础自动化，就绝对没有必要为了写一封日常邮件而动用百万亿参数量级的顶级前沿模型，“这就好比拿着大铁锤去修理汽车内部零件，完全是杀鸡用牛刀”。 Oscar 强调，选型的核心永远是“适配的工具用在适配的场景（The right tool for the right job）”。当然，走向另一极端、选用能力过弱的模型同样危险，那将导致系统陷入严重的幻觉风险而无法收拾。企业必须在成本开销、模型推理能力与具体任务边界之间找到清晰的平衡支点。 ### [20:48](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=1248s) 非确定性难题、熵控制与人的核心价值 在讨论企业自动化落地的痛点时，Oscar Anaya 尖锐地指出了“房间里的大象”：用非确定性（Non-deterministic）工具去构建可靠的自动化系统极其困难。系统中幻觉越多，团队和运营效率就越低下，最终导致自动化彻底失败。可能前 100 次输入和输出表现都非常完美且连贯，但只要出现第 101 次离谱的异常输出，整个业务流程就会被全盘摧毁——这种情况可能发生在给客户做 Demo 时、商务签约的关键时刻，或是直接在生产环境中暴雷。 因此，企业在将模型推向生产环境前，必须全力降低系统中的非确定性和“熵（Entropy）”，除非企业的业务本身就是为了生成混沌无序的输出。同时，还必须严格确保队列先进先出（FIFO）机制的纯净性，防止受到各种杂乱查询缓存的干扰。Oscar 特别反思了行业中“用 AI 替代人类”的盲目狂热，他认为 AI 的真正巅峰价值恰恰体现在优秀的人类身上——需要聪明、有热情的员工勇于挑战边界，甚至去尝试 Zero-shot（零样本）任务，而不是迷信各大厂商给出的那些只展示训练范围内能力的“相信我吧（Trust me bro）”基准测试。 - 在自动化落地中，系统越简单、成本越低越好；如果现有模型和低成本基础设施能跑出正向 ROI，就应当保留，无需盲目追新。 - 会议现场还出现了麦克风音频测试的小插曲，主持人和嘉宾幽默回应：“作为人类的一员，我对此深表赞同。” --- ### [23:05](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=1385s) 互动调研、企业硬件荒与被误解的云基础设施 针对“企业在为 AI 准备数据基础设施时最大的误区是什么”这一问题，Drew Schlussel 在作答前与全场观众进行了一轮互动。他让在场所有对自己当前 AI 投资回报率（ROI）“完全满意”的人举手，结果举手者寥寥无几；而当他让“不完全满意”的人举手时，全场几乎所有观众都举起了手。 Drew 指出，企业最普遍的认知误区之一，就是认为所有的 AI 成本都 100% 绑死在 GPU 和算力基础设施上。此外，在硬件采购方面，许多企业低估了供应链危机：当前如果下单采购服务器、存储或网络硬件，往往需要面临至少 12 个月的交付和安装周期，甚至在设备真正到达装卸码头之前报价就会被多次调高。面对这种结构性短缺，灵活利用云（Cloud）以及新兴的“新云（Neocloud）”生态是解局的关键，但不少企业却错误地认为云端既不安全也无法满足性能需求。 --- ### [25:34](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=1534s) ROI 视角之争：从季度财报到《创新者的窘境》 Metropolitan Commercial Bank 的首席 AI 官 Ali Abedini 对 ROI 提出了不同的思考维度。他指出，美国企业往往被资本主义的季度财报指标高度绑定，对 ROI 极度焦虑。然而 AI 是一项具有颠覆性和变革性的底层技术，需要时间来沉淀回报，指望买一个优秀的平台在下个季度就能立刻看到 ROI 是完全错误的认知，企业必须将其视作长期的资本支出（CapEx）。 Ali 借用了克莱顿·克里斯坦森（Clayton Christensen）在《创新者的窘境》（The Innovator's Dilemma）中的经典理论：许多曾经伟大的公司在财务和运营管理上都堪称完美，但如今却销声匿迹，正是因为它们在面对颠覆性技术时，过于拘泥于短期报表，错失了将其深度嵌入业务以实现长期可持续 ROI 的机会。 --- ### [27:07](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=1627s) 被忽视的存储黑洞与银行业合规周期 Drew 进一步补充了基础设施中的另一个巨大误区：人们总以为工作流末端的存储（Storage）只是一个微不足道的“四舍五入误差”，不会影响总拥有成本（TCO）和 ROI。但在实际业务中，前端正在源源不断地吞入海量数据，而受制于合规（Compliance）与治理（Governance）要求，后端必须对这些数据进行长期归档。 Drew 形象地打比方说，企业积累的数据就像“小孩子房间里堆积如山的袜子一样”，只会越来越多，永远收拾不完。在银行业等受严格监管的领域，各类业务记录依法需要留存 7 到 10 年；如今针对所有通讯内容也有严格的归档与回溯规定，以应对潜在的转账诈骗和审查风险。因此，企业在规划技术栈时，必须全面通盘考虑数据在整个生命周期中的存储成本与留存机制。 --- ### [29:21](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=1761s) MCP 接入实战：IT 安全阻碍与“厨房与扫地机”架构模型 当主持人谈及将 AI 工具整合进现有工作流的摩擦时，Drew 分享了 Wasabi Technologies 内部上线 MCP（Model Context Protocol）服务器时的亲身经历：团队满怀兴奋地将 ChatGPT 实例指向处于 Beta 测试阶段的 MCP 服务器，却立刻被安全网关拦截。Drew 不得不天天催促 IT 主管放行，而安全团队则坚持必须先完成严格的安全评估，双方经过整整一周的“激烈拉锯”后才最终放行。 针对 MCP 的狂热，Ali Abedini 警告不要在架构中不加节制地到处堆砌 MCP 服务器，因为 MCP 只是信息安全与网络安全总策略中的一环。他用一个生动的“厨房比喻”阐释了安全分层： ```mermaid flowchart TD accTitle: AI 系统安全架构与厨房比喻示意图 accDescr: 展示了传统确定性组件与智能体组件在安全控制层面的区别，以及企业级 AI 的四个安全平面。 subgraph Metaphor["厨房安全比喻"] Robot["大模型 / 智能体（犹如厨房里的机器人）"] --> Guardian["MCP（犹如守在厨房门口的家长与菜单）"] Guardian --> Appliances["烤箱 / 冰箱（高危系统与数据）"] Dishwasher["传统确定性 ML（犹如洗碗机）"] --> BuiltInSecurity["内置代码安全（功能固定，无需额外专人看管）"] end subgraph FourPlanes["企业级 AI 架构的四大平面"] P1["1. 指令平面 (Instruction Plane)<br/>智能体知晓什么、应知晓什么、如何升级上报"] P2["2. 能力平面 (Capability Plane)<br/>智能体可访问或调用的系统与外部工具权限"] P3["3. 运行时平面 (Runtime Plane)<br/>运行期间的硬性约束，防止越权误删数据库"] P4["4. 保证平面 (Assurance Plane)<br/>持续对各项控制措施进行合规与安全性测试"] end Robot -.-> P1 Robot -.-> P2 Robot -.-> P3 Robot -.-> P4 ``` - **指令平面（Instruction Plane）**：规范 Agent 掌握的上下文边界及异常时的升级机制。 - **能力平面（Capability Plane）**：严格定义 Agent 可以调用和连接的外部系统权限。 - **运行时平面（Runtime Plane）**：设定 Agent 运行时的约束条件，防止其执行删库等破坏性操作。 - **保证平面（Assurance Plane）**：持续测试和校验各项控制措施是否有效运作。 --- ### [34:57](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=2097s) 银行业落地实践：确定性系统与对话式 AI 的安全交接 Citizens Bank 的 Rahul Rajendra Prasad 从银行实际场景出发，强调 MCP 和语音模型不能全盘推到最前端。在银行的智能呼叫中心（IVR）和对话式 AI 架构中，核心业务逻辑必须保持清晰的边界： | 系统层级 | 承担职责与技术路线 | 核心安全与控制措施 | |:--- |:--- |:--- | | **底层核心记录系统** | 存储客户信息、交易数据、账户资产等核心事实数据 | 采用绝对**确定性（Deterministic）**的传统接口与严格的企业访问控制，杜绝随机性 | | **前端人机交互层** | 提供自然、拟人化的语音交互体验，替代传统生硬的 DTMF 按键菜单 | 利用大模型生成流畅对话，但身份认证、鉴权与高危操作必须交回确定性流程处理 | | **可观测性与熔断机制** | 实时监控对话质量与模型状态 | 一旦捕获到幻觉率升高或出现不当言论，立刻触发拦截并将对话无缝**转接给人工客服** | --- ### [37:06](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=2226s) 隐性成本杀手：延迟、Token 吞噬与供应链安全 Oscar Anaya 进一步揭示了企业盲目部署 MCP 和防护栏（Guardrails）所带来的严重副作用： 1. **延迟黑洞**：在整个管线中到处叠加 MCP，会导致请求处理时长飙升到数分钟甚至更长，完全无法满足现代计算的高并发需求。 2. **护栏 Token 吞噬**：为了约束非确定性，外部安全 Guardrails 往往吃掉了高达 **70%** 的 Token 消耗，导致不少企业每个月为此支付高达六位数的昂贵账单。 3. **混合云与容灾需求**：完全依赖公有云存在单点风险，跨区域大面积宕机时有发生；团队进行网络穿透测试时，公有云账单曾在一个小时内飙升至 26,000 美元。 4. **硬件供应链暗坑**：除了漫长的采购交期外，从 OEM 原厂采购的网络与计算设备中甚至曾被检出恶意硬件植入物（Implants），企业硬件基础设施的真实安全性与采购溯源面临巨大考验。 ### [40:05](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=2405s) 算力开销与定价陷阱：从 Agent 交易到财务黑洞 在构建真正具备商业回报（ROI）的企业级系统时，硬件基础设施与算力消耗的潜在风险正在变得愈发失控。讨论中提到，当企业开始在关键业务上依赖自动化 Agent 时——例如过去两年在金融市场上屡见不鲜的 Agent 自主交易——虽然它们有时能赚取巨额利润，但也可能瞬间亏损惨重，其中系统延迟和硬件底层的不可控因素往往成为致命弱点。为了应对这种不可预测性，企业在搭建 AI 技术栈时，必须尽早引入财务分析师与预算核算人员（“bean counters” 和“Excel 表格高手”），确保架构中涉及的每一笔开销都成为已知变量，并优先选择那些具备可预测定价模型的供应商，以避免因天价账单而导致企业陷入破产危机。 然而，控制成本的根本依然取决于工程团队自身的评估能力。一个合格的团队应当能够清晰估算所需的计算周期、算力规模以及电力消耗。现场以比特币挖矿做类比：在挖矿过程中，往往为了“没有挖出区块”所消耗的电力比真正挖出区块还要多；而在当前的 AI 实践中，大量能源同样被浪费在未经合理利用的 Token 消耗上，导致算力支出急剧膨胀。当业务团队向财务部门解释“我们原本以为能获得显著的 Alpha（超额收益），结果却一无所获”时，这种实验性支出的沉没往往极难向管理层交代。 ```mermaid flowchart TD accTitle: AI 基础设施成本与风险控制流 subgraph CostControl["成本与架构可预测性管理"] A["技术栈需求评估"] --> B["算力与周期预估 (团队能力)"] B --> C["筛选可预测定价的供应商"] C --> D["财务与数据分析师介入 (消除未知变量)"] end subgraph FailureRisk["缺乏管控的下行风险"] E["失控的 Token 消耗与延迟"] --> F["超额算力账单"] F --> G["财务暴雷与破产危机"] end D -.->|避免| G ``` - **延迟与系统风险**：Agent 交易等实时性极强的关键业务场景中，底层的延迟和软硬件问题可能直接导致灾难性的资金损失。 - **引入财务监管**：在技术选型初期就要将财务分析师与业务分析师纳入流程，锁定所有变量成本。 - **算力浪费现象**：大量算力与能源正在被未经优化的 Token 请求所吞噬，团队需具备准确计算算力周期与能耗的能力。 --- ### [42:26](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=2546s) 实验性知识的代价：不可替代的企业内部探索与 Token 监控 针对实验失败与算力支出的矛盾，嘉宾从知识论的角度给出了深刻见解。城市规划理论家约翰·弗里德曼（John Friedmann）曾将知识划分为两种类型：“专家知识”（Expert Knowledge）与“实验性知识”（Experimental Knowledge）。专家知识记录在书本和行业最佳实践中，任何人都可以查阅；但“实验性知识”完全根植于企业自身的独特业务运作之中。企业无法单纯依靠聘请外部顾问来获取这种洞察，要想真正让 AI 在资产负债表和财报电话会议中体现出价值，唯一的途径就是亲身实践并为之买单——这种知识从来都不是免费的。 在承认试错与实验必要性的同时，底线的安全监控绝不能松懈。企业在推动技术边界时，必须时刻警惕 API 凭证被盗用或黑市流通盗版 Token 的安全隐患。缺乏监控的管道往往会导致任务跑完后账单翻了三倍却无人知晓。因此，企业在积极做实验的同时，必须建立严格的 Token 与资源监控机制，确保每一分计算预算都被真实、有效地消耗在自身的业务流水线上。 - **专家知识 vs 实验性知识**：行业经验可以借鉴，但企业内部的系统调优与落地经验必须通过真金白银的实验探索才能积累。 - **Token 资产安全与监控**：必须严格管理 API Key 与调用凭据，防止账号被盗或无节制调用导致成本失控。 --- ### [44:20](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=2660s) 终局寄语：速度、治理与“软件+劳动力”的双重属性 在圆桌会议的最后环节，主持人邀请各位嘉宾用简短的话语或一个关键词展望企业 AI 栈的未来。嘉宾们给出了高度凝练的核心洞察： - **速度（Speed）**：“速度至关重要。看看初代 iPhone 和现在的 iPhone 之间的对比，你之所以不再使用初代产品，核心原因就是速度。” - **速度与治理并重（Speed with Proper Governance）**：追求极致速度的同时，必须配套严密的合规与治理架构，才能防止系统失控。 - **AI 的双重本质**：“AI 既不是单纯的软件，也不是单纯的劳动力，它两者皆是。请以此为准则进行构建（AI is not a software, it's not a labor, it's both. So build for that）。” - **人文底色**：嘉宾们以幽默而坚定的“人类万岁（Long live the humans）”为讨论画上句号，强调在高度智能化的技术栈中，人的判断力与核心价值始终不可替代。 ## Notable quotes > [05:44](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=344s) AI 既不是单纯的软件，也不是单纯的劳动力，它两者皆是。一旦意识到这一点，你就明白首席 AI 官的职责必须兼顾深厚的技术底蕴与敏锐的人力组织视角。