Dataiku Deployments Replace Fragmented AI Tools to Drive ROI

Dataiku 部署取代碎片化 AI 工具以提升 ROI 将专用检索工具整合为单智能体系统消除了工作流摩擦，并为企业 AI 部署带来了可衡量的财务回报。 Arthy Krishnamurthy 概述了 Agent AI 从试点项目向企业级规模化生产的过渡。她分享了 Roche 一位专利律师的案例，该律师通过将 10 个专用 RAG 工具整合为名为“Lex”的单一自主 Agent，成功解决了“工具蔓延”问题。演讲梳理了从简单 RAG 系统到多 Agent 架构的演进历程，指出了 Agent 衰退和碎片化治理等企业摩擦点，并提供了一个聚焦于业务主导价值、集中化平台和劳动力演进的战略框架。 Dataiku 成立超过 12 年，服务 700 多家客户并带来 4 倍投资回报率；其将 AI Agent 定义为依托大语言模型、拥有自主代理权并能代用户多步骤完成任务的系统。 (0:04) 罗氏制药（Roche）专利团队最初通过 RAG 开发了“Chart White Book”，每起案件节省 5 至 10 小时，并为 50 名律师的试点团队年均节约 25 万美元。 (2:18) 为解决多达 10 个独立 RAG 应用带来的“工具蔓延”并走出“生成式 AI 悖论”，罗氏构建了单 Agent 系统“Lex”，可同时检索数万份法律文献直接合成完整起诉书。 (5:44) 企业 AI 架构演进分为四个阶段：静态 RAG 检索、单 Agent 行动执行、多 Agent 协同以及自我优化的自主 Agent 系统。 (8:28) 企业在实现 AI 价值时面临缺乏投资回报率、Agent 蔓延衰减及统一治理缺失等挑战，需通过业务驱动、统一平台调度治理以及积极推进劳动力技能转型来破局。 (11:22)

Arthy Krishnamurthy of Dataiku argues that capturing enterprise value from artificial intelligence requires moving beyond isolated retrieval tools to orchestrated agentic systems governed by unified platforms and clear business metrics. Resolving Tool Sprawl: The Roche Case Study (2:18) Patent attorneys face strict compliance requirements and a low 15% industry-standard application acceptance rate. At pharmaceutical firm Roche, a team of roughly 200 patent attorneys led by an attorney named Sarah initially addressed research bottlenecks by building the 'Chart White Book.' Using retrieval-augmented generation (RAG) and semantic search across legal corpus data, the application saved 5 to 10 attorney hours per case, generating roughly $250,000 in annual savings across a pilot group of 50 attorneys handling 50 cases each year. However, expanding the data corpus to prior art, technical specifications, and court decisions resulted in 10 disconnected, specialized RAG applications. This fragmentation created tool sprawl and what Krishnamurthy terms the 'GenAI Paradox,' where potential efficiency is lost as users struggle to navigate multiple point solutions. To resolve this, the team deployed 'Lex,' a single-agent system that queries tens of thousands of documents concurrently from a single prompt to synthesize complete legal briefs and required filing citations. The Evolution from RAG to Agentic Systems (8:28) Krishnamurthy outlines a four-stage architectural progression for enterprise AI adoption, illustrated through sales operations. The initial phase relies on basic RAG systems that query static knowledge bases for information, such as analyzing customer win/loss reasons. The second stage introduces single agents capable of executing discrete tasks, such as combining customer satisfaction scores and interaction histories to draft sales proposals. The architecture advances into multi-agent systems, where specialized software agents—such as pricing optimization, competitive intelligence, and CRM agents—collaborate to assist human staff. The final stage consists of fully agentic systems that operate autonomously across the organization, identifying multi-deal trends, dynamically revising process documentation, and coaching staff in real time. Enterprise Scaling Hurdles: Sprawl, Decay, and Governance (11:22) While AI delivers value through worker augmentation, process automation, and integrated business chains, Krishnamurthy identifies three primary structural barriers to enterprise scaling. First, organizations suffer from an absence of ROI when they select ill-suited use cases, over-engineer simple tasks, or fail to secure user adoption. Second, enterprises face agent sprawl and performance decay. Point solutions from software-as-a-service vendors, hyperscalers, startups, and internal developers create unmanaged ecosystems prone to cascading operational failures and context drift over time. Third, the lack of centralized governance results in unmonitored token consumption, ballooning infrastructure costs, security risks, and an absence of audit trails across business units. A Blueprint for Governed Enterprise Deployment (15:00) To counteract deployment friction, Krishnamurthy outlines a three-pillar framework. First, AI investments must be led by business units rather than IT alone, establishing clear operational problem statements and quantifiable targets before development begins. Second, organizations need a centralized platform to manage agent interoperability, monitor performance decay, and enforce universal security and financial governance. Finally, enterprises must actively address workforce transformation. Leaders are urged to maintain transparency regarding changing job responsibilities and support employee upskilling, while individual professionals should proactively utilize organizational development budgets to learn emerging AI workflows. While transitioning from fragmented RAG tools to agentic systems can yield measurable productivity gains, long-term enterprise viability depends on establishing centralized orchestration and auditability to prevent agent decay and uncontrolled infrastructure expenses.

