IBM Unlocks $3.5 Billion in Savings Through Enterprise AI Agents

IBM通过企业级AI智能体节约35亿美元 只有当企业从清晰的自上而下战略而非技术试点出发，并从简单的任务自动化转向多智能体协同编排时，企业级AI部署才能产生经过验证的商业价值。 IBM与Oracle展示了一份战略蓝图，旨在超越AI实验阶段，利用AI智能体实现可衡量的企业价值。通过分享IBM作为“零号客户”（Client Zero）推动实现35亿美元年度运营生产力节约的实践经验，演讲者阐明了从基础自动化迈向多智能体协同编排的转变。演讲内容涵盖演进式自动化框架、多模型微调策略、通过Oracle Cloud Infrastructure上的watsonx Orchestrate保障数据主权，以及跨行业的真实落地应用案例。 IBM 与 Harris Poll 联合开展的调查显示，仅有 15% 的企业高管确信其实模化部署的 AI 带来了可量化的业务价值，这形成了从业务战略出发的“AI 领导者”与仅做技术试点的“AI 学习者”之间的显著鸿沟。 (0:04) IBM 作为自身的“零号客户”，在约 150 亿美元的总支出预算中，通过流程简化及 AI 自动化实现了超过 35 亿美元的年化生产力节约，且其技术驱动重心已由最初的传统自动化转向占 75% 的 AI 方案。 (2:35) IBM 提出了流程自动化从确定性 RPA、LLM 任务增强，迈向单一 AI Agent 及多 Agent 协同编排的四个演进阶段，并利用 watsonx Orchestrate 平台构建类似“运动员名单”的模块化 Agent 混合调度体系。 (7:28) IBM 与 Oracle 推进深度生态合作，将 watsonx 与 Granite 模型直接部署于 Oracle 云基础设施（OCI）上，通过联邦查询避免非必要的数据迁移，同时凭借领域微调小模型在降低 95% 运营成本的基础上实现更高的任务准确率。 (12:52)

IBM Consulting and Oracle executives argue that enterprise artificial intelligence generates measurable return on investment only when firms lead with end-to-end business strategy rather than technology-driven experimentation. The AI Value Gap and Strategy Direction (0:04) According to an IBM-commissioned Harris Poll of approximately 2,000 C-suite executives, only 15% feel confident they are deriving quantifiable business value from AI deployments at scale. Shobhit Varshney categorizes this top 15% as "AI Leaders" and the remaining 85% as "AI Learners" stuck in pilot loops. The primary difference stems from the sequence of implementation. Leaders adopt a "left-to-right" strategy, defining a primary business goal, mapping data exhaust, assessing workflow impacts, and embedding governance before deploying AI tools. Conversely, learners take a "right-to-left" path, launching free pilots and trying to alter employee routines before constructing a retrospective business case—a method Varshney terms a recipe for perpetual experimentation without enterprise value. Internal Operational Proof: IBM as Client Zero (2:35) IBM used its internal operations as "Client Zero" to validate AI integration at scale, responding to CEO Arvind Krishna's target of achieving $2 billion in run-rate productivity savings exiting 2024. The firm reported generating over $3.5 billion in annual productivity savings against an overall expense envelope of roughly $15 billion. The savings derived from $2.0 billion in automation and AI and $1.5 billion in process simplification. Over two years, IBM's internal savings mix evolved from an even split between basic automation and AI to 75% AI and 25% traditional automation. Specific initiatives included a unified financial management dashboard that IBM reports unlocked $200 million in value, alongside grassroots participation from 170,000 employees in an internal challenge to build localized workflow automations across its nearly 300,000-person global workforce. Four Levels of Enterprise Process Automation (7:28) Varshney outlines a progression through four distinct automation tiers. Level 1 relies on deterministic, rule-based robotic process automation that provides low cost and low latency within rigid parameters. Level 2 integrates large language models into existing nodes to interpret visually complex documents. Level 3 introduces single autonomous agents that take over localized "ask to task" execution, independently carrying out multi-step research or data-gathering requests. Level 4 reaches multi-agent orchestration, where an orchestrating model breaks down an overarching problem and delegates tasks to specialized sub-agents. While an IBM Institute for Business Value survey of 800 C-suite executives found 78% viewing agentic AI as transformative, 76% recognized that data security, privacy, and organizational overhauls remain unaddressed prerequisites. Enterprise architectures will likely remain hybrid, balancing fast, low-cost deterministic trees with resource-heavy, non-linear reasoning agents. Enterprise Deployment Architecture and Model Selection (12:52) Deploying agents across legacy infrastructure requires solving data gravity and governance constraints. Federico Torreti of Oracle and Varshney note that moving proprietary data across firewalls presents security risks; platforms such as IBM watsonx Orchestrate running on Oracle Cloud Infrastructure instead query systems of record directly using federated access models. To counter model selection fatigue, the speakers recommend choosing curated, domain-specific models over general frontier benchmark leaders. Varshney reports that shifting from large frontier models to fine-tuned open-source or IBM Granite models on targeted tasks can improve domain accuracy by 5% to 10% while reducing operating costs by 95%. Organizations are also increasingly framing these agent deployments around employee enablement—such as shortening construction contract bidding or accelerating hiring workflows—rather than treating automation purely as headcount reduction. While the architectural transition to multi-agent ecosystems is underway, long-term enterprise adoption hinges on operational governance. Leaders must balance latency, compute expense, and data sovereignty while shifting attention from authoring individual agents to continuously managing and auditing complex agent networks.

