Wells Fargo VP Links Agentic AI Scale To In-House Governance

富国银行副总裁将Agentic AI规模化与内部治理挂钩 代理式人工智能的企业级部署需要在对外推广前，利用内部小语言模型和人类监督控制来建立组织信任。 本场会议探讨了从传统自动化和生成式AI向具备推理、规划和执行复杂企业工作流能力的自主Agentic AI的转变。Srinivasa Manapragada概述了一个整合小语言模型（SLM）、API、企业数据和模型上下文协议（MCP）工具的多层参考架构。他强调，金融企业的成功部署依赖于解决具体业务问题，同时嵌入严格的治理、安全以及人机协同（Human-in-the-Loop）控制。最终，通过负责任的AI原则建立内部信任，将赋能更快速的决策和运营可扩展性。 Agentic AI代表了继传统自动化、生成式AI和AI副驾驶（Co-pilot）之后的下一个范式转变，实现了从回答问题到自主完成复杂业务目标的跨越。 (0:04) 企业AI战略必须从解决具体的业务问题入手，而不是仅仅关注技术本身。 (0:04) 系统架构必须从第一天起就融入治理、安全、合规、可扩展性以及人类监督机制。 (0:04) 金融机构可以利用针对特定场景定制的内部小语言模型（SLM），在预算限制内开展工作，而不是依赖外部专有模型。 (5:18) 企业级部署依托于安全调用内部企业工具、API和数据库，而不是完全依赖第三方商业AI服务。 (5:18)

Srinivasa Manapragada outlines an enterprise framework for agentic AI at financial institutions like Wells Fargo, asserting that sustainable scaling requires moving beyond reactive chatbots to goal-driven autonomous systems anchored by rigorous internal governance. The Evolution Toward Autonomous Workflows (0:04) According to Manapragada, enterprise artificial intelligence is moving through a structured evolution: progressing from traditional automation to generative AI, AI co-pilots, and now agentic AI. Rather than functioning as simple prompt-and-response interfaces, agentic systems are designed to reason, plan, retain contextual memory, and execute multi-step workflows to achieve defined business outcomes. To make this transition viable in an enterprise context, organizations must prioritize concrete business problems over technology deployment for its own sake. Building on modular architectures and standard application programming interfaces (APIs) provides the foundation for these agents to operate as the core operating model for next-generation enterprises. Tailoring Models to Financial Enterprise Constraints (5:18) Deploying agentic systems in highly regulated financial environments requires navigating distinct cost and infrastructural constraints. Rather than relying exclusively on costly external proprietary models, Manapragada argues that enterprises should prioritize internal Small Language Models (SLMs) optimized for specific business contexts and funding limits. Operational success depends on connecting these internal models safely to enterprise systems. These agentic capabilities are exposed to internal and external users across three primary access channels: web applications, mobile interfaces, and conversational chatbots. Multi-Layered Architecture and Tool Integration (7:31) Manapragada presents a multi-tiered reference architecture designed to coordinate agentic actions securely across an organization. Above the foundation models and SLMs sits an orchestration layer that manages multi-agent collaboration, enabling distinct agents to communicate, delegate tasks, and execute multi-step processes. Beneath the models, an enterprise integration layer connects agents to existing databases, knowledge repositories, internal business applications, APIs, and Model Context Protocol (MCP) tools. This setup allows autonomous systems to safely invoke enterprise utilities and retrieve live data strictly within organizational policy boundaries. Centralized Governance and Human Oversight (10:06) Surrounding the entire architecture is an end-to-end governance layer that Manapragada identifies as a mandatory prerequisite for deployment. He stresses that an enterprise must 'win our house' internally by establishing security, identity management, audit compliance, observability, and leadership reporting before extending agentic capabilities externally. A central component of this governance framework is mandatory Human-in-the-Loop oversight. Maintaining human intervention points across enterprise workflows ensures regulatory compliance, provides system explainability, and builds internal trust, ultimately enabling accelerated decision-making and productivity without sacrificing data security. The presentation concludes that while agentic AI offers significant productivity gains and process automation, its enterprise viability rests on treating governance and human oversight as architectural fundamentals rather than retrospective controls.