Dataiku 业务价值与转型业务负责人 Arthy Krishnamurthy 指出，企业人工智能应用正在从离散的知识检索系统迈向端到端执行的智能体系统；若要跨越工具蔓延与投资回报缺失的瓶颈，企业必须坚持业务价值导向，建立统一的编排治理平台，并协同推进员工技能转型。 专利律所案例：从工具蔓延到智能体整合 (2:18) 在跨国制药企业罗氏制药（Roche），专利律师主管 Sarah 带领一支约 200 人的律师团队，负责战略性专利组合管理、法规遵从及商业价值变现。由于行业内初始专利申请的通过率仅约 15%，团队需在海量既有判例和专利文献中核查现有技术。为提升效率，该团队初期开发了名为 Chart White Book 的检索增强生成（RAG）语义搜索工具，通过提取包含上下文的完整文本引用，使每起案件节约 5 至 10 个律师工时；按律师年均处理 50 起案件计算，50 名律师的部署组每年实现约 25 万美元的成本节约。 然而，随着语料范围进一步扩充至判决书与技术规范，团队相继衍生出 10 个各自独立的专用 RAG 工具，引发了严重的“工具蔓延”（Tool Sprawl），律师被迫在多个工具间切换摸索，陷入生成式 AI 潜力难以转化为实际产能的困境。为此，团队联合 Dataiku 构建了名为 Lex 的单一智能体系统。Lex 能够同时检索并协调数万份多类型文档，律师输入单一提示词即可自动生成包含完整支持引用的法律摘要，完成了从碎片化搜索工具到端到端任务执行的跨越。 AI 成熟度曲线：从 RAG 到自主系统 (8:28) Krishnamurthy 以企业销售运营场景为例，阐述了企业迈向智能体架构所经历的四个发展阶段：第一阶段为 RAG 检索系统，针对客户定价反馈或赢单原因等静态知识库执行问答检索；第二阶段为单智能体系统，能够调取客户互动历史与满意度数据，自主执行起草个性化提案或外发邮件等具体行动。 第三阶段演进为多智能体协作系统，通过编排竞品情报智能体、定价优化智能体及客户关系管理（CRM）智能体，协同辅助客户经理促成复杂交易；第四阶段则升级为自主优化的智能体系统，不仅能跨多笔交易自主分析潜在业务模式，还能动态迭代销售手册，并为一线代表提供实时销售辅导。 规模落地瓶颈：回报、蔓延与治理风险 (11:22) 演讲分析认为，企业 AI 的价值主要体现在员工能力增强、流程自动化以及将二者串联的智能业务链三个层面。然而，当前企业在规模化落地过程中面临三大结构性阻碍：首先是投资回报率（ROI）缺失，常由业务场景选型失误、简单流程过度工程化或一线员工采纳度不足所致。 其次是智能体蔓延与性能衰减，随着各类 SaaS 厂商、垂类初创公司、云服务巨头以及企业内部自研智能体的大量涌现，系统间互操作困难且极易引发级联故障；同时，随着企业业务背景发生变化，缺乏维护的智能体还会出现能力衰减。最后是统一治理体系的缺失，碎片化的业务尝试与未受监控的 Token 消耗，正引发基础设施成本激增、合规审计困难和安全漏洞，给首席信息官带来了严峻的治理危机。 规模化实施框架：价值驱动与平台治理 (15:00) 为应对上述落地障碍，Krishnamurthy 提出了三位一体的企业推进框架：第一，坚持业务价值先行，由业务部门而非纯技术团队确立核心业务问题与前期量化考核指标，促成 IT 与业务线的深度对齐。 第二，投资统一平台基础设施，通过中心化架构实现跨智能体的企业级编排、实时衰退监测以及关于成本与安全的全局通用治理。第三，重视人才队伍建设，管理层应对智能体引入带来的岗位角色变化保持透明，员工也应主动利用企业专业培训预算提升 AI 协作技能，实现组织模式与生产力工具的同步演化。 Krishnamurthy 强调，智能体系统落地的核心挑战不仅是算法与架构演进，还在于组织能否建立敏捷的学习机制与统一治理体系。不过，面对异构多智能体系统中因业务背景变动引起的“智能体衰减”现象，演讲尚未提供具体的长效量化监测标准与故障追责判定方案。

## Background 本视频由 Ai4 于 2025 年 8 月 27 日发布在 YouTube 平台。演讲由 Dataiku 美洲区业务转型负责人 Arthy Krishnamurthy 主讲，重点探讨 Agentic AI 从试点迈向规模化生产环境的落地路径，并通过制药巨头罗氏（Roche）专利律师工作流转型的真实案例，解析帮助企业摆脱“生成式 AI 悖论”并实现组织进化的核心要素。 ## TL;DR 演讲以大型药企专利律师团队从“工具蔓延”的离散 RAG 演进至端到端单一智能体“Lex”的实际案例切入，展现了 Agent 赋能知识工作者并创造量化回报的潜力。随后系统梳理了技术从 RAG、单/多 Agent 迈向自主优化 Agentic 系统的进阶路径，以及其对员工赋能、流程自动化与智能业务链的重塑。最后，讲者深入剖析了阻碍企业规模化落地的三大核心痛点——ROI 缺失、代理无序蔓延与治理真空，并提出了以业务价值为引领、构建统一企业级平台底座与以人为本赋能人才团队的应对战略。 ## Atmosphere & style 整体氛围务实、专业且极具战略洞见，展现出成熟科技商业领袖清晰利落的演讲风格。讲者摒弃了空洞的技术炒作，通过一线律师的鲜活故事与层层递进的架构图表，将复杂的技术演进转化为直观生动的业务场景。从个体提效的兴奋到面对 CIO 级治理困境的严肃审视，整场分享节奏明快、切中要害，让听众在感受到 Agent 落地潜力的同时，也能深刻体会到企业转型所需的严谨与审慎。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=tYREmZVqhKk&t=4s) AI Agent 的本质是由大语言模型（LLM）驱动、具备多步骤目标达成能力，并拥有代表人类行动之“代理权（Agency）”的系统，其自主权高低取决于业务场景与企业风险偏好。 - [00:04](https://www.youtube.com/watch?v=tYREmZVqhKk&t=4s) Dataiku 平台通过对底层数据与基础设施的中立性（Agnostic）赋能企业，在推动 AI 与 Agent 规模化落地的同时帮助客户实现了平均 4 倍的投资回报率（4x ROI）。 - [02:36](https://www.youtube.com/watch?v=tYREmZVqhKk&t=156s) 罗氏（Roche）制药基于 RAG 构建的专利检索工具 "Chat White Book" 单案为律师节省 5-10 小时并年省 25 万美元，证实了生成式 AI 辅助复杂专业尽职调查的初期价值。 - [05:48](https://www.youtube.com/watch?v=tYREmZVqhKk&t=348s) 针对不同专业场景盲目上线离散的 RAG 工具会引发严重的“工具蔓延（Tool Sprawl）”，导致工作流割裂并增加员工在工具切换上的心智负担。 - [05:48](https://www.youtube.com/watch?v=tYREmZVqhKk&t=348s) 单一智能体系统 "Lex" 能够统一调度海量文档并直接输出包含法条引用的完整辩护简报，通过无缝嵌入实际工作流打破了“生成式 AI 悖论（GenAI Paradox）”。 - [08:28](https://www.youtube.com/watch?v=tYREmZVqhKk&t=508s) 企业 Agent 落地呈现清晰的四阶演进路径：从被动问答的 RAG 引擎，迈向具备行动力的单 Agent，再扩展至多 Agent 协同作业，最终演进为全自主演练与闭环迭代的 Agentic 系统。 - [10:35](https://www.youtube.com/watch?v=tYREmZVqhKk&t=635s) 高阶 Agentic 系统能够跨越孤立交易进行全局模式识别，自主发现有效策略并自动更新业务手册或提供实时辅导。 - [10:35](https://www.youtube.com/watch?v=tYREmZVqhKk&t=635s) Agentic AI 释放企业深层转型的关键，在于将“员工赋能”与“流程自动化”深度融合成端到端的“智能业务链”，进而重塑客户体验与商业变现模式。 - [12:35](https://www.youtube.com/watch?v=tYREmZVqhKk&t=755s) 阻碍企业释放 Agent 价值的首要痛点在于投资回报率（ROI）缺失，其根源往往是业务选题失误、简单场景被过度工程化，以及员工缺乏场景理解。 - [12:35](https://www.youtube.com/watch?v=tYREmZVqhKk&t=755s) 多方自研与外购引发的“代理蔓延（Agent Sprawl）”导致异构孤岛林立并易发级联故障；同时数据与环境变化带来的“代理衰退（Agent Decay）”亦亟待主动检测与拦截。 - [12:35](https://www.youtube.com/watch?v=tYREmZVqhKk&t=755s) 多团队各自为政缺乏统一治理会导致 Token 算力开销与预算迅速失控，并在数据安全合规、操作审计及链路追溯上形成巨大盲区。 - [15:01](https://www.youtube.com/watch?v=tYREmZVqhKk&t=901s) 实现 Agent 成功落地的第一大支柱是以业务价值为引领（Lead with Business Value），由业务与 IT 深度协同定义痛点与可量化的业务回报指标。 - [15:01](https://www.youtube.com/watch?v=tYREmZVqhKk&t=901s) 企业必须投资具备编排能力的统一平台底座，以解决跨系统互操作难题，并在模型性能衰退前提供主动监控与全链路通用治理。 - [15:01](https://www.youtube.com/watch?v=tYREmZVqhKk&t=901s) 组织变革的核心在于以人为本（Focus on Workforce），管理层需透明沟通岗位职责演变，员工也需主动借助培训预算积极完成技能升级。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=tYREmZVqhKk&t=4s) 介绍与 AI Agent 的概念对齐 演讲者 Arthy Krishnamurthy 作为 Dataiku 的美洲区业务转型负责人，开场简述了当天的分享主旨：通过实际案例讲述企业在构建与扩展 AI Agent（智能体）时所面临的规模化难题，以及克服这些问题的核心模式。在 Dataiku 内部，她的工作主要是帮助客户找准 AI 能够解决的具体业务痛点，并在客户看到成效后，协助其搭建相应的人员与流程运营模型。Dataiku 已成立超过 12 年，服务了 700 多家客户，并帮助客户实现了平均 4 倍的投资回报率（4x ROI）。其平台的核心优势在于对底层数据和基础设施层完全保持中立（Agnostic），从而帮助企业资产具备面向未来的适应性（Future-proof）。 ```mermaid flowchart LR accTitle: Dataiku 赋能 AI 与 Agent 规模化落地的运作闭环 subgraph DataikuPlatform["Dataiku 平台赋能"] A["识别业务痛点与场景"] --> B["构建、部署与治理 AI / Agent 产品"] B --> C["沉淀人员与流程运营模型"] C --> D["实现 4x 投资回报率 (ROI)"] end ``` 在深入故事之前，演讲者先进行了语义层面的概念统一，将 Agent 定义为：以大语言模型（LLM）为驱动、旨在通过多步骤方式达成既定目标，并且拥有代表人类采取行动之“代理权（Agency）”的系统。