Oura, Redwood Wrap Legacy Stacks in MCP to Deploy Autonomous Agents

Oura 与 Redwood 通过 MCP 封装传统架构部署自主智能体 将平台治理与易变的基础模型解耦，并用标准协议封装确定性传统系统，可在削减运营成本的同时保护核心业务逻辑。 企业架构领军人物探讨了如何跨复杂企业技术栈，将 AI 从被动推理工具转变为自主、可审计的运营工作流。专家们主张，与其对确定性传统系统进行高风险重写，不如将既有业务逻辑封装在模型上下文协议（MCP）服务器中，并在易变的基础模型之上抽象出治理和遥测层。他们概述了应对智能体泛滥的实用方法，包括提示词路由、自动化许可证审计智能体，以及可防止供应商锁定并优化云支出的模型无关角色画像。专家组强调采用“以需求为导向”的准入模式，并搭配当日出结果的快速原型沙盒，从而将每项 AI 举措与可衡量的业务 ROI 直接挂钩。 企业 AI 架构必须从咨询式推理转向在多元技术栈中具备严格治理的主动、可审计运营执行。 (0:04) 通过 MCP 等协议编排的自主 AI 智能体可执行站点可靠性工程（SRE）及服务台任务，解决多达 50% 的 IT 工单并提交自动化代码拉取请求。 (3:13) 多智能体工作流会带来死循环和输出冲突等运营风险，因而需要操作员级编排、可审计的决策图以及人工在环（human-in-the-loop）检查点。 (5:36) 访问控制和合规治理必须与快速实验层解耦，并通过声明式自然语言策略由平台原生处理。 (9:01) 鉴于基础模型的条款和能力变化迅速，遥测、日志记录与审计规则必须抽象到编排层，而非绑定至特定 LLM。 (12:32)

To build resilient AI-first organizations, enterprise architects are wrapping deterministic legacy systems rather than replacing them, decoupling governance layers from volatile third-party models, and enforcing outcome-focused financial controls to curb tool sprawl. Wrapping Deterministic Cores with Model Context Protocol (18:32) In high-volume transactional environments like financial services, which process approximately 10 billion transactions daily, architects cannot replace settled business logic with probabilistic models. Even a 99% accuracy rate would yield 100 million failed transactions per day. Panelists agree that discarding legacy platforms risks losing decades of embedded edge-case handling and institutional intelligence. Instead, modern architectures position probabilistic AI to handle semantic intent and natural-language queries while leaving deterministic systems to execute analytical operations. Redwood Software and Oura employ the Model Context Protocol (MCP) to wrap legacy codebases and non-AI software. This approach exposes existing systems to autonomous agents as callable tools without altering foundational code. Oura uses MCP wrappers to isolate technical debt from SaaS tools lacking AI development roadmaps, while enterprises in traditional sectors like insurance use generative models to map end-to-end workflows, exposing undocumented logic before writing new automation code. Decoupling Platform Governance from Volatile Models (12:32) Because foundation model vendors rapidly adjust terms—such as abruptly ending automatic retention of prompt and response history—binding access rules, audit logging, or business logic directly to a specific model version creates enterprise risk. Governance, observability, and compliance tracking must instead live in an abstraction layer above the models. Michael Maher notes that platforms are replacing manual role configurations with natural-language security constraints that automatically map into role-based and attribute-based access controls. To manage shifting compliance standards without slowing down developers, Christopher Lundy recommends pairing technical and legal AI counsel to translate regulatory requirements into concrete operational guardrails. His team runs an automated agent every Monday morning that crawls documentation, terms of service, and privacy policies across enterprise SaaS tools and open-source repositories like Hugging Face. The agent flags unannounced underlying model switches and intellectual property risks before they impact production workloads. Orchestrating Autonomous Agents and Shadow Deployments (3:13) Enterprises are scaling autonomous agent systems into core operations. The Friedkin Group deployed an IT help desk agent that autonomously resolves 50% of incoming support tickets. Oura deploys 10 to 20 new AI agents each week and supports over 4,000 active AI-generated applications, including Site Reliability Engineering agents orchestrated via MCP that analyze multi-product telemetry and submit automated pull requests to fix codebase errors. Multi-agent interaction introduces operational failure modes, including circular reasoning and conflicting instructions. Redwood addresses this by modeling agent workflows as work graphs with full decision audits and human-in-the-loop checkpoints. To curb shadow AI, Oura maintains a dual architecture: 400 software engineers work from standardized templates pre-configured with guardrails, while business users generate agents with built-in telemetry streaming directly to IT operations. In addition, supervisory agents monitor external low-code platforms like Glean and Lovable to detect non-compliant agent creations. Transitioning from Token Maxing to AI FinOps (29:46) As SaaS vendors embed native AI studios into their products, enterprises face redundant costs and fragmented data pipelines. Lundy cautions that paying separate AI fees for the same integration logic across multiple applications—such as Workato and Zip—penalizes enterprise budgets, driving the need for formal AI FinOps frameworks. Redwood controls internal spending by routing queries through prompt routers and specialized fine-tuned agent platforms, which handle 60% to 70% of engineering workflows using cost-effective task-specific models rather than expensive frontier LLMs. Maher and Ajit Singh emphasize shifting enterprise culture from unbounded experimentation to rigorous return on investment. The Friedkin Group adheres to a "lead with need" model, where candidate AI projects must be vetted by a joint finance and architecture committee against two- to three-year ROI projections. Similarly, Oura combats multi-hundred-thousand-dollar implementation delays by conducting AI-assisted discovery sprints, turning morning business requirement sessions into fully governed sandbox prototypes by the afternoon. While rapid agent prototyping and automated infrastructure factories demonstrate clear operational velocity, long-term architectural stability relies on strict modularity. As model capabilities, provider policies, and SaaS integrations fluctuate, enterprises face the ongoing challenge of maintaining model-agnostic abstraction layers and outcome-driven intake processes without stifling workforce innovation.

来自Pave、Redwood Software、Oura与The Friedkin Group的企业架构主管们指出，构建AI优先型企业的关键在于将确定性的既有系统与概率性AI解耦，通过MCP协议封装保留核心业务资产，并在模型层之上构建统一编排架构，以平衡开发创新、合规治理与商业回报。 确定性系统与概率性AI的MCP封装 (18:32) 架构设计必须严格区分确定性传统系统与概率性AI。在金融服务等高频交易场景中（日均处理约100亿笔交易），即便顶尖模型达到99%的准确率，完全依赖概率性系统仍会导致每天产生1亿笔错误交易。因此，已稳定运行5至20年的核心业务逻辑不可盲目废弃。现代AI的核心优势在于语义推理和意图理解，而底层精确计算与合规账本仍需由确定性系统支撑。 Redwood Software与Oura均明确反对对既有系统采取激进的推倒重写模式。遗留代码库沉淀了企业数十年的领域规则与边缘异常处理经验，重写往往伴随极高失败率。行业通行的应对方案是通过模型上下文协议（MCP）服务器将老旧系统及缺乏AI路线图的第三方SaaS应用进行封装，将其作为工具暴露给自主Agent调用。这既保护了既有业务逻辑的完整性，又赋予了上层Agent灵活的操作能力。 平台层抽象与内嵌合规治理架构 (10:51) 在生命科学和数字健康等强监管领域，FDA法规等合规标准要求严格执行标准操作程序（SOP）并具备完备审计链路，传统验证周期常长达6个月，无法匹配AI的迭代速度。为此，企业需将合规与治理从实验层解耦，直接下沉至底层基础设施。Michael Maher指出，治理手段正在从繁琐的人工权限配置演进为自然语言策略，系统将声明式安全约束自动编译为基于角色或属性的访问控制（RBAC/ABAC）与审计日志。 由于底层大语言模型（LLM）版本及服务条款频繁变动（例如Anthropic和OpenAI曾调整条款，停止自动留存提示与响应历史），直接将合规与业务逻辑绑定至特定模型已成为架构反模式。Jason Kehl强调，必须在模型层之上建立独立的编排与遥测机制，通过工作图（work graphs）记录决策链路并在关键节点设置人工介入检查点，以有效防范多Agent交互时产生的冲突、死锁与幻觉风险。 避免厂商锁定与监管影子Agent扩散 (24:09) 保持模块化和“模型无关（model-agnostic）”是架构韧性的核心。The Friedkin Group首席架构师Ajit Singh分享经验称，某关键应用曾因供应商突变商业模式而被迫解绑，整个迁移耗时三个月。为防范被单一厂商锁定，团队为不同Agent角色配置动态模型路由，避免昂贵的前沿模型被滥用于全天候背景任务；同时引入专职AI法律与技术顾问，将碎片化监管条款转化为可执行的工程护栏。 面对企业内部激增的应用与影子Agent风险（例如Oura内部运行着4000多个AI生成应用及每周新增10至20个Agent），架构团队建立了自动化全生命周期监控。Oura每周一通过审计Agent抓取所有企业SaaS及Hugging Face开源模型的条款与隐私声明，实时识别静默的模型切换与知识产权风险。同时，内部推行双轨制管理：针对约400名软件工程师提供预置安全护栏的标准化模板，对业务人员则通过监督Agent监控Glean、Lovable等无代码平台，在违规配置扩散前发出主动预警。 从代币消耗到ROI验证与AI FinOps (29:46) 多供应商SaaS工具普遍内置集成组件，极易导致基础设施重复建设与成本失控。Chris Lundy指出，若员工在不同SaaS工具中重复生成相同的自动化逻辑，企业将为同一逻辑向多家AI供应商重复付费，这催生了取代传统TCO模型的AI FinOps管理框架。Redwood Software通过提示词路由器将轻量任务分流至高性价比模型，并将60%至70%的工程流收敛至统一编排平台以遏制预算超支。 企业管理文化正从早期的无节制实验（“Token Maxing”）全面转向以业务成效为核心的投资回报率评估（“AI ROI”）。Ajit Singh强调“以需求为导向”，Friedkin集团由财务与架构委员会对候选项目进行2至3年的ROI测算后方可立项。在交付层面，Oura通过设计冲刺在上午利用Agent梳理业务痛点并量化自动化收益，下午便在受控沙盒中交付可用原型；Friedkin集团则利用自动化“基础设施工厂”，仅由2名平台工程师在3小时内即可按需交付隔绝数据泄漏的多租户沙盒，实现了敏捷创新与财务审慎的平衡。 与会专家一致认为，企业AI落地的核心瓶颈已从模型本身转向架构解耦、动态合规与业务价值变现。然而，在放手让各业务单元快速构建成千上万个轻量Agent的同时，如何长期监控多平台间的隐性重复成本并规避知识产权侵权，仍是企业架构持续面临的治理难题。

## Background 该内容由行业峰会频道 Ai4 于 2026 年 8 月 21 日发布在 YouTube 平台上。视频记录了一场题为《Enterprise Architecture for AI-First Enterprises》的圆桌论坛，汇聚了来自 Oura、Quickbase/Pave、Redwood Software 及 The Friedkin Group 的企业架构高管与技术负责人，共同探讨如何重构底层系统以支持模块化、可扩展且符合合规要求的 AI 原生架构。 ## TL;DR 四位技术领袖深入分享了企业从传统架构向 AI 优先转型的实战路径，核心强调不可盲目推倒重构遗留系统，而应利用 MCP 等协议进行中间层封装与解耦。讨论涵盖了利用 SRE 智能体军团实现自动化运维、通过工作图谱（Work Graph）与人机协同解决多 Agent 冲突与审计黑盒，以及将治理能力静默内置于平台底层以兼顾生命科学等强监管行业的合规与创新。此外，嘉宾们还针对 SaaS 工具蔓延、动态提示词路由控本、AI FinOps 财务评估以及应对智能体激增的双轨监控机制提供了切实可行的架构指导方案。 ## Atmosphere & style 整场圆桌讨论务实而干货密集，技术领导者们摒弃了空洞的行业炒作，以极具一线实战感的坦诚视角直击架构痛点。现场气氛既有探讨遗留系统“盲目重写必翻车”等工程宿命时的会心幽默，又有在面对模型频繁更迭、合规条款突变时的冷静与从容。四位嘉宾与主持人的互动节奏紧凑、观点互补，从多智能体死锁到 Token 预算失控的拆解均直切要害，展现出深厚的技术定力与架构智慧。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=sJ0-fOALML0&t=4s) 企业 AI 的核心挑战已从算法模型本身转向“可信赖度与合规治理”，以及推动 AI 从单纯提建议的“推理顾问”演变为具备审计与问责机制的“实际运营执行者”。 - [03:13](https://www.youtube.com/watch?v=sJ0-fOALML0&t=193s) 智能体劳动力（Agentic Workforces）正成为打破企业人手瓶颈的核心抓手，如 Oura 通过 MCP 服务器编排的 SRE 智能体集群实现从异常检测到自动化提交 PR 的运维闭环。 - [05:36](https://www.youtube.com/watch?v=sJ0-fOALML0&t=336s) 针对多 Agent 协同容易出现的死循环与冲突死锁，需引入工作图谱（Work Graph）精细定义通信边界，并在关键决策节点挂载 Human-in-the-Loop 人机协同卡点。 - [10:51](https://www.youtube.com/watch?v=sJ0-fOALML0&t=651s) 将合规治理、SOP 追踪与权限控制下沉为底层平台的内置服务，使研发人员在前端享有充分创新自由的同时免去冗长繁琐的传统合规验证周期。 - [12:08](https://www.youtube.com/watch?v=sJ0-fOALML0&t=728s) 为了应对外部大模型频繁更迭与单方面修改服务条款的风险，企业必须坚持模型中立并构建更高维度的平台编排层，严禁将权限与治理逻辑直接绑死在单一模型上。 - [15:51](https://www.youtube.com/watch?v=sJ0-fOALML0&t=951s) 传统遗留系统是高度可靠的“确定性”执行基座，而现代 AI 系统是擅长意图理解的“概率性”推理大脑，架构演进的核心在于通过 MCP 等封装协议实现两者协作，而非推倒重写。 - [25:31](https://www.youtube.com/watch?v=sJ0-fOALML0&t=1531s) 针对 SaaS 工具蔓延与开源授权陷阱，可部署专门的自动化监管 Agent 持续巡检供应商条款变更及生产环境知识产权风险。 - [27:18](https://www.youtube.com/watch?v=sJ0-fOALML0&t=1638s) 推行 AI FinOps 成本管控，通过提示词路由器（Prompt Router）按任务复杂度分流模型，并践行“以需为先”（Lead with Need）原则，以明确的北极星指标和长期 ROI 评估立项。 - [34:52](https://www.youtube.com/watch?v=sJ0-fOALML0&t=2092s) 通过敏捷研讨会（Sprint）结合 AI 实时旁听评估，企业能够在半天内完成商业价值测算，并在沙箱中快速交付功能完备的工作原型。 - [36:35](https://www.youtube.com/watch?v=sJ0-fOALML0&t=2195s) 经受时间检验的架构资产包括主动生成代码修复 PR 的架构评估框架，以及能在极短时间内为多租户交付完全隔离环境的“基础设施工厂”。 - [40:55](https://www.youtube.com/watch?v=sJ0-fOALML0&t=2455s) 面对企业内部智能体泛滥，应采取研发模板内置护栏、企业侧开箱即用可观测性以及针对第三方低代码平台的监管 Agent（Meta-Agent）反向巡检等多层次治理策略。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=sJ0-fOALML0&t=4s) 圆桌开场与嘉宾背景：AI 架构的核心优先级 圆桌讨论在探讨企业架构如何在 AI 时代实现敏捷、快速与灵活应变中拉开序幕。主持人 Jennifer Conrad 首先邀请四位企业技术掌门人介绍各自的业务背景以及当前最为紧迫的技术路线重心。来自 Quickbase 旗下 Pave 的战略与业务分析总监 Michael Maher 率先发言，他带领的团队主要负责将内部的运营数据、产品数据和市场拓展（GTM）数据转化为全公司的 AI 驱动决策。在他看来，当下最关键的任务并非单纯堆砌算法，而是让这些 AI 生成的推荐具备“可信赖度与合规治理”（trusted and governed）。 紧接着，Redwood Software 的工程高级副总裁 Jason Kehl 阐述了其作为全球后台系统编排者的核心使命。Redwood 致力于将传统系统与现代技术融为一体，横跨 SAP、Oracle、JD Edwards 以及各大主流云平台。Jason 指出，当前行业在认知上存在巨大断层：人人都在调用 AI 强大的推理与建议能力，但 AI 往往只停留在“顾问”角色，并未真正参与系统的实际操作。Redwood 当前无论是服务内部还是赋能超大规模客户，最大的技术抓手就是推动 AI 从单纯的“推理思考”跨越到具有严格审计追踪、问责机制和受控环境的“运营执行”。 智能穿戴设备巨头 Oura 的首席企业架构师 Christopher Lundy 则带来了超高速成长型企业的视角。他打趣道如果现场还有谁没戴 Oura Ring，现在“正打折促销，赶紧入手”。幽默之余，他直言 Oura 现阶段的最高准则是如何在激进创新的同时安全防范工具蔓延（contain the sprawl）。Oura 目前拥有 1,800 名员工，企业内部同时管控着超过 18 种企业级 AI 工具，架构团队的核心职责就是为终端用户建立起安全可控的创新通道。随后，业务横跨众多行业的多元化企业集团 The Friedkin Group 首席架构师 Ajit Singh 总结了集团的两大战略方向：一是将 AI 深度融入企业文化，提升全员的 AI 认知素养；二是打造通用的 Agent 与应用开发底层平台，让各个业务子公司在解决实际商业痛点时不必重复造轮子。 - **Michael Maher（Pave / Quickbase）**：聚焦打通运营与业务数据，建立可信且受控的 AI 决策闭环。 - **Jason Kehl（Redwood Software）**：专注于后端系统编排，推动 AI 从只提建议的“推理顾问”演进为可审计、可问责的“实际运营执行者”。 - **Christopher Lundy（Oura）**：在 1,800 名员工规模的高速成长架构中，安全纳管 18 款以上企业级 AI 生产力工具。 - **Ajit Singh（The Friedkin Group）**：推动跨行业财团的 AI 文化普及，搭建企业级 Agent 基础设施底座以杜绝重复建设。 --- ### [03:13](https://www.youtube.com/watch?v=sJ0-fOALML0&t=193s) 落地实践验证：服务台智能体与 Oura 的 SRE 自动化大军 在谈及已经见效的具体 AI 项目时，Ajit Singh 分享了 The Friedkin Group 平台验证的标杆案例。为了检验新建 AI 平台的实用性与健壮性，他们首先打造了一个面向全集团的服务台智能体（Help Desk Agent）。这一试点项目带来了三大立竿见影的收获：首先是主动暴露了平台底座原本存在的架构断层，便于团队及时填补缺陷；其次是通过实战沉淀出哪些方案可行、哪些不可行，避免旗下各大业务线在后续落地中重蹈覆辙；最直观的商业价值在于，该智能体已经独立承接并解决了超过一半的 IT 支持工单，将人工服务台的工单负荷直接腰斩。 ```mermaid flowchart TD accTitle: Oura SRE Agentic Workforce Architecture subgraph Detection["多维监测与智能分析"] A["跨产品日志与运行环境"] --> B["SRE 智能体群"] B --> C["异常模式识别与横向比对"] end subgraph Orchestration["编排管理"] C --> D["MCP 服务器统一编排"] end subgraph Remediation["修复与交付"] D --> E["向技术团队输出修复建议"] D --> F["直接生成并提交修复代码 PR"] end ``` 转到 Oura 这边，Christopher Lundy 坦言他们面临着截然不同却极为典型的成长烦恼：公司业务爆发式增长，而团队编制极为精简，人员招聘的速度根本赶不上业务拓展的步伐。为了打破人手瓶颈，企业架构团队全力押注于“智能体劳动力”（Agentic Workforces）。他们构建了一套由 MCP（Model Context Protocol）服务器进行统一编排的 SRE 智能体集群。这些智能体超越了单纯的日志排障，它们能够跨越多条产品线横向比对底层运行环境的异常模式，主动向工程团队提供修复建议；在特定场景下，智能体甚至可以直接生成并提交 PR（Pull Request）来自动修复代码缺陷，彻底免去工程师的手动干预。Christopher 披露，目前 Oura 内部每周有 10 到 20 个新 Agent 部署上线，全公司利用 AI 构建并投入日常使用的应用数量已经超过 4,000 个，如果没有自动化的智能体部队，日常运维支持系统早就难以维系。 - **The Friedkin Group 服务台成效**：通过构建自研 Helpdesk 智能体，直接将人工支持工单量减少了 50%，同时完成了集团 AI 平台底座的能力验证。 - **Oura 智能体规模数据**：每周新增并投产 10 至 20 个智能体；全员构建的 AI 原生应用总量已突破 4,000 个。 - **Oura 运维革新**：SRE 智能体借助 MCP 服务器协同，实现从环境异常检测、跨产品关联分析到自动化提交代码 PR 的全链路闭环。 --- ### [05:36](https://www.youtube.com/watch?v=sJ0-fOALML0&t=336s) 编排与反馈：工作图谱协同与销售推荐决策闭环 随着智能体的大量涌现，协同失控与决策黑盒成为最棘手的挑战。Jason Kehl 强调，人类做决策原本节奏缓慢且有迹可循，而 Agent 的决策速度极快，系统往往很难捕捉决策链路。在多 Agent 协同场景中，智能体经常会出现逻辑陷入死循环、彼此冲突甚至“当场互相争吵”（literally argue with each other）的荒唐局面。针对这一痛点，Redwood Software 正在重构一套“运营调度者”（Operator）机制。该方案通过绘制清晰的工作图谱（Work Graph），精细定义 Agent 之间的通信边界与依赖路径，明确在哪些关键决策节点必须暂停并引入“人机协同”（Human-in-the-Loop）干预。这种审计机制能够准确回答“系统为什么做出此决策”、“具体是谁触发的”以及“这是否为最优解”，从而为高合规行业客户提供真正的治理保障。 与此同时，Michael Maher 介绍了 Quickbase 与 Pave 内部正在运转的智能推荐引擎——“Next Best Call”。该工具旨在根据客户多维信号，为销售代表自动匹配最佳跟进策略。团队最初借助 Snowflake MCP 搭建 Snowflake 智能体将底层数据对接至 Claude，并通过 Workato 等自动化流将每日建议推送到销售最熟悉的 Slack 中。然而，团队很快遭遇了工具瓶颈：Slack 这类聊天应用缺乏后续执行追踪能力，更无法收集高质量的业务反馈，仅仅依靠简单的“点赞或点踩”（thumbs up or thumbs down）根本无法使模型掌握销售未采纳推荐的深层原因。 为了解决这一断层，团队将这套推荐系统深度整合进全新的 Pave 产品中，打造出一体化工作台与强反馈闭环。销售人员在系统内处理日常业务的同时，能够清晰标记推荐是否切中要害。一旦一线销售捕捉到全新的客户触达模式，系统还能直接把业务人员的自然语言构想交给 Agent 进行第一轮代码编写与数据清洗，迅速生成可运行的 MVP，彻底省去了漫长的传统工程研发排期。在此之上，针对权限分配过于依赖管理员手动配置角色的历史弊端，他们正依托 Quickbase 积淀 20 年的权限体系，让构建者能够直接用自然语言描述合规要求（例如“业务代表仅能查看自己名下的客户名录”、“财务人员只能查看账单数据且严禁触碰敏感隐私”），系统便会自动将其编译为底层角色权限与审计规则。 - **多 Agent 交互困境**：决策链路不透明、智能体间死锁对抗、缺乏审计记录。 - **Redwood 编排架构**：引入 Work Graph（工作图谱）管理多智能体协作，在关键节点挂载 Human-in-the-Loop 人工干预卡点。 - **Pave “Next Best Call” 演进**：放弃仅依赖 Slack 推送与简易点赞反馈的模式，转入统一的 Pave 应用闭环，实现从一线销售反馈到 Agent 自动编写 MVP 代码的原型快速落地。 - **自然语言权限治理**：以自然语言策略直接驱动企业级 RBAC 权限与审计规则生成，继承 Quickbase 20 年的底层管控沉淀。 --- ### [09:52](https://www.youtube.com/watch?v=sJ0-fOALML0&t=592s) 护栏与自由：在严苛合规与开发者创造力之间寻找支点 在迈向下一个议题前，现场围绕“创新自主权”与“集中合规控制”的博弈展开了热烈探讨。Jason Kehl 与 Christopher Lundy 补充指出，企业在技术架构演进中最常陷入的纠结，就是如何在激发生命力的同时守住底线。管理层希望每一位开发者和一线员工都能发挥天马行空的创造力，绝不能让死板的规章制度与繁冗流程成为绊脚石；但另一方面，现实运行中又必须有不可逾越的安全护栏。理想的架构方向是将治理能力下沉为由中央统一托管的底层平台服务，使一线团队既享有决策权和所有权，又无需背负沉重的治理运维负担。 这种静默而坚固的后台治理在强监管行业尤为生死攸关。Christopher Lundy 以 Oura 所在的健康操作系统（Health OS）为例补充强调，医疗健康领域有着极其严苛的合规指南。令人深思的是，很多现行的行业监管标准其实是在上世纪 90 年代末到 2000 年代初制定的，这些条文在制定时根本无法预见当下由大模型驱动的 AI 时代。然而，现代企业别无选择，必须通过能够自我证明、严密可追溯的现代企业架构，在不打扰开发者创新的前提下，全方位满足那些古老而严谨的法定合规要求。 ### [10:51](https://www.youtube.com/watch?v=sJ0-fOALML0&t=651s) 生命科学领域的合规壁垒与内置化平台治理 在生命科学领域推进技术创新往往面临极其严苛的现实约束。Christopher Lundy 指出，该领域必须遵循如美国 FDA（食品药品监督管理局）等机构的合规要求。虽然监管机构的政策正在逐步向 AI 开放，但企业依然受到必须严格按字面执行的标准作业程序（SOP）的约束，每一步操作都必须保留完整的追溯证据，而这背后最核心的支柱就是“治理”（Governance）。这带来了一个极为尖锐的矛盾：团队希望通过捕捉人体生理特征与信号来开发具有突破性的创新医疗或健康产品，但如果按照传统生物制药的研发周期，动辄陷入长达六个月的系统验证（Validation）流程中，技术创新根本无法按时落地。这两种工作节奏完全无法契合。 破解这一矛盾的关键在于思维方式的转变：必须在前端给予研发人员充分的思考与创新自由，让他们放手尝试 AI，并将治理控制移至流程或后端，以此确保在安全可控的前提下快速推进。Ajit Singh 随即分享了他们的解法——直接将治理机制深度内置在底层平台中。通过在平台原生嵌入合规控制能力，业务人员在探索和释放创造力时便无需为复杂的治理规则分心，因为合规防线早已在基础设施层静默运行。 ```mermaid flowchart LR accTitle: 平台级内置治理与业务创新的解耦架构 subgraph DevLayer["研发创新层（高灵活性）"] A["生理信号与人体数据挖掘"] --> B["创新 AI 模型 / Agent 开发"] end subgraph PlatformLayer["底层统一治理平台（内置合规）"] C["平台级内置 SOP & 证据链追踪"] D["全局权限控制 (RBAC / ABAC)"] end B --> PlatformLayer PlatformLayer --> E["免去 6 个月人工验证的快速产品交付"] ``` - **验证周期痛点**：传统生物医药及生命科学产业的标准验证流程通常长达 6 个月，如果将 AI 创新直接套入旧有的审计路径，创新周期将被严重拖垮。 - **治理内置化**：将 SOP 证据收集、审计跟踪等治理需求沉淀为平台底层能力，使前端开发与合规合流，兼顾速度与合规。 --- ### [12:08](https://www.youtube.com/watch?v=sJ0-fOALML0&t=728s) 抽象解耦：应对模型频繁更迭与外部服务商条款变更 当主持人询问这种内置治理是否会导致系统开发交付的瓶颈时，Ajit Singh 直言确实遭遇过阻碍，其中最典型的瓶颈就是基于角色的访问控制（RBAC）与基于属性的访问控制（ABAC）的权限配置。对此，Jason Kehl 强调了多层架构抽象的紧迫性与必要性。当下 LLM 和底层 AI 技术每天都在推陈出新，团队也在持续创建性能大幅优于旧版的新一代 Agent；如果把权限、合规与治理逻辑直接绑死在某一个具体的模型或智能体上，简直是“不可理喻的疯狂之举”（"tying any of that at that level to the agents LLM... is insane"）。企业必须在模型之上构建更高维度的编排与平台层，保持治理逻辑的绝对稳定，从而在底层模型瞬息万变的外部环境下，依然能够确保系统的可靠性。 随后，Christopher Lundy 补充了一个所有依赖第三方大模型厂商的企业都会深有感触的现实痛点。Anthropic 和 OpenAI 等供应商经常会在一夜之间单方面修改服务条款（Terms and Conditions），突然停止替企业存储提示词（Prompt）和响应历史数据。这种变更来得极其突然：“我们一上班就看到了新条款，只能立刻问自己接下来该怎么做”。这意味着数据沉淀与追溯的责任被瞬间甩回给了企业自己。 更致命的是，企业虽然可以自行记录提示词和模型响应文本，但随模型原生返回的那些极为关键的丰富元数据（Rich Metadata）——例如系统运行表现、模型评估指标、幻觉（Hallucination）发生率等指标数据——彻底消失了，必须单独向厂商发起特殊申请才能获取。面对这种外部冲击，团队不得不全面重构内部系统设计：将治理机制提升到超越单个用户乃至业务平台的企业级流程高度，集中捕获和挖掘所有交互数据，以此建立更加稳固可靠的管控屏障。 --- ### [14:15](https://www.youtube.com/watch?v=sJ0-fOALML0&t=855s) 全球合规下的数据治理重构：从人工脱敏到 AI 智能分类 在跨国企业的实际运作中，治理策略必须直面全球各地法律法规的割裂现状。Christopher Lundy 解释道，跨国企业的数据往往分散驻留在全球数十个不同的主权国家中，每个国家都拥有截然不同的数据合规标准。企业不能被动受制于这些规则而导致业务瘫痪，必须将跨国合规逻辑内生化至整体系统架构中。 在过去的传统做法中，如果想要把欧洲本土存储的数据用于美国分支机构的业务，通常只能采取粗暴的数据脱敏（Data Obfuscation）方式，彻底屏蔽敏感字段，但这往往导致数据可用价值急剧缩水。如今，借助现代 AI 技术栈，企业能够重塑治理模型：AI 能够在合规边界内直接读取并聚合底层数据，提取出有价值的业务信号和高阶洞察，仅将聚合后的洞察跨国境共享，并与其他系统进行比对得出结论。同时，企业彻底颠覆了以往依赖人工做数据分类打标的旧机制，转而让 AI 自动分析并识别数据层级，再由 AI 向数据管理专员（Data Stewards）和业务负责人给出分类建议及所需元数据的规范建议，大幅提升了数据合规运营的流转效率。 - **多国合规支持**：AI 架构需原生理解跨国合规差异，通过生成聚合洞察而非搬运原始敏感明细的方式合法合规地实现全球业务联动。 - **智能化数据分类**：改变人工定义标签与数据治理的老方法，由 AI 主动推荐数据分类等级和必填元数据属性。 --- ### [15:51](https://www.youtube.com/watch?v=sJ0-fOALML0&t=951s) 确定性与概率性之辩：传统系统不能轻易推倒重来 谈及大型复杂遗留系统（Legacy Systems）的现代化改造时，Ajit Singh 从计算本质的维度对传统系统与 AI 系统做出了极具深度的剖析。他指出，两类系统在运行逻辑上存在根本性的分野： | 维度 | 遗留系统（Pre-AI Systems） | 现代 AI 系统 | |:--- |:--- |:--- | | **计算范式** | **确定性（Deterministic）** | **概率性（Probabilistic）** | | **可靠性表现** | 经过 5 年、10 年乃至 20 年严密验证，执行结果绝对稳定、非黑即白 | 基于概率预测生成 Token，即便是顶尖模型，原生正确率往往也只有 60%~70% | | **业务契合度** | 极度适合金融结算、账务流水等对容错率为零的核心业务基础设施 | 擅长模糊语意推理、意图理解与自然语言生成 | | **输入与解析** | 需要开发人员硬编码将模糊需求逐层转换为数学公式、代码管线与结构化数据 | 能够直接解析人类自然语言的真实意图并自主规划执行 | Ajit 以全球金融服务行业举例：全球每天发生的高达 100 亿笔金融交易全部建立在严密的确定性系统之上。即使假设未来 AI 模型的准确率提升到了不可思议的 99%（即每 100 次调用只犯 1 次错），应用在 100 亿笔交易场景下，每天也将产生整整 1 亿笔错误交易——这足以让整个世界金融基础设施彻底崩溃。因此，稳定运转了数十年的确定性遗留系统必须继续执行其底层职责。 AI 带来的根本颠覆在于它赋予了软件“推理能力”（Reasoning）和“理解意图”的能力。以往业务人员提出一句简单的需求（例如“我想看看目前有多少客户正在流失”），研发人员需要耗费数天编写全套数据清洗 Pipeline，建表跑批后再丢进模型出结果；如今 AI 能够瞬间读懂业务意图。因此，现代架构演进的终极命题并非“用 AI 替换旧系统”，而是探索如何让确定性的遗留基座与概率性的推理大脑无缝协作。 ```mermaid flowchart TD accTitle: 确定性遗留系统与概率性 AI 架构的协作关系 subgraph IntentLayer["意图与交互层（概率性推理）"] User["业务自然语言诉求: '查看流失客户'"] --> AI["AI 推理引擎 (理解业务意图)"] end subgraph LegacyLayer["企业核心底层（确定性执行）"] MCP["中间封装层 (MCP 协议)"] Core["运行 10~20 年的遗留业务系统"] end AI -->|"规范化指令调用"| MCP MCP --> Core Core -->|"高可靠性计算数据输出"| MCP MCP -->|"结果汇总交付"| AI ``` --- ### [18:32](https://www.youtube.com/watch?v=sJ0-fOALML0&t=1112s) 用 MCP 封装对抗重构冲动，借助大模型重塑陈旧业务流 Jason Kehl 会心一笑，坦言作为软件工程师，面对看不懂或者老旧的代码时，本能的冲动永远是“把它们全推倒重写”。但现实中每一个试图重写老系统的人最终都会碰得头破血流（"get burned"），因为那些看似丑陋的代码和陈旧的业务流程中，实际凝结了企业多年踩坑总结出的核心业务智能（Business Intelligence）与实战经验，盲目重构只会陷入巨大的痛苦之中。 Redwood Software 对此采取的破局方案是“只封装，不重写”：保持客户原有的遗留业务系统和处理逻辑不动，在其外部通过模型上下文协议（MCP, Model Context Protocol）服务器搭建包装层（Wrapper）。这相当于为老系统穿上一层现代接口外衣，让各类 AI Agent 能够轻松调度老系统的功能、向其传递参数并接收确定性的执行结果。这不仅规避了丢失核心业务逻辑的风险，也解除了大家因系统年代久远、无人通晓内部代码细节而不敢触碰的恐惧。 ```mermaid flowchart LR accTitle: MCP 封装遗留系统架构 Agent["新一代 AI Agents"] -->|"标准化 MCP 协议调用"| MCP["MCP Server 封装层"] MCP --> Legacy["无人敢动的陈旧遗留系统 / 确定性业务逻辑"] Legacy -->|"可靠执行结果"| MCP MCP -->|"反馈结构化上下文"| Agent ``` 随后，Christopher Lundy 介绍了 Oura 截然不同的云原生技术栈：Oura 全公司拥有 500 多个应用程序，其中仅有区区 1 套本地部署（On-prem）系统，绝大多数都是现代云端 SaaS 应用。但他们面临的另一重挑战是外部 SaaS 供应商的掉队问题——如果某些供应商的 Roadmap 上根本没有 AI 规划，Oura 会直接拒绝采购合作；对于硬件研发等少数必须依赖的非 AI 软件，他们同样会通过 MCP 进行封装，强行将其技术债务与核心架构隔离开来。 最后，曾有多年保险行业从业经历的 Christopher 为深陷极度古老系统泥潭的企业支招（他还风趣地向台下致歉：“如果你在保险行业工作，真的很抱歉，那里的系统确实太陈旧了”）。很多落后企业不敢变革，往往是因为“既定业务流程本身”被神化成了不可动摇的教条，但这些流程大多是 5 到 10 年前技术手段匮乏时妥协的产物。与其盲目重构，不如借助 Anthropic 或 OpenAI 等前沿大模型对现存业务流程进行逆向工程分析： 1. **行业对标诊断**：将复杂的现行业务流程直接喂给大模型，让 AI 结合当前最新的行业最佳实践指出流程中的冗余与缺陷； 2. **自动化可行性设计**：让大模型自主设计并输出全套自动化重构方案； 3. **领域专家介入校对**：将 AI 提出的重组方案交由企业业务专家（SME）核对纠错。 Christopher 强调，哪怕 AI 提出的流程诊断方案中只有 70% 是正确的，这对于企业也是极具价值的突破——它的核心收益往往不是自动化工具本身，而是逼迫企业终于把糊里糊涂运行了多年的底层业务逻辑彻底梳理得一清二楚。 ### [21:42](https://www.youtube.com/watch?v=sJ0-fOALML0&t=1302s) 保持模型与应用中立：在技术迭代与监管浪潮中保持敏捷 在系统现代化改造中，单纯推倒重建（rip and replace）往往代价巨大，更合理的做法是采用封装（wrap）或双轨并行策略，借助最新工具重新构想业务流程。面对 AI 技术极快的迭代节奏（新模型每两周甚至每月都在更新）以及潜在的地缘监管风险（例如某些工具在美国被禁），企业架构必须具备极高的敏捷性。首要原则就是**保持模型中立（Model Agnostic）**，确保上层应用可以在不同供应商的模型之间无缝切换。 此外，企业内部的合规体系也必须与工程紧密协同。团队通过引入专门的 AI 法律顾问和技术法务专家，将海量复杂的 AI 新规提炼并整合为可落地的内部合规红线，确保工程实践不触及法律边界。在具体工程实现上，团队为通用 Agent 角色（Personas）配置了首选模型切换能力，不仅能按需动态换用更适配的模型，更能避免 7x24 小时运行带来的账单失控问题。 ```mermaid flowchart TD accTitle: Model Agnostic Enterprise Architecture subgraph Governance["合规与法律护栏"] LAW["AI 法律与技术法务顾问"] --> RULES["合规标准整合与红线制定"] end subgraph Orchestration["Agent 与调度层"] RULES --> AGENT["通用 Agent (Persona)"] AGENT --> ROUTER{"模型路由与策略切换"} end subgraph Providers["多元模型生态 (即插即用)"] ROUTER --> M1["轻量/成本优化模型 (日常任务)"] ROUTER --> M2["前沿通用大模型 (复杂推理)"] ROUTER --> M3["自研/开源微调模型 (专用领域)"] end ``` - **架构解耦教训**：遵循架构基础原则（Architecture 101），确保每个组件都是即插即用（Plug and Play）。曾有企业因为某核心 SaaS 供应商大幅变更商业模式与产品形态，花费约 3 个月时间才彻底将其从系统中剥离。 - **动态模型分配**：根据任务轻重匹配模型，日常高频任务优先选用轻量模型以控制成本，避免“大炮打蚊子”（Overkill）。 --- ### [25:31](https://www.youtube.com/watch?v=sJ0-fOALML0&t=1531s) SaaS 蔓延与开源授权陷阱：利用 AI Agent 进行全天候监控 随着各家 SaaS 软件争相内置 AI 能力，其背后的模型变更与条款调整带来了巨大的隐性风险。例如，当 SaaS 供应商将其底层模型从 OpenAI 悄然切换为 Claude、Grok 或 Mistral 时，可能会对企业现有的业务和数据安全产生不可预估的冲击。为此，架构团队构建了专门的自动化 Agent，每周一早晨自动巡检所有在用应用的官网条款、隐私政策及 AI 使用规范，一旦检测到变动立即向团队告警。 开源模型同样潜藏着严重的商业化授权危机。如果团队使用开源模型进行 NLP 研发并上线生产环境，某些托管在 Hugging Face 等平台上的模型在细则中会附带知识产权主张条款（Claim IP），一旦涉及数十亿美元估值的大型企业，这类授权陷阱将直接导致产品合规失效或巨额收入损失。因此，利用 Anthropic Claude 协作工具或 Lovable、Pave 等自动化构建工具来管理企业集成资产，已成为降低此类风险的必要手段。 - **每周一自动审计**：自动化 Agent 持续抓取所有 SaaS 站点的 Terms & Conditions 与 Privacy Policy，比对 AI 策略变化。 - **开源许可证风险**：严格排查带有“进入生产环境即主张衍生知识产权”的开源 AI 模型条款，规避潜在的法律纠纷。 --- ### [27:18](https://www.youtube.com/watch?v=sJ0-fOALML0&t=1638s) 成本与算力博弈：提示词路由、调度平台与企业级 Token 预算 在平衡云端与本地（In-house）算力时，企业通常仅在本地保留极少数专用 AI 工作负载（或用于高性能计算实验），绝大多数流量依然跑在云端。在预算管理策略上，许多企业在前期往往将**鼓励创新与业务实验**置于严格治理和成本削减之上，导致企业级 Token 预算屡屡超标。然而，随着组织迈向成熟期，“预算爆仓”已不再被容忍，团队正通过工程化手段重塑成本管控。 在具体落地方案上，两大抓手起到了关键作用：一是引入**提示词路由器（Prompt Router）**，根据任务复杂度智能分流，普通摘要与文档处理自动分配轻量模型，仅关键任务调用前沿大模型；二是打造统一的**编排平台（Orchestration Platform）**。目标是将 60% 至 70% 的工程师日常任务统一接入该平台，通过微调专用 Agent 和集中化监控，防止成本无序失控，同时允许企业客户自带模型（BYO LLM）以实现商业化变现。 ```mermaid flowchart LR accTitle: Enterprise AI Cost Control and Prompt Routing REQ["企业用户 / 内部调用请求"] --> PR["Prompt Router (提示词路由器)"] PR -->|"日常摘要 / 简单文本"| LM["轻量成本优化模型"] PR -->|"复杂逻辑 / 深度推理"| FM["Frontier 前沿基座大模型"] LM & FM --> ORCH["统一编排平台 (Orchestration Platform)"] ORCH --> COST["集中式 Token 预算与用量监控"] ``` - **面向外部客户**：支持客户自带模型（Bring Your Own LLMs），或由平台提供统一模型服务并按需计费。 - **内部调用治理**：通过编排平台收拢绝大部分研发工作流，实现对用量、Agent 行为和模型调用的端到端可视化。 --- ### [29:46](https://www.youtube.com/watch?v=sJ0-fOALML0&t=1786s) 集成平台蔓延与“AI FinOps”：技术与业务的“以需为先” 当前许多内置 AI 的新型软件（如各类带有 App Studio 或连接器平台的 SaaS 工具）都试图直连企业数据。这导致了严重的连接器蔓延与“重复付费”困境：用户在 Workato 中用 AI 生成连接 Recipe，又跑到 Zip 中使用同类提示词构建工作流，不仅造成逻辑碎片化（类似过去“某员工十年前写的孤岛代码无人能懂”），还导致企业就同一业务需求向不同模型重复支付多次调用成本。这种经济学压力促使企业必须建立超越传统 ROI 与 TCO 框架的 **AI FinOps（AI 财务运营）** 体系。 针对技术团队追求前沿酷炫（Flashy）与业务团队死磕投资回报率（ROI）之间的天然矛盾，企业推行了**“以需为先”（Lead with Need）**的核心信条。任何 AI 项目立项的前提必须是明确的业务痛点，并由专门的财务评审团队对每个用例进行多维度 ROI 测算——不仅评估上线当下的即时效益，更着眼于未来 2 至 3 年的长期运营价值与可持续回报。 - **多重付费陷阱**：在 Workato、Zip 等多个系统内重复用 AI 生成自动化工作流，造成算力与工具成本的无形浪费。 - **Lead with Need 原则**：所有 AI 投资必须由业务真实需求驱动，并结合 2 到 3 年维度的长期财务模型进行立项审核。 ### [32:43](https://www.youtube.com/watch?v=sJ0-fOALML0&t=1963s) AI 投资回报率（ROI）与业务需求驱动的快速见证 在探讨企业如何落地 AI 时，Ajit Singh 首先强调了需求驱动与长期视角的重要性。在项目初期，由于前期构建和基础设施的持续投入，ROI 往往是负面的，但核心在于“它随着时间的推移能否收回成本并创造价值”。当进入技术实现阶段时，团队倾向于寻找能快速见证成效的“速赢”（Quick Win）切入点。这不仅能向业务部门快速证明概念并展示潜在 ROI，也能在技术侧建立信心。如果现有平台无法完美支撑这些快速实现，技术团队会同步识别差距并在后续迭代中将其拉回统一平台体系，采取多维度的推进策略。 Michael Maher 接过话茬，风趣地指出：“今年是‘Token 最大化’（Token maxing）的一年，但我坚信明年必须是‘AI ROI 之年’。”他解释道，由于 AI 让“一切皆有可能”，企业面临着无休止的需求积压（Infinite Backlog）。为了解决优先级问题，Pave 建立了一套严密的需求准入流程（Intake Process），要求业务端在立项之初就必须明确北极星指标（Northstar Metric）与预期的 ROI，以此迫使业务方在动手前进行全盘思考。 Jason Kehl 则补充了 Redwood Software 的视角：无论是面对客户还是内部协作，关键都在于“聚焦最终成果（Outcomes）”，而不是把 AI 当作“看什么都是钉子的锤子”。谈论实际成果与业务目标，远比空谈技术流程或“我们在用 AI”更能与业务伙伴达成共识。 ```mermaid flowchart TD accTitle: AI需求准入与ROI优先级评估流程 subgraph Intake["需求准入与业务定义"] A["业务挑战 / 痛点提出"] --> B["定义北极星指标与预期 ROI"] B --> C{"通过准入优先级筛选?"} end subgraph Evaluation["技术实施与架构对齐"] C -->|否| D["退回重构或放入无限积压库"] C -->|是| E["快速见证方案 (Quick Win)"] E --> F{"是否符合统一平台架构规范?"} F -->|是| G["投入生产并持续监控业务成果"] F -->|否| H["识别平台差距并逐步重构收敛"] H --> G end ``` --- ### [34:52](https://www.youtube.com/watch?v=sJ0-fOALML0&t=2092s) 适配版 Google Sprint：现场研讨与智能体实时原型构建 Christopher Lundy 分享了 Oura 内部 AI 生产力团队的特色实践。他们借鉴了 Google 经典的 5 天设计冲刺（Google Sprint）理念，将其改造为敏捷的现场研讨会（On-site Workshop）。因为团队高度分布式，他们会把不同业务团队召集到现场： - **上午痛点分析与智能体协同评估**：研讨会的前几个小时，业务团队详细阐述其在流程、人员配置或执行速度上面临的挑战。此时，Christopher 与架构师团队在旁协同，背后的数个 AI Agent 正在实时“旁听”对话。 - 一个 Agent 会根据讨论内容自动撰写商业案例（Business Case），精确量化自动化带来的独立业务价值，甚至会在不可行时直接提示“不要走这条路”； - 另一个 Agent 则会预测落地所需的人力与资源投入。 - **化解业务部门的“影子软件采购”隐患**：很多业务负责人常会宣布“我刚买了套新软件，两周内必须上线”，却完全没有与 IT 或 AI 团队对齐。Christopher 团队会直接用数据算账：“按照常规流程，实施这套系统将耗费 30 万美元，你想知道怎么优化吗？”业务方通常会欣然接受引导。 - **下午交付沙箱可用原型**：上午完成潜力识别后，下午团队就能在沙箱环境中生成一个功能完备的工作原型（Working Prototype）。在内置 Agent 的全流程治理保障下，该原型甚至可以直接具备推向生产环境的条件。 --- ### [36:35](https://www.youtube.com/watch?v=sJ0-fOALML0&t=2195s) 系统自动化与经受时间检验的架构决策 进入现场问答环节，观众询问在不使用现成工作流产品时，企业如何构建底层智能体平台。Jason Kehl 说明，Redwood Software 的核心专长在于“系统级自动化”（针对 SAP、Oracle 等底层大型企业系统），而非单纯的“人类操作自动化”；Ajit Singh 则表示，其企业在处理工作流时的首要原则是“重构与消除”——先审视该流程是否真正必要，能否直接优化删减；如果确实需要，再采用简洁纯粹的软件代码来实现。 主持人随后提问：在技术瞬息万变、工具层出不穷的环境下，有哪些技术决策能经受住时间的检验？ - **扩展的良好架构框架（Well-Architected Framework）与主动代码修复（Christopher Lundy）**：Oura 在主流云厂商的架构框架基础上进行了深度扩展，融合了 6R、CMMI 等体系。该方案的核心在于“关注员工为什么做（Why）以及对组织的影响，而非具体怎么做（How）”。系统自动扫描代码仓库后，会输出覆盖 8 个维度的评分报告。更关键的是，他们不只是把报告扔给架构评审委员会去批评团队，而是由 Agent 直接生成修复用的 Pull Request（PR）提交回仓库，并根据团队迭代节奏（Sprint Cadence）对整改任务进行优先级排序，帮助业务工程团队无痛整改。 - **基础设施工厂（Infrastructure Factory）与多租户极速隔离（Ajit Singh）**：作为拥有众多子公司的集团财团，The Friedkin Group 在 2024 年 1 月启动 AI 战略时，做出了构建“基础设施工厂”的关键决策。当子公司提出环境需求时，工厂通过全自动化软件流水线在 3 小时内即可交付一套完全隔离、杜绝任何数据泄露风险的环境（涵盖 Sandbox、Dev、Test），供其训练模型和运行应用。目前，该架构仅靠 2 名平台工程师就支撑了超过 50 个相互隔离的业务环境，并已无缝扩展至智能体 AI（Agentic AI）的支撑中。 --- ### [40:55](https://www.youtube.com/watch?v=sJ0-fOALML0&t=2455s) 应对 Agent 激增：分层治理、全链路可观测性与反向巡检 针对现场观众关于“企业内部智能体泛滥（Proliferation of Agents）如何管理、编目与度量影响”的提问，Christopher Lundy 详细拆解了 Oura 的双轨治理机制： ```mermaid flowchart LR accTitle: 企业级 Agent 双轨治理与监控架构 subgraph DevStream["工程研发侧 (约400名工程师)"] A1["工程开发模板"] --> A2["内置安全护栏<br/>(AI委员会/EA/IT规范)"] A2 --> A3["合规 Agent 交付"] end subgraph EntStream["企业业务侧 (企业应用/无代码平台)"] B1["企业 Agent 部署"] --> B2["出厂自带可观测性<br/>(Observability Out of the Gate)"] B2 --> B3["集中式 IT 运维分析域<br/>(统一监控与告警)"] end subgraph ShadowAgent["第三方/低代码平台监控 (如 Glean, Lovable)"] C1["低代码/零代码构建"] --> C2["监管 Agent 反向巡检<br/>(Supervisor Agent)"] C2 -->|偏差信号| C3["触发偏差告警并人工介入对齐"] end ``` - **研发侧模板内置护栏**：针对公司约 400 名软件工程师，提供标准化的开发模板。模板在初始化时就已经由 AI 委员会、企业架构部（EA）及 IT 部门固化了安全护栏（Guardrails）。 - **企业侧开箱即用的可观测性**：在企业运营侧构建的 Agent，从上线第一天起就内置了全套可观测性组件（Observability Out of the Gate）。无论部署在何处，其运行监控与告警数据均全部集中回流至 IT 运维分析平台。 - **针对外部低代码/SaaS 平台的监管 Agent**：对于员工在 Glean 或 Lovable 等第三方平台自行搭建、脱离标准流水线的 Agent，团队专门在其上层构建了“监管 Agent”（Meta-Agent）。该智能体会持续监测第三方平台中的构建活动，一旦检测到偏离企业治理规范的行为，立即向架构团队发送信号并触发人工介入沟通。 研讨在热烈的掌声中顺利结束。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=sJ0-fOALML0&t=4s) contain the sprawl — Christopher Lundy > [05:36](https://www.youtube.com/watch?v=sJ0-fOALML0&t=336s) literally argue with each other — Jason Kehl > [05:36](https://www.youtube.com/watch?v=sJ0-fOALML0&t=336s) thumbs up or thumbs down — Michael Maher > [12:08](https://www.youtube.com/watch?v=sJ0-fOALML0&t=728s) tying any of that at that level to the agents LLM... is insane — Jason Kehl > [12:08](https://www.youtube.com/watch?v=sJ0-fOALML0&t=728s) 我们一上班就看到了新条款，只能立刻问自己接下来该怎么做 — Christopher Lundy > [18:32](https://www.youtube.com/watch?v=sJ0-fOALML0&t=1112s) get burned — Jason Kehl > [18:32](https://www.youtube.com/watch?v=sJ0-fOALML0&t=1112s) 如果你在保险行业工作，真的很抱歉，那里的系统确实太陈旧了 — Christopher Lundy > [25:31](https://www.youtube.com/watch?v=sJ0-fOALML0&t=1531s) Claim IP — Christopher Lundy > [29:46](https://www.youtube.com/watch?v=sJ0-fOALML0&t=1786s) Lead with Need — Jason Kehl > [32:43](https://www.youtube.com/watch?v=sJ0-fOALML0&t=1963s) 今年是‘Token 最大化’（Token maxing）的一年，但我坚信明年必须是‘AI ROI 之年’。 — Michael Maher > [34:52](https://www.youtube.com/watch?v=sJ0-fOALML0&t=2092s) 我刚买了套新软件，两周内必须上线 — Christopher Lundy > [34:52](https://www.youtube.com/watch?v=sJ0-fOALML0&t=2092s) 按照常规流程，实施这套系统将耗费 30 万美元，你想知道怎么优化吗？ — Christopher Lundy > [36:35](https://www.youtube.com/watch?v=sJ0-fOALML0&t=2195s) 关注员工为什么做（Why）以及对组织的影响，而非具体怎么做（How） — Christopher Lundy ## Mentioned Ai4, Christopher Lundy, Michael Maher, Jason Kehl, Ajit Singh, Jennifer Conrad, Inc. Magazine, Oura, Oura Ring, Quickbase, Pave, Redwood Software, The Friedkin Group, SAP, Oracle, JD Edwards, MCP (Model Context Protocol), Snowflake, Claude, Workato, Slack, FDA, Anthropic, OpenAI, Grok, Mistral, Hugging Face, Lovable, Zip, Glean, Google Sprint ## Video description Christopher Lundy - Principal Enterprise Architect, Oura Michael Maher - Director of Strategy & Business Analytics, Pave Powered by Quickbase Jason Kehl - Senior Vice President, Engineering, Redwood Software Ajit Singh - Chief Architect, The Friedkin Group Jennifer Conrad - Senior Writer, Inc. Magazine Enterprises can’t bolt on AI—they must build for it. This session explores how forward-thinking organizations are re-architecting systems to support modular, scalable AI capabilities. Panelists will share how they’ve enabled cross-functional AI workflows, modernized legacy stacks, and built platforms that evolve with the pace of innovation. Learn what makes a truly “AI-ready” architecture and how to design for speed, scale, and sustainability. [Source](https://www.youtube.com/watch?v=sJ0-fOALML0)

[0:04] Christopher Lundy: contain the sprawl [5:36] Jason Kehl: literally argue with each other [5:36] Michael Maher: thumbs up or thumbs down [12:08] Jason Kehl: tying any of that at that level to the agents LLM... is insane [12:08] Christopher Lundy: 我们一上班就看到了新条款，只能立刻问自己接下来该怎么做 [18:32] Jason Kehl: get burned [18:32] Christopher Lundy: 如果你在保险行业工作，真的很抱歉，那里的系统确实太陈旧了 [25:31] Christopher Lundy: Claim IP [29:46] Jason Kehl: Lead with Need [32:43] Michael Maher: 今年是‘Token 最大化’（Token maxing）的一年，但我坚信明年必须是‘AI ROI 之年’。 [34:52] Christopher Lundy: 我刚买了套新软件，两周内必须上线 [34:52] Christopher Lundy: 按照常规流程，实施这套系统将耗费 30 万美元，你想知道怎么优化吗？ [36:35] Christopher Lundy: 关注员工为什么做（Why）以及对组织的影响，而非具体怎么做（How）

[5:36] Jason Kehl: literally argue with each other [29:46] Ajit Singh: We have a mantra in our enterprise. We call it lead with need. So we lead with when there is a business need. [32:43] Michael Maher: this year has been the year of token maxing [32:43] Michael Maher: next year is going to be the year of AI ROI. [32:43] Michael Maher: infinite backlog of things you can do [32:43] Jason Kehl: hammer and making everything a nail. [36:35] Jason Kehl: the easy problem

[0:04] We're going to be talking about how enterprise  architecture can be adapted for the age of AI. How [0:09] to be quick, how to be nimble, how to be flexible.  Um, we're going to try to save about 10 minutes at [0:14] the end for questions. So, you can start thinking  about those. Um, but to start, let's go down the [0:18] line. I'd love to hear a little bit about your  backgrounds and also what your priorities are [0:23] right now. Hey everyone, Michael Mahar, uh  director of strategy and business analytics [0:28] at PAVE and more broadly known as Quickbase  uh where I lead a team that's responsible [0:34] for transforming our operational data, product  data, go to market data into AIdriven decisions [0:42] uh across the business and right now my biggest  focus is making those AI recommendations trusted [0:48] and governed. Uh Jason Keel. Um I run engineering  for Redwood Software. Redwood Software. Um we kind [0:57] of orchestrate the whole back end of the world.  So like your SAPs, your oracles, your JD Edwards, [1:02] your cloud platforms and also your legacy systems.  We bring like the new and the old all together. [1:07] And our priority right now on the AI front  is both for ourselves internally but also for [1:12] our customers creating like helping to create the  transition from being AI is reasoning and advising [1:19] and helping but AI is not operating to moving from  basically the abundance of like AI reasoning that [1:24] we can all access today into how do we actually  make this operational in a way that is governed, [1:30] makes sense, it's auditable, it's accountable,  it has the appropriate controls on it and that's [1:34] really kind of what our AI push is right now both  internally and helping our customers to do that [1:39] um at extremely large scales. And Chris  Lendy, I'm the chief architect at Aura and my [1:47] uh my assignment, my role there is to help people  innovate. So Aura, as you know, is very large [1:53] company, a startup. And if you don't have a ring,  shameless plug, um they're they're on sale. Uh [1:59] anyway, the uh the big part the big priority  that we have right now is when we innovate, [2:05] how do we make sure that we're doing that in  a safe way? Um how do we contain the sprawl? [2:10] And then also as new tools and things are coming  out, how do we enable our users, our end users, [2:15] our 1,800 employees to be able to use those? Um we  have to manage over 18 enterprise class AI tools [2:23] today and we'll talk a little bit more about that.  Ajit Singh. Uh I'm chief architect at uh Fredken [2:32] Group. It's a private consortium of many different  companies in many different industries. And [2:39] uh our priorities are following. One is how can we  have our employees more AI aware? So like putting [2:50] AI into the culture of the enterprise that is  one. Second thing is building a platform where [2:57] agents or AI applications they can be developed  to solve the business problem and the building [3:05] the infrastructure or the platform that is not  done again and again. So those are our two top [3:13] priorities. Great. And um we'll stay with you  for a second. And um I'd love to go back down [3:17] the line and hear about an AI project that each  of you have worked on and um what kind of problem [3:23] it's addressing. Okay. So the AI project which we  have worked on and it is already paying off. We [3:31] started building the platform that okay how we  can help our many different companies to solve [3:38] their business problem and as part of validation  of the platform we are building we built a help [3:46] desk agent and what it has done is it has done the  following. Number one it has identified the gaps [3:53] which we had in our platform so that we can close  those gaps. It has helped us to learn what works [4:01] what doesn't work so that our business businesses  they do not repeat the same mistake. And the third [4:07] one is that now our help desk tickets are cut  in half. Half of those are already resolved by [4:13] agents. Yeah. And for us uh one of the one of the  biggest uh challenges but also most rewarding is [4:22] uh Aura is a very tight company with regards to  the number of employees because we are growing [4:27] um and we're growing so fast that we can't hire  people fast enough to do the work. So um our big [4:34] push right now with enterprise architecture  helping the different organizations is to [4:38] create agentic workforces. So we developed an  architecture um that creates uh agents and those [4:45] agents are orchestrated through an MCP server to  do the role of S sur and not just support but also [4:52] looking at um not just the logs anomalies it's  comparing this product to that product to that [4:57] product to that product to look for um look for  patterns within our architecture and within our [5:03] environment our operating environment and then  it's recommending back to the teams remediation [5:08] and sometimes even submitting PRs to fix the  things so that the people don't have to go and [5:12] do that. Uh it's it's important for us to do  this because we have on on average between 10 [5:19] and 20 agents being built per week into deployed  into production. Um we have over 4,000 apps that [5:26] have been created with AI and that are currently  being used across our organization. And so we have [5:31] to manage that effectively and we just can't hire  support people fast enough to do that. So we build [5:36] agentic workforces. I think uh a similar um with a  slight different bent to what's just kind of being [5:43] discussed is the big project which we're we're  really pushing on well we got a lot of projects [5:47] um is really around what I kind of mentioned  earlier which is like how do you orchestrate [5:50] these things how do you record every decisions  made that's made right like it was so slow when [5:54] it was us doing it as people right and now these  decisions happen so quickly with agents so how do [6:00] you make those agents record those decisions where  do you put those in in like in your docs and and [6:05] make it auditable how do you create a platform in  which you can insert human in the loop. How do you [6:10] improve reliability? Because like you know agents  get stuck or agents conflict with each other or [6:15] they literally argue with each other, right? And a  lot of what we're doing is creating this like kind [6:19] of operator concept around all that which helps  to identify okay what should the work graph look [6:23] like? How should these agents talk to each other?  Where do they relate to each other? Where do they [6:27] make decisions? Where do you have to pause and  say hey I need a human in the loop to make a call [6:30] here? Um and and kind of orchestrating that.  