若无法在这两点之间取得平衡，系统架构必然走向失败。 — Ali Abedini > [05:44](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=344s) 宁可把少数几件事做到极致，也绝不试图烧干整片大海。 — Ali Abedini > [14:17](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=857s) 最终的合规与法律责任永远必须由具名的人类员工承担。在强监管行业中，系统里必须有一个明确的‘问责对象’。 — Ali Abedini > [18:23](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=1103s) 这就好比拿着大铁锤去修理汽车内部零件，完全是杀鸡用牛刀。 — Oscar Anaya > [18:23](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=1103s) 适配的工具用在适配的场景。 — Oscar Anaya > [20:48](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=1248s) 作为人类的一员，我对此深表赞同。 — Ilyse Liffreing / 嘉宾 > [27:07](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=1627s) 小孩子房间里堆积如山的袜子一样，只会越来越多，永远收拾不完。 — Drew Schlussel > [44:20](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=2660s) 速度至关重要。看看初代 iPhone 和现在的 iPhone 之间的对比，你之所以不再使用初代产品，核心原因就是速度。 — 嘉宾 > [44:20](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=2660s) AI 既不是单纯的软件，也不是单纯的劳动力，它两者皆是。请以此为准则进行构建。 — Ali Abedini > [44:20](https://www.youtube.com/watch?v=Ntf2mLcaXRo&t=2660s) 人类万岁。 — 嘉宾 ## Mentioned Ilyse Liffreing, The Trade Desk, The Current, Drew Schlussel, Wasabi Technologies, Rahul Rajendra Prasad, Citizens Bank, Ali Abedini, Metropolitan Commercial Bank, TD Bank, Oscar Anaya, Lucien, Ai4, Clayton Christensen, 《创新者的窘境》（The Innovator's Dilemma）, John Friedmann, Model Context Protocol (MCP), ChatGPT, iPhone ## Video description Ilyse Liffreing - Editorial Lead/ Head of Growth, The Trade Desk Drew Schlussel - Vice President, Product & Solutions Marketing, Wasabi Technologies Rahul Rajendra Prasad - Senior Vice President of Technology, Citizens Bank Ali Abedini - Chief AI Officer, Metropolitan Commercial Bank Oscar Anaya - Co-founder, CTO, Lucien As organizations accelerate their AI adoption, building the right software stack becomes a critical enabler of speed, control, and business value. This holistic panel brings together technology and platform leaders to explore how they’re architecting flexible, scalable AI environments using best-in-class platforms, tools, and orchestration layers. Panelists will discuss how they evaluate AI solutions, manage cloud-native services, integrate AI tools into preexisting workflows, and balance innovation with enterprise goals. Whether you’re selecting vendors, stitching together custom workflows, or scaling existing systems, this session will help you build a modern AI stack that fits your strategy. [Source](https://www.youtube.com/watch?v=Ntf2mLcaXRo)

[5:44] Ali Abedini: AI 既不是单纯的软件，也不是单纯的劳动力，它两者皆是。一旦意识到这一点，你就明白首席 AI 官的职责必须兼顾深厚的技术底蕴与敏锐的人力组织视角。若无法在这两点之间取得平衡，系统架构必然走向失败。 [5:44] Ali Abedini: 宁可把少数几件事做到极致，也绝不试图烧干整片大海。 [14:17] Ali Abedini: 最终的合规与法律责任永远必须由具名的人类员工承担。在强监管行业中，系统里必须有一个明确的‘问责对象’。 [18:23] Oscar Anaya: 这就好比拿着大铁锤去修理汽车内部零件，完全是杀鸡用牛刀。 [18:23] Oscar Anaya: 适配的工具用在适配的场景。 [20:48] Ilyse Liffreing / 嘉宾: 作为人类的一员，我对此深表赞同。 [27:07] Drew Schlussel: 小孩子房间里堆积如山的袜子一样，只会越来越多，永远收拾不完。 [44:20] 嘉宾: 速度至关重要。看看初代 iPhone 和现在的 iPhone 之间的对比，你之所以不再使用初代产品，核心原因就是速度。 [44:20] Ali Abedini: AI 既不是单纯的软件，也不是单纯的劳动力，它两者皆是。请以此为准则进行构建。 [44:20] 嘉宾: 人类万岁。

[3:39] Drew Schlussel: AI doesn't sleep and it's persistent [3:39] Rahul Rajendra Prasad: Reimagine the Bank [3:39] Oscar Anaya: benevolent battle [3:39] Ali Abedini: boil the ocean [20:48] Oscar Anaya: trust me bro benchmarks [25:34] Ali Abedini: The Innovator's Dilemma [29:21] Drew Schlussel: model of the week [37:06] Oscar Anaya: Why are you paying for something that should already be predetermined in the model itself? [40:05] Drew Schlussel: bring in the bean counters, bring in the business analysts, bring in the Excel spreadsheet wizards [41:30] Oscar Anaya: it takes more power to not find a block than to find one [41:30] Ali Abedini: experimental knowledge is unique to your operation [44:20] Oscar Anaya: Look at your iPhone. Compare it to the first iPhone to the iPhone you have now. There's a reason you're not using that first one. Speed. [44:20] Ali Abedini: AI is not a software, it's not a labor—it's both. [44:20] long live the humans

[0:04] So hi everyone, my name is Eile. I am a senior  editor at the current which is an editorial [0:11] publication owned by the trade desk and I am  welcoming everyone to this wonderful panel today [0:19] which is called building the enterprise AI stack.  So as everybody knows over the past few years the [0:26] conversation around AI has really shifted  dramatically. It's no longer asking whether [0:33] companies should simply adopt AI. It's about  asking how to build an AI ecosystem and make it [0:41] actually bring value to the business. So figuring  out the right stack has become one of the biggest [0:47] strategic decisions of them all. So today's panel  is exploring just that how leading organizations [0:54] are evaluating AI into including in um including  them into existing workflows and creating really [1:03] scar scalable flexible architectures. So, we're  fortunate today to have this fantastic group with [1:10] us. And um first we have Ali Abinini and sorry  if I really destroy your name, but um he is the [1:20] chief AI officer and Metropolitan Commercial Bank  where he's building the bank's AI strategy from [1:26] the ground up after previously leading one of the  largest AI organizations in banking at TD Bank. [1:33] Next we have Oscar Ana, the co-founder and CTO  of Lucien, a cyber security expert specializing [1:41] in ethical hacking and securing missionritical  systems against emerging threats. Joining us also [1:48] is Drew Scholsh, the vice president of product  and solutions marketing at Wasabi Technologies. [1:55] He has decades of experience spanning storage,  data protection, engineering, security, [2:00] and helping enterprises solve complex technology  challenges. And finally, we have Rahul Rashad, [2:08] senior vice president of technology at Citizens  Bank, bringing deep expertise in enterprise [2:13] technology strategy and scaling solutions  for a leading financial institution. So, [2:19] I want to kick it off with a question for everyone  here. Can each of you share the single biggest AI [2:26] priority on your desk for this year? And can you  also dig into a little bit of what problems does [2:31] it address and what would you say is the biggest  measurable impact that it serves? Drew, would you [2:38] like to start? Sure. Absolutely. There we go. Hey,  first thank you all for coming. Really appreciate [2:44] the turnout. Great to see so many people in the  first session. I was a little nervous that'd be, [2:49] you know, like my three cousins and that would  be it. So appreciate you being here. Uh for me, [2:55] I'm I'm here in two roles, right? I'm here as a  vendor exhibiting at the show, but I'm also here [3:00] as a a user, somebody who is embracing AI. So  in my role in marketing at Wasabi Technologies, [3:07] um I'm really exploring how to leverage AI in  our sales motion. Um we're looking at how to, [3:15] you know, cultivate opportunities which you all  are going to become. Sorry. Um, I'll make