AstraZeneca Halts Autonomous AI Agent Meshes Over Cost, Security Risks

阿斯利康因成本与安全风险叫停自主AI Agent网格架构 首席架构师Wayne Filin-Matthews警告称，真正的企业级Agent走向成熟仍需五年时间，且需要对运营模式进行彻底重构，而非任由不受约束的跨平台自主运行。 阿斯利康首席企业架构师Wayne Filin-Matthews阐述了这家跨国制药企业如何在不增加固定IT人员编制的前提下，利用Agentic AI在五年内实现营收翻倍并推出20款新药。阿斯利康并未放任失控的跨平台Agent网格发展，而是专注于在现有核心企业系统内部构建Agent，同时解决可组合性、记忆生命周期以及难以预测的Token成本挑战。讨论重点指出，应利用认知行为科学设计多Agent团队，将合规监管与数据主权直接嵌入自动化工作流，并务实地预期组织运营模式围绕企业级Agent走向成熟仍需五年时间。 企业 AI 战略正经历根本性转变，从部署通用生成式大模型迅速转向由自主与半自主系统构成的“智能体经济”（Agent Economy）。 (0:04) 阿斯利康计划在不扩充现有约 10,000 名 IT 人员的前提下，于 5 年内推出 20 种新药并将营收从 500 亿美元翻倍至 1000 亿美元，智能体自动化成为实现非线性增长的核心驱动力。 (3:00) 面对组合层不成熟、身份委托未决及链式调用引发的算力成本激增，企业应限制失控的跨平台 A2A（智能体间）交互，优先在各核心业务平台内推行智能体开发。 (6:06) 多智能体系统设计应融入认知行为科学与组织动力学（如引入唱反调或“拖延者”角色以避免过早形成草率结论），而非仅视其为 RPA 流程的线性延伸。 (11:34) 自主智能体框架预计需要约 5 年才能在企业中真正成熟，其成功不仅取决于技术沉淀，更依赖于企业运营模式与治理架构的根本性重塑。 (16:36)

AstraZeneca's Chief Enterprise Architect, Wayne Filin-Matthews, argues that meeting aggressive enterprise growth and sustainability targets requires shifting to an agent-driven architecture rather than expanding headcount. However, he cautions that real business value depends on resolving foundational governance and composability gaps rather than rushing toward unconstrained agent-to-agent networks or near-term artificial general intelligence. Scaling Pipeline Velocity Under Fixed Headcount (3:00) AstraZeneca is pursuing an aggressive five-year strategy to launch 20 new medicines and double company revenue from $50 billion to $100 billion. Operating across 126 international markets, including market leadership in China and a $50 billion manufacturing commitment in Virginia, the company must compress the traditional 10-year drug development cycle while maintaining a fixed IT division of roughly 10,000 personnel and meeting a net-zero carbon commitment by 2030. Filin-Matthews explains that linear scaling through IT hiring is unviable. Instead, business priorities centered on patient impact and operational sustainability require using agentic automation to accelerate delivery times without expanding the organizational carbon footprint or staffing overhead. Platform Democratization Versus Uncontrolled Agent Meshes (6:06) Rather than standardizing on a single dedicated agent builder, AstraZeneca focuses on democratizing agent capabilities across established core platforms, including SAP Joule, ServiceNow, Microsoft AI Foundry, and AWS Bedrock Agent Core. Filin-Matthews notes that virtually every enterprise software provider is embedding agent functionality directly into its existing stack. Concurrently, Filin-Matthews placed an enterprise hold on broad, unconstrained agent-to-agent (A2A) communication across frameworks such as the Model Context Protocol (MCP). He attributes this restriction to critical deficits in the composability layer, specifically unresolved challenges in delegated identity access, lifecycle governance for short- and long-term agent memory, and unpredictable compounding costs when multi-agent chains traverse multiple external APIs, search tools, and inference endpoints. Cognitive Multi-Agent Design and Automated Governance (11:34) Rejecting the idea of agents as mere robotic process automation (RPA) workflows, Filin-Matthews highlights the necessity of incorporating cognitive behavioral science into multi-agent architectures. In collaboration with Stanford University, AstraZeneca is exploring specialized scientific personas—such as a contrarian or 'procrastinator agent' designed to challenge early consensus and prevent premature conclusions during hypothesis generation. This structural reimagining extends to regulatory compliance. Rather than deploying retrieval-augmented generation (RAG) tools like Amazon Q simply to help staff read bureaucratic checklists, Filin-Matthews advocates redesigning governance so agent teams autonomously validate data quality, assess security risks, and enforce cross-border restrictions like China's Personal Information Protection Law (PIPL). Humans are positioned 'in the loop' or 'on the loop' exclusively at decisive governance gates. Historical Cloud Lessons and the Five-Year Horizon (16:36) Reflecting on his 38 years in enterprise technology, Filin-Matthews compares current expectations to early cloud forecasts 15 years ago during Microsoft's BPOS era, when industry observers prematurely declared that hybrid infrastructure would disappear within five years. He rejects claims that artificial general intelligence will arrive in the next two to three years. Instead, Filin-Matthews projects that enterprise agency will require approximately five years to achieve full maturity. In his assessment, this timeline will involve continuous iteration, operational missteps, and, most crucially, a comprehensive restructuring of enterprise operating models to properly support agentic workflows at scale. Filin-Matthews concludes that enterprise value from AI agents depends on foundational architectural discipline rather than speculative milestones like AGI. The primary hurdle remains organizational: companies must resolve composability, memory governance, and regulatory sovereignty while adapting internal operating models to absorb multi-agent systems.