演讲者强调，系统被赋予多少自主代理权，取决于具体要解决的问题类型、所在行业标准以及企业的风险偏好。随后，演讲者现场互动调查了听众中正在构建 Agent、带领构建团队以及来自业务部门的人员分布。 --- ### [02:36](https://www.youtube.com/watch?v=tYREmZVqhKk&t=156s) 专利律师 Sarah 的困境与 RAG 工具的最初成功 为了生动展示 Agent 的价值，演讲者讲述了西欧一家大型制药企业专利律师 Sarah 的真实案例。Sarah 领导着一个约 200 人的专利律师团队，日常承担三项核心职责：一是专利组合战略管理（识别可申请专利的研发成果）；二是确保药物符合监管政策以顺利获批；三是协助公司探索商业化盈利路径。 专利申请的前期尽职调查（Due Diligence）工作极其繁重： 1. **现有技术检索（Prior Art Search）**：团队必须深入排查已有的所有公开专利，以证实自身发明的真正新颖性。 2. **判例法与合规审查（Case Law Review）**：提交专利申请后，律师团队需要研读海量判例文件与指导方针，确保申请符合极高规范。全行业的专利初次申请通过率通常仅在 15% 左右，律师必须在追求工作效率的同时做到极致详尽。 为了缓解检索压力，Sarah 团队基于大语言模型与检索增强生成（RAG）构建了一款名为 **"Chat White Book"** 的语义检索应用。与普通仅返回 token 片段的检索工具不同，该应用能够提取出完整的对应上下文及关联引用，为律师组织有力的法律依据。 - **试点规模**：在约 50 名专利律师中率先推广。 - **效率提升**：单起案件平均为律师节省 5 至 10 小时。 - **业务收益**：按每位律师每年处理近 50 起案件计算，该工具每年预计可直接节省 25 万美元（$250K），同时明显拉升了专利申请的通过率。 --- ### [05:48](https://www.youtube.com/watch?v=tYREmZVqhKk&t=348s) 工具泛滥的瓶颈与智能体 "Lex" 的破局 看到 Chat White Book 的显著收益后，Sarah 团队决定进一步扩展语料库，针对现有技术、研究文献、技术规范、法院判决等不同专业领域，陆续开发了 10 个功能各异的专用 RAG 工具。然而，这随即引发了严重的**“工具蔓延/泛滥（Tool Sprawl）”**问题：专利律师们不得不耗费大量宝贵时间去思考“应该在什么阶段按何种顺序使用哪个工具”，流程反而被割裂。 ```mermaid flowchart TD accTitle: 从分散的专业 RAG 工具演进为统一的单 Agent 系统 Lex subgraph ToolSprawl["工具蔓延困境 (Tool Sprawl)"] T1["现有技术检索工具"] T2["判例法检索工具"] T3["技术规范工具"] T4["法院判决工具"] end subgraph LexAgent["单一智能体系统: Lex (罗氏 Roche 实践)"] L["Lex Agent 统一调度"] DS[("数万份底层源文档")] L --> DS L --> OB["自动串联工作流并生成法律摘要简报 (Brief)"] end ToolSprawl -->|"替代离散工具序列"| L ``` 为了彻底解决工具割裂，团队构建了名为 **"Lex"** 的单智能体（Single Agent）系统： - **全流程贯通**：根据律师输入的特定需求，自动串联起数十万份底层源文档。 - **输出完整简报**：不仅提供内容摘要，更一并打包所有关联法条与案卷引用，直接生成专利申请所需的完整辩护简报（Brief）。 - **打破“生成式 AI 悖论（GenAI Paradox）”**：许多企业投入 GenAI 却无法将其转化为实际利润或生产力，原因在于部署与消费模式脱节；而 Lex 之所以成功，是因为它由专利律师团队专为专利律师群体设计，并借助 Dataiku 的 RAG 与对话式 UI 能力无缝嵌入了实际工作流。 - **案例真实背景**：Sarah 的故事并非虚构，该案例来自知名跨国药企罗氏（Roche）。 --- ### [08:28](https://www.youtube.com/watch?v=tYREmZVqhKk&t=508s) 从 RAG、单 Agent 到多 Agent 系统的进阶路径 企业从基础的 RAG 应用迈向真正的 Agent 系统是一个清晰的演进过程。演讲者以销售（Sales）领域的实际场景为例，梳理了不同阶段的技术形态与能力跨越： | 进阶阶段 | 销售场景具体形态 | 核心能力与行为模式 | |:--- |:--- |:--- | | **基础 RAG 应用** | 客户满意度洞察工具 | 仅用于被动回答简单业务问题（如分析客户对报价的看法、去年丢单的主要原因等）。 | | **单 Agent 系统 (Single Agent)** | 个性化销售辅助智能体 | 兼具“辅助分析”与“主动行动”能力；结合客户满意度洞察与历史交互，自动预测并生成定制化提案或外发邮件（企业可自行配置其是否具备全自动发送的 Agency 权限）。 | | **多 Agent 系统 (Multi-Agent)** | 动态协同销售团队 | 调度多个专业智能体协同作业（例如：竞对情报智能体 + 定价优化智能体 + 客户关系管理智能体），各司其职共同辅助客户经理（AE）促成交易关闭。 | | **Agentic 系统 (Agentic Systems)** | 虚拟销售培训师 (Virtual Sales Trainer) | 深度集成于业务环境中的全自主演练与进阶辅导系统。 | ### [10:35](https://www.youtube.com/watch?v=tYREmZVqhKk&t=635s) 自主优化系统与 Agentic AI 的业务价值演进 从单笔交易分析迈向全局自主优化，标志着智能体系统（Agentic Systems）的成熟。在高度优化的自主系统中，AI 代理不再仅仅审视孤立的个案，而是跨越多个交易进行综合模式识别，自动分析哪些策略行之有效、哪些陷入停滞，并主动执行后续步骤——例如自动更新销售手册，甚至为一线销售代表提供实时辅导。这一演进路径呈现出清晰的技术叠加逻辑：首先是通过检索增强生成（RAG）构建生成引擎；接着引入单一代理（Single Agent）以赋予行动能力；随后扩展为多代理协同（Multi-Agent Collaboration）；最终演进为能够自主调优与闭环迭代的代理系统。 ```mermaid flowchart LR accTitle: Agentic AI 演进路径与业务影响 subgraph TechEvolution["技术演进层级"] direction LR A["RAG 生成引擎"] --> B["单代理 Action 行动"] B --> C["多代理协作 Collaboration"] C --> D["自主优化 Agentic 架构"] end subgraph BusinessImpact["业务转型释放"] direction TB E["员工赋能<br/>(Worker Augmentation)"] F["流程自动化<br/>(Process Automation)"] G["智能业务链<br/>(Intelligent Business Chain)"] E & F --> G end D -.-> BusinessImpact ``` 在企业价值的兑现层面，生成式 AI 与代理技术的业务影响主要分布在三大维度，并且三者之间能够串联融合： - **员工赋能（Worker Augmentation）**：侧重于大幅提升知识工作者的日常生产力与决策效率。 - **流程自动化（Process Automation）**：彻底消除企业内部繁琐、重复的手工与日常流程。 - **智能业务链（Intelligent Business Chain）**：将员工赋能与流程自动化无缝串联进统一的端到端工作流中。 当这三种能力深度交融，一名被技术赋能的员工能够随时调用自动化流程，并通过全链路工作流驱动业务时，企业迎来的将不仅是渐进式的运营改善，而是深层次的业务转型。这种根本性变革将重塑客户体验，开创全新的商业模式，并解锁截然不同的产品与服务变现路径。 --- ### [12:35](https://www.youtube.com/watch?v=tYREmZVqhKk&t=755s) 阻碍企业落地的三大挑战：ROI 缺失、代理蔓延与治理真空 尽管代理技术的潜力巨大，但许多企业在推进业务转型时仍被一系列严峻现实所束缚。阻碍企业释放业务价值的核心痛点集中在三个方面： ```mermaid flowchart TD accTitle: 企业落地 Agent 面临的核心阻碍 subgraph Challenges["三大核心阻碍"] C1["投资回报率缺失<br/>(Lack of ROI)"] C2["代理无序蔓延<br/>(Agent Sprawl)"] C3["缺乏统一治理<br/>(Lack of Governance)"] end C1 --- R1["• 选题错误<br/>• 过度工程化<br/>• 员工缺乏理解"] C2 --- R2["• 异构孤岛与互操作困难<br/>• 级联故障与根因分析难<br/>• 代理衰退 (Agent Decay)"] C3 --- R3["• Token 消耗与成本失控<br/>• 缺乏单一可见性视窗<br/>• 安全、审计与追踪缺失"] ``` - **投资回报率缺失（Lack of ROI）**：很多项目无法验证成效，其根源往往在于选错了业务问题、对原本简单的场景进行了过度工程化设计，或是企业员工根本不理解该项技术及其具体应用场景。 - **代理蔓延（Agent Sprawl）**：企业内部正面临来自四面八方的代理碎片化涌入——Salesforce、Workday 等 SaaS 应用自带内置代理，专注垂直场景的初创公司交付专用代理，各大云厂商（Hyperscalers）推出平台代理，同时企业内部还在自研定制代理。各代理之间缺乏互操作性，无法顺畅通信，极易导致重复造轮子。随着多代理系统链路延长，一旦某个环节失效，将引发级联故障，其实时根因定位与修复极其困难。此外，“代理衰退（Agent Decay）”已成为现实问题，随着业务数据和环境上下文的变化，模型表现会持续退化，如何及时检测衰退并维持系统信任度仍待解决。 - **缺乏统一治理（Lack of Unified Governance）**：多团队各自为政进行代理实验是 CIO 的“终极噩梦”。如果缺乏全局视窗来统一监控每日 Token 生成量及相关算力开销，项目预算将迅速失控；同时，数据安全合规、操作审计与全链路可追溯性等核心治理诉求也将面临巨大盲区。 --- ### [15:01](https://www.youtube.com/watch?v=tYREmZVqhKk&t=901s) 破解之道：业务价值导向、平台底座与以人为本的组织转型 要突破上述困境，企业必须在战略推进与落地实践中确立明确的行动纲领。成功的关键不仅在于技术架构的选型，更在于业务与 IT 的协作模式，以及对组织中“人”的持续投资。 ```mermaid flowchart TD accTitle: 代理落地战略破解框架 subgraph Strategy["落地实施三大支柱"] S1["以业务价值为引领<br/>(Lead with Business Value)"] S2["投资正确平台底座<br/>(Enterprise Platform)"] S3["重塑与赋能人才团队<br/>(Focus on Workforce)"] end S1 --> |业务与 IT 协同定义指标| P["可量化的业务回报"] S2 --> |编排、监控与统一治理| P S3 --> |透明沟通与技能升级| P ``` - **以业务价值为引领（Lead with Business Value）**：在实现高价值回报的标杆企业中，普遍采用业务与 IT 深度协同的模式。由业务部门提出核心痛点、界定正确的问题陈述并制定明确的成功衡量指标，确保代理技术在交付后能够产生可被精准量化的商业成果。 - **投资正确的平台底座（Invest in the Right Platform）**：要快速保障 ROI 并化解异构蔓延，企业需要引入具备企业级编排能力的统一平台，以解决跨系统互操作难题；平台还需具备持续监控与优化机制，在发生衰退前主动拦截性能降级，并提供全链路的通用治理与成本管控。 - **聚焦人才与组织赋能（Focus on Your Workforce）**：企业往往在系统、平台和技术维护上投入巨资，却容易忽视最核心的资产——员工。作为企业管理者，必须承担起公开透明沟通的责任，明确告知代理系统的引入对具体岗位职责意味着什么，帮助员工预判未来角色变化并掌握所需新技能；而对于职场个人而言，不应被动等待，而要主动利用企业内部的培训与发展预算（L&D Budget），积极提升 AI 时代的新型职业能力。 在演讲尾声，讲者总结了 Dataiku 在支持企业代理化转型与治理方面的能力，并邀请对实际代理落地案例、企业规模化扩展挑战感兴趣的与会者前往 406 号展位交流探讨。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=tYREmZVqhKk&t=4s) Dataiku 已成立超过 12 年，服务了 700 多家客户，并帮助客户实现了平均 4 倍的投资回报率（4x ROI）。 — Arthy Krishnamurthy > [00:04](https://www.youtube.com/watch?v=tYREmZVqhKk&t=4s) 将 Agent 定义为：以大语言模型（LLM）为驱动、旨在通过多步骤方式达成既定目标，并且拥有代表人类采取行动之“代理权（Agency）”的系统。 — Arthy Krishnamurthy > [02:36](https://www.youtube.com/watch?v=tYREmZVqhKk&t=156s) 全行业的专利初次申请通过率通常仅在 15% 左右，律师必须在追求工作效率的同时做到极致详尽。 — Arthy Krishnamurthy > [02:36](https://www.youtube.com/watch?v=tYREmZVqhKk&t=156s) 按每位律师每年处理近 50 起案件计算，该工具每年预计可直接节省 25 万美元（$250K），同时明显拉升了专利申请的通过率。 — Arthy Krishnamurthy > [05:48](https://www.youtube.com/watch?v=tYREmZVqhKk&t=348s) 专利律师们不得不耗费大量宝贵时间去思考“应该在什么阶段按何种顺序使用哪个工具”，流程反而被割裂。 — Arthy Krishnamurthy > [05:48](https://www.youtube.com/watch?v=tYREmZVqhKk&t=348s) 许多企业投入 GenAI 却无法将其转化为实际利润或生产力，原因在于部署与消费模式脱节；而 Lex 之所以成功，是因为它由专利律师团队专为专利律师群体设计。 — Arthy Krishnamurthy > [10:35](https://www.youtube.com/watch?v=tYREmZVqhKk&t=635s) 当一名被技术赋能的员工能够随时调用自动化流程，并通过全链路工作流驱动业务时，企业迎来的将不仅是渐进式的运营改善，而是深层次的业务转型。 — Arthy Krishnamurthy > [12:35](https://www.youtube.com/watch?v=tYREmZVqhKk&t=755s) 很多项目无法验证成效，其根源往往在于选错了业务问题、对原本简单的场景进行了过度工程化设计，或是企业员工根本不理解该项技术及其具体应用场景。 — Arthy Krishnamurthy > [12:35](https://www.youtube.com/watch?v=tYREmZVqhKk&t=755s) 随着业务数据和环境上下文的变化，模型表现会持续退化，如何及时检测衰退并维持系统信任度仍待解决。 — Arthy Krishnamurthy > [12:35](https://www.youtube.com/watch?v=tYREmZVqhKk&t=755s) 多团队各自为政进行代理实验是 CIO 的“终极噩梦”。如果缺乏全局视窗来统一监控每日 Token 生成量及相关算力开销，项目预算将迅速失控。 — Arthy Krishnamurthy > [15:01](https://www.youtube.com/watch?v=tYREmZVqhKk&t=901s) 企业往往在系统、平台和技术维护上投入巨资，却容易忽视最核心的资产——员工。 — Arthy Krishnamurthy ## Mentioned Arthy Krishnamurthy, Dataiku, Sarah, Roche (罗氏), Chat White Book, Lex, Salesforce, Workday ## Video description Arthy Krishnamurthy, Senior Director, Americas Business Transformation Lead, Dataiku Abstracts As agentic AI moves from pilot to production, a clear shift is underway—it’s driving workforce evolution. This session explores real-world deployments, including a major pharma company, where AI agents autonomously make decisions and collaborate with humans. Through the lens of a patent attorney’s transformed workflow, we’ll highlight key success factors that distinguish impactful deployments from the 78% stuck in the “GenAI paradox.” [Source](https://www.youtube.com/watch?v=tYREmZVqhKk)

[0:04] Arthy Krishnamurthy: Dataiku 已成立超过 12 年，服务了 700 多家客户，并帮助客户实现了平均 4 倍的投资回报率（4x ROI）。 [0:04] Arthy Krishnamurthy: 将 Agent 定义为：以大语言模型（LLM）为驱动、旨在通过多步骤方式达成既定目标，并且拥有代表人类采取行动之“代理权（Agency）”的系统。 [2:36] Arthy Krishnamurthy: 全行业的专利初次申请通过率通常仅在 15% 左右，律师必须在追求工作效率的同时做到极致详尽。 [2:36] Arthy Krishnamurthy: 按每位律师每年处理近 50 起案件计算，该工具每年预计可直接节省 25 万美元（$250K），同时明显拉升了专利申请的通过率。 [5:48] Arthy Krishnamurthy: 专利律师们不得不耗费大量宝贵时间去思考“应该在什么阶段按何种顺序使用哪个工具”，流程反而被割裂。 [5:48] Arthy Krishnamurthy: 许多企业投入 GenAI 却无法将其转化为实际利润或生产力，原因在于部署与消费模式脱节；而 Lex 之所以成功，是因为它由专利律师团队专为专利律师群体设计。 [10:35] Arthy Krishnamurthy: 当一名被技术赋能的员工能够随时调用自动化流程，并通过全链路工作流驱动业务时，企业迎来的将不仅是渐进式的运营改善，而是深层次的业务转型。 [12:35] Arthy Krishnamurthy: 很多项目无法验证成效，其根源往往在于选错了业务问题、对原本简单的场景进行了过度工程化设计，或是企业员工根本不理解该项技术及其具体应用场景。 [12:35] Arthy Krishnamurthy: 随着业务数据和环境上下文的变化，模型表现会持续退化，如何及时检测衰退并维持系统信任度仍待解决。 [12:35] Arthy Krishnamurthy: 多团队各自为政进行代理实验是 CIO 的“终极噩梦”。如果缺乏全局视窗来统一监控每日 Token 生成量及相关算力开销，项目预算将迅速失控。 [15:01] Arthy Krishnamurthy: 企业往往在系统、平台和技术维护上投入巨资，却容易忽视最核心的资产——员工。

[0:04] Arthy Krishnamurthy: holding the "agency" to act on behalf of the user [2:18] Arthy Krishnamurthy: Chart White Book [5:44] Arthy Krishnamurthy: tool sprawl [5:44] Arthy Krishnamurthy: GenAI Paradox [11:22] Arthy Krishnamurthy: a CIO's worst nightmare [11:22] Arthy Krishnamurthy: Worker Augmentation [11:22] Arthy Krishnamurthy: Process Automation [11:22] Arthy Krishnamurthy: Intelligent Business Chains [17:36] Arthy Krishnamurthy: Go figure out what that budget is and go start learning and have these conversations with your teams.