在Ai4大会上，富国银行（Wells Fargo）的Srinivasa Manapragada指出，企业级人工智能正在从被动问答的副驾驶（Co-pilots）升级为具备推理、规划与自主执行能力的智能体AI（Agentic AI）；而在金融机构中实现规模化应用的核心，在于将严格的端到端治理、内部小语言模型（SLM）与人机协同机制深度融入系统架构。 从生成式对话向自主智能体范式演进 (0:04) Manapragada将企业人工智能的发展划分为四个连续阶段：传统自动化、生成式AI、AI副驾驶，以及当下的智能体AI。在他看来，企业级AI的应用范式不能仅停留在简单的“提示-响应”式聊天交互，而必须演变为能够理解商业目标、开展逻辑推理、自主规划并执行复杂工作流的系统。这种转变标志着技术角色从单纯回答问题升级为主动作业以达成既定商业目标。 他进一步强调，推进企业智能体战略必须坚持“业务问题先行”的原则，即技术部署必须始于具体的实际业务挑战，而非对前沿技术的盲目跟风。企业应依托模块化架构和规范的API设计，将智能体AI视作下一代智能企业的核心运营模型，以此确保技术迭代能够切实转化为组织生产力的提升。 金融约束下的分层架构与小模型适配 (5:18) 针对金融机构在落地过程中面临的现实瓶颈，Manapragada指出，昂贵且黑盒化的外部商业大模型并不完全契合金融场景的严苛要求。受限于资金预算与数据合规等现实约束，企业应当转向构建专有内部小语言模型（SLM）。这些小模型针对特定业务语境进行优化，能在受控前提下安全调用企业内部工具与业务流。 在落地形态上，系统通常通过Web应用、移动端界面和对话机器人三个入口面向终端用户。Manapragada提出了一套清晰的分层参考架构：从上至下依次为用户入口、AI体验层、负责跨智能体通信与任务委派的编排层、基础模型与内部SLM层，以及整合了API、数据知识库、内部应用和模型上下文协议（MCP）工具的企业集成层，从而实现不同模型与既有数字资产的协同作业。 以治理为核心构筑内部信任壁垒 (7:31) 围绕金融系统对安全与合规的极端敏感性，Manapragada提出“先安内，后拓外”（winning our house）的原则，即在向外部客户开放能力之前，必须首先建立起严密的内部治理体系以赢得组织内部的信任。治理层并非边缘外挂组件，而是作为基础底座贯穿于用户体验、智能体编排、模型计算和系统集成的全流程，涵盖身份认证、安全策略、审计合规审查、全链路可观测性以及管理层分析报告。 在治理与合规执行中，“人机协同”（Human-in-the-Loop）被界定为不可妥协的核心支柱。Manapragada主张，负责任AI（Responsible AI）必须在所有企业级智能体应用中保留人工监督与干预渠道。这一机制不仅是满足金融监管合规和保障决策可解释性的必要手段，也是消除算法不确定性、确保系统安全运行的基本前提。 执行能力落地与企业战略价值释放 (10:06) 在具体操作层面，智能体系统的高效运转依赖于四项核心技术能力：多步逻辑推理、情境化任务规划、长效记忆保留，以及在遵循企业既定策略前提下的自动化API调用与内部知识库检索。这些能力使系统能够跨越孤立系统调取资源，将抽象的业务指令分解为可执行的系统操作。 Manapragada总结认为，在严密的治理防护与数据安全框架下，成功部署智能体AI将为企业带来切实的商业回报。这包括加快管理决策进程、实现业务流程的智能化自动化、直接提高员工整体产出，并最终支持企业在风险可控的前提下，安全地将服务能力拓展至外部合作伙伴网络。 Manapragada勾勒了金融机构在安全合规边界内规模化落地智能体AI的战略路径，确立了治理先行与人机协同的基调；但该演讲主要集中于顶层架构设计与原则倡导，对于复杂业务场景中多智能体协同冲突裁决、长周期记忆管理以及极端错误场景下人工接管的时效性等深层工程挑战，尚未提供具体的应对细节。

## Background 该视频由 YouTube 频道 Ai4 于 2026 年 8 月发布，主讲人为富国银行（Wells Fargo）副总裁 Srinivasa Manapragada。视频简介指出，本次演讲探讨了企业如何结合大模型、模型上下文协议（MCP）、API 与智能编排，在兼顾治理、安全与合规的前提下，构建并规模化部署生产级智能体 AI（Agentic AI）。 ## TL;DR 演讲梳理了企业 AI 从传统自动化、生成式内容到自主协同型智能体（Agentic AI）的跃迁，指出其核心差异在于从“回答问题”转变为“达成业务目标”。主讲人提出了一套涵盖接入交互、体验定义、智能编排、基础/小语言模型（SLM）与企业集成的分层技术架构。同时，他强调治理（Governance）与人机协同（HITL）绝非事后补丁，而是保障企业信任、全流程审计并实现规模化落地的核心中枢。 ## Atmosphere & style 整场演讲兼具金融架构师的严谨务实与前沿布道者的从容。开场虽遭遇了播放故障的小插曲，讲者通过幽默自嘲迅速化解尴尬，营造出轻松且专注的现场氛围。叙述逻辑紧密、层层递进，避开了空洞的概念炒作，以极具工业级质感的系统图解和落地法则，展现了强监管行业在推进前沿 AI 变革时的稳健风范。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=1qH3irPXI5Q&t=4s) 企业 AI 技术经历了从传统自动化、生成式 AI、AI 助手/副驾驶到 Agentic AI（智能体 AI）的四阶段演进，系统正从被动响应 Prompt 转向具备主动规划与自主执行能力。 - [01:46](https://www.youtube.com/watch?v=1qH3irPXI5Q&t=106s) 智能体 AI 与传统聊天机器人的本质区别在于：“传统机器人负责回答问题，而 Agentic AI 负责达成业务目标”。 - [01:46](https://www.youtube.com/watch?v=1qH3irPXI5Q&t=106s) Agentic AI 依托三大核心支柱能力：从全局语义洞察真实意图的理解力、自主拆解复杂工作流的推理与规划力、以及联动企业系统闭环执行的行动力。 - [02:47](https://www.youtube.com/watch?v=1qH3irPXI5Q&t=167s) 在企业落地中不必盲目堆叠昂贵的基础大模型，应根据成本、安全与性能权衡，构建结合基础模型与自建小语言模型（SLM）的协同分层架构。 - [02:47](https://www.youtube.com/watch?v=1qH3irPXI5Q&t=167s) 企业级智能体系统由交互层、体验层、编排层、模型层和集成层构成，其中编排层作为中枢大脑负责多智能体调度（Multi-Agent Collaboration）与记忆同步。 - [02:47](https://www.youtube.com/watch?v=1qH3irPXI5Q&t=167s) 集成层借助自研 API 网关与模型上下文协议（MCP），为智能体提供安全、标准化的知识库检索与内部微服务读写接口。 - [08:43](https://www.youtube.com/watch?v=1qH3irPXI5Q&t=523s) 治理机制（Governance）绝非部署后追加的补丁，而是贯穿整个系统架构、确保合规与安全不可分割的核心底座。 - [08:43](https://www.youtube.com/watch?v=1qH3irPXI5Q&t=523s) 人机协同（HITL）机制处于治理体系的中心，在关键决策与敏感操作环节保留人类审批把关，是智能体实现规模化落地的必要前提。 - [08:43](https://www.youtube.com/watch?v=1qH3irPXI5Q&t=523s) 智能体架构必须具备端到端的可观测性与可审计性，全链路记录推理链、记忆调用及 API 交互以满足金融强监管要求。 - [12:23](https://www.youtube.com/watch?v=1qH3irPXI5Q&t=743s) 企业推进 Agentic AI 需坚持业务问题导向与 API 优先的模块化设计，将其作为下一代智能企业的核心运营模式而非单纯的技术实验。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=1qH3irPXI5Q&t=4s) 开场故障与演进历程：从传统自动化走向 Agentic AI 演讲伊始，现场出现了一点音视频播放的技术小插曲（播放了预先录制的虚拟讲解视频片段），演讲者 Srinivasa Manapragada（富国银行副总裁兼 AI 传播者）幽默地打趣说，既然遇到技术故障，今天不妨让现场变得更“具智能体特性（Agentic）”，甚至开玩笑说“大家别介意我直接溜出这个会场”。随后，他正式切入本次大会的主题——“如何在规模化场景下构建受治理的企业级智能体 AI（Building Governed Agentic AI at Scale）”。他特别澄清，虽然自己来自富国银行（Wells Fargo），但本次分享不仅限于富国银行自身，而是面向广义金融及大型企业的普适性架构探索。 