And uh that's that's a big push we have both [6:35] internally and also to enable for our customers  because everyone seems to be building agents [6:39] at such a rapid clip but no one but everyone you  talk to wants actually to have governance around [6:44] those because they just want to know like like why  did it do that? What did it do and who made that [6:48] decision and like was the right one? Um especially  for the industries that we're in. And so um that's [6:53] that's where we're really focused as far as an  AI project. Awesome. And uh yeah, internally uh [6:59] at Quickbase and PAVE, we have a tool we refer to  as next best call. Um trying to position for our [7:05] sales reps uh AI recommended plays uh to reach  out uh based off of account signals. And we do [7:14] that in ways through like a Snowflake agent we've  set up with like a Snowflake MCP to get that that [7:20] data into Claude. Um, we also have a way of  uh leveraging work to get that data into um [7:27] Slack daily for for reps to see that and get those  suggestions. Um, but there like we're putting it [7:34] in the place where reps are working. But I think  more of the issue is that like those platforms [7:38] weren't built for the next steps. They weren't  built for like how do we track if there's actually [7:42] actions being taken on these plays. Um, there's  also we're missing that ability for reps to give [7:48] us an idea of like this suggestion was offtarget  or it was missing entirely. And so what um what we [7:56] did what we worked to implement was we're starting  to put this into our new PAVE product. It's an [8:01] application where our our sales teams are working  in those daily and it's a way we're consolidating [8:08] all the data that they need to see uh to take  their day-to-day actions. But that gives us a [8:14] feedback loop uh better than just a thumbs up or a  thumbs down uh whether this thing is working. And [8:20] so our agent can actually learn and and get more  accurate. But it also gives a mechanism for them [8:27] if they come up with new ideas of like there these  are new patterns we're seeing that we're where we [8:31] want to reach out to customers. um we can actually  take that from their idea and and have our agent [8:37] take a first pass at the code and and the pull  together the data and the criteria required [8:42] for that to actually serve that up and have go  from an idea to a working MVP quickly without [8:49] uh a engineering cycle in between. So um that's  I think that's helped a lot and we're also just [8:55] fine-tuning the criterias we have in place.  Uh, but I think that's giving us speed, that's [9:01] giving us velocity. But I I'd be remiss to talk  about speed and velocity without talking about [9:05] governance. And so, not to end on a dull note on  this question, but like I think the governance [9:10] there is important. Um, right now most tools we  use is like you actually have an admin has to go [9:16] in and manually adjust the roles and permissions  for these things. Where it's going is more that a [9:21] builder can explain in plain language like here's  account reps should only have visibility into the [9:28] accounts in their book of business and finance  should only have visibility into billing data or [9:34] and they shouldn't have access to sensitive data.  Um and the system can actually take a lot of that [9:38] and build out the right roles permissions and  audit criteria to track that. So uh the good news [9:44] there is that's something that we have um and  QuickBase has been doing for 20 years now. So, [9:48] it's a matter of us implementing that into the  AI native technology we have in our PAVE product. [9:52] Great. And I Jeff, can I make a quick comment?  Um, a previous discussion before we started here [9:56] that we were having, I think, is is really apt to  what what Michael was just saying and also Chris [10:00] was just saying, which is we want all of our  our developers, our employees, everything like [10:04] that to be super creative. And so, we don't want  governance and controls to get in the way, but [10:09] like we're realizing more and more they have to  be there. And so, I feel like um that's like a key [10:13] topic that all of us are kind of wrestling with  is like how do we let people be creative and run? [10:17] Let the let teams have ownership, let them make  decisions, but at the same time the controls kind [10:22] of moving more to be centrally managed so they  don't have that burden in that workload and yet [10:26] the company and and the business and everything  like that can do what it needs to do as far as [10:30] um as far as its side of governance goes  without bothering the teams. So, and that's [10:34] really important because when you're dealing  with regulated environments like for instance, [10:38] we do health health OS and healthc care has a  specific set of guidelines that were written in [10:44] the late 90s early 2000s and they require you to  fulfill those guidelines in whatever way you can. [10:51] On the flip side of that, when you deal with life  sciences, life sciences has like FDA compliance. [10:57] FDA compliance is starting to move towards AI,  but they still are bound by SOPs that you have [11:03] to follow to the letter. You have to show all the  evidence and a big part of that is governance. So, [11:08] how do you enable somebody to innovate, create  their own AI that's doing some very novel things [11:13] with uh your physiology, for instance, and the  signals your body is giving off and take that [11:19] and turn it into something that is a product  that will benefit people, but then not being [11:26] tied up in a six-month validation process. If  you're familiar with the bioarma life cycle, [11:31] being stuck in six months of validation,  that doesn't work. They don't coincide. [11:35] So, so to the two points here, what you have  to do is be creative. Let people innovate, [11:41] let them think through the problem, let them use  AI and then we can use governance on the back end [11:46] or in the process to help them do it safely. Yeah.  So the way we are solving governance problem is we [11:53] are implementing it in the platform so that the  creativity and the innovation they are enabled [12:00] they are need not to worry about the governance  because that comes built in. Have you found uh [12:08] bottlenecks with that or or situations where the  governance has inhibited the speed to develop and [12:14] deliver? Yes. So uh there are those situations and  very easy one or the many of us would have faced [12:23] is the arbback and aback. I think that's really  important too because like uh you don't like we're [12:32] changing LLMs. We're changing the AI all the time.  We're creating new agents. They're better than our [12:36] old agents. The tying any of that at that level  to the agents LLM like is insane. Like you've got [12:42] to have something which is one level higher or  multiple levels higher about the orchestration [12:46] through the through our platform or through  whatever it is where that remains steady right [12:50] all this other stuff can change because it is like  every day right and um but that orchestration that [12:55] governance and that like you know making sure  that things are reliable regardless of the fact [12:58] that the LM's just changed I think is all really  important and also I to I think too on on if we [13:05] extend that I don't know if you've all had the  experience with um anthropic or open AI where [13:10] they change their terms and conditions and they  don't store your history from the prompt and the [13:15] response anymore. That was literally something  that happened overnight. We came in new terms and [13:19] conditions. What do we do? Well, now the onus  has been put back on us as a company to track [13:24] that. Now, we did we track the prompt and the  response, but all that rich metadata that came [13:30] back with the AI that told us how well or how bad  it worked, how many hallucinations it had, uh, is [13:36] gone. You have to request that specifically from  the company. So when we think about governance, [13:42] how do you apply the right governance and where  do you get the data that I mean that required [13:47] us to change how we did all of our systems.  So to your point, a level above the user, [13:52] a level even above the platform as a company or  a corporate process enabling the the collection [13:58] of all that information and then mining that data  to make sure that we're putting the right controls [14:03] in place. that that is that that was a real  challenge, but I think we're working through it by [14:07] again allowing the person that's doing their work,  GRC, whatever, to be creative using AI. Great. And [14:15] um while we're on the topic of governance, are  there any specific strategies that you found have [14:19] been particularly effective for making sure that  people have that room to experiment to bring in [14:24] the new tools that are just constantly coming at  us while also making sure that you are following [14:29] the regulations and making sure that you know  your sites don't break, your systems don't break. [14:35] I think for us um changing policy and not just  policy within uh within our company but within the [14:42] industry um there's an expectation we are a global  company so our data lives in many many countries [14:49] and that data every single country has their own  standards so we have to build that into the model [14:55] and what that means is that the AI has to be aware  of it but cannot be constrained by it for instance [15:01] if we have data that is sitting in Europe and we  want to use that data In the United States, one of [15:06] the ways that we would do that in the past is we'd  offiscate data and not allow you to get sensitive [15:10] information. Well, now we have different tools  where we could see the data. We can we can use [15:15] the data, aggregate it, give signals and insights  and share those insights across the globe and then [15:21] compare that with other systems and other data  and then come up with an answer for something. We [15:27] we changed the model. So instead of the the old  way of doing governance and data classification [15:32] in that we allow AI to classify and then it  recommends to our our sponsors and to our data [15:38] stewards what that data should look like and then  what data is necessary metadata is necessary in [15:44] order for that AI to function properly. Great. Um  and I'd love for us to talk a little bit about how [15:51] you manage legacy