[0:04] So today I'm here to talk about um [0:06] agents in action. So I figured it's [0:09] because it's kind of end of day, it [0:10] might be easier if I start off by [0:12] telling a story and then um also share [0:16] some scaling issues enterprises are [0:18] facing as they're starting to build [0:20] these agents and trying to scale them. [0:23] And as we go through this, we'll also [0:24] talk about some of the key patterns on [0:27] how to overcome these issues. [0:29] So quick intro about me. I lead the [0:31] business value and transformation [0:32] practice at data. So what that means is [0:35] I work with our customers to help the [0:38] right problem to help uh identify the [0:40] right problem that AI can solve and once [0:43] they start seeing value from it and they [0:45] want to scale the value then I work with [0:47] them to establish the right people and [0:49] the process operating model. [0:52] Data IQ, the company that I'm a part of, [0:54] for those unfamiliar, we uh are we've [0:57] been in existence for 12 plus years and [1:00] across 700 plus customers, we've helped [1:02] deliver forex ROI. How did we do that? [1:05] We help them create, deploy and govern [1:08] data and AI products at scale. And now [1:10] with agents, we are pretty much [1:12] extending the same philosophy. The cool [1:15] thing about us uh our platform is we are [1:17] agnostic to the underlying data and [1:19] infra layer because of which customers [1:22] um are able to futureproof their [1:23] investments through us. [1:26] Now before I jump in um what I want to [1:29] do is um for the purpose of semantic [1:32] alignment just define what I mean by an [1:34] agent. So again, most of you are [1:36] probably aware an LLM powered system [1:39] designed to achieve objectives typically [1:41] through a multi-step approach and the [1:43] key part it has the agency to act on [1:45] your behalf. The amount of agency that [1:48] you provided is really a function of the [1:49] problem statement that you're trying to [1:51] solve and each industry is different and [1:53] the risk appetite is different but you [1:55] get the idea. A quick show of hands [1:58] here. How many people um are building or [2:01] have built an AI agent? [2:05] Okay. How many of you are leading a team [2:07] of such builders? [2:10] Okay. Do we have any folks here from the [2:13] the business side? Any business [2:14] stakeholders? [2:16] Okay. None in the room. All right. So, [2:19] what I want to do now is uh tell you a [2:21] story. And in case you were wondering [2:23] why I asked for a show of hands, uh, [2:25] yes, I I was planning to fine-tune some [2:27] of some portions of my presentation, but [2:29] more importantly, wanted to make sure [2:31] you guys were listening. So, you guys [2:32] passed. Great. So, let me start off by [2:36] telling you a story about Sarah. [2:39] So, Sarah is uh a patent attorney and [2:42] she works for one of the largest [2:43] pharmaceutical companies in Western [2:45] Europe. She leads a team of around 200 [2:48] patent attorneys. So Sarah's job is [2:51] multiaceted. [2:52] Three things she's responsible for. One, [2:55] she's responsible for strategic [2:57] portfolio management. What that means is [2:59] she works with her company to identify [3:01] inventions that can be patented. Second, [3:05] she has to make sure the the company's [3:07] able to go through all the regulatory [3:08] policies etc. that's needed to get an [3:12] approval for a drug. And the third [3:14] thing, she wants to you the company [3:16] wants to make money. So she's there to [3:18] figure out how. [3:20] Now let's talk about the day in the life [3:22] of Sarah. So the key thing is she and [3:25] her team literally have to do a lot of [3:27] due diligence. So what does that mean? [3:30] In order for you to file a patent, you [3:32] first have to make sure that your [3:33] invention is truly novel. No one has [3:36] ever invented such a thing before. So [3:38] you look at something called prior art, [3:40] which really you're looking through [3:42] existing patents to see if you know your [3:44] idea uh was there before or not. [3:47] Once you've identified that, now you're [3:49] going to submit a patent application. [3:51] And guess what? You think it'd be easy, [3:53] but no. You have to go through a lot of [3:56] case files, case laws to make sure [3:58] you're following certain guidelines. So, [4:00] she and her team are going through pages [4:03] and pages of documents, understanding [4:05] these case laws to make sure when is [4:07] when they submit a patent, the [4:09] acceptance rate is actually high and it [4:10] goes through. industry standard [4:12] acceptance rate for uh patent [4:14] application stands close to 15 [4:15] percentage. So she wants to make sure [4:17] that when she submits a patent it better [4:19] go through. [4:21] So there's quite a bit of documents and [4:24] I'll talk about some of them in a [4:25] minute. But the whole pain point here is [4:28] how does she manage efficiency but also [4:32] be very thorough in her process. So [4:34] that's kind kind of the pain that she's [4:36] dealing with. So when she heard about [4:38] the promise of um large language models [4:41] and uh retrieval all the rag based um [4:44] applications that everyone else was [4:45] building, Sarah decided to give it a [4:48] shot. She built what was called as the [4:50] chart white book which is basically a [4:52] rag powered application does semantic [4:54] search not just retrieving tokens but it [4:57] goes beyond that and is able to extract [4:59] the full text corresponding to those [5:02] tokens. So what that means is it's not [5:05] just a specific section of the text, but [5:08] any related references that Sarah and [5:10] her team need in order to make a [5:11] compelling case. When she started [5:14] rolling this out across uh around like [5:17] 50 attorneys, what she started to see [5:19] was the value Chad White Book was [5:21] starting to deliver. It was able to [5:23] quickly save 5 to 10 hours attorney [5:25] hours per case. And typically these [5:28] people handle close to 50 cases per [5:30] year. All of a sudden, they're looking [5:32] at 250K in estimated annual savings. [5:36] So clearly there's value and she's also [5:39] able to increase the acceptance rate for [5:41] these patents. Now if you're Sarah, what [5:44] would you do? You see this value. What [5:46] are you thinking? [5:48] You decide to do more. You decided to [5:51] build more tools. And that's what she [5:53] did. She kind of expanded the corpus to [5:56] include all different kinds of documents [5:58] that they would need in order to support [6:00] their prosecution. Everything from prior [6:02] art, research documents, technical [6:04] specifications, code decisions, you name [6:06] it. And before you know, she had 10 [6:09] different tools. And the reason it's 10 [6:10] different is because each has an area of [6:12] specialtity. And boom, what happened? [6:15] The patent attorneys now had a different [6:18] problem. [6:19] they were spending precious time trying [6:22] to figure out which tool to use in what [6:24] sequence in order to make a compelling [6:27] case. [6:28] So now she had what we call as a tool [6:30] sprawl. [6:32] So she decided to solve this problem. [6:35] How? By introducing Lex. So what is Lex? [6:39] Lex is a single agent system and it's [6:42] basically designed to replace entire [6:45] workflows, string together multiple [6:48] documents based upon a specific query [6:50] from a patent attorney. Under the hood, [6:53] it's really looking at tens and [6:54] thousands of uh source documents and it [6:58] compiles a brief for the patent attorney [7:00] to go file a patent application. Now, [7:03] the key thing about the brief, it's not [7:05] just a summary, but it's also putting [7:07] together related references. everything [7:08] a patent attorney needs in order to go [7:12] uh file a compelling case. [7:15] Now, just like the thing about Lex that [7:19] you just heard, it basically helped [7:21] Sarah break out what uh break out of [7:23] what we call as a genai paradox. What is [7:26] that? It's this observation that genai [7:29] basically has a lot of potential but it [7:32] almost it doesn't always translate into [7:34] business value specifically pertaining [7:36] to profitability or productivity [7:38] improvement. Why is that? It's because a [7:41] lot is dependent on how you deploy these [7:44] applications how you consume these [7:46] applications and even the right use case [7:47] that you pick. [7:50] So I'm here to tell you that the story [7:52] that you heard about Sarah is not [7:54] hypothetical. [7:56] Lex is very much a real single agent. It [7:59] was built by one of our customers, Ro. [8:03] And the cool thing about Lex is it was [8:06] built by a team of patent attorneys for [8:09] patent attorneys. [8:11] How do they do that? Using data's um rag [8:15] capabilities and conversational UI. So I [8:18] talked about the business value just [8:19] from one of the tools. The real value of [8:22] Lex is still being computed. Um we [8:24] should be sharing more details about it [8:25] in the coming month. [8:28] Now you just heard about Lex right and [8:31] the thing about RO but in reality the [8:34] question that often we get asked is what [8:36] does it look like as enterprises level [8:39] up from rag based applications through [8:41] to single agents to multi- aents etc. I [8:44] want to take the example of um sales and [8:47] kind of walk you through how that whole [8:50] journey looks like. So let's look at [8:52] maybe um let's start with sales and [8:55] within sales let's start with a rag [8:57] based customer satisfaction insights [8:59] application. What's the idea here? So [9:02] the system is really trying to help an [9:04] account exec answer a simp answer answer [9:06] some of the simple questions like okay [9:08] what does my enterprise customer think [9:09] about the pricing or looking at deals [9:12] last year? What was the main reason we [9:14] lost? [9:16] So in this case it's really existing to [9:18] answer some of those simple questions. [9:21] Now when you move from rag to a single [9:23] agent personalized selling uh