IBM咨询与Oracle在Ai4大会上提出，企业AI应用正面临仅有15%领军者能获得可量化回报的“价值鸿沟”；要实现商业突破，企业必须将重心从零散的技术试点转向由业务战略驱动的智能体编排与工作流再造。 跨越15%价值鸿沟与战略方向抉择 (0:04) IBM咨询数据与AI业务负责人Shobhit Varshney指出，企业落地AI必须兼顾“选择正确的AI”与“合规地实施AI”，在追求投资回报率的同时内嵌治理与安全机制。IBM委托哈里斯民调对约2000位企业高管进行的调查显示，仅有约15%的高管确信其规模化AI部署带来了可量化的业务价值，这部分企业被称为“AI领跑者”，而其余85%的企业仍停留在“AI学习者”阶段。两者的根本分歧在于执行路径：学习者采取“从右到左”模式，从免费技术试点出发，试图事后补充商业合理性，往往沦为无效实验；领跑者则采取“从左到右”模式，先确立业务北极星目标，在评估数据沉淀、流程影响与人才治理后，再针对性部署AI以重塑工作流。 零号客户实践与35亿美元生产力释放 (2:35) IBM将自身作为“零号客户”验证企业级AI效能。IBM首席执行官Arvind Krishna曾设定在2024年底前实现20亿美元年化生产力节约的目标，而财报数据显示，IBM最终在约150亿美元的年度总支出基础上，实现了超过35亿美元的年化生产力节约。其中20亿美元直接来自自动化与AI技术，15亿美元源自流程简化与整合；节约构成的技术配比也从两年前自动化与AI各占一半，转变为AI占据75%、传统自动化占25%的格局。通过节省成本建立的资金飞轮被重新投入研发与增长领域。 在面向全球近30万名员工的组织推进中，端到端财务绩效管理平台的整合贡献了2亿美元的业务价值。在运营机制上，IBM打破传统职能孤岛，组建跨职能团队追踪客户旅程，并通过高管指导与基层赋能相结合的模式推广应用。全球共有17万名员工参与了内部watsonx挑战赛，在日常岗位中主动构建AI工作流，配合两周一轮的敏捷冲刺周期，推动生产力收益大幅超出最初设定的基准线。 业务自动化的四层演进与智能体编排架构 (7:28) IBM商业价值研究院对800位高管的调研显示，78%的受访者认为智能体AI将从根本上重塑其业务，但76%的高管强调必须首先解决内部数据安全、隐私与组织变革问题。Varshney提出了企业流程自动化的四层演进模型：第一层为基于固定规则树的确定性RPA，具备低成本与高速度；第二层为引入大语言模型处理文档解析的任务增强；第三层为单智能体执行，即由人类拆解任务后，智能体自主完成跨网页查阅、表格整理与历史比对等具体目标；第四层为多智能体编排，系统可自动拆解复杂高阶需求并协调专门的智能体分工协同。 针对多智能体协作，IBM提出采用watsonx Orchestrate作为混合工作流的中央控制层，将其比喻为组建包含四分卫等专业位置的“运动员团队”。在该架构下，中央编排智能体负责评估需求，并按需将公司建档、运营开通以及法律合规等工作分发给专用智能体。这种模块化设计允许企业随时替换个别模型或工具，同时避免对底层数据安全和企业防火墙造成冲击。Varshney强调，未来的企业系统将是混合形态：确定性工作流保障响应速度、低成本与确定性输出，智能体架构则负责处理复杂的非线性推理。 生态合作落地：员工赋能与专用模型微调 (12:52) Oracle人工智能与机器学习高级总监Federico Torreti与Varshney结合两家公司的联合方案指出，推进AI落地正从单纯的成本削减转向员工赋能。例如一家工程咨询公司通过部署Oracle Fusion HCM智能体优化招聘流程，实现了约56%的生产力提升；户外媒体巨头Outfront Media利用智能体在客户交互现场提供情境数据支持，传统建筑企业则通过自动化文档处理显著缩短投标周期。面对分散在本地或多系统中的核心数据资产，双方主张避免不必要的数据迁移，依托watsonx Orchestrate在原始记录系统中开展联邦查询，以保障数据主权与合规。 在模型选型策略上，双方指出无节制地接入前沿模型容易引发“决策疲劳”并推高运营成本。实践证明，相较于盲目追求通用前沿模型，在Oracle云基础设施（OCI）数据科学平台上选用经过企业合规审计的IBM Granite等较小模型并进行领域微调更为务实。这种结合垂直场景的微调模式不仅能够将特定任务的准确率提升5%至10%，还能将运行成本大幅削减95%，使分层多智能体架构在经济与算力层面具备规模化落地的可行性。 两位发言人强调，企业AI的应用重心已从前期的模型构建转向多智能体生态的长期治理与统一运维。如何在兼顾数据主权、低延迟与算力预算的前提下，管理好日益庞大的专业子智能体网络，将是决定企业能否真正跻身前15%领跑者行列的核心考验。