阿斯利康首席企业架构师Wayne Filin-Matthews指出，面对业务翻倍与严格的资源限制，企业级人工智能正加速转向“智能体经济”；然而，企业必须警惕过热预期，通过重塑组织运营模式与治理层，在未来约五年的演进周期中构建安全、可控的多智能体系统。 业务扩张与资源约束驱动的技术转向 (3:00) 阿斯利康目前在全球126个市场开展业务，是中国市场领先的跨国药企，近期还承诺向美国弗吉尼亚州的生产基地投资500亿美元。根据公司的五年战略规划，其目标是将新药研发上市周期从传统的10年大幅缩短，在5年内推出20种新药，并在未来4年内将年营业收入从500亿美元翻倍至1000亿美元。然而，在实现这一目标的同时，阿斯利康面临着严格的运营边界：其全球IT团队规模将维持在约10,000人不变，且必须在2030年前达成净零碳排放目标。 Filin-Matthews指出，单纯依靠线性增加技术人员的扩张模式已不可行。面对“以患者为中心”与“可持续发展”的双重诉求，技术团队必须依托自动化与AI来打破产能瓶颈。从今年年初与AWS、微软等头部云厂商的交流中可以明显观察到，企业的技术叙事正在迅速跨越基础通用大模型，转向以自主和半自主智能体为核心的“智能体经济”，以智能化手段支撑极速扩张与减排目标。 核心平台赋能与智能体交互管控 (6:06) 在企业智能体架构设计上，Filin-Matthews认为行业起初争论“应该使用哪个单一平台构建智能体”属于伪命题，因为当前主流企业级软件都在加速原生集成智能体能力，包括ServiceNow、SAP Joule、Microsoft AI Foundry以及AWS Bedrock等。阿斯利康因此确立了在现有成熟核心平台内部普及智能体创建的原则，以激发各业务线与科研部门的技术民主化。 与此同时，Filin-Matthews对大规模放开跨平台的自主智能体间网络化交互（A2A）持保留态度，即便面对支持模型上下文协议（MCP）的框架也暂未全面开放。他强调当前智能体的“可组合层”存在显著的企业级治理缺失：首先是跨平台身份委托与权限最小化难以统一追踪；其次，若智能体链条串联起多个大模型、搜索工具与外部API，将导致推理费用与调用成本呈指数级复杂化，预算极难预测；此外，智能体长短期记忆的生命周期管理，也亟须建立类似企业数据留存规范的合规体系。 融合认知行为科学的多智能体协作 (11:34) 针对目前行业将智能体简单视为线性机器人流程自动化（RPA）升级版的做法，Filin-Matthews提出了明确批评。他认为高价值的多智能体应用不仅是算法管道的搭建，更需要借鉴组织动力学与认知行为科学，模拟真实人类团队的交互模式。 为此，阿斯利康正与斯坦福大学等学术机构展开合作，尝试在科研团队中引入特定的人格化智能体角色。例如，在科研探索场景中设计“拖延者智能体”或反共识角色，专门负责挑战团队过早达成的共识，减缓草率下结论的过程，迫使系统对科学假设进行更深层次的推演与论证。这种行为科学导向的智能体设计，正成为突破传统流程自动化的关键方向。 数据主权合规与治理流程重构 (13:42) 在医药这类强监管行业中，Filin-Matthews警告企业切勿利用现代AI去巩固陈旧的官僚机制。例如，许多企业仅将检索增强生成（RAG）技术用于解答复杂的内部合规清单，本质上仍是让人类去适应繁琐的审批流程。真正的变革应当是彻底重构治理本身，让自主智能体团队直接承担风险评估、安全合规检查与数据质量验证，仅在关键决策节点保留人类的监督或干预（Human-in-the-loop / Human-on-the-loop）。 这一框架在应对跨国数据合规时尤为紧迫。面对诸如中国《个人信息保护法》（PIPL）及严格的数据出境评估机制，不受控的自主智能体极易导致敏感信息越界泄露，造成不可逆的合规风险。通过将合规检查、数据驻留约束与主权要求直接内嵌于智能体底层框架，跨国企业才能在确保合规与数据安全的前提下，维系全球科研与商业运营的敏捷交付。 五年成熟周期与组织运营模式重塑 (16:36) 对于未来2至3年的技术布局，Filin-Matthews明确表示看好自主智能体及其框架演进，但明确否定了通用人工智能（AGI）将在2到3年内落地的激进论调。回顾自己在微软从事早期云服务BPOS（Office 365前身）等长达38年的IT从业经历，他指出15年前行业曾预言“5年内传统基础设施将彻底消亡、全面进入云时代”，但至今绝大多数企业仍在使用混合架构。 基于这一历史规律，Filin-Matthews预测企业级智能体要达到真正的成熟可靠并建立完备治理，大约需要5年的试错与演化周期。他强调，智能体落地的最大阻碍并非技术本身，而是企业组织层面的适应能力：企业必须彻底重构其内部运营模式，使其能够适配并管理自主智能体工作流，否则仅将AI强行拼贴于旧体制之上将无法实现预期的商业价值。 Filin-Matthews认为，企业级智能体代表了未来IT架构的核心方向，但其成功落地绝非依靠单一技术工具的堆砌或短期内对AGI的盲目押注。企业在推进智能化演进时，核心挑战在于如何在保持数据主权合规、运行成本可控以及权限安全的前提下，重塑底层的组织运作与治理模式；这一深层变革仍需经历数年的试错迭代与经验沉淀。

## Background 本期内容由 Ai4 于 2025 年 8 月 24 日发布在 YouTube 平台上。视频记录了 Dapple 联合创始人兼首席执行官 Tricia Martinez-Saab 与阿斯利康（AstraZeneca）首席企业架构师 Wayne Filin-Matthews 的炉边对话，深入探讨了阿斯利康如何设计下一代 AI 与智能体基础设施，以支持大规模科学发现、跨国监管治理及可持续发展目标。 ## TL;DR 阿斯利康首席企业架构师 Wayne Filin-Matthews 阐释了公司如何依托 Agent 架构应对激进增长挑战，计划在不盲目扩充人力的情况下实现 4 年内营收翻倍并在 5 年内推出 20 款新药。他指出当前应警惕过早开放大规模跨平台 Agent-to-Agent（A2A）通信，需优先解决可组合层缺失与多跳 Token 成本预测难题；同时主张跳出线性自动化思维，将认知行为学引入多 Agent 科研协作体系。此外，他强调应以自动化合规 Agent 替代传统问答 Chatbot 处理跨国数据主权问题，并基于历史技术周期预判智能体生态需经历约 5 年沉淀与组织重塑方能真正成熟。 ## Atmosphere & style 整场对话沉稳、理性且极具工业实战底蕴，充盈着资深架构师面对技术狂热时的清醒与定力。Wayne 结合其 38 年横跨微软、汇丰与戴尔的工程经验，以极富洞察力的历史镜像（如早期云计算发展路径）从容解构当前的 AGI 炒作与 Agent 落地误区。主持人的精准引导与 Wayne 坦率务实的行业真知相互交织，使整场对话在展现制药巨头宏伟科技蓝图的同时，透露出对企业级护栏、成本治理与人机协同深思熟虑的严谨底色。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=4s) 阿斯利康计划在4年内实现营收翻倍（500亿至1000亿美元）并在5年内推出20款新药，在1万名IT人员的规模限制下，Agent架构是实现非线性生产力跨越的核心路径。 - [00:04](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=4s) 制药研发面临比汽车制造业更严峻的碳足迹挑战，阿斯利康将AI提速与患者生命挽救及2030年净零排放（Net Zero）承诺紧密绑定。 - [06:06](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=366s) 企业无需试图构建单一统一的Agent平台，而应允许各业务部门在SAP Joule、ServiceNow、Azure、AWS等原生平台中实现Agent民主化落地。 - [06:06](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=366s) 现阶段坚决管控大规模跨平台Agent-to-Agent（A2A）通信，核心瓶颈在于企业级可组合层（统一Agent目录、委托授权、可追溯性）尚未成熟。 - [06:06](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=366s) 多Agent在复杂推理链路中涉及搜索、LLM与API的多跳调用，Token与会话成本叠加使得跨Agent预算极难预测，易导致IT成本失控。 - [11:25](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=685s) 落地Agent不能停留在RPA的线性流程延伸思维，必须将认知行为学融入架构，设计具备不同认知特性的Agent协作团队。 - [11:25](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=685s) 在科学探索团队中特意配置“审慎质疑/拖延者 Agent”（Procrastinator Agent），通过放慢节奏和审视前置假设，提升科研成果的严谨性与质量。 - [13:42](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=822s) 告别让人类向Chatbot问答查询规则的低效模式，转向由后台治理Agent团队自主执行合规审计、数据主权核查与人机协同把关。 - [16:10](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=970s) 针对严格的属地化监管（如中国PIPL与跨境数据流动规则），必须依托底层的Agentic架构自动化、规范化地执行安全流转策略。 - [16:22](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=982s) 未来2到3年坚定加倍押注Agent生态，但参照以往云计算基础设施的演进规律，Agent达到真正的企业级成熟与治理至少需要5年试错周期。 - [16:22](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=982s) 企业级AI成功的关键不仅在算法与模型本身，更在于配合Agentic系统对企业内部组织架构与底层运营模型（Operating Model）进行深层重塑。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=4s) 嘉宾介绍与阿斯利康的宏大 AI 战略愿景 对谈在 Ai4 大会开场，Dapple 联合创始人兼首席执行官 Tricia Martinez-Saab 向观众介绍了阿斯利康（AstraZeneca）首席企业架构师 Wayne Filin-Matthews。Wayne 在基础设施与企业级架构领域深耕多年，此前曾担任戴尔（Dell）首席企业架构师、汇丰银行（HSBC）全球数字架构负责人、埃森哲（Accenture）技术咨询资深总监，并在微软（Microsoft）拥有 11 年的企业服务工程与技术转型领导经验。作为福布斯技术委员会与 Gartner 研究委员会成员，他负责掌舵这家全球制药巨头横跨 126 个市场的全球技术与架构战略。 谈及阿斯利康的 AI 成熟度与宏观趋势，Wayne 直言，如果有人说在大型企业落地 AI 很简单，"那绝对不是真的"（"if anybody ever tells you this is easy... it's not"）。AI 本身并不是新鲜事物，但今年 1 月初他在西海岸与 AWS、微软会晤时深刻感受到范式转移——整个行业的对话重心已经全面从传统模型转向了 Agent 经济（Agent Economy）与代理能力（Agency）。对阿斯利康而言，这正切中其激进的发展目标：作为中国市场排名第一的跨国药企，阿斯利康近期宣布在弗吉尼亚州投资 50 亿美元建设美国制造基地，并计划在未来 4 年内将公司营收从 500 亿美元翻倍至 1000 亿美元。传统上研发一款新药通常需要 10 年，而阿斯利康计划在未来 5 年内推出 20 款全新药物。 在庞大的业务扩张压力下，阿斯利康目前拥有约 10,000 名 IT 员工。Wayne 强调，公司不可能通过再雇佣 10,000 名 IT 人员或简单翻倍业务人手来实现规模扩张，Agent 架构是实现这种非线性生产力飞跃的唯一路径。制药行业长期面临着比汽车制造业还要沉重的碳足迹挑战，阿斯利康将核心价值观锚定在"患者至上"与"可持续发展"（目标在 2030 年实现净零排放）之上。Wayne 动情地提到，在研发和药物上市过程中，"分秒必争"（"every moment matters"），新疗法上市越快，在亚洲、非洲等全球各地就能挽救越多患者的生命。 - **业务增长目标**：在 4 年内将企业营收从 500 亿美元翻倍至 1000 亿美元。 - **药物上市提速**：打破传统 10 年一款新药的研发周期，计划在未来 5 年内推出 20 款新药。 - **规模限制与挑战**：IT 团队维持在 10,000 人左右，通过 Agent 架构和 AI 规模化赋能研发与运营，杜绝人力线性扩张。 - **ESG 承诺**：推动制药全流程脱碳，确保在 2030 年达成 Net Zero 净零碳排放目标。 --- ### [06:06](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=366s) 破解 Agent 架构难题：拒绝过早跨平台 A2A，聚焦可组合性与成本治理 面对快速演进的 Agent 生态，拥有 38 年 IT 架构经验的 Wayne 提出了深刻的反思。他分享了自己最喜欢的箴言——"一切都在变，但一切又都没变"（"everything changes but nothing changes"），底层技术的本质逻辑仍在延续，但技术迭代的速度已被极大压缩。在阿斯利康这样崇尚科研探索、业务技术团队高度自治的企业中，今年初许多高管经常问他："阿斯利康构建 Agent 的统一平台是什么？" Wayne 的回答是，这是一个错误的问题，因为所有核心平台都在原生嵌入 Agent 能力，包括 ServiceNow 的工作流、SAP Joule 的财务优化、微软 Azure AI Foundry、AWS AgentCore / Strands 等。 Wayne 采取的架构原则是：**在核心平台内推进 Agent 民主化，但现阶段坚决对大规模跨平台的 Agent-to-Agent（A2A）通信说"不"**。即使相关方案支持 MCP（Model Context Protocol）等协议，跨平台直接互通在当下企业级场景中依然风险巨大。这背后的核心瓶颈在于企业级**可组合层（Composability Layer）**极度缺失：统一的 Agent 目录（Catalog）、委托访问控制（Delegated Access）、可追溯性以及监管合规层尚未成熟。 更为严峻的挑战是**成本预测与控制**。在多 Agent 复杂推理链路中，一个请求可能从 Microsoft Agent Builder 发起，经由 AI Foundry 调度，调用必应检索，再访问大型语言模型，最后触发 MuleSoft 的 API。这一过程中交织着会话成本、Token 成本和 API 费用，导致企业财务和架构团队几乎无法精准预测跨 Agent 交互的预算，极易导致 IT 预算失控。此外，Agent 长期记忆（Long-term Memory）与短期记忆的保存周期亦牵涉严格的数据生命周期合规，必须建立严密的技术护栏。 ```mermaid flowchart TD accTitle: AstraZeneca Agent Governance Strategy accDescription: Architectural diagram showing AstraZeneca enabling domain agents within native SaaS/PaaS platforms while managing the composability bottleneck and preventing unmanaged cross-platform A2A communication. subgraph NativePlatforms["业务原生 Agent 民主化层 (允许自治落地)"] MS["Microsoft 生态<br/>(Agent Builder / Foundry)"] SAP["SAP Joule<br/>(财务与流程优化)"] SNOW["ServiceNow<br/>(自主工单与运维决策)"] AWS["AWS 生态<br/>(Bedrock / AgentCore)"] end subgraph Bottleneck["企业级可组合层 (Composability Layer) 核心挑战"] Identity["Agent 身份与委托授权<br/>(Identity & Delegated Access)"] Catalog["统一 Agent 资产目录<br/>(Agent Catalog & Traceability)"] CostControl["多跳成本与 Token 预测<br/>(Cost & Token Forecasting)"] MemoryGov["记忆生命周期合规<br/>(Long/Short-term Memory)"] end subgraph Blocked["当前受控/限制环节"] A2A["大规模跨平台 A2A 通信<br/>(Agent-to-Agent at Scale)"] end MS --> Bottleneck SAP --> Bottleneck SNOW --> Bottleneck AWS --> Bottleneck Bottleneck -.