回顾企业智能的发展历程，AI 经历了几个极为清晰的技术跃迁：最初阶段是基于固定规则和脚本的“传统自动化（Traditional Automation）”；随后迎来了“生成式 AI（Generative AI）”，机器开始能够生成内容和回答复杂问题；紧接着是“AI 助手与副驾驶（AI Assistants / Copilots）”，这一阶段极大地辅助了员工个人的日常生产力。如今，企业正全面跨入崭新的第四阶段——“智能体 AI（Agentic AI）”时代。系统不再仅仅是被动响应人类的 Prompt，而是具备主动推理复杂任务、拆解目标并自主协同执行的能力。 对于大型企业而言，部署 Agentic AI 的成功标准远不止于模型能力本身，必须在系统设计之初就将治理（Governance）、安全性（Security）、合规性（Compliance）、可扩展性（Scalability）以及人类监督机制（Human Oversight）深度植入底层架构中。 ```mermaid flowchart LR accTitle: AI技术演进四阶段 A["传统自动化<br/>(Traditional Automation)"] --> B["生成式 AI<br/>(Generative AI)"] B --> C["AI 助手与副驾驶<br/>(Copilots / Assistants)"] C --> D["智能体 AI 时代<br/>(Agentic AI)"] ``` --- ### [01:46](https://www.youtube.com/watch?v=1qH3irPXI5Q&t=106s) Agentic AI 的三大核心能力与本质区别 演讲中深入剖析了 Agentic AI 与当下企业普遍使用的传统问答机器人（Chatbots）及简单助手之间的本质断代差异。Srinivasa 将智能体 AI 的关键能力归纳为三大支柱： 1. **深度意图理解（Understanding）：** 不再只是机械地匹配 Prompt 关键词生成文本，而是能够从全局语义中深入理解用户背后真正想要达成的最终业务目标。 2. **推理与规划（Reasoning & Planning）：** 智能体能够自主将复杂的业务工作流拆解为若干逻辑步骤，判断当前环节需要调取哪些上下文与信息，智能挑选最合适的工具，并协调多项活动以达成目标。 3. **行动与工具调用（Action & Workflow Execution）：** 智能体不会在给出一段文字回答后就戛然而止，它能直接与企业内部的业务系统、API、数据库、知识库以及各类微服务进行安全交互，端到端推动业务流程的闭环执行。 正如演讲者所强调的核心差异：“传统聊天机器人负责回答问题，而 Agentic AI 负责达成业务目标（Traditional chatbots answer questions, agentic AI helps accomplish goals）。”这一根本性质的转变，直接将企业 AI 的命题从“如何生成好答案”升级为“如何在大型企业内部安全、受控地部署自主执行系统”。 --- ### [02:47](https://www.youtube.com/watch?v=1qH3irPXI5Q&t=167s) 企业级智能体分层架构设计与技术栈 针对真实企业环境下的落地挑战，Srinivasa 结合金融行业的实际场景指出，并非所有机构都拥有无上限的预算去盲目堆叠市场上昂贵的大型专有模型。因此，企业需要建立自主可控的内部模型体系，特别是由小语言模型（SLMs / Small Language Models）与基础大模型协同驱动的分层架构： ```mermaid flowchart TD accTitle: 企业级受治理智能体系统分层架构 subgraph Layer1 ["1. 接入交互层 (User Entry Points)"] direction LR U1["Web 应用"] ~~~ U2["移动端 App"] ~~~ U3["企业聊天终端"] end subgraph Layer2 ["2. AI 体验层 (AI Experience Layer)"] EXP["端到端业务目标定义与情境交互"] end subgraph Layer3 ["3. 编排协同层 (AI Orchestration Layer)"] ORCH["多智能体协作 (Multi-Agent) / 工具分发 / 规划与记忆"] end subgraph Layer4 ["4. 模型能力层 (Foundation & SLMs)"] MODELS["基础模型 / 企业自建小语言模型 (SLM)"] end subgraph Layer5 ["5. 企业集成层 (Enterprise Integration)"] direction LR INT1["自建 API 服务"] ~~~ INT2["知识库 / 数据库"] ~~~ INT3["MCP 协议工具集"] end subgraph Gov ["横向治理与防护中枢 (Governance Backbone)"] GOV_ALL["安全认证 / 策略合规 / 审计监控 / 人机协同 (HITL)"] end Layer1 --> Layer2 Layer2 --> Layer3 Layer3 --> Layer4 Layer3 --> Layer5 Gov -.- Layer2 Gov -.- Layer3 Gov -.- Layer4 Gov -.- Layer5 ``` - **多触点交互层（User Entry Points）：** 覆盖 Web 端、移动 App 以及嵌入式聊天组件，作为全场景的入口。 - **AI 体验层（AI Experience Layer）：** 定义具体的业务场景最终诉求，承接用户会话并将意图转化为结构化任务。 - **AI 编排层（AI Orchestration Layer）：** 整个架构的中枢大脑，负责内部与外部多个智能体之间的对话调度（Multi-Agent Collaboration）、动态选择工具、协调工作流并在多智能体间同步记忆与规划。 - **模型基础设施层（Foundation Models & SLMs）：** 根据成本、合规与性能权衡，综合调用基础大模型或内部轻量化的小语言模型（SLMs）来进行逻辑推理与决策。 - **企业集成层（Enterprise Integration）：** 包含自研 API 网关、企业知识库（Knowledge Bases）、内部业务系统，以及支持模型上下文协议（MCP, Model Context Protocol）的企业工具集，确保智能体能够安全读写数据。 --- ### [08:43](https://www.youtube.com/watch?v=1qH3irPXI5Q&t=523s) 治理机制（Governance）与负责任 AI：构建企业信任基石 在整个架构全景图中，贯穿所有层级的核心是**治理（Governance）**。Srinivasa 旗帜鲜明地指出：“治理绝不是部署完成之后才追加的补丁，治理本身就是系统架构不可分割的一部分。”对于金融等强监管行业而言，必须坚持“先赢得内部信任，才能赢得外部世界（First if we try to win our house, then we can try to win the external world）”。 治理与负责任 AI（Responsible AI）的落地框架主要涵盖以下核心维度： - **身份认证与权限策略（Identity & Policy Enforcement）：** 智能体访问任何内部知识库或调用 API 时，必须严格受限于企业既有的安全访问控制与合规策略。 - **人机协同（Human-in-the-Loop, HITL）：** Srinivasa 将人类干预机制置于治理体系的正中心。他强调“没有人类介入，规模化落地将无从谈起”。在关键决策和敏感操作节点，必须保留人类审批与把关环节，这是建立机构信任的决定性力量。 - **全面可观测性与审计（Observability & Auditability）：** 全生命周期跟踪智能体的推理链条、记忆检索路径、API 调用记录及输出结果，生成面向管理层与监管机构的可解释性合规报表。 - **高效与安全并存：** 当这套治理机制成熟运转时，企业不仅不会被规则拖累，反而能够在确保资产安全的前提下，大幅提速决策流程、释放自动化生产力。 --- ### [12:23](https://www.youtube.com/watch?v=1qH3irPXI5Q&t=743s) 规模化落地 Agentic AI 的五大核心法则 在演讲的尾声部分，Srinivasa 梳理并重申了面向下一代智能企业的五大战略行动指南，为正在探索智能体落地的企业组织提供了清晰的路线图： - **从业务问题出发，而非追逐技术本身（Start with business problems, not technology）：** 技术的迭代日新月异，切忌为了使用新技术而立项，必须紧扣企业最真实的痛点和商业目标。 - **坚持模块化架构与 API 优先设计（Modular architectures and well-designed APIs）：** 构建解耦的模型、插件与接口能力，确保系统能够按需灵活拼装、平滑升级以满足业务诉求。 - **将治理作为构建信任的核心前提（Governance is essential for trust and enterprise scale）：** 治理是企业自身数据安全及对外输出服务信誉的生命线。 - **始终保留人类监督机制（Human oversight remains critical）：** 在所有自主工作流的设计中将人类置于关键环路中，兼顾效率与风控合规底线。 - **将 Agentic AI 确立为下一代智能企业的运营模式（Operating model for intelligent enterprises）：** 智能体协作不仅是对内赋能提效的引擎，也将成为企业对外交付新型智能服务的基础设施。 在分享的最后，Srinivasa 再次向观众致歉并幽默化解了开场的技术小插曲，随即宣布进入现场 Q&A 互动环节。 ## Notable quotes > [01:46](https://www.youtube.com/watch?v=1qH3irPXI5Q&t=106s) 传统聊天机器人负责回答问题，而 Agentic AI 负责达成业务目标（Traditional chatbots answer questions, agentic AI helps accomplish goals）。 — Srinivasa Manapragada > [08:43](https://www.youtube.com/watch?v=1qH3irPXI5Q&t=523s) 治理绝不是部署完成之后才追加的补丁，治理本身就是系统架构不可分割的一部分。 — Srinivasa Manapragada > [08:43](https://www.youtube.com/watch?v=1qH3irPXI5Q&t=523s) 先赢得内部信任，才能赢得外部世界（First if we try to win our house, then we can try to win the external world）。 — Srinivasa Manapragada > [08:43](https://www.youtube.com/watch?v=1qH3irPXI5Q&t=523s) 没有人类介入，规模化落地将无从谈起。 — Srinivasa Manapragada > [12:23](https://www.youtube.com/watch?v=1qH3irPXI5Q&t=743s) 从业务问题出发，而非追逐技术本身（Start with business problems, not technology）。 — Srinivasa Manapragada > [12:23](https://www.youtube.com/watch?v=1qH3irPXI5Q&t=743s) 坚持模块化架构与 API 优先设计（Modular architectures and well-designed APIs）。 — Srinivasa Manapragada > [12:23](https://www.youtube.com/watch?v=1qH3irPXI5Q&t=743s) 将治理作为构建信任的核心前提（Governance is essential for trust and enterprise scale）。 — Srinivasa Manapragada > [12:23](https://www.youtube.com/watch?v=1qH3irPXI5Q&t=743s) 始终保留人类监督机制（Human oversight remains critical）。 — Srinivasa Manapragada > [12:23](https://www.youtube.com/watch?v=1qH3irPXI5Q&t=743s) 将 Agentic AI 确立为下一代智能企业的运营模式（Operating model for intelligent enterprises）。 — Srinivasa Manapragada ## Mentioned Srinivasa Manapragada, 富国银行 (Wells Fargo), Ai4, Agentic AI (智能体 AI), 大语言模型 (LLMs), 小语言模型 (SLMs), 模型上下文协议 (Model Context Protocol / MCP), 人机协同 (Human-in-the-Loop / HITL), 负责任 AI (Responsible AI), Copilot / AI Assistants, 生成式 AI (Generative AI) ## Video description Srinivasa Manapragada - VP, Wells Fargo Building Governed Agentic AI at Scale The next evolution of enterprise AI is not simply generating responses, it is orchestrating intelligent AI agents that can reason, plan, collaborate, and execute business processes securely at scale. This session explores how Agentic AI combines Large Language Models (LLMs), Model Context Protocol (MCP) servers, APIs, enterprise data, and intelligent orchestration to autonomously understand user intent, invoke enterprise tools, and deliver contextual business outcomes while maintaining governance, security, and compliance. Attendees will learn practical strategies for designing scalable AI agent architectures, implementing Human-in-the-Loop (HITL) controls, and applying Responsible AI principles to build trusted enterprise solutions. Join this session to discover an enterprise blueprint for deploying secure, governed, and production-ready Agentic AI that accelerates innovation, reduces operational complexity, and delivers measurable business value. [Source](https://www.youtube.com/watch?v=1qH3irPXI5Q)

[1:46] Srinivasa Manapragada: 传统聊天机器人负责回答问题，而 Agentic AI 负责达成业务目标（Traditional chatbots answer questions, agentic AI helps accomplish goals）。 [8:43] Srinivasa Manapragada: 治理绝不是部署完成之后才追加的补丁，治理本身就是系统架构不可分割的一部分。 [8:43] Srinivasa Manapragada: 先赢得内部信任，才能赢得外部世界（First if we try to win our house, then we can try to win the external world）。 [8:43] Srinivasa Manapragada: 没有人类介入，规模化落地将无从谈起。 [12:23] Srinivasa Manapragada: 从业务问题出发，而非追逐技术本身（Start with business problems, not technology）。 [12:23] Srinivasa Manapragada: 坚持模块化架构与 API 优先设计（Modular architectures and well-designed APIs）。 [12:23] Srinivasa Manapragada: 将治理作为构建信任的核心前提（Governance is essential for trust and enterprise scale）。 [12:23] Srinivasa Manapragada: 始终保留人类监督机制（Human oversight remains critical）。 [12:23] Srinivasa Manapragada: 将 Agentic AI 确立为下一代智能企业的运营模式（Operating model for intelligent enterprises）。