sure [3:21] you don't get too many emails, okay? Um, but the  fact is that, you know, AI doesn't sleep and it's [3:27] persistent and it doesn't forget that it needs to  follow up or it doesn't uh forget that there are [3:33] clear signals that someone either wants to be a  buyer or couldn't care less and leave me alone and [3:39] don't bother me anymore. Um, AI is fantastic at  doing that for us. So we're definitely looking at [3:45] how to leverage that in our pursuit of growing the  company. And the real challenge is right finding [3:52] the right technology and finding the right way to  implement it and to stay in compliance with all [3:59] the laws and rules that govern our international  business. So there you go. There's truly so many [4:06] ways. So uh Rahul. Yeah. Hey, thanks everyone for  coming. I'm Rahul Prasad and uh one of my biggest [4:14] air priorities this year is in uh creating a  unified customer servicing experience for our [4:19] customers. Um and that involves combining the  uh digital mobile and online banking channels [4:26] as well as uh customer service and colleague  uh assisted experiences for our customers. [4:33] Um and in in relation to that we have started a  very critical initiative which we call reimagine [4:39] the bank and that is pretty much trying to address  any kind of fragmentation that exists today in our [4:45] banking experiences. Right? So we want customers  to be having a very seamless experience across [4:50] self-service channels whether they are um you know  uh interacting or doing money transfers on their [4:56] mobile app or online banking channels. uh but at  the same time also if they want to have assisted [5:02] uh experiences by walking into a branch or  calling into the contact center or IBR channels [5:08] the experience should not be fragmented so they  don't need to start from scratch they can continue [5:13] from where they left off in a different channel so  that's been our critical um investment this year [5:18] and and for the next few years in ensuring  that that happens um especially as you have [5:23] accumulated a lot of um um systems and processes  over a period of time it becomes very important [5:30] to ensure that the ecosystem is set up so we can  really leverage the power of AI uh to get to the [5:37] next level. So that's been our primary uh focus  this year. And how about you Oscar? Oh let's see [5:44] testing. Awesome. Sorry I was a little bit late  to test this. Um name is Oscar part of security [5:50] doing a lot of crazy stuff with AI. um biggest  priority. Geez, it's really tough truthfully [5:57] trying to fight the I guess the benevolent battle  with AI. It's interesting the level of automation [6:04] attacks have received nowadays specifically with  financial institutions which I'm very blessed [6:08] to be with with this group specifically uh in  the financial sector. uh we've we've seen and [6:14] identified a critical amount of vulnerabilities  that have been coming up that we were just unaware [6:20] of 10 years ago, four years ago or 5 years ago.  And so it's very difficult currently to try to [6:27] pick up everyone and put them at the same level in  terms of awareness of attacks, awareness of scams, [6:33] awareness of all these other data exfiltration  mechanisms that are happening specifically in in [6:38] institutions that require um regulated industries.  So it's very very difficult and so that's been [6:43] the battle that we've been pursuing with AI.  Obviously there is a good way to use AI to a [6:48] lot of these things. Um but our biggest effort  is trying to make sure that AI is not the only [6:53] tool we use. Trying to make sure that the human  mind is is the most capable one that we're trying [6:58] to reinforce and leverage to use artificial  intelligence in a specific way that actually [7:04] benefits the organizations themselves. And so it's  been an interesting one. uh a lot of discovery, [7:08] a lot of very dark stuff that's been happening  on online. So, uh it's definitely an interesting [7:14] avenue that we've been pursuing, but uh but  yeah, that's a quick rundown. Last but not least, [7:21] good morning everybody. Uh so, um I'm I'm trying  to um uh share with you um two major important [7:29] things. So, first one was personal. Um I've  been in I've been doing this for for so long and [7:36] um I feel I figure out still I don't know what is  uh emerging. So the first thing that it helps me [7:45] um was understanding what actually I'm doing what  does it mean to lead AI strategy governance and [7:54] and delivery. Um I have priv privilege to oversee  the end to end um and you know my my recent [8:04] uh concept about what is this AI because if we  don't understand it personally we cannot really [8:11] help our organization to develop a strategy  governance and security around it. So it was a [8:18] big big thing. It's ongoing. Uh but just to share  with everybody my current understanding is AI is [8:25] not software. Uh it's not labor. It's both. So  the moment that we realize that you understand [8:33] that the role of chief AI officer, it should be  high on technology and high on people. If you [8:40] don't balance that and don't design a strategy  and governance accordingly, you set yourself [8:47] for failure. So the second one was the actual  governance and strategy uh having that conception. [8:54] So the biggest and single uh most important  initiatives that we have done was twofold [9:02] uh designing and implementing the governance it's  very important and also shaping the strategy. So [9:10] the how how we design uh we are regulated by one  of the most demanding regulators uh in New York [9:19] and it it's very important you know people you  know argue it's it's a it's it's a drawback but I [9:25] see it as an advantage. Uh so we we embed AI into  all policy and procedures and it took us almost [9:36] a year to really have every line of business  own that I should not own their policy. They [9:43] are in driving seat and my role is just helping  them enable them facilitate that. The second one [9:50] was around a strategy. So I spent a decade in  a GA bank highly matrix organization globally [9:59] I always have like you know dual reporting  structures I learned a ton how to navigate [10:05] that but in order to be successful in the era  of AI you got to be nimble so that's why I think [10:14] you know the majority of the largest organization  they won't survive at least in the same shape and [10:22] operating model that they are working right  now. So we were trying to redesign the work [10:27] uh flow you know we we retire the scattered you  know AI use case portfolio and we are focusing on [10:35] end to end domain transformation and we are just  getting fewer things done better not just going [10:41] boil ocean yeah I mean none of these jobs are easy  especially with such an emerging technology and [10:48] suddenly you have to make sure it's also regulated  so Um so this next one is for Rahul. Um on that [10:55] note because as I just said financial services  has some of the highest regulations of the world [11:03] um and careers. How has that changed the way  Citizens Bank really evaluates AI vendors and then [11:10] maybe even compared with a typical enterprise?  