systems and these big complex  systems that you already have in place. um how can [15:56] you modernize while also making sure that you're  maintaining smooth day-to-day operations? Uh I can [16:03] take that one. Okay. So the way we look at legacy  systems is and AI systems in a very two different [16:12] way. Legacy systems they are deterministic systems  they are solving the problem. They will continue [16:19] to solve the problem. I will give you the example  of we are also in financial services. Now 10 [16:27] billion transactions are made every day across the  world. These models are probabilistic. If we look [16:36] at the token tokens that how correct they are,  it is in somewhere 60 to 70%. Now let's assume [16:45] that these models are very good and but they are  still probabilistic. Let's assume they are just [16:51] one one time wrong in 100 times. Now if we look  at the transactions out of 10 billion 100 million [16:59] transactions are wrong financial infrastructure  cannot run on probabilistic. So now with that [17:07] background with that example so what we are  saying is the legacy system or the systems which [17:14] are deterministic which have been tested and they  do what they are supposed to do they are doing it [17:19] for uh 5 year 10 year or even 20 year they will  continue to do it. Now what has changed after [17:26] the AI? So these are the when we say legacy I call  them preAI systems they will continue to work. Now [17:33] what has changed after the AI is that now software  can region it can also understand the intent. So [17:43] previously if someone tells a developer that hey  I want to see how many customers are churning [17:51] the developer will take this one sentence need to  build the whole pipeline create the data and then [17:58] feed that data into a model and model will tell  that how many customers are churning. Now that [18:04] sentence can be fed and software understands  that what's happening. So the way we take it [18:10] is uh that we are bringing these two systems work  together and the challenge here is that how they [18:20] can work together together collaborate with each  other and not only answer the question which are [18:25] given in mathematics or in analytical form but  also answer the question which are given as an [18:32] intent. I I'm I have to get in here because I I  smile a lot as a software developer. I mean the [18:37] temptation to rewrite anything which you don't  understand or is old is always there right but [18:42] every single one of us who've gone through this we  get burned because you realize that that code or [18:46] that business process or that business system has  actual business intelligence and has experience [18:51] built in and man doing that you're just you're  you're going to go through a world of pain. Um, [18:55] so just like what what Redwood does for our  customers right now, um, we tell them, hey, keep [19:00] your legacy system running, keep your your legacy  business process running, right? We're actually [19:04] going to wrap it and we're actually going to make  it accessible through through our MCP server to [19:08] agents. So agents can now call it and and give  it information and run it and stuff like that. [19:13] So you get kind of the hybrid world where like  people are developing agents, but the business [19:16] logic that's already settled just simply gets  called and orchestrated and run and the outputs [19:21] are are are executed. Um, and I think wrapping is  a good way of thinking about that because um, that [19:26] way you don't lose all that logic and you don't  run into the like, oh, I'm going to rewrite this [19:29] and it's going to be better and then you realize  that there's a reason why people did things which, [19:33] you know, at first blush didn't make sense. So  that's what we're doing and our customers seem [19:36] to seem to latch on to that as pretty reasonable  because we have a lot of customers that have [19:40] really old legacy business processing system  and nobody in the company knows what they do, [19:44] right? But everyone's afraid of them and so we're  like, hey, don't touch them. Let us let us help [19:47] you here and make them available to agents but we  won't replace them. So that's an opportunity. So [19:52] Aura is unique. We have one onrem system and we  have 500 applications. The rest are all in the [19:59] cloud or SAS applications. So we don't necessarily  have that problem. Where we're finding issues is [20:04] when we go out and we need to we need to uh sign  up for a SAS solution or a software. Some of those [20:11] providers are not up to date. Like I just talked  to someone. They don't even have AI in their [20:16] platform. They have no nothing on their roadmap.  So what we do is we look at it and we say, "I'm [20:19] sorry, we can't do business with you." Um, if it's  something absolutely necessary, say for product [20:24] development or hardware or something, we might  bring it in. But to your point, we wrap it with an [20:29] MCP. And that what that what that does is it takes  the technical debt that comes with that system [20:35] kind of out of the picture. On the flip side of  that though, I think about all the people because [20:40] my past has been in antiquated systems. I worked  in insurance for a long time. Sorry if you're [20:45] working in insurance. It is very antiquated. um  how do you how do you make those organizations [20:51] uh drive towards excellence? The one thing that  I see again is a lot of decisions are being made [20:57] because of the the um the process. Well, remember  a lot of those processes were developed 5 10 years [21:04] ago when some of the functionality we have in  SAS solutions or cloud wasn't even in play. So [21:10] the idea is use anthropic, use open AI to ask it  to analyze the process, tell you what you should [21:17] be doing based on current industry practices and  then ask AI how it would automate that pro that [21:24] process. then take that back to your your subject  matter experts and find out what part of that is [21:29] wrong because I can guarantee you if you even  get 70% that's value not from the automation [21:35] perspective but from the process rewrite so  you know what your business is doing. Yeah, [21:42] it it fully on board with the wrap methodology.  It can't just be a rip and replace of of the tool. [21:47] It has to at least run in parallel if you're not  wrapping it. So fully agree with you. Yeah. Now I [21:52] also actually to build on what Chris said is he's  that this uh instead of just uh using the same [22:01] process re-imagine it and as he was describing  that now there is a tool available which can [22:08] help reim re-imagination of the process easier and  faster. Yeah. Great. And um on the other side of [22:18] that is that AI moves very quickly. And so you  want your systems to be nimble enough that you [22:25] can adapt to new tools, pivot if something  comes on the horizon or say if a particular [22:31] tool gets banned in the United States. Um and  so I wonder how you're thinking through that. [22:38] [gasps] I mean the easy answer here is staying  model agnostic with whatever tools you're using. [22:44] making sure that tool you're using has the  options of the different models to use across [22:48] the different providers. But yeah, we stick close  with our legal team. In fact, our legal team just [22:54] hired a new council specifically for AI. We have  a technical uh council as well. She she does a [23:01] great job helping us navigate the problems. And  then what they've done is they've taken all these [23:05] standards uh that are that apply to AI all these  new these new regulations coming out and what they [23:12] did is they consolidated them into a few that  we can apply when you know through our thought [23:18] process we can apply them uh I can't think of  the word anyway we can apply them in a good way [23:25] um so that they they don't allow us to go close  to the edges of those those legal u boundaries. [23:31] This reminds me a small thing which we did um a  couple months ago was uh for our generic agent [23:36] personas like the more generic agents we have  we actually have the ability to specify um their [23:41] preferred model to use and so we can update that  model kind of at will for that particular persona [23:47] that agent because what we're finding is a lot of  the different technologies especially when you're [23:50] trying to cost optimize are good at particular  things and you don't want to overkill because [23:55] when you're running things like 24/7 like oh my  gosh that bill can go up if you're not watching [23:59] it right and So having the ability to do that is  is one thing which we've done. Um there's a lot [24:04] more to do but that's just an idea. Let me add to  that question because that's a very good question [24:09] and the reason I say very good is I will share  the story but my suggestion is and that's what [24:15] we do. We try to use architecture oneonone. Let's  make sure that every component plug and play. And [24:21] we as an organization we have just gone through  one application which we were using and it was [24:28] it had a big use across the enterprise across many  of our uh business companies and then that uh that [24:37] application changed its commercials. They changed  the whole product and it just did not fit with the [24:44] new competition which has come. it just did not  fit what we where we want to go and then only [24:52] solution was to just like sever the relationship  with them and uh and it did take us around 3 [25:00] months to get off of it but again it is for us it  was a great