Srinivasa Manapragada: The next evolution of enterprise AI is not simply generating responses, it is orchestrating intelligent AI agents that can reason, plan, collaborate, and execute business processes securely at scale. [0:04] AI Avatar / Srinivasa Manapragada: Agentic AI transitions the paradigm from merely answering questions to actively accomplishing complex business goals. [0:04] AI Avatar / Srinivasa Manapragada: Always start with concrete business problems, not just technology. [0:04] AI Avatar / Srinivasa Manapragada: Build on modular architectures and well-designed APIs. [0:04] AI Avatar / Srinivasa Manapragada: Treat governance as essential for enterprise scale and organizational trust. [0:04] AI Avatar / Srinivasa Manapragada: Retain human oversight as a non-negotiable critical component. [0:04] AI Avatar / Srinivasa Manapragada: Adopt agentic AI as the operating model for next-generation intelligent enterprises. [5:18] Srinivasa Manapragada: Execution hinges on invoking internal enterprise tools rather than relying solely on external commercial AI services. [7:31] Srinivasa Manapragada: winning our house

[0:04] So, I'll make it very easy for you  because of the technical glitch. [0:08] And rather we wanted to make  it more agentic today. So, [0:16] don't mind if I run away from this hall. [0:21] Building governed agentic AI at scale. So, this  is the topic what uh the conference guys gave me. [0:28] And I'm generally speaking I'm not speaking about  the enterprise Wells Fargo level, but uh it's [0:34] rather a kind of a generalized enterprise  level. But, let me run away from this hall. [0:41] Witness the next major evolution. This is agentic  AI, secure, governed, and responsible. Welcome [0:48] Srinivasa Mana future of enterprise intelligence.  I'm Srinivasa Mana Pragada. Today we'll explore [0:52] building governed agentic AI. Agentic AI  moves beyond chatbots. It actively helps [0:56] accomplish goals. Governance must be part of the  architecture. Let's discuss agentic AI. We must [1:01] build trusted systems at an enterprise scale. It  drives real business value while remaining secure. [1:06] Over the past few years, we've experienced a  move from traditional automation to generative AI [1:11] where machines could create concert questions.  Then came AI assistants and help people become [1:15] more productive. Today we're entering a new  phase. Instead of simply responding to prompts, [1:19] AI is becoming capable goals, reasoning through  complex tasks. This new paradigm is known as [1:25] agentic AI. However, for enterprises, success  is not just about deploying with governance, [1:30] security, and security, compliance, scalability,  and human oversight built into the architecture [1:35] from the beginning. Today's session is about  how organizations can build trusted agentic AI [1:41] that delivers business value while remaining  secure, governed, and responsible. So, [1:46] let's talk more and understand what makes  agentic AI different from the AI assistants [1:53] many organizations are using today. What is  agentic AI is much more than a chatbot. The first [1:59] capability is understanding instead of simply  reacting to a prompt, an AI agent understands [2:03] what the user is actually trying to accomplish.  