## Background 该内容由 YouTube 频道 Ai4 于 2025 年 8 月 26 日发布，由 IBM 数据与 AI 负责人 Shobhit Varshney 与 Oracle 资深 AI/ML 总监 Federico Torreti 联合演讲。视频主要介绍了 IBM 如何告别单纯的 AI 实验阶段，通过在 HR、销售、IT 与采购等全业务链落地 AI Agent（智能体），在两年内释放出高达 35 亿美元的生产力价值，并展示了如何结合 Oracle 生态快速搭建多 Agent 协同体系。 ## TL;DR 演讲深入剖析了 IBM 作为“零号客户”通过 AI 与自动化实现 35 亿美元年化生产力跃升的转型历程，揭示了 15% 的“AI 领跑者”如何通过以战略和北极星目标为核心的自左向右路径拉开竞争差距。内容系统梳理了从传统 RPA 规则流演进到 Watsonx Orchestrate 驱动的多 Agent 协同编排架构，并结合双方在工程、建筑与招聘等真实行业场景中的落地实践，强调了尊重“数据重力”与利用联邦查询避免盲目搬迁数据的重要性。最后，演讲指出企业规模化部署的核心在于采用分层架构与场景化微调小模型（如 Granite），在提升 5% 至 10% 准确率的同时降低 95% 的推理成本，并深化对 Agent 全生命周期的长效治理。 ## Atmosphere & style 整场演讲充满务实、严谨且极具说服力的企业级商业气息，褪去了空泛的技术炒作，全程以详实的财务回报率（ROI）、架构拓扑与量化指标为支撑。主讲人 Shobhit 叙述逻辑严密、自信沉稳，借助“四分卫组队”和“带新员工”等生动比喻化繁为简；随后与 Oracle 嘉宾的协同对话自然默契，将前沿的 Agent 理论扎实地锚定在企业防火墙、数据重力与严苛的生产成本控制等现实挑战上，给人以极具落地参考价值的厚重感与启发感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=4s) IBM 提出企业级 AI 的双轮驱动战略：必须兼顾“选择正确的 AI”（聚焦高价值场景与持续 ROI）与“以正确的方式做 AI”（内置治理、信任与价值观对齐）。 - [01:07](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=67s) 调研显示仅 15% 的企业成为“AI 领跑者”，其关键在于确立市场制胜的北极星目标并“自左向右”重构流程，而非陷入从零散工具试点出发的“自右向左”实验陷阱。 - [02:35](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=155s) IBM 将自身作为“零号客户”，通过组织精简与 AI 自动化释放了超 35 亿美元年化生产力价值，其中 AI 驱动的增效占比已由两年前的 50% 跃升至 75%。 - [02:35](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=155s) 生产力跃升释放的大额自由现金流被重新注入核心研发与业务拓展，形成了“增效—再投资—业务增长”的良性飞轮，直接支撑了资本市场表现。 - [04:17](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=257s) 规模化推动 AI 变革需打破部门纵向孤岛，围绕端到端客户旅程组建横向敏捷团队，并结合双周冲刺与全员基层创新激发组织活力。 - [07:28](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=448s) 流程自动化正从基于确定性规则的传统 RPA 阶段，大步迈入具备非结构化数据理解、自主规划与跨系统协同能力的 Agentic AI 阶段。 - [09:20](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=560s) 企业落地的最佳架构是将“固定确定性工作流”与“Agent 动态工作流”相融合，利用编排中枢将高层需求拆解并分发给专精 Agent 协同执行。 - [12:52](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=772s) AI Agent 的认知定位已发生根本转变：领跑企业不再单纯将其视作“削减成本”的工具，而是作为“赋能员工能力边界扩张”的增长引擎。 - [16:36](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=996s) 践行“AI 追随数据重力”原则，企业应采用联邦查询直接在底层源头记录体系中调度计算，避免盲目迁移核心数据资产带来的安全与合规风险。 - [16:36](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=996s) 面对层出不穷的大模型，企业应以“场景化精选模型”打破决策疲劳，采用具备强审计性与企业合规性的模型作为生产级部署基石。 - [18:54](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=1134s) 相较于依赖通用的超大模型，针对垂直场景微调的专精小模型能够将运行推断成本降低近 95%，同时将任务准确率提升 5% 至 10%。 - [18:54](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=1134s) 多 Agent 规模化架构应采取分层设计：顶层主控 Agent 负责全局统筹与路由调度，底层垂直领域 Agent 则由低成本、高响应的小模型承担。 - [19:58](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=1198s) 企业 AI 的核心演进正从前期的“如何构建 Agent（Agent Authoring）”，转向涉及持续运维、合规监控与调度的“Agent 全生命周期管理（Agent Management）”。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=4s) 践行 AI 战略：做对的 AI 与用对的方式做 AI IBM Consulting 数据与 AI 负责人 Shobhit Varshney 开场介绍了他在企业级 AI 领域的实践经验，并深入分享 IBM 如何通过 AI 和 AI Agent（智能体）实现自身的业务转型。在 IBM 内部，AI 战略的核心理念被概括为一个阴阳互补的双轮驱动模型：**“选择正确的 AI”（Doing the right AI）** 与 **“以正确的方式做 AI”（Doing AI right）**。 ```mermaid flowchart LR accTitle: IBM Doing the Right AI vs Doing AI Right subgraph RightAI["选择正确的 AI (Doing the Right AI)"] A1["业务场景识别"] --> A2["AI 素养与技能普及"] A2 --> A3["推动全员采纳与落地"] A3 --> A4["持续实现季度 ROI 价值"] end subgraph DoingRight["以正确方式做 AI (Doing AI Right)"] B1["健全的治理框架"] --> B2["内置全流程信任与安全"] B2 --> B3["与企业核心价值观对齐"] B3 --> B4["负责任的高效部署"] end ``` 一方面，企业需要识别出真正产生价值的业务用例，提升员工的 AI 素养，推动自下而上的技术采纳，确保所选场景能持续产生可量化的投资回报率（ROI）；另一方面，必须在技术落地伊始就植入强大的 AI 治理体系与信任机制，确保 AI 战略与组织的核心价值观深度契合，以负责任且高效的方式部署每一项用例。 ### [01:07](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=67s) AI 领跑者与摸索者的本质差异：自左向右 vs 自右向左 IBM 联合哈里斯民意调查（Harris Poll）对约 2,000 名企业高管（C-suite）展开了调研，探究有多少企业能够将真实的业务价值直接归因于其规模化落地的 AI 部署。结果显示，仅有约 15% 的高管明确表示对 AI 投资所带来的巨大商业回报充满信心，IBM 将这部分企业定义为“AI 领跑者”（AI Leaders）；而其余 85% 的企业则被称为“AI 摸索者”（AI Learners）。 ```mermaid flowchart TD accTitle: AI Strategy Direction Comparison subgraph Leaders["15% AI 领跑者（自左向右：以战略为核心）"] L1["明确战略与北极星目标"] --> L2["价值编排与数据沉淀挖掘"] L2 --> L3["重塑端到端业务工作流"] L3 --> L4["赋能人才队伍"] L4 --> L5["技术构建与数据治理"] end subgraph Learners["85% AI 摸索者（自右向左：以实验为核心）"] R1["发起免费概念验证/试点 (PoC)"] --> R2["局部改变人类工作方式"] R2 --> R3["倒推拼凑商业可行性案例"] end ``` 这两类企业在推进 AI 战略的路径上存在根本性分歧： - **AI 领跑者（自左向右路径）**：首先从核心战略出发，确立市场制胜的“北极星指标”。