experience and that's why I'm sharing [25:07] with you all because uh not only models every  other month if not every other week there is a [25:14] better one but even the applications they will  change the new one will come which are better [25:20] products. So ensure that you stay as nimble as you  can and do not bolt onto one specific application. [25:31] I would say too on that one if you don't have a  registry of which of your SAS applications for [25:37] instance has AI and what models they're using  I would highly recommend that you do that. Um, [25:43] we created an agent and all it does every Monday  morning is it goes out to every single one of our [25:47] apps, all of the sites that they have. We look  at the terms and conditions, privacy policy, [25:52] AI usage. It comes back, it analyzes what we  currently have, and then if if it's changed, [25:57] it alerts us that there was a change and what  that change was. The other reason we do that is [26:03] because when those SAS providers change their  model from OpenAI to Claude or Claude to Grock [26:10] or Mistl to whatever the case is, we know the  impact uh and the risk that is bringing to our [26:16] company. I'll give you an example. So let's  say we're doing research and we're using an [26:21] open- source uh LLM or an open source AI model  and that AI model is doing NLP for us and we're [26:27] trying to use it to let's say communicate  with people who speak a different language. [26:31] If that license changes, that could invalidate our  product and we lose that revenue. In fact, there [26:38] are some AI models that are on HuggingFace today.  If you read the fine print, it says if you go to [26:44] production with that model working, they claim IP.  I mean, when you're talking about a multi-billion [26:51] dollar company, that could be a major impact.  So, look at that. You can use automation to do [26:56] that today. And if you don't have it, just just  create that. Let let AI let anthropic co-work or [27:02] cloud co-work rather something like that go out  and do it. Another one is uh lovable and I know [27:07] you guys just came out with pave. Um Lovable is  great at that because you can just automate it. [27:12] You tell it what you want it to do and it'll go  and connect and do that as long as it's connected [27:16] in your enterprise. Great. Everyone has a little  homework. Um, while we're on the topic of cost, [27:23] I think another thing that people would love to  know about is how are you balancing hosting your [27:27] AI workloads in the cloud versus inhouse and what  are the considerations as far as both cost and [27:33] security there? We only have two in-house um AI  workloads. Uh we do have lots of experimentation [27:42] on high performance computing. Um but the the  biggest thing for us is the the costs are managed [27:47] at the enterprise level. So we we have a certain  budget allocated towards tokens. Uh and honestly [27:54] we go past that all the time and the reason that  is is because we're promoting innovation and [27:59] experimentation over governance and cost cutting.  The reason why is because we know somebody out [28:05] there in in our our population of people are going  to use AI to come up with something very unique [28:11] and so we let them do that. At the same time, uh  we also look at people who are using um some of [28:18] the lighter models. Maybe they're just doing  uh summarization or changing their documents. [28:23] We don't necessarily give them access to like the  the heavy foundational models or the um you know [28:29] like the frontier models that might be out there.  Um from a from a customer product perspective um [28:35] we allow our customers to bring their own LLMs if  they want, right? Or else we provide the service [28:39] as well and then we we we charge for it. Um and so  on the product side we really haven't run into too [28:44] much of a cost issue. Internally we absolutely  have and um the two things we've done that have [28:49] significantly changed the game for us is the the  first one is we introduced a prompt router which [28:53] has been useful and the second thing has been um  basically our using our own platform is when we [29:00] talked about like uh like agents like and and  how agents can consume like this whole concept [29:04] of having an orchestration platform which there's  lots of talk about that and there's been more open [29:09] source has been coming out like ideally I would  like most like 60% 70% of my engineers work to all [29:15] be funling through that orcha orchestration  platform so that we have agents which are [29:19] fine-tuned to the particular models. We also have  visibility into the cost. We make sure things [29:23] don't run away. So it's about it's about I think  getting there now. We've all blown our budgets [29:27] like multiple times already. And now people are  getting serious about like you can't actually do [29:31] that anymore. And so it's now is kind of the time  where I think all of us are kind of evolving to [29:36] we like to make use of a platform either in-house  built or external or open source or whatever. and [29:41] also things tools such as like a prompt router to  help us to to control cost too. I have a question [29:46] for you guys all three of you actually. Uh so  we we use workcato as an integration platform [29:53] orchestration engine but a lot of the new software  that have AI built into them also have a workesque [30:00] connector platform app studio and they want they  want them they want their app to connect to our [30:07] data and use use their app. Now there's obviously  advantage to that because they have access to the [30:13] data that we would not necessarily give them.  How are you how are you all dealing with that? [30:18] I mean like pave I mean you guys are on the  forefront of that but how are you dealing with [30:21] that sprawl? Because I can name five major systems  that we use that have like rcado built into them. [30:28] I mean, rcado can definitely lead to I mean  a lot of different people who are managing [30:33] the different rcado recipes and I think with pave  we're we're leveraging more we're going into more [30:39] of a partnership to support our connections. I  think where the there was initial I think number [30:43] I saw of like 700 plus connections to different  tools which can help. Um but yeah I mean that's [30:49] the biggest thing is getting from data from  A to B is is the most important thing. Um, [30:54] so without those connections, it's it's tough. But  you have to do it in a way where it's built in and [30:59] it's not something on the fringes like, you know,  Joe built this 10 years ago and we don't know how [31:04] this recipe is built. Yeah. Because I'm I'm  thinking you're paying for it multiple times, [31:08] right? So if somebody's using AI to create  a prompt for work to build a recipe, well, [31:12] now they go into Wiccato and they build a recipe  using their AI and then they go into Zip and they [31:16] use that same version of that prompt to build the  Zip workflow. you just now paid for it three times [31:22] with three different models and in the past it  would have been okay I got six months I got to [31:27] go build this thing right so it's kind of hard  the economics of it is hard which is causing AI [31:33] fin to emerge differently than your past ROI and  TCO evaluation of value and you know I think we [31:40] have a mix in the room and and this conference of  people who are on the business side and people who [31:44] are on the tech side and I think sometimes your  priorities can be a little bit different maybe [31:48] the tech side wants to create very modern, very  flashy things. Business side is very focused on [31:53] the return on investment. Um, so how I'd love to  hear from each of you how you're working together [31:59] and if there are specific amples examples where  the tech and the business side have successfully [32:03] collaborated. Uh, I can start with this. So  we have a mantra in our enterprise. We call [32:09] it lead with need. So we lead with when there is  a business need. And when I say business need, [32:18] business also need to have a very well- definfined  ROI. So when we started the AI initiative few [32:26] years few few years back in our finance we have  a team which looks at each use case and looks at [32:35] its ROI not at the time of implementation but how  it is going to work out 2 years 3 years or so now [32:43] because in the beginning ROI will be negative  because we are investing in building it but [32:49] does it pay over the time so that is one thing  we lead with the uh need. Now when it comes to [32:57] the technology then at that point we explain  and that's why in the beginning I was talking [33:03] about the platform that sometimes there is a  need and we can say we can see that yes it's [33:08] a quick win which also quick win not only for  business but also for technology because we [33:15] can prove we can prove the concept and also show  that okay what kind of ROI it has so quick win [33:25] And uh we will we will implement a quick win. We  will implement it fast. But at the same time if [33:31] our existing platform doesn't support it then  we identify those gaps and then we continue [33:37] to work that how we close those gaps and the  quick win which was uh not 100% uh conformed [33:44] to the platform how to bring it back into the  platform. So we take this multiacet approach. [33:51] on that. Yeah, I I'd say this year has been the  year of of token maxing. I have to imagine next [33:56] year is going to be the year of AI ROI. I think  it really has to be. And I think what we've done [34:01] internally is we've built more of like an intake  process where you kind of have to define what that [34:05] northstar metric is, what that ROI is going to  be, what you're going to try to track there. And [34:10] that helps us set the priority of like what we're  working on, what we're pulling through. I mean, [34:13] with AI, everything is possible now. It it opens  the infinite backlog of things you can do. So how [34:18] you prioritize that, how you prioritize that with  ROI, it it makes the business fully think that [34:23] through ahead of time. Yeah. Our customers always  like to talk about outcomes and I think working [34:27] internally also. So both our customers when we  talk about outcomes and how Redwood enables them [34:30] to have the outcomes rather than AI which is  the hammer and making everything a nail, right? [34:35] Um and then internally using that language with  our business partners has helped also like this [34:38] is the outcome we're trying to achieve. This is  what we think we're going to get from it and ROI [34:42] etc. um I think has been the key just talking  about outcomes rather than the process or the [34:47] fact that we're using AI and things like that.  