The second capability is reasoning and planning. [2:08] The agent breaks complex work into logical steps,  determines what information is needed, selects the [2:13] appropriate tools, and coordinates activities  to complete the task. The third capability is [2:17] rather than stopping with an answer, the agent  can interact with enterprise applications, APIs, [2:22] knowledge repositories, and business services to  help complete the workflow. The key difference [2:26] is simple. Traditional chatbots answer questions,  agentic AI helps accomplish goals. That shift from [2:32] answering becomes how do we safely deploy inside  a large enterprise? The key difference is simple. [2:37] Traditional chatbots answer questions, agentic AI  helps accomplish goals. That shift from answering [2:43] becomes how do we safely deploy inside a large  enterprise? Everything begins with the user. [2:47] Interaction can be through a web application,  mobile application, or conversation. The [2:51] experience layer provides the interaction between  determines what switch tools should be used and [2:56] manages execution across the enterprise. Those  agents leverage foundation models and large [3:01] language models to reason, generate responses,  and make decisions. The enterprise services. They [3:07] then connect securely to enterprise capabilities  through APIs, knowledge repositories, but perhaps [3:13] the most important layer is governance.  Governance is not something we add after [3:17] deployment. Governance is part of the architecture  itself. Security, identity, policy enforcement, [3:23] human oversight, auditing, observability, and and  compliance must be embedded throughout the entire [3:29] AI life cycle. This is what enables organizations  to build AI systems that people can trust. [3:36] Second is orchestration. Multiple agents work  together, coordinate tasks, select tools, [3:42] and execute workflows across enterprise systems.  Third is integration. AI creates value only when [3:49] it can securely interact with enterprise  applications, APIs, knowledge repositories, [3:55] and business data. Finally, responsible AI  brings everything together. Security, privacy, [4:01] governance, explainability, and compliance  ensure AI remains in line with organizational [4:05] objectives. When these four pillars work together,  organizations can achieve faster decision-making, [4:10] higher productivity, intelligent automation,  and responsible AI adoption at enterprise scale. [4:16] As we look ahead, the evolution of enterprise AI  becomes very clear. We started with traditional [4:20] automation. Then came generative AI. Next  came AI co-pilots that assist people. Today, [4:26] we are entering the era of agentic AI, where  intelligent agents collaborate, reason, and [4:33] execute work toward business goals. This is the  foundation of the autonomous enterprise. It's not [4:39] about replacing people. It's about empowering them  with trusted AI capabilities. Now, I'd like to [4:44] leave you with five key takeaways. First, always  start with business problems, not technology. [4:49] Second, modular architectures and well-designed  APIs. Third, governance is essential for trust [4:55] and enterprise scale. Fourth, human oversight  remains a critical component. And finally, agentic [5:01] AI is becoming the operating model for the next  generation of intelligent enterprises. Thank you [5:06] very much for your time. I'm Srinivas Manapragada,  your AI evangelist. As we look forward, the era of [5:12] let's build trusted autonomous together. I  look forward to shaping the future with you. [5:18] So, what do you think? [5:24] If there was this kind of technical glitches,  do you think that agentic can help this kind of [5:30] presentations? This kind of recording, what I  pre-planned? I know there's a lot of events I [5:36] participated, a lot of conferences I participated.  