Yeah. So thanks for that question. Um so when it [11:18] comes to evaluating AI vendors um they they pretty  much become they are not just seen as a technology [11:27] partner um but they become once once they are part  of the customer journey and our um and our banking [11:34] systems they pretty much become price um control  stack right so uh everything that we do as an [11:41] enterprise that has to follow a certain regulatory  standards compliance and you know model risk [11:47] validations and things of that nature um would  be extended to the vendors as well. Uh that's why [11:53] it's very important that it's not just a a shiny  new demo that we see uh typically um as vendor [11:59] presentations and and uh the you know the model  response and things of that nature but also one of [12:05] the my remits has been in terms of evaluating how  the vendor scales for our enterprise stack right [12:11] um if we have practices within our within  within the bank um that that you know that [12:17] where there's herence to certain regulatory  compliance needs security needs for that matter [12:23] um how does the vendor and their platform  uh accustom get accustomed to that those [12:28] practices. Luckily because um because this is not  this is something that is uh developed over time [12:36] um many vendors are starting to think in that in  that regard right so uh the good thing is um some [12:42] of these regulatory practices not just for banking  but also healthcare and government and other areas [12:48] as well. So how how does data get transferred  from you know one location to other where data [12:53] at rest and transit and things of that nature. Um  as well as like security uh controls and whatnot [12:59] that we typically embed into our pipelines  and infrastructure. Um all of that stuff when [13:06] we evaluate a vendor um we do very thorough inter  interrogation in terms of like how they scale for [13:13] many of those enterprise requirements that that we  put in front of them. Um and luckily like I said [13:18] a lot of the vendors uh especially in the field of  AI uh because there's so many things like you know [13:24] um an AI could hallucinate when we put it in front  of a customer. Um so how do we have mechanisms [13:30] kill switches and things of that nature where we  can quickly at least in my space uh the where I [13:35] deal with conversational AI um how do I quickly  um change uh switch a customer's experience [13:41] from an IVR uh that's front-facing AI to a a  human assisted experience. So whenever there [13:48] is hallucination and things of that nature that  come into the picture. So a lot of those things [13:52] are being handled by many of the vendors as we do  evaluations uh including things like observability [13:58] uh security and standards around uh how how  monitoring and observability works as well. [14:04] Um so that's how we evaluate from end to end  before we choose a vendor in this particular [14:10] space at least in banking. And there's so many  options too out there. There is. Yeah. Um so [14:17] for ALI your bank really treats AI agents like a  new class of workforce. Every agent gets its own [14:26] identity at least privilege access a named human  owner and a tested kill switch. So what does [14:34] everyone else get wrong when they bolt agents onto  an enterprise stack built for humans? Thank you. [14:42] So uh I think know I like the u the concept  of being less wrong. I don't think that any [14:49] human being be ever be right. We are all always  less wrong. So what what we are working on is uh [14:58] again going back to that concept which AI is both  software and a labor these days is moving toward [15:07] labor but it's never human employee. So let's not  confuse ourselves. So accountability always sit [15:14] with human employee. That's why you need to have  you know name human basically you need to have [15:19] somebody to blame right especially in a highly  regulated uh industry. So I I think uh what most [15:30] people get that wrong is that definition. So they  think you know it's a software so let's build a [15:36] more restrictive uh platform. They don't miss  the the labor part and you know the knowledge. [15:43] I I really like the the keynote uh of uh data IU  uh that they talk about you know labor own that [15:51] risk they they are responsible and accountable and  we are enabling that so most organization that I [15:59] had you know there are you know good organization  out there uh they are focusing on governing the [16:05] model and they don't forget that agents they have  hands right so how we how we regulate that how [16:11] we govern that And you know past few weeks we saw  you know what happens um you know an um a company [16:19] a frontier company uh evaluation AI model just  you know escape uh from the sandbox and you know [16:27] compromise another organization production uh uh  platform. So you know not to name names some of [16:35] yeah I don't I don't want to um name any companies  here but you know part of it is just you know they [16:41] are trying to just you know uh increase the the  IPO the potential IPO because nobody shipped like [16:48] that nobody shipped an evaluation model without  safety and guard rate. So that was intentional. [16:54] So they wanted to you know beat the evaluation  metrics. So they just unleashed the model and [17:00] uh they were expecting and the way that they  market it. But one lesson that we can learn is [17:06] the autonomy of this technology like um they are  truly like you know labor agent. So they can do [17:15] um you know something that for example every  time that you open up you know Excel sheet [17:20] you see menu file you always expect that but  with you know LLMs and generative AI you know [17:28] uh there's a lot of surprises and that's why you  need to have um identity the way that you know [17:35] we we hire and fire our employee you know don't  confuse with that but you know as I said they are [17:41] both software and labor. So treat them as such.  