Yeah. And ours uh we have a little bit different [34:52] approach. If you're familiar with like Google  sprints, it's a five-day design to development [34:56] phase. Um we've adapted that. Our AI productivity  team has adapted that and what they do is they [35:01] go from uh team to team. They invite them to an  on-site workshop because we're pretty distributed. [35:06] um they go through and the first I don't know  first two hours or the first half of the day [35:11] they're explaining some of the challenges that  they're having within their business structure [35:16] whether it's process or people or just how fast  things are moving as they're doing that myself [35:21] and my solutions architect we have several agents  that are listening to the conversation we have one [35:27] that will go and build a business case based on  the input from those those teams it will identify [35:32] the distinct uh business value that we could  get if we automate something or not. Um it'll [35:38] tell us don't even go down that road. Another  one might be um that it tells us like how many [35:44] people is it going to take to actually implement  this? I don't know if you have this problem, [35:47] but business users sometimes or business partners  uh will say, "Yeah, I just bought this software. [35:52] We're going to implement it and it needs to be up  in two weeks." Okay, great. Did you talk to it? [35:58] Did you talk to the AI team? No. Okay. Well, the  average process is going to cost you $300,000 to [36:05] implement that. Would you like to know how to do  that better? And most of the time they'll say yes. [36:09] And so we lead one of these squads. The output of  that though is really interesting because in the [36:14] morning we identify the the uh the potential. In  the afternoon we have a working prototype that's [36:20] sitting in our sandbox and could be promoted  to production because uh it's it's uh governed [36:25] by all of our agents. So that's that's one way  that we do it. There's several other ways too, [36:30] but yeah, it's just it's just giving them the  ability to see what's possible. Great. Um, [36:35] so we have about seven minutes. If there  are any questions from the audience, [36:38] I think you have you're using work, you're using  pave, but what about the other two? What are you [36:43] using as your agent platform? Are you building it  internally or are you buying? So, Redwood Redwood, [36:47] sorry. So, Redwood software actually does a lot of  this. Um, just we aren't about we don't focus on [36:53] human automation. We focus on systems automation.  And so the SAP the oracles really we happen to do [36:57] what I call the easy problem human automation  also but we try to use our own stuff too. We [37:02] are not using work and uh depending on what kind  of workflow is uh we are doing reimagining it. Do [37:11] we really need that process? Can we optimize and  uh just not do it? That is our number one goal. [37:19] Number two is if we need it then how can we write  plain and simple software to make it happen. Yeah. [37:26] So um obviously on the one hand we have a fast  evolving tech landscape you know new tools new AI [37:32] um that's out of our control and then you know  we have people uh who make decisions right that [37:39] we hope will stand the test of time. So can maybe  each of you share one example of a decision or you [37:46] know an initiative that you made that has proven  to to be relevant like you know a couple years [37:52] down the road. I can do one uh so in architecture  we use uh modified well architected frameworks. [37:59] So you know the cloud providers provide that kind  of standard framework. We've enhanced that quite a [38:04] bit to do analysis and use different frameworks  like 6R CMMI. Um I mean there's a handful that [38:12] are in that are part of that. But the reason it's  stood the test of time is because it doesn't look [38:16] at how a person does what they do. It looks at why  they do what they do and the impact it's going to [38:22] have on the organization. So we might scan your  repository and it comes out with eight pillars, [38:28] eight different scores and a consolidated score.  And then we go beyond that. We just don't go [38:33] throw it to the architecture review committee and  say, "Yeah, you're not doing a good job." We say, [38:37] "Look, here's some PRs that we wrote for you that  we're going to submit back to your repository. We [38:42] know you probably don't have time to do it." Um,  because those are the ones that are critical to [38:46] our organization. The other ones, they the uh this  agent stack ranks them and it provides them with, [38:53] okay, for the next two or three sprints based  on your cadence, um, this is what you should be [38:58] doing. So, that that actually works really well  and we've gotten really good adoption feedback [39:02] from that. So in our case actually what we did is  uh in uh January of 2020 uh 2024 when we started [39:16] all this AI initiative. So instead of trying to  build work space or environment because we are [39:23] a consortium we have many different companies and  then every company they need their own work spaces [39:30] like sandbox dev test and all. So what we did is  that let's build a infrastructure factory where [39:39] we get the information and the requirements and  we use software to build those environments for [39:46] them and uh so now in 3 hours when a company  comes to us that hey we need this environment [39:56] so in 3 hours we can build them totally separate  isolated no no possib ability of data leak from [40:04] one company to another a complete environment  they where they can build their models they can [40:10] run their applications so on and that's where  we are enhancing it also for the agentic AI so [40:18] that whenever someone comes to us what we need to  do is we just run the software and they will have [40:24] uh their own all the environments they need if  they need more here you go so and another uh Good [40:32] thing about it which we did not expect at that  time but now working very well as of now we have [40:38] more than 50 different environments for different  businesses and two platform engineers are [40:44] supporting it. So that is one decision uh worked  out to be right. Yeah, a question. Oh, sorry. Are [40:55] you guys done? Okay, thank you. Uh thanks. This is  very helpful. Chris, this is a question for you. [41:00] Looks like you are very agent friendly. Your  organization is going gung-ho on agent. Kudos [41:05] on that. Can you speak about how are you managing  this prolification of agents in your organization, [41:10] right? How are you making sure that you have all  the catalogs and all the registries and you know [41:16] how are you safeguarding and how are you measuring  the impact of those agents? Thank you. Yeah. So [41:21] that well architected framework I talked about um  when a person goes to create an agent we have two [41:28] ways. If you're on the software engineering  side, which we have about 400 engineers, [41:32] uh they have a template that they follow. That  template has the guardrails built into it based on [41:36] what the software engineering team has said, our  AI council as well as enterprise architecture and [41:41] IT. Um so when they start, the agent already has  the guardrails. On the enterprise side, uh we go [41:49] a little bit further than that. We have the guard  rails and put in the observability right out of [41:53] the gate. So as soon as that agent is deployed, it  doesn't matter where it's going. It always the the [41:58] the monitoring and alerting always homes back to  the centralized IT operations domain where we do [42:04] analytics and and things on that. Um but that does  pose a problem because we can create an agent in [42:11] glean for instance or in lovable or something like  that that doesn't follow that process. So what we [42:17] had to do is we had to build essentially another  agent on top of those two platforms um in order [42:23] to make sure that when people are building those  agents that they're doing it properly. It sends [42:28] us signals, tells us what's out of alignment and  then we go have a conversation with them. Well, [42:34] great. And with that we're pretty much at time.  So thank you all for your time and attention and [42:40] questions. Um and if you have any additional  questions we'll be in the corner over there. [42:45] [applause]