But, you know, getting into this kind of mode of [5:42] creating your enterprise level of understanding  like how you can try to present yourself. [5:50] And thinking how you can make use of this  agentic. But enterprise level when we see [5:57] about the financial industry like from  Wells Fargo or any kind of financial area. [6:03] The main we need to understand what is our  goal and mainly and what is our context. [6:14] The reason and we need to plan in a such a  way that how we can try to make sure that [6:22] how our enterprise uh companies can  be benefited with this agentic AI. [6:28] I see a lot of people came out with a new  terminology here in the conferences like [6:34] from the traditional automation it moved to gen  uh uh gen AI then it came to the agentic AI. [6:42] But where where we wanted to execute the  workflows that is what important. And [6:48] and the basically when you execute the workflows  what is that you try to invoke yourself here is [6:57] the enterprise tools. Not many companies have  that kind of finances or that kind of funding [7:04] that they can try to utilize which are in the  market the tools or the models that they are [7:10] there. But they need to envision themselves to  create their own internal models that is small [7:16] language models. So that is what we are trying  to uh focus from the enterprise level and where [7:24] we take this architecture in a such a way  that the entry point is nothing but the users. [7:31] Where we have the three entry points is web mobile  and the chat chat the chatbots what we have it. [7:38] And where we say the second layer is a AI  experience layer. Experience layer is like what is [7:46] what what is the end goal that you wanted to build  your uh agent kind of stuff in your enterprise. [7:53] And then lays in the architect AI orchestration  layer where this is has the major part in your [8:00] enterprise level where agents talk with different  agents and internal and external kind of agents [8:06] and try to make an orchestra layer that can build  your workflow. And the third and the fourth one is [8:12] the the foundation models that the foundation  models what you are trying to use it in that. [8:18] And the fifth one is the enterprise integration  where you have the APIs where we have to build [8:25] our own APIs and where we have the databases, the  knowledge bases what we are trying to collect, [8:31] how to integrate the third stuff and then we  have the business application internal business [8:36] applications and there are MCP enabled tools in  our organization. But what do you see on the whole [8:43] orchestration or architecture behind, the main key  role is the governance. The governance because we [8:49] need to get the trust from internal from our  you know the the the people who work with us, [8:57] we need to get the trust. Then first if we  try to win our house, then we can try to win [9:03] the external world. So the governance is the main  important thing about the whole architecture what [9:10] we are taking as a reference here. This considers  as a security, the identity, the policy and the [9:17] main thing which is I put it in the middle is the  human intervention. Without human intervention, [9:23] nothing is possible. So we need to have that  kind of capability and we need to comply with the [9:30] audit, we need to have the compliance and we need  to have the observability, the overall things and [9:35] how we are trying to report, how we are trying  to generate our analytics and reports to the [9:40] higher management here. So, at our  organization level or any enterprise level, [9:45] this governance plays a main role. And if  we are if we are trying to consider this [9:51] governance as a key input to our whole agentic  stuff, I think so we have won. We have won in [9:58] you know, getting the things how efficiently we  can try to get this work done in our enterprise. [10:06] The framework where the deployment, whenever  you are trying to deploy, when you have these [10:13] capabilities of the governance and everything,  the intelligence plays the key part. Like [10:19] how are you trying to plan? How are you trying  to reason? How are you trying to memorize that [10:25] particular stuff? And then it comes to  the orchestration. That what I said is in [10:30] the previous slide is the orchestration is when  talking to different agents. So, multiple agents [10:35] collaboration and