So the platform that you built for your software [17:49] is not going to work for for this era. The  platform that you built for your human employee, [17:55] it's not going to work either. So we all need to  figure out how those intersection mean for our [18:02] own organization. And I cannot give you a single  answer. Everybody should figure out on their own. [18:10] Yeah, there's I mean there's so much that can go  wrong and could maybe will go wrong in the future. [18:17] So, everybody's trying to figure it out. Um and  that leads into the next question pretty well. Um [18:23] Oscar, this one's for you. Um many organizations  are very overwhelmed by the pace of AI innovation. [18:30] How do you decide when a new model or framework  actually is worth adopting? It's an act. That's [18:38] actually a really interesting question because  [clears throat] I feel like we get bombarded [18:41] constantly with newer and newer and newer models,  yet we've seen the exact same capabilities over [18:47] and over that are actually utility to the  organization that's actually employing them. [18:51] And so it's really really interesting because they  all get more and more expensive. And so my take is [18:57] specifically in this subject, right? I think  the capability of the team has to be at least [19:04] at 10% of what the model's capability is. that  includes either it's intellectually uh in terms [19:10] of automation, in terms of discovery, in terms of  research and development. It's really important [19:14] for the team to be and I hate to use this word,  but as smart or close to as smart as what the [19:20] base model would be. Now, if your team is using it  and and has high token usage and things like that, [19:25] well, there's a there's an whole subject on that  in terms of token efficiency. There's people that [19:30] actually use it and simple, you know, prompts  generate complex outputs. And then there's people [19:36] who have really complex inputs, but it generates a  very simple output. And so I think the token usage [19:42] itself is also relative to how the model itself  responds. And the truth is if it doesn't have [19:48] an ROI, I don't think it should be implemented.  If if there is something simple about it and all [19:53] you need is automation, you know, implement  that. There's no need to have the hottest, [19:59] newest, 100red trillion parameter model just to  do emails. That makes zero sense, you know. And [20:05] not just that, the cost of it. The cost of it  is so insane that we're not using these models [20:10] for what they're supposed to be used. I mean, I've  seen, unfortunately, I've seen some employees use, [20:14] you know, the hottest Fable 5 to write an email.  And I'm like, what are you doing? Like, why are [20:19] you doing that? This is not what that's meant  for. And I And so, in essence, what I'm trying [20:23] to get to is the right tool for the right job. you  know, you can use a hammer to break everything and [20:27] and put a car together, but in reality, there are  specialized tools for specialized um task. And so, [20:34] in essence, that's really where I'm going. Now,  the the other end of this, right, it's really [20:38] really critical and really important. You don't  want to get a a model that's too dumb because then [20:43] you fall into the idea of like hallucinations.  How do you even deal with that? And so the more [20:48] hallucinations you get, the less efficient your  team becomes, the less efficient your operation [20:52] becomes, and thus the automation fails. I think  it's really important also to to address like the [20:57] elephant in the room. It's very hard to automate  with non-deterministic tools. It's incredibly [21:03] difficult. How do you do that? It's impossible.  You may have a 100 inputs and 100 outputs that [21:08] are coherent and it's that one output that's  incoherent that completely trashes everything. [21:13] It could happen during a demo. It could happen  while closing a client. It could happen while [21:18] in production. The truth is you want to minimize  the level the level of non-determinism inside. Oh, [21:25] that's interesting. Test. There we go. Thank  you so much. Mixing. Um, you want to make sure [21:31] you test your models and you lower the amount of  non-determinism in your systems. You want to lower [21:36] the amount of entropy in the actual system itself  before you push it into production. Unless entropy [21:42] is your business where you're generating just  chaotic outputs. Truthfully, you got to be careful [21:48] with that. It's very very important to make sure  that um first in first out is actually first in [21:53] first out and you're not getting some sort of  cached output from you know different queries [21:57] around. So it's a very complex subject. The truth  is the team is really what pushes the model. You [22:04] have to make sure your team is up to par. And  and I really I'm not very happy with the way [22:09] the industry has been going about replacing humans  with AI. I think the peak value of AI is in humans [22:15] itself. You know, you need good humans that are  really willing to push the boundary to actually [22:20] push the models themselves. You can't do the trust  me bro benchmarks. You know, that's just not going [22:26] to really put it there. Uh the benchmarks  only tell what the model is trained to do. [22:30] You want very smart and intelligent people that  are passionate about what they're doing to say, [22:35] you know what, this model has never done it, but  let's see if it could zero shot it, you know, [22:39] and that's when you get true discovery. So that's  when you start using the model for what it's meant [22:43] to be. But in terms of automation, the cheaper  the better, honestly. Like if your model works, [22:48] keep it. Don't upgrade it. If your infrastructure  is cheap and you're actually at an ROI, keep it. [22:53] That's worth, you know, so much more. But uh but  yeah, that's pretty much my take on it. As a human [22:59] myself, I vouched for that. [sighs] And yeah,  sometimes simple really is better. You don't [23:05] have to get fancy all the time. Uh, so this last  one is for Drew. What would you say is the biggest [23:11] misconception companies have about preparing  their data infrastructure for enterprise AI? All [23:18] right. So before I answer that, it's it's audience  participation time for a moment. All right. look [23:26] up from your Instagram or Instaface or whatever  you're up to. The question I have for you all is [23:31] who here, raise your hand, is completely satisfied  with the ROI they're getting on AI today. [23:41] Wow. Wait, let's do that again. Put your hand up  all the way if you're completely satisfied with [23:49] the ROI. Oh my goodness. Okay, thank you. And  those of you who are not completely satisfied, [23:57] go ahead and raise your hands. That would be the  rest of you. Okay. Excellent. Surprised. Yeah. [24:03] Interesting. Right. And that's part of our our  spiel here. So, I think the biggest misconception, [24:10] well, there's a lot. Um, I like the let's not  replace all the humans, uh, personally. Um, [24:17] but I think it's also about the fact that all of  your cost is 100% tied up in your GPUs or in your [24:26] in your compute infrastructure. Okay. Now, little  bit of a sidebar here. Um, good luck to anybody [24:33] who is putting in an order today for any kind of  hardware, whether it's servers, storage, network, [24:41] because you're going to be told that it's going  to be at least 12 months to ship and install, [24:48] and they're probably going to repric your quote  multiple times before it