->|"架构设卡：暂不开放直接互通"| A2A ``` - **架构选型策略**：不押注单一平台，允许各业务部门在 SAP Joule、ServiceNow、Azure、AWS 内利用原生 Agent 工具进行赋能。 - **架构护栏（Guardrails）**：现阶段严格管控跨系统 Agent-to-Agent 直接调度，优先解决身份认证与语义层治理。 - **成本治理盲区**：单次任务涉及多跳链路（搜索、LLM、中间件 API），Token 与会话调用成本叠加，预测难度极高。 - **合规与记忆周期**：明确区分短期上下文与长期记忆，对接企业数据生命周期管理（DLM）合规标准。 --- ### [11:25](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=685s) 超越 RPA 线性思维：引入认知行为学与 Agent 科学协作团队 Wayne 指出，当前许多企业在落地 Agent 时犯的最大错误，就是**以纯技术视角看待 Agent，把它们当作 RPA（机器人流程自动化）的线性延伸**——仅仅设计一个 Agent 去调用接口或执行单向自动化决策。他强调，企业必须跳出单纯的技术堆栈讨论，将**认知行为学（Cognitive Behavioral Science）**和组织行为学融入 Agent 架构中，研究如何组建由多个具备不同认知属性的 Agent 构成的"团队"（Teaming Structure）。 以阿斯利康与斯坦福大学（Stanford）的合作为例，研发部门的目标不是简单扩招人类科学家，而是构建由 Agent 组成的科研协作团队，协助进行深层次科学探索、分子预测与实验结果验证。在真实的顶尖科研团队中，成员并不是步调一致的执行机器；相反，团队中往往需要特定行为特征的角色。Wayne 举了一个生动的例子：在 Agent 科学团队中，他们特意配置了一个**"拖延者 / 审慎质疑 Agent"（Procrastinator Agent）**，因为行为科学证明，当团队中存在这种放慢节奏、审视假设的成员时，科学家反而能产出更严谨的高质量研究成果。 Wayne 呼吁架构师和企业领导者向埃森哲（已赋能 70 万名员工及 Agent 协作系统）等先行者学习，在设计多 Agent 系统时深入考量团队动态与行为交互机制。唯有将认知科学与技术架构紧密结合，才能真正释放 Agent 协同的高阶价值。 > "每个人都在谈技术、技术、技术。但你必须退后一步思考：什么样的 Agent 团队认知行为要素才能最终驱动最佳结果与最高价值？" > —— Wayne Filin-Matthews --- ### [13:42](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=822s) 强监管与数据主权：从问答式 Chatbot 转向自治的合规与治理 Agent 团队 针对制药行业最为棘手的数据主权、合规性与监管治理挑战，Wayne 剖析了传统思维与 Agentic 思维的本质区别。他分享了一个内部真实场景：有团队利用 Amazon Q 构建了一套基于 RAG（检索增强生成）的知识库问答助手，用来回答"企业合规审批的下一步流程是什么？"之类的问题。虽然这被包装成高大上的"AI 流程导航器"，但 Wayne 直截了当地指出，这种模式本质上依然停留在过去，是在浪费技术潜力。 真正的现代化思路，不应该是让人类员工一遍遍去向 Chatbot 提问以了解合规流程，而是**直接用 Agent 重构整个治理、风险管理、安全与数据主权架构**。在新的体系中，由专门的治理 Agent 团队在后台自主完成数据主权校验、合规审计与数据质量核查，并在关键决策节点引入"人机协同"机制（Human-in-the-loop / Human-on-the-loop）进行把关。 Wayne 特别强调了全球本地化合规的复杂性。例如，阿斯利康是中国制药市场的领头羊，如果盲目放开 Agent 自主交互而缺乏内置的合规与主权控制机制，极有可能触碰中国在过去 18 个月内强化实施的跨境数据传输安全评估以及《个人信息保护法》（PIPL）监管红线。因此，架构团队正在建立智能治理团队结构，确保 Agent 在合规的前提下自主运转，保护过往在监管治理上积累的全部资产。 - **去 Chatbot 化思维**：告别依靠人工向问答机器人查询规则的低效模式，转向 Agent 全自动执行合规流水线。 - **人机协同架构**：在关键风险节点保留 Human-in-the-loop 审批机制，平衡系统自主性与监管审慎性。 - **严格遵守属地法规**：特别针对中国《个人信息保护法》（PIPL）与跨境数据流动规则设立 Agent 专属控制层，规避合规风险。 ### [16:10](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=970s) Agentic 架构在跨境合规与复杂场景中的作用 在探讨企业级 AI 的治理落地时，Wayne Filin-Matthews 强调了智能体架构（Agentic Framework）在解决实际合规与数据流转挑战中的核心价值。以跨国企业涉及中国等地区的跨境数据传输（cross-border data transfer）为例，这种高度复杂且敏感的任务绝不应该依赖人工操作去逐一处理，而必须交由底层的 Agentic 架构自动化、规范化地执行与把控。 ```mermaid flowchart LR accTitle: Agentic 架构驱动的合规数据流转流程 subgraph Governance["Agentic 治理架构"] A["跨境数据传输请求"] --> B["Agentic 智能体框架"] B --> C["合规策略与数据流转控制"] end C --> D["安全跨国交付 (如中国跨境场景)"] ``` ### [16:22](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=982s) 企业级 AI 的未来下注：智能体成熟度、历史镜像与组织架构重塑 当主持人 Tricia Martinez 询问如果作为“下注者”，在未来两到三年内最看好并会加倍押注（double down）哪些能给企业级 AI 带来颠覆性变革的领域时，Wayne 明确指出了“智能体（Agency and Agents）”。他认为智能体及其周边的解决方案将在企业中大规模激增并逐步走向成熟。然而，谈及通用人工智能（AGI）的具体落地时，他抱持着冷静而务实的态度，认为两到三年内不可能真正实现。 为了说明这种技术演进的周期规律，Wayne 分享了自己在微软时期的亲身经历。他早在微软 BPOS（Business Productivity Online Suite，即 Azure 和 Microsoft 365 的早期前身）时代就开启了云原生旅程。他回忆道，15年前业内就充斥着“传统基础设施将在5年内消亡”、“混合云只是一个存活5年的过渡形态”的激进论调，但时至今日，现场依然几乎没有任何一家企业能够做到完全脱离混合架构而纯粹跑在云端。基于自己长达 38 年的技术从业经验，Wayne 直言智能体要达到真正的成熟水平至少还需要 5 年时间。如果 5 年后大家再次坐在一起复盘，必然会看到行业经历大量的经验教训与试错失败，而且在智能体的治理（governance）与控制权交接方面，届时也依然处于攻坚阶段。 除了技术本身，Wayne 强调当下正在发生的更深层变革是企业组织结构的重组。为了真正释放并赋能 AI 与智能体的生产力，各大企业正在对其底层的运营模型（operating model）进行全面重塑。Tricia 对此总结并确认了“加倍押注智能体未来”这一核心方向，随后宣布讨论结束并号召全场为 Wayne 的精彩分享鼓掌致谢。 - **智能体演进周期**：从当前的爆发式涌现到真正具备企业级成熟度与可靠治理，预计需要大约 5 年的迭代与试错周期。 - **历史技术镜像**：15 年前 BPOS 时代关于“5 年消灭基础设施”的预言与当下的 AGI 狂热形成对比，证明技术成熟与架构转型需要更长期的沉淀。 - **组织运营重构**：企业级 AI 成功的关键不仅在算法，更在于配合 Agentic 系统对企业内部组织架构与运营模式进行系统性重塑。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=4s) if anybody ever tells you this is easy... it's not — Wayne Filin-Matthews > [00:04](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=4s) every moment matters — Wayne Filin-Matthews > [06:06](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=366s) everything changes but nothing changes — Wayne Filin-Matthews > [06:06](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=366s) 阿斯利康构建 Agent 的统一平台是什么？ — 企业高管 > [11:25](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=685s) 每个人都在谈技术、技术、技术。但你必须退后一步思考：什么样的 Agent 团队认知行为要素才能最终驱动最佳结果与最高价值？ — Wayne Filin-Matthews > [13:42](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=822s) 企业合规审批的下一步流程是什么？ — 员工 > [16:22](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=982s) 传统基础设施将在5年内消亡 — 行业舆论 > [16:22](https://www.youtube.com/watch?v=1sSVXHi_aVU&t=982s) 混合云只是一个存活5年的过渡形态 — 行业舆论 ## Mentioned AstraZeneca (阿斯利康), Wayne Filin-Matthews, Tricia Martinez-Saab, Dapple, Ai4, Dell (戴尔), HSBC (汇丰银行), Accenture (埃森哲), Microsoft (微软), AWS, Forbes Technology Council (福布斯技术委员会), Gartner, Stanford University (斯坦福大学), ServiceNow, SAP Joule, Azure AI Foundry, AWS AgentCore / Bedrock / Strands, Amazon Q, MuleSoft, Microsoft Agent Builder, Model Context Protocol (MCP), BPOS (Business Productivity Online Suite), Microsoft 365, PIPL (个人信息保护法) ## Video description Tricia Martinez-Saab, CoFounder & CEO, Dapple Wayne Filin-Matthews, Chief Enterprise Architect, AstraZeneca Abstracts AstraZeneca is leading the charge in applying AI to accelerate scientific discovery, operational excellence, and global collaboration. But enabling AI at enterprise scale requires more than just models—it demands a new approach to infrastructure. In this fireside chat, Wayne Filin-Matthews (Chief Enterprise Architect, AstraZeneca) joins Tricia Martinez (Founder & CEO, Dapple) for an inside look at how AstraZeneca is designing the next generation of AI infrastructure to support complex, high-impact workloads across the organization. They’ll dive into: AstraZeneca’s vision for scaling AI across R&D and beyond Infrastructure choices that support both performance and governance How global enterprises are approaching sustainability, security, and control What it takes to build a future-proof AI stack This is a rare opportunity to hear directly from a global enterprise shaping the future of AI—from strategy to execution. [Source](https://www.youtube.com/watch?v=1sSVXHi_aVU)

[0:04] Wayne Filin-Matthews: if anybody ever tells you this is easy... it's not [0:04] Wayne Filin-Matthews: every moment matters [6:06] Wayne Filin-Matthews: everything changes but nothing changes [6:06] 企业高管: 阿斯利康构建 Agent 的统一平台是什么？ [11:25] Wayne Filin-Matthews: 每个人都在谈技术、技术、技术。但你必须退后一步思考：什么样的 Agent 团队认知行为要素才能最终驱动最佳结果与最高价值？ [13:42] 员工: 企业合规审批的下一步流程是什么？ [16:22] 行业舆论: 传统基础设施将在5年内消亡 [16:22] 行业舆论: 混合云只是一个存活5年的过渡形态

[6:06] Wayne Filin-Matthews: everything changes but nothing changes [3:00] Wayne Filin-Matthews: every moment matters [3:00] Wayne Filin-Matthews: patient first [0:04] Wayne Filin-Matthews: agent economy and agency [13:42] Wayne Filin-Matthews: in the loop [13:42] Wayne Filin-Matthews: on the loop [16:36] Wayne Filin-Matthews: I can remember listening to a conversation... they were saying, 'Oh, you know, infrastructure is dead in five years; hybrid basically cloud will last five years.' Anybody basically purely cloud in the room that has no hybrid solution? Yeah, this was 15 years ago they were making this promise.