the workflow kind of execution  what we are trying to see. And there is a [10:40] API invocation. So, that is plays a key part like  how you try to subscribe to your internal APIs [10:48] and how you try to orchestrate that  and integrate to the multiple agent [10:52] and how do you collaborate with the multiple  agencies? And then comes to the integration part. [10:59] What are the things like you you have the  knowledge repository databases in your [11:04] How do you integrate them? And there is lot of  technicalities and there is lot of you know, [11:10] the perception that we need to follow certain  rules in the company where when we are trying to [11:15] access these kind of repositories. So, we need to  also play that kind of you know, key things that [11:22] we need to make sure that those capabilities are  being also addressed here. And as I mentioned that [11:30] human intervention, that is a responsible AI.  So, that is a part like where human is the main [11:36] person like you know, you need to put him as a  human in the loop for all the applications what [11:42] we are developing and have a kind of you know, uh  that will address your kind of compliance and the [11:50] how you can expand your you know, things to the  organization that will may that will give you kind [11:56] of a key kind of a person where they can create  a trust for you. And this whole things will help [12:03] to make us a faster decisions in a company and try  to improve the productivity or you know, you can [12:11] you can intelligently you can try to automate  the stuff and you can try to adopt and you can [12:17] also try to patronize this particular  agents for the external use of people. [12:23] And let's talk about the four you know, takeaways [12:28] and sorry, the five takeaways what we are trying  to plan and let's you know, lay the road map here. [12:36] Let's just start not with the you know, start  with the business problem not with the technology [12:41] because technology is a world a day day to  day but let's start with the kind of you know, [12:47] tackling with the business problem. That is the  most important thing for anybody to you know, [12:52] succeed in this agentic AI stuff. And we need  to also build the AI kind of using the API [13:01] uh and modular architecture models there where  we can try to make sure that that will try to [13:07] you know, suffice whatever the needs  that we are trying to attend the goal. [13:11] And as I mentioned I pushed the thing into the  governance into the middle part here. The third [13:17] takeaway is governance is very essential to create  a trust in any enterprise level. So, that's what [13:24] we are trying to build in our organizations  internally and trying to talk to the external [13:31] vendors who tries to, you know, come  to our organization and try to use our [13:37] kind of services. So, we are trying to make sure  that the governance is the main key thing that we [13:42] are your data or your kind of information is very  secure in us. And the fourth takeaway, as I said, [13:50] everywhere we need to have the human oversight.  That is a very critical part in the organization [13:55] where we need to whatever we are trying  to address through the agentic stuff, [13:59] that has to have a human intervention. And the  final thing is like, you know, it's it's in a [14:05] kind of a operating model that we need to have a  kind of a intelligent kind of enterprises to the [14:12] to the to the organizations what we  are trying to give and explain them [14:17] that how we we are trying to use our agentic kind  of stuff internally as well as we are trying to [14:23] suffice this kind of needs and  services to the external stuff. [14:29] Thank you and I'll open for the questions. I  think so I I made little bit clear like with [14:34] the all the technical I'm very sorry about  that, but any any questions from you guys.