system, [9:26] you're now looking at an agent that's [9:29] actually assisting uh an account exec [9:32] but can also act if it decides to. What [9:35] does that look? So for example, it could [9:37] be looking at customer satisfaction [9:39] insights, previous customer [9:41] interactions, and it could predict or [9:43] maybe even generate a custom proposal. [9:46] It could generate an outbound email. [9:48] Again, do you want to make it completely [9:50] autonomous and have it send it or not? [9:52] That's a decision uh you as an [9:54] enterprise can take multi- aents. As we [9:57] move on here, we are looking at [9:58] orchestrating multiple agents. What does [10:00] that look like? So, when we have a team [10:02] of dynamic selling agents, you basically [10:05] have let's say three different agents. [10:06] For example, one could be doing [10:08] competitive intelligence. Another could [10:11] be looking at pricing optimization. A [10:13] third could be looking at customer [10:14] relationship management. The key thing [10:16] is all three of them are working [10:19] together to help the account exec close [10:22] deals. [10:23] Now once we move into the realm of [10:25] agentic systems, what is that look like [10:28] from a sales standpoint? Here we're [10:31] looking at an example of a virtual sales [10:33] trainer. What is that? Uh let me show [10:35] you how that interaction looks like. So [10:38] here you're no longer looking at one [10:40] single deal, but rather the agent is [10:42] analyzing multiple deals, looking for [10:45] patterns that are working, not working. [10:47] And guess what? It takes the next step [10:49] of automatically updating your sales [10:51] manual and even coaching your reps in [10:54] real time. That's the optimized [10:57] autonomous system that we're looking at. [10:59] So the key takeaway here is [11:03] each is really building upon what came [11:05] before, right? So with RA you started [11:07] with a generative engine and then you [11:10] added the notion of action with a single [11:12] agent and then you added the notion of [11:14] collaboration with multi- aents and then [11:16] you're going into autonomous [11:17] optimization with agentic systems. [11:22] So that's that's all great. Now what are [11:25] enterprises actually seeing in terms of [11:27] business impact? So there's really three [11:30] different categories that we see [11:31] business impact getting unlocked today. [11:34] One it's around worker augmentation the [11:37] whole notion of improving productivity [11:38] of your worker. [11:40] The second is around process automation [11:43] uh eliminating manual mundane processes. [11:46] And the third is intelligent business [11:48] chain which is really stringing together [11:49] worker augmentation and process [11:51] automation together into a workflow. [11:55] Now what happens as things mature and [11:57] there all these three capabilities start [11:59] to [12:01] before you know let's imagine you have a [12:04] an augmented worker invoking an [12:07] automated process and stringing all of [12:09] this together using a workflow. All of a [12:12] sudden you're not just looking at [12:14] changing or making an operations better. [12:16] You are transforming the way things [12:18] really work. And that's really where new [12:20] things start to happen. You start to [12:22] reimagine customer experience. You start [12:25] to unlock new business models. New you [12:27] have uh completely new different way [12:30] different ways in which you can monetize [12:31] your product or service. [12:35] So today if you look at enterprises and [12:37] kind of see okay where are they with [12:38] respect to unlocking business impact or [12:40] business transformation [12:42] there's a few challenges that's really [12:44] holding them back. Well there might be [12:46] several challenges. I just want to focus [12:48] on and anchor on three uh unique ones [12:51] that we typically hear. First is lack of [12:54] ROI. Why is that? It could be different [12:57] reasons as to why there's lack of ROI. [12:59] Picking the wrong problem statement or [13:01] overengineering the solution to a simple [13:03] problem statement or um maybe your the [13:06] employees don't even understand what the [13:08] technology is or how to apply it. [13:12] Second challenge, agents bro. [13:15] You have agents coming out of your um [13:17] applications like Salesforce and [13:19] Workday. You have applications coming [13:21] out of key startups which are very fit [13:23] for purpose. You have agents coming from [13:26] uh different hyperscalars and you you [13:28] might decide to build custom agents as [13:30] well. How are we going to ensure all [13:33] these different agents interoperate, [13:35] talk to each other, do not build [13:37] duplicative stuff and kind of share [13:39] learnings? We we don't we haven't solved [13:41] for that just yet. [13:43] Now imagine as