ever hits your dock. All [24:54] right, there is a structural shortage going  on. The good news is right, myself included, [25:00] there's this thing called cloud and you can  leverage that as part of your infrastructure. [25:06] The bad news is everybody thinks that cloud is  insecure, that cloud cannot keep up. That's a big [25:12] misconception. You can leverage cloud and you can  also find vendors especially in the Neocloud space [25:19] that can help you realize a real uh satisfying  return on your investment. Right. I wish we [25:28] could do a Q&A cuz I would go right over to the  people who are totally satisfied with their ROI. [25:34] So I have um just you know thinking out loud  uh about the concept of ROI uh about this [25:41] uh transform transformative and disruptive  technology you know capitalism American [25:48] corporations are so optimized for quarterly you  know financials and you know they are talking [25:55] about ROI but we all forget this is a disruptive  technology takes some time to see the ROI So [26:03] I'm I'm trying to balance out that vision. I  don't reject to you got to be cautious about [26:10] um your your financials, but you need to carve  out this as a capex as you know something that [26:16] it takes sometimes and you're not going to see it  next quarter. You know I just bought this great [26:21] platform. I'm going to see the ROI next quarter.  That's that's a wrong um you know concept. You [26:27] know again I may be less wrong about this or  completely wrong but I want to refer to Klay [26:33] Christensen book you know innovator dilemma you  know they are all good companies they are doing [26:40] really great managing their financials perfectly  but we don't hear their names anymore because they [26:47] miss this disruptive uh technology and what it  takes to really embed that and see that ROI in [26:55] in a more sustainable manner in the future. So  that's just you know one comment I want to make [27:00] uh for the ROI concept here. Excellent. Um I just  want to add on to what I was saying earlier that [27:07] another misconception is that uh again all your  cost is tied up in compute and GPU and there's [27:13] this little thing at the end of the workflow  called storage where I have lived for 30 years [27:19] plus and everyone thinks that it's a rounding  error that you know it's not going to affect my [27:26] total cost of ownership it's not going to affect  my ROI except for the fact that I'm collecting [27:31] as much massive massive amounts of data on the  front end and then guess what? I have to archive [27:36] it and keep it for compliance and governance  reasons on the back end and it never ends. It [27:44] just keeps piling up. It's like all the socks in  your kids' room. It will just continue to grow and [27:49] grow and grow. All right? There's nothing you can  do about it. You can't clean it up. Um and again, [27:57] you need to look at look at your stack, right?  Look at your hardware stack. look at your software [28:02] stack but then also consider how the life cycle of  a data or datim will live in your environment. I [28:10] believe in the banking environment you have what 7  to 10 years required to maintain various different [28:17] records. Um I believe now you also have rules and  regulations around communications that have to be [28:24] archived and maintained so that you can wind back  any conversation you have. you know, when someone [28:29] calls and says, you know, give all my money to  XYZ, you're like, I don't know about that, right? [28:35] That whole the whole scam thing. So, you know,  for anybody here, one more question if I may. [28:41] All right. Who here has come to this conference to  figure out what stack to build? Okay, good. Wow, [28:51] fantastic. And everybody else, you've got it all  figured out and you're just trying to optimize [28:56] it. Is that fair? Put your hands up. I want to see  your hands, too. Okay. Excellent. Well, you're in [29:02] the right place, right? You're in the right place.  Um, I think you will learn not only from this [29:10] panel, but from all the various conversations this  week, right? There are things to avoid and there [29:15] are things to go after. And I think I've probably  spoken long enough about what I think is a a [29:21] challenge. So, I'll hand that. Fantastic. Um, on  that note though, integrating those AI tools into [29:29] your stack or your workstream is rarely seamless,  I would say. Can you share maybe a story of a [29:37] hurdle in doing so and how your team navigated it?  And Drew, I I feel like you would be the best to [29:45] answer this. I just I just went through this, so I  I'll I'll kick it off and throw it over to to you [29:50] all. But um we we we implemented an MCP server  for Wasabi storage and we're all excited about [29:57] it. It's in beta. Everyone can try it. And we  immediately you like ran to our cloud instance [30:03] or our chat GPT instance and pointed it at the  MCP server and got immediately rejected because [30:09] my security team had not cleared a path for us to  actually use the connector. And I called my head [30:18] of IT and screamed and yelled and said, "How am  I supposed to demo this for customers if I can't [30:25] connect to it?" And he said, "We're still doing  our security evaluation. You're going to have to [30:30] wait a while." And I said, "How long is a while?"  And he said, "A." And we just went back and forth [30:36] every day for, I don't know, [snorts] a week  until I beat him senseless. And he finally said, [30:40] "Okay, you're fine. Go ahead and do what you  want." But and and I get it. I've I've studied. [30:45] I I'm a practitioner of good security and  good process, good procedure. But you know, [30:52] you talk about the model of the the flavor of the  day, right? The model of the week, the connector [30:57] of the day, right? We all want to go off and try  these new things. And yet there's this important [31:02] function of security to make sure we don't do  something so that we set loose this monster who [31:08] goes off and rapes and pillages bad things on  the rest of the world. So yeah, there you go. [31:15] Does anybody else have a story? I can I can  share something uh really specific and also [31:21] something very high level. So about the concept  of MCP I think you know if you just you know [31:27] um search it or um use AI assistant for that a lot  of us researching through that and you know I had [31:37] um you know some good conversation with our you  know second line and and our platform about about [31:43] that. So you know the conception is we need  to throw out MCP everywhere and sometimes we [31:49] forget MCP server. This is one little piece  of overarching you know uh uh information [31:58] security and cyber security strategy that you  have your you have AI gateway you have your [32:04] um you know observability platform and you know  the the example that I use is you know uh I use [32:11] kitchen you know in the kitchen um think about  MCP as a parent with a menu uh in the kitchen and [32:20] you know in in your house, you may have a robot.  