[0:04] All right. Well, thank you so much for [0:05] joining us this afternoon. I know [0:06] there's so many other events going on. [0:09] Um, but I will have to say we have one [0:12] of the most interesting ones. I've known [0:13] Wayne for a number of years, and I'm [0:16] more and more excited about the work [0:18] that Astroenica is doing in advancing [0:21] the AI field for enterprise, but also on [0:24] the infrastructure sovereignty and [0:26] intelligence side. So just some quick [0:29] background on who Wayne is. He is a [0:32] distinguished chief architect at [0:34] Astroenica where he leads global [0:36] strategy and architecture for one of the [0:38] world's largest life sciences companies. [0:42] He's bringing a track record of [0:44] delivering enterprise transformation at [0:46] scale. Having previously been the chief [0:48] enterprise architect at Dell, the global [0:50] head of digital architecture at HSBC, [0:53] technology and consulting senior [0:55] director at Accenture, and a senior [0:58] transformation leader at Microsoft. He's [1:00] a Forbes technology council and Gartner [1:02] research board member, a BCS chartered [1:05] fellow, and a winner of ICMG's global [1:07] chief enterprise architect of the year. [1:10] It's safe to say that he's an expert in [1:12] the infrastructure and architecture [1:14] space. [1:17] So, um I know we only have 20 minutes [1:19] and I do want to leave time for some [1:21] questions, but we want to cover today [1:23] everything from where enterprise AI is [1:26] really headed, Agentic AI platforms, [1:28] orchestration at scale, sovereignty, [1:31] multicloud infrastructure, all of that [1:34] good stuff. So, I just want to hop in [1:36] really quickly. Um let's start by [1:38] looking at where enterprise AI is headed [1:41] today. So Wayne, how would you describe [1:45] um the maturity of AI adoption at [1:47] Astroenica today? And like how do you [1:49] think that relates overall to the larger [1:51] Fortune 1000 landscape? [1:54] >> Yeah, everyone can hear me. Okay. Uh [1:55] good to see everybody. Hope you're [1:56] enjoying first day of uh AI4. Um good to [2:00] meet you all. Uh if anybody ever tells [2:01] you this is easy in all of the [2:03] presentations today, it's not as you [2:05] probably know. uh especially when you're [2:06] working in a large enterprise like [2:08] Astroenica uh across 126 markets. It's a [2:11] great question. Everyone in the room I [2:13] think knows like AI is not necessarily [2:16] new. Yeah, it's we've been doing this [2:18] for 500 years in terms of basically the [2:20] journey that we've been on to this [2:21] point. I think I woke woke up on uh New [2:25] Year's Day uh and we were doing a trip [2:27] to the west coast and we were sort of [2:29] meeting with AWS and Microsoft and uh I [2:32] I spent 11 years at Microsoft as the AI [2:35] uh uh engineering interface into the [2:37] engineering teams uh in MCS. Um and what [2:41] what we noticed in January is the world [2:43] had changed significantly that everyone [2:45] was talking about agents. Yeah, it was [2:47] an agent economy and agency that sort of [2:50] has now trumped the AI conversation a [2:53] little bit in terms of um what the [2:56] promise of of agency and agent uh gives [2:58] to us. For Astroenica, [3:00] um we've been doing machine learning and [3:02] modeling and basically the normal AI [3:04] stuff for a number of years. We have a [3:07] really bold ambition over the next 5 [3:09] years. um as one of the largest [3:11] pharmaceutical uh um sort of [3:13] organizations globally. We are the [3:15] number one pharmaceutical in China. Um [3:17] and uh you probably may or may not have [3:20] seen announcements that we're investing [3:21] about $50 billion in a US manufacturing [3:24] site uh in Virginia as well. So a big [3:27] shift, but we're bringing 20 new [3:28] medicines to market um which is a lot [3:31] for a pharmaceutical. Normally it takes [3:33] 10 years to bring one medicine. We're [3:35] bringing 20 in the next 5 years. We'll [3:38] also double the size of the [3:39] organization. So we'll double it from 50 [3:41] billion revenue to 100 billion revenue [3:44] in the next four years. Uh so it's no [3:46] longer 5 years. And with all of that [3:48] comes obviously a lot of complexity in [3:50] terms of that growth. Agents are really [3:53] the only way we can do it without [3:55] growing the IT function. and we have [3:57] about 10,000 people in IT today uh to [4:00] really basically support what we're [4:02] trying trying to do from a science [4:03] perspective in R&D and operations and [4:06] enabling units in our rare disease [4:07] business if that makes sense. [4:09] >> I didn't know those stats. That's [4:10] amazing. So, is AI driving all of that [4:13] growth largely? [4:14] >> Um I I think it's AI is not driving the [4:17] growth. We sort of um we're we're an [4:19] organization that sort of our our um [4:22] sort of values basically sit on two [4:24] elements, patient first and [4:26] sustainability. Uh the pharmaceutical [4:28] industry is traditionally not well known [4:30] for its uh carbon footprint. Uh probably [4:32] even worse than if you don't know than [4:34] the the motor industry. Um um so that [4:37] there's two elements. um every moment [4:39] matters which means the longer it takes [4:42] us to bring a lifesaving drug to market [4:45] a new modality to market uh especially [4:48] in areas that we operate in you know in [4:50] Africa in in Asia etc. Um people [4:54] actually die um and not to to paint a [4:57] grim picture and you know everyone to to [4:59] feel a little bit down but speed is [5:02] ultimately our driver. the quicker we [5:04] can bring new modalities to market um [5:06] the the more patients that that we can [5:08] help um and more uh more um sort of uh [5:12] uh illnesses that we can cure. At the [5:14] same time, we're also trying to bring [5:16] down that footprint from a [5:17] sustainability perspective to to get to [5:19] net zero by 2030 as well. Um and AI is [5:23] basically the driving force. We're not [5:25] going to do that by adding another [5:26] 10,000 people in it and doubling the [5:28] size of the pe uh the people in the [5:29] business. Um but really it's about [5:32] driving those economies of scale through [5:34] agency and AI. [5:35] >> So I think mo most of the audience knows [5:37] that pharma is one of the industries [5:39] that will be largely impacted by AI. So, [5:42] um, all of this incredible growth, [5:45] life-changing, uh, solutions you're [5:47] implementing, what are the biggest [5:49] challenges from an enterprise [5:50] perspective and and implementing AI [5:52] because obviously you're using AI for [5:54] drug discovery, but there are so many [5:56] other use cases and from a a corporate [5:58] perspective like bringing technology [6:01] from the bottom from the top down is [6:03] just so difficult for you on the [6:05] architecture side. [6:06] >> Yeah. And again, I've been in the IT [6:07] industry for 38 years. uh you know I [6:10] started off at IBM and sort of was at HP [6:12] and I've been around the block so I've [6:14] been through the journeys that um when [6:16] you've been in the IT industry that long [6:17] the interesting thing on this journey [6:19] specifically and and my favorite tagline [6:21] is everything changes but nothing [6:23] changes the technology is still doing [6:25] what the technology is supposed to do [6:27] the the difference on this is basically [6:29] the speed of change and what I mean by [6:31] that in terms of in January I was being [6:34] asked as the chief architect so I you [6:36] know I'm responsible for strategic [6:38] technology strategy. When you have an [6:40] organization that's structured the way [6:42] Astroenica is that the the organization [6:44] has been deliberately structured to to [6:46] let people experiment, you know, you [6:48] can't tell a scientist not to experiment [6:50] and test out a thesis. You know, sort of [6:52] goes against the grain. Um, you also [6:54] basically the way the organization [6:56] structured is you have a lot of autonomy [6:58] in the business technology groups to to [7:00] go off and and experiment and do their [7:02] own thing. So if you work in if you've [7:04] worked in banking before, you know, [7:06] there's a lot of control around [7:07] regulation and compliancy. We're still [7:09] heavily regulated, but there's a more [7:11] drive to basically how you enable [7:13] experimentation and innovation early on. [7:16] And the reason I'm mentioning this um is [7:19] the questions I was being asked in [7:21] January were questions like, "Hey Wayne, [7:24] what is our agent uh our platform for [7:26] building agents?" That was the question [7:28] I was being asked by by by senior uh IT [7:32] individuals. And I said that's the wrong [7:34] question to ask because every every [7:37] platform will be agent enabled. Yeah. [7:39] You've got Service Now and now, you've [7:41] got SAP jewel, you've got AI foundry, [7:44] you've got AWS and strands and uh agent [7:47] core etc etc etc. The list goes on. So [7:50] the way that we approached the problem [7:51] and I'll try to keep this short and we [7:53] only got 20 minutes but um we I changed [7:56] the sort of conversation a little bit [7:58] and I said first of all the com and this [8:00] was before some announcements that our [8:02] uh key providers like AWS and Microsoft [8:04] were announcing. We said we want to [8:06] change the conversation and the [8:08] conversation is look you want to [8:09] democratize AI and agent building. How [8:12] do you give AI to the hands of the [8:14] individuals that that basically will get [8:17] the most value out of basically the [8:19] enablement of AI? To do that, you've [8:21] obviously got to identify platforms, but [8:23] you've got to put some sort of, you [8:25] know, milestone guard rails in place [8:27] that you don't enable. For example, you [8:29] probably heard and you will hear over [8:30] the next couple of days all of these [8:32] solutions to enable agent to agent. At [8:34] the moment, I've sort of said no to [8:37] agent to agent. Even if you basically [8:39] support MCP and support 80A, I've said [8:42] no to agent to agent at scale. The [8:44] reason I've said that is we want to [8:46] basically create the democratization [8:48] layer. So what I've said is as a [8:49] principle, enable agents within core [8:52] platforms. You want to enable agent [8:54] building in Microsoft and basically use [8:56] the Microsoft ecosystem for agents, go [8:58] ahead and do it. You want to basically [9:00] use Sabjule for basically doing finance [9:02] uh optimization and use productivity [9:04] gains with Sabjul, go ahead and do it. [9:06] you want service now to reroute tickets [9:08] and automate basically autonomous [9:10] decision-m with service now go ahead and [9:12] do it the reason I've said this is [9:15] because what was lacking was the [9:16] composability layer or the or the [9:18] semantic kernel which I'm not allowed to [9:20] use anymore if anybody from Microsoft uh [9:22] I'm not allowed to use that because [9:23] Microsoft have trademarked semantic [9:25] kernel so I'm not allowed to use that [9:27] term um but the composability layer was [9:30] where the trickiness has come out so for [9:32] example you know where do you put your [9:34] agent catalog how do to ensure that [9:37] you've got agent identification. It's [9:38] great that even the previous session [9:40] they were talking about delegated access [9:42] for agents etc. But the composability [9:45] layer is the weakest layer at the moment [9:47] especially for an enterprise that needs [9:49] to manage uh the complexity of agent to [9:52] agent communication. What I mean by that [9:54] is security you know identity delegation [9:58] you know traceability uh uh regulatory [10:01] and cost. This is the other challenge I [10:04] think we're all facing at the moment is [10:06] the ability to forecast accurately [10:09] agent-to-ag communication within memory [10:11] reasoning because you're enabling agents [10:14] to potentially communicate in in [10:17] multiple sources. So what we've done in [10:19] terms of our testing and modeling out is [10:21] you know you may have an agent that [10:23] starts off in agent builder with [10:25] Microsoft. that goes through AI foundry, [10:27] talks to Bing, then decides I'm going to [10:30] basically go out to an LLM, and then all [10:33] of a sudden you've got these issues that [10:35] basically say, well, half of it is [10:37] conversation cost, half of it is [10:38] tokenization cost. If I use an API on [10:41] Muleoft, it's now a theme cost. Yeah. [10:44] And the ability to forecast accuracy [10:47] cost of agent to agent is a is a big [10:49] problem. um where the the the vendors [10:52] are obviously the big sort of you know [10:54] cloud providers are trying to tackle [10:56] that problem is how do you forecast [10:58] agentto agent enablement at speed um to [11:02] to to make sure that you're doing this [11:04] safely and you don't blow your IT [11:05] budgets um and then there's a whole [11:07] bunch of complexity around things like [11:08] in like how long should you keep [11:10] long-term memory for in an agent there [11:12] are compliancy and regulatory reasons [11:14] the same as we do for data life cycle [11:16] management for regulatory concerns [11:18] around uh agent long-term and short-term [11:21] memory. So, there's a whole bunch of [11:23] complexity that sits within this. But [11:25] the the reason that I've sort of [11:26] suggested we do it this way is we start [11:29] building agents, we start running at [11:32] speed to get all of those learnings. And [11:34] then the other thing that's missing out [11:35] of the conversation in large enterprises [11:38] is people are not taking the cognitive [11:41] behavior of of the people side into the [11:44] agent conversation. And what I mean by [11:47] that is I still see, you know, [11:48] individuals, teams that have been given [11:51] the technology to enable agents, but [11:53] they're still doing it in a very linear [11:55] way. They're sort of creating it as an [11:57] extension to RPA, an agent talking to [11:59] something to make autonomous decision [12:01] making. That's not the conversation that [12:03] we should be having. The conversation we [12:05] should be having is what types of teams [12:07] of agents need to be put into place with [12:10] new structure and what is the behavioral [12:13] science of agents that are needed to [12:15] deliver value. So the work that we do [12:17] with uh organizations like Stanford I'll [12:19] give you quick one quick example is we [12:22] don't