例如，一家传统抵押贷款审批需要 3 周的房贷公司，若能将审批周期压缩至 24 到 48 小时，便能颠覆行业竞争格局。确立这一北极星目标后，领跑者会分析自身数据沉淀（Data Exhaust）与价值捕获能力，进而重构业务工作流、赋能人才队伍，最后才落地具体技术选型与数据治理。 - **AI 摸索者（自右向左路径）**：往往直接从免费的局部工具试点（PoC）起步，试图先改变某一部分员工的工作习惯，随后再艰难地倒推拼凑商业价值案例（Business Case）。这种“自右向左”的模式往往沦为无休止的“AI 实验陷阱”。 ### [02:35](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=155s) IBM 作为“零号客户”：35 亿美元生产力跃升的飞轮效应 IBM 将自身定位为各项技术方案的“零号客户”（Client Zero），信奉“喝自己的香槟”（Drink our own champagne）。IBM 全球整体运营支出盘子约为 150 亿美元，公司最初由 CEO Arvind Krishna 设定了在 2024 年底前实现 20 亿美元生产力节省的宏伟目标，而实际成效远超预期，最终实现了年化运行率（Annual Run Rate）超过 35 亿美元的生产力价值释放。 - **收益构成细分**：在 35 亿美元的总收益中，约 20 亿美元直接来源于自动化与 AI 技术的赋能；其余约 15 亿美元来自流程简化、组织合并与架构精简。 - **技术结构演进**：两年前，IBM 内部自动化与 AI 的比重约为 50/50；如今，纯自动化占比已降至 25%，而 75% 的降本增效均由 AI 深度驱动。 - **财务与资本市场回响**：35 亿美元节省释放出的自由现金流被重新投入核心业务研发与市场拓展，形成了“生产力提升 $\rightarrow$ 利润再投资 $\rightarrow$ 驱动业务增长 $\rightarrow$ 释放自由现金流”的良性飞轮，直接支撑了 IBM 股价在过去几年跑赢大盘与竞争对手。 ### [04:17](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=257s) 跨领域赋能与横向协同运营架构 在落地层面，IBM 逐个业务领域深入挖掘价值潜力。例如在财务绩效管理（Financial Performance Management, FPM）领域，IBM 打造了单一管理视图（Single Pane of Glass），将全公司所有财务决策集中整合呈现，仅此一项就释放了 2 亿美元的商业价值；此外，在客户支持、IT 现代化、数字劳动力（Digital Labor）及覆盖全球近 30 万名员工的 HR 转型中均取得了显著突破。 ```mermaid flowchart TD accTitle: IBM Transformation Structure and Horizontal Alignment subgraph GovStructure["多层级转型组织架构"] G1["指导委员会 (Steering Committee)"] --> G2["生产力探索团队 (Productivity Discovery Teams)"] G2 --> G3["全员基层创新 (Grassroots Workforce)"] end subgraph HorizontalShift["业务运营视角转变"] H1["原纵向部门隔离视角\n(财务/营销/合同各自割裂)"] --"打破壁垒，转向横向贯通"--> H2["端到端客户旅程\n(线索 -> 账户 -> 定价 -> 合同签署)"] end ``` 为了在大规模组织中推动变革，IBM 建立了一套严谨的文化与治理体系： - **外部基准对齐**：对每个职能部门（如采购规模、IT 工程师与工单比例）进行全球行业基准测算，明确提升目标。 - **从垂直孤岛转向横向贯通**：打破各部门各自为政的局部优化，围绕客户从市场、销售、定价到签约的“横向端到端旅程”组建跨职能敏捷团队，设立统一的综合 KPI。 - **双周敏捷冲刺（2-Week Sprints）**：快速验证用例成效，形成标准化操作手册（Playbook）。 - **自下而上的全员参与**：各业务部门骨干组成的生产力探索团队贡献了约 10 亿美元的价值挖掘；而通过面向全员的 watsonx 挑战赛，全球多达 17 万名员工参与设计日常工作流自动化方案，形成了坚实的基层群众基础。 ### [07:28](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=448s) 从 RPA 到 Agentic AI：流程自动化的演进阶段 在当前的 AI 浪潮中，AI Agent（智能体）已成为重构端到端业务流程的核心焦点。根据 IBM 商业价值研究院（IBV）针对 800 余位高管的专项调研： - **78%** 的企业高管高度看好 Agentic AI（智能体 AI）对企业业务的重塑能力。 - **76%** 的高管清醒地意识到，要释放智能体的最大价值，企业内部仍需在数据安全、隐私保护以及底层流程重构上进行大量深度转型。 从流程自动化演进的视角来看，行业历经了不同阶段： - **第一阶段：RPA（机器人流程自动化）时代**。针对汽车保险理赔等固定场景，领域专家（SME）可以在白板上画出清晰的决策树与规则流（“如果满足 A 则执行 B”）。此类自动化成本低、执行速度快、输出确定一致，但其本质完全受限于人类预先定义的规则框架。 - **演进阶段（进阶智能）**：随着技术深入，系统逐步具备理解非结构化数据、自主规划与协同决策的能力，跨入真正由数据与 Agent 驱动的智能化新阶段。 ### [09:20](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=560s) 从任务拆解到多 Agent 协同：企业自动化工作流的进阶之路 企业在推进自动化流程时，正经历着从单一环节大语言模型（LLM）赋能向高阶 Agent 进化的关键转折。早期阶段通常只是在既定流程中将某个具体节点（例如视觉识别读取特定文档并进行快速分析）替换为大模型；而进入第三阶段后，AI Agent 开始展现出自主执行独立子任务的能力。这就像团队中迎来了一位新入职的员工或实习生——当你交代任务：“能否帮我对比一下我们与竞争对手的财报？”该 Agent 能够自主打开浏览器检索竞品、下载数据、生成对比表格并参照过往格式整理，最终在得出满意结果后反馈给负责人。现阶段，高层的需求拆解（Ask to Task）依然主要由人类完成，人类管理者接收上级需求后拆解为具体子任务下发给 Agent，并在获得反馈后进行验证与推进。 随着企业内部积累了大量的“需求-任务拆解”（Ask to Task）知识库，便可以进一步训练核心模型来承接宏观需求并自动进行任务分解与 Agent 调度。此时，系统演进为多 Agent 协同编排架构，能够应对如“为什么我这次的报销理赔金额与上次不同？”这类涉及多个底层系统的复杂业务问题。然而，多 Agent 框架虽然大幅拓宽了解决复杂问题的边界，但也带来了更高的延迟、计算成本以及更严格的结果治理与准确性验证要求。相比之下，传统的确定性工作流虽然必须穷举每一步路径，却具有确定可控的优势。因此，企业落地中最具价值的做法是将“固定工作流”与“Agent 动态工作流”深度结合。IBM 内部正是通过 Watsonx Orchestrate 来实现这种跨固定与动态工作流的统一智能编排，无缝调遣来自 IBM 自身或生态伙伴的各类 Agent 与工具集。 ```mermaid flowchart TD accTitle: Watsonx 多 Agent 编排架构 subgraph Ingestion["接入与分发层"] A["多渠道请求 (Slack / 呼叫中心)"] --> B["Watsonx Orchestrate 编排中枢"] end subgraph Agents["专业 Agent 协作层"] B --> C["企业画像分析 Agent"] B --> D["生态伙伴入职 Agent"] B --> E["IBM AI 法务顾问 Agent"] end subgraph Integration["执行与支撑层"] C & D & E --> F["底座系统与记录体系 (System of Record)"] end ``` - **模块化组队机制**：在企业入职（Onboarding）等复杂场景中，编排中枢会将宏观需求拆解为企业画像、生态入职对接、法务合规审核等多项步骤，分别指派专精的 Agent 协作完成。这如同组建一支高水平运动队，根据比赛战术灵活引入特定技能的“四分卫”，实现工作流的敏捷组装与持续替换。 --- ### [12:52](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=772s) IBM 与 Oracle 的生态协作：从概念验证走向生产落地 为了避免客户被技术绑定，IBM 坚持开放的生态合作战略，确保企业能够自由置换组件，并遵循“AI 追随数据重力”（AI follows data gravity）的原则，在企业防火墙后安全处理核心资产。