Robot is, you know, LLMs, all those, you know, [32:26] super intelligence. So, it makes sense that you  have a parent, guardian parent uh standing at [32:33] the door of kitchen and check with the robot what  it want to do. Uh you never touch you know oven, [32:42] you never, you know, mess up with the fridge. You  just go and do that and it just check and and go [32:46] there. But you don't have that guardian parent for  your bolt-in uh dishwasher because it always does [32:55] exactly the same thing. So you need to you don't  need multiple parents to watch for your dishwasher [33:01] and you know cabinet and everything. So a single  you know inrompt in code security help with uh the [33:09] dishwasher you know that what what is a dishwasher  is traditional machine learning right but when [33:14] you have agents you got to have MCP in your  architecture but again MCP is not just a total [33:21] architecture that you threw everything you're  just going to overbuild and you're going to even [33:26] increase your cyber security because you have some  MCP which you know you don't know at some point [33:32] how to observe the more that you have the more  complex uh you need to deal with your platform. [33:38] So what I recommend is always make sure that you  have four plane in your architecture. The first [33:43] one is instruction plane. Uh what does the agent  uh know, what should it know and [snorts] how to [33:52] escalate. The second plane is your capability  plane. So what um capability you know what [34:00] agent what system it has access to or invoke or  um uh and and and the third one is is runtime [34:08] runtime plane you know when your agent and up and  running uh what constraint it it operates within [34:14] it doesn't go and you know delete your entire you  know database and the the last one which we are [34:20] investing time is assurance how we make sure they  are doing the way that they supposed to do it's [34:26] about like testing control testing your controls  and and all of that. If you apply these four plane [34:32] in any case you know if you have cloud you you do  it with subscriptions different vendors they have [34:36] accounts you know um uh you can figure out even  if you're going to be uh on prem as long as you [34:43] follow that for plane I think hopefully you will  be safe. Yeah, I mean I just I one thing first, [34:50] the best anal analogies on this panel for sure.  [laughter] Um sorry, Ro, did you have a story? [34:57] I was just going to say um one of the ways we to  your point like MCPS um we cannot throw it around [35:05] everywhere, right? So um one of the ways we have  done um things in uh at least in the IVR and the [35:11] chat and the conversational AI space is how do we  get to a point where we do have integrations with [35:17] all our record source record systems right where  there's customer information there's transaction [35:23] information account information and things of  that nature. um still keep them safeguarded and [35:28] you know like the the the proper enterprise um you  know controls around it um in more deterministic [35:34] forms but at the same time also when when  it comes to customerf facing intelligence [35:39] that we're trying to put out there uh rather than  traditional conversational uh uh you know DTMF or [35:46] things of that nature as you call into a contact  center um have that natural human human like [35:52] voice capabilities um facing the customer, right?  So, um that hand off between uh the LLM and the [36:01] deterministic flows is going to be very critical  and how we safeguard between those in terms of [36:06] identity, authentication, quoting and things of  that nature um is going to be very critical. So it [36:13] it is not just throw everything um in front of the  customer or the colleague um creative MCPs and you [36:19] know voice models and things of that nature but  have a very deterministic path where it's needed [36:24] uh where we need to safeguard the data uh but at  the same time um where there's opportunities to [36:29] you know have meaningful conversation humanlike  conversation with the customer and safeguarding [36:36] and guardrailing it right like where where there's  a point where as part of your observability [36:41] um you you're capturing that there's hallucination  rates is going high or the ground and it's totally [36:46] being inappropriate and things of that nature  where you can control it in time and be able [36:50] to move it move the conversation over to a  human agent that would definitely help. So [36:56] those are some of the safeguards at least from  a banking perspective we are putting in place [36:59] um so so we are compliant and you know keeping up  with the regulations and things of that nature. [37:06] I agree to some extent. Um, latency like it's the  one thing we all brush under the rug. Latency, [37:16] latency, latency. Like you start implementing  MCP left, right, left, right, you know, you start [37:21] doing it everywhere. Next thing you know, your  entire pipeline is a minute. Like I don't think [37:25] that's feasible. And in the new modern age of  computing, that doesn't make any sense. You know, [37:30] especially when it comes to guardrails consuming  70% of your entire token use. That literally [37:35] doesn't make any sense. Why are you paying for  something that should already be predetermined [37:39] in the in the model itself? And so that's the one  big secret that most of the time you don't see [37:44] when it comes to connecting all these agents into  MCP servers. All the guardrails, you're paying for [37:48] them, you know, and I'm pretty sure some of you  in here are paying for those guardrails on the [37:52] three figures, you know, I mean, six figures. It's  it's pretty ridiculous. And so latency guardrails [37:59] and infrastructure, you put everything in the  cloud, which I think is great. We've launched [38:03] several attacks against our own infrastructure  when it comes to pentesting. Last time we did [38:07] about $26,000 in an hour. I don't think that's  that cool. That's that's pretty buming. So, [38:14] um it's really important to try to figure out  the good intermediate between latency, you know, [38:18] having the determinism in the pipeline itself that  can definitely save you money on the long run and [38:23] having the proper cloud architecture that's not  just full cloud or full onrem but a good hybrid [38:28] model. I mean, we've seen EC2s fall left and  right. We've seen the entire stack from the whole [38:35] left continent, you know, west coast, east coast  go down. The whole world starts burning. And so, [38:41] do you want to be in that party or do you want to  be in a party that's self, you know, sufficient [38:45] and resilient to any type of attack or any type of  thing that's happening currently in the