want to basically grow the number [12:24] of scientists that we have in research [12:25] and development. We want to, you know, [12:27] we'll grow some, but we want to [12:29] basically have teams of agent scientists [12:32] that can do the work with access to that [12:35] data much more quickly to basically get [12:37] results and predict new molecules, [12:39] predict new accuracy to to to do the [12:41] deep science. To do that, you have to [12:44] understand how a science team works. A [12:46] science team works when the agents come [12:48] together. You have to have a [12:49] procrastinator agent because scientists [12:52] basically work better when there's a [12:53] procrastinator in their team. Yeah. So [12:56] it's this this cognitive behavioral [12:58] element that basically is being slightly [13:01] overlooked from a technology [13:02] perspective. Everyone's talking [13:03] technology, technology, technology. But [13:06] you also have to take a step back and [13:08] say what is the cognitive behavioral [13:10] element of teams of agents that will [13:13] ultimately drive the best results and [13:14] the highest value. And I don't see [13:16] enough of that in the industry being [13:18] conversed and and conversations around [13:20] it. Accenture is doing a good job in [13:22] terms of how they've enabled their [13:24] 700,000 agents because they took that [13:27] step back. There are pockets of sort of [13:29] good stuff happening. Um but but I would [13:32] recommend that you know if you if you [13:34] work in an enterprise and you're on the [13:36] agent journey specifically think about [13:38] the cognitive behavioral elements of [13:40] those individuals as well in the teams [13:41] of agents. [13:42] >> Okay. So technically [13:44] we are very far away from a thriving [13:47] enter agentic AI future for enterprise [13:49] right from the technical perspective [13:51] right there's just so many missing [13:52] layers so much to build so much to be [13:54] thought out [13:55] >> but taking a step back you're in one of [13:57] the most regulated industries compliance [13:59] governance everything around data [14:02] sovereignty how are you thinking about [14:04] this how does this impact your [14:06] infrastructure buildout um everything [14:09] and how you operate globally [14:10] >> and this this comes back to the comments [14:12] So, so I I'll give an example. We've got [14:14] a we've got a team and it's not not to [14:16] criticize the team. The team's taking [14:19] Amazon Q and they're but building [14:21] basically [14:23] a ragified architecture to ask questions [14:26] of what our governance processes. So, so [14:29] let me give you the example. Yeah. [14:30] They're using the latest and greatest [14:32] AI. They've got LLMs basically all the [14:34] good stuff. Yeah. And they're asking [14:36] questions of this chatbot type thing to [14:39] say, what is the next step in our [14:40] governance process? basically to get [14:42] something approved and it's just the [14:45] example. Now some of you might be [14:46] sitting and go that sounds okay Wayne [14:48] you've created you know an AI navigator [14:50] to basically help the organization. Same [14:52] with a a research assistant spent a lot [14:54] of money basically on an AI research [14:56] assistant to help reproducibility of [14:58] science. The challenge with that though [15:00] is you shouldn't be doing that anymore. [15:02] You should be thinking well why do I [15:04] need an individual to ask questions of a [15:06] chatbar to understand the governance [15:08] process? Why don't I take the [15:10] opportunity to say what is the team of [15:12] governance and risk and security look [15:14] like in the new world of of Agentic and [15:17] how do I reimagine not humans doing this [15:19] and asking questions but how do I [15:22] restructure my governance risk data uh [15:26] sovereignty data quality elements that [15:28] agents are doing it uh on on our behalf [15:32] with the elements of human in the loop [15:33] or human on the loop basically at the [15:35] right decision points to make sure we're [15:37] compliant. So I'll give an example. You [15:39] know, we're we're um number one in [15:41] China. Um if I enabled agents today um [15:45] and without these controls and [15:47] conversations in and the team dynamics, [15:49] I potentially could break our crossber [15:51] data transfer and pipple rules that [15:53] China introduced sort of 18 months ago. [15:56] So all of a sudden basically all the [15:57] work that we've done previously could be [16:00] eroded in terms of that that that [16:01] compliance and regulatory governance [16:03] that's required. Whereas actually what [16:05] we're trying to do now is basically say [16:07] this is the teaming structure to handle [16:10] things like crossber data transfer with [16:12] China and it's not human beings that [16:14] should be doing that. It's basically [16:15] your agentic framework that that enables [16:17] you to do it. [16:19] >> Does that make sense? [16:22] >> Okay. So if you if you were a betting [16:23] man thinking about everything that [16:25] you're dealing with in enterprise AI, [16:28] what would you double down in? What are [16:29] the areas that you think in the next two [16:32] to three years are going to be [16:33] gamechanging for enterprise AI? I [16:36] >> I think agency and and agents will start [16:40] to to to proliferate more and more and [16:42] the solutions that fill that space um [16:44] will be will become more and more and [16:46] more mature. We're still on this journey [16:48] to AGI specifically. I I don't see that [16:51] becoming truly real two or three. I [16:54] honestly I was I remember being at [16:56] Microsoft and I can remember uh uh [16:58] listening to a conversation. So I [17:00] started my Microsoft cloud journey with [17:01] BPOS. So if anyone remembers BOSS was [17:04] the predecessor for Azure um and there [17:06] were these conversations that that there [17:08] was they were saying oh you know [17:09] infrastructure is dead in 5 years hybrid [17:11] basically cloud will last 5 years. [17:14] Anybody basically uh just purely cloud [17:16] in the room that has no hybrid solution. [17:19] Yeah this was 15 years ago they were [17:21] making this promise. when you've been [17:23] doing this as long as I have now for 38 [17:25] years, I don't think we're going to get [17:26] to a level of maturity in agency and [17:28] agents until probably five years. That's [17:31] when that if we sit back here together [17:33] in 5 years, we'll probably have a load [17:36] of lessons learned, a load of failures. [17:39] Um, and basically we still won't be [17:41] where we need to be in 5 years in terms [17:43] of giving that governance and control [17:44] over. Um, so I would say that's going to [17:48] happen. Um, and then it's going to be a [17:50] restructuring of basically just [17:52] organizations. [17:54] That that's what I'm seeing at the [17:55] moment is restructuring of the operating [17:57] model in organizations that that [17:59] ultimately leverage and enable AI and [18:01] agents uh specifically. [18:03] >> Okay. So, doubling down on obviously the [18:04] agentic feature. Yeah. [18:06] >> Okay. I think our our time's up. Let's [18:08] give Wayne a round of applause. Thank [18:10] you so much, Wayne, for your input.