these agents get built, [13:46] these are the failures that get [13:48] cascaded. The issue is how do you [13:50] identify root cause when an agent fails [13:53] and how are you going to remediate those [13:55] actions in real time? And finally, the [13:59] idea of agent DK is pretty real. As data [14:03] changes, business contexts change, how [14:06] are we going to start to identify the [14:08] agent performance as a DK case, [14:10] remediate them and make sure the trust [14:12] is still retained. These are still [14:15] problems out there that I haven't seen [14:18] and approached just yet, but all these [14:20] things we need to solve for. And [14:22] finally, lack of unified governance. [14:25] This is a CIO's worst nightmare as you [14:30] have different teams experimenting with [14:32] different agents. Pretty much costs are [14:35] going to go out of control. If you don't [14:38] have a single pane of visibility into [14:40] the kind of um the number of tokens that [14:42] you're, you know, generating every day [14:44] and the cost associated with it, pretty [14:46] soon the budgets are going to be out of [14:48] control. I haven't even talked about [14:51] security and uh auditability, [14:53] traceability, none of that stuff here, [14:55] but you guys get the picture. [14:59] So now the question is how do we [15:01] overcome um some of these challenges? [15:04] What sort of approach do we take? [15:07] So here the first thing is it really to [15:09] lead with business value. The companies [15:12] or the enterprises that are succeeding [15:14] in generating business value, the model [15:17] that we see work best is where the [15:19] business and the IT are working [15:20] together. It's businessdriven. Business [15:23] is asking the questions or coming up [15:24] with the right problem statements and [15:26] the success metrics so that when an [15:28] agent is implemented, there is success [15:31] at the end of it and you you can [15:32] actually quantify it. [15:35] The second key thing is around investing [15:37] in the right platform. For all the [15:40] reasons that I described previously with [15:42] respect to the challenges enterprises [15:44] are facing, the quickest thing that you [15:46] can do to secure your ROI is to invest [15:48] in the right platform. A platform that [15:50] can help you with um enterprise [15:52] orchestration so you can solve the [15:54] interoperability problem. something that [15:56] can um help you with continuous [15:58] monitoring and optimization so you can [16:00] detect those agent u decay before things [16:03] happen and and finally the notion of uh [16:06] universal governance. [16:09] The last thing I want to call out and [16:10] this is something I feel is is often not [16:14] u focused or it's not something u [16:17] enterprises tend to lead with or focus [16:19] on is to focus on your workforce. I read [16:22] this interesting article uh from from [16:24] one of the leaders that really talks [16:26] about this quite a bit. It's this [16:28] observation that as leaders as [16:31] enterprises we tend to invest a whole [16:33] lot on systems and platforms and [16:35] technologies and trying to figure out [16:37] how to maintain them etc. Yet we often [16:40] overlook the single most important thing [16:42] that matters to the company the people. [16:46] So what do leaders do? As leaders in [16:48] this room you have a responsibility. [16:51] What is it? Be really transparent out [16:54] there and talking about okay yes we're [16:56] introducing agents but what does it mean [16:58] for a certain individual's role? How is [17:01] his or her role going to change and what [17:04] kind of skills do they actually need in [17:06] order to you know keep up with the [17:08] workforce or the role of the future and [17:11] how how are these different learning and [17:13] development teams and capabilities [17:15] changing to support your workforce as [17:18] leaders. I think the onus is on you to [17:20] make sure you you have these [17:21] conversations. [17:23] And how about the individuals in the [17:24] room? Don't wait. I don't think you need [17:27] to wait. I think you need to take [17:28] control, right? U there's so many [17:30] learning programs out there that you can [17:32] sign up for or maybe every every company [17:34] that I know has a learning and [17:36] development budget for their employees. [17:38] Go figure out what that budget is and go [17:40] start learning and have these [17:42] conversations with your teams. And [17:44] that's pretty much it with that. I want [17:46] to wrap with a quick uh overview of data [17:48] IU, but if you have if you're interested [17:50] in learning more about data or how um [17:54] you want to see examples of some of [17:56] these agents in action or even just talk [17:58] about your specific scaling challenge, [18:01] feel free to stop by 406 booth number [18:03] 406. And thank you for your time.