world? Um, [38:50] another thing is how many of you do business  internationally? You know, how many of you [38:54] sell things internationally? How many of you have  agents deployed internationally that are trading [38:58] in multiple exchanges that are aggregating data  that are analyzing PDFs left and right? I mean, [39:04] it's crazy. The pipeline requires a really  fast throughput and I think that's the thing [39:09] that we haven't reached yet. We have the level of  automation. These things are very, very smart. I [39:14] mean, don't get me wrong, they're amazing. But I  don't know if it's good when it takes 3 minutes, [39:19] 20 minutes, or an hour to get something done, you  know, all to see something come back with a very [39:24] crazy hallucination that you're like, that's  just not what I was talking about. That's just [39:27] simply not it. And so really critical point to  uh to the cloud stuff. Uh we haven't seen or at [39:35] least it hasn't been made public, but a lot  of the stuff that's getting purchased as as [39:40] uh um someone mentioned in the panel is is taking  forever. Once it gets there, who's installing it? [39:46] Who knows what's getting put on your cloud  architecture? Is it actually being produced [39:51] and procured by the right parties? I mean,  it's crazy. Just look up how many implants [39:56] have been found in devices purchased by  OEMs from OEMs. The actual OEM, you know, [40:01] you go and you purchase it from their website  and next thing you know, your entire router, [40:05] your switch has implants on it. It's ridiculous.  It's getting out of control. And so from a from a [40:10] perspective at least in terms of building a proper  ROI or building a proper business, you have to [40:15] take all these things into account, especially  if you're actually putting something critical [40:19] together. You start trading on agents, which we've  all seen in the past two years, and we've we've [40:24] seen several agents start trading and make a lot  of money and then lose an insane amount of money. [40:30] Latency is going to become an issue. And so, yeah,  that that's just my key point to that. And I don't [40:34] know, you have you had a comment. I I just want  to make one thing. you you brought up just now and [40:39] that is you know the unpredictable nature of your  spend right so folks here who are trying to build [40:45] a stack bring in the bean counters bring in the  business analysts bring in the Excel spreadsheet [40:51] wizards okay and make sure that there is a known  variable for every cost that you're going to incur [41:00] and I also recommend that you look at the vendors  who have predictable pricing because it's going [41:07] to make a huge difference in at the end of the day  whether you get a surprise bill that theoretically [41:13] can put you into bankruptcy as a company. I think  we heard that story in the last uh couple months [41:19] or you know you really are getting the right value  from the right investment that you're making in [41:25] your AI stack. Thanks. But one thing to that  comment real quick, it comes back to the team. [41:31] It really comes back to the team. your team is  capable of estimating the amount of compute that [41:36] will be required. It it's not it shouldn't be that  complicated to try to figure out, hey, how much is [41:41] this going to run me in cycles? How much compute  are we going to run? How much power are we going [41:46] to burn? You know what I mean? I mean, it's crazy  and and I hate to skew so so vaguely, but it's the [41:50] best example that I have. I mean, when it comes  to like Bitcoin mining, it takes more power to [41:55] not find a block than to find one. And so it's  ridiculous. The amount of energy we're throwing [42:00] into tokens that are not being consumed the right  way is getting out of control. And so the compute [42:05] spend is relative to that. And then you bring the  bean counters and they're very mad at that, you [42:10] know, but how do you excuse that? How do you even  tell someone, well, we made an experiment because [42:15] we thought we were going to have a significant  alpha and next thing you know, you don't. And [42:19] so it's very difficult, but um so I mean we know  that's that's right. But you know uh in order for [42:26] all of us to really uh make it happen and feel the  difference in our balance sheet in our you know [42:33] earning calls we got to generate experimental  knowledge you know I'm I'm city planner by [42:39] training I don't know how I ended up in banking  but uh one of one of the uh great researcher that [42:46] I had great respect was John Friedman says there  is two type of knowledge expert knowledge and [42:51] experimental knowledge expert knowledge is on the  books you know we we have access to experimental [42:58] knowledge is unique to your operation you cannot  generate that experimental knowledge by actually [43:04] doing it so again going back to the question  of ROI and you know balancing out in order for [43:11] your organization to really scale you got to do  it don't wait for that person ex consultants or [43:19] you know to to generate that information for you  you can always leverage the best practices and you [43:25] know uh the lesson learned from other organization  but in order for your organization you got to do [43:32] it to generate that experimental knowledge and  you got to pay for it. So it's not it's not free. [43:40] That's a tough one. It it's definitely not free.  And uh how many of you had had your token account [43:45] stolen? How many of you had had purchased you  know or per or seen in the market stolen tokens? [43:52] be careful app proving those. I agree,  experimentation is key. It really is. You got [43:56] to push the boundary, but at the end of the day,  you got to monitor it really well. Make sure that [44:01] your tokens are actually yours and that you're  consuming them. Don't just send them down the [44:05] pipeline and then say, "We finished, but somehow  it costs three times more." And so, very, very [44:10] difficult, very interesting, but at the end of  the day, we we really just got to hone in on what [44:15] we're using, you know? But, uh, but yeah, great,  great point there. Great. Well, we're almost out [44:20] of time, so I I do want to end on a positive  note. Um, as you guys look towards the future, [44:27] um, what are like one quick word or sentence that  you're looking forward to in figuring all this [44:34] out? Speed. Speed. Very critical. Look at your  iPhone. Compare it to the first iPhone to the [44:40] iPhone you have now. There's a reason you're not  using that first one. Speed. Very nice. Anybody [44:46] else? I can just wait but with proper governance  like that nice just you know the the sentence that [44:54] I start with AI is not a software it's not a labor  it's both so built for that true long live the [45:03] humans long live the humans yes let's give a round  of applause what a great panel thank you so much.