作为深耕数十年的重量级合作伙伴，Oracle 的基础设施与 IBM 的咨询服务及软件能力（包括原生运行于 Oracle 平台的 Watsonx Orchestrate 和各类大模型）形成了强力互补。Oracle 资深 AI/ML 总监 Federico Torreti（Feddy）受邀登台，共同探讨企业如何跨越从 2024 年的试点探索（PoC）阶段，迈入 2025 年以 15% 领跑企业为代表的全面生产部署阶段。 双方的联合实践证明，AI Agent 的普及不仅限于科技企业，传统实体行业同样获得了巨大的效能跃迁。在建筑工程领域，传统企业利用自动化与 AI Agent 大幅缩短了招投标周期（Time to Bid）；户外媒体巨头 Outfront Media 则通过为一线销售代表提供交互现场的富信息赋能，显著提升了潜在客户触达与业务转化质量。此外，知名工程咨询服务企业 Wood 部署了 Oracle Fusion HCM Agent 优化其端到端招聘流程（Source-to-Hire / IR-to-Hire），成功实现了约 56% 的整体生产力提升。这些领先企业的共性在于完成了认知层面的根本转变：不再将 AI Agent 仅仅视作“削减成本”的手段，而是作为“赋能员工创造更多增量价值”的引擎，去攻克过去无法解决的业务难题。 - **56% 生产力提升**：工程咨询公司 Wood 借助 Fusion HCM Agent 优化端到端招聘流程，消除了多处传统流转瓶颈。 - **价值导向转变**：领先的 15% 企业摆脱了单纯的 Cost-out 视角，转向赋能员工能力边界扩张。 --- ### [16:36](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=996s) 应对数据重力与模型泛滥：联邦查询与场景化模型策略 针对大量非科技原生企业数据分散在本地机房（On-Premise）或多处异构系统、难以集中迁移的现实痛点，Federico 强调企业应尽量避免盲目进行大规模数据搬迁。最佳的实践心智模型是“让数据保留在源头的记录体系（System of Record）中”，通过 Watsonx Orchestrate 等平台所具备的联邦查询（Federated Queries）能力，将计算和推理分发到数据所在之处，兼顾数据合规、安全隐私与架构灵活性。 面对业内每周层出不穷、性能不断刷新的前沿大模型，企业普遍面临严峻的选择焦虑与决策疲劳（Decision Fatigue）。单纯追求“提供最广泛的模型目录”往往适得其反，更高效的策略是“按需逆向推导”——针对特定的业务场景，精选经过严格验证与优化的模型组合。在此背景下，IBM Granite 系列模型原生接入了 OCI Data Science（Oracle 云基础设施数据科学平台），以其出色的可审计性（Auditability）、企业合规性与可控成本，成为了生产级部署的理想基石。在企业落地旅程中，团队通常会先用最前沿的模型进行小范围概念验证以确认可行性，但一旦进入严苛的生产环境，高确定性、高合规性的精选模型才是企业级架构的核心支撑。 - **数据重力架构**：坚持数据不动、智能下沉，通过联邦查询保护核心数据资产与企业知识产权。 - **精选模型（Curated Choice）打破决策疲劳**：相比盲目追求前沿大模型，具备审计与合规属性的模型（如 IBM Granite 在 OCI 上的应用）更能满足企业生产级部署的实际诉求。 ### [18:54](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=1134s) 规模化落地的核心：模型微调、成本控制与分层 Agent 架构 在将 AI Agent 真正推向规模化企业级部署时，企业必须面对一系列工程与运营层面的现实约束，包括成本控制边界（Cost Envelope）、响应延迟（Latency），以及模型本身的知识范围——模型究竟是该依赖覆盖全网的通用世界知识，还是针对特定业务领域进行深度微调（Fine-tuning）。演讲者指出，IBM 在实践中通过采用自家的 Granite 模型以及其他开源模型，针对特定业务场景进行定向微调，从而大幅收敛模型任务范围，取得了显著成效。从通用的超大模型实验阶段过渡到“专精于特定任务的小模型”，客户能够看到令人瞩目的投资回报率（ROI）：任务准确率不仅提升了约 5% 到 10%，而运行这些模型的推断成本更是骤降了近 95%。 这种向小型、高效模型转型的趋势在多 Agent（Multi-Agent）体系架构中体现得尤为明显。在企业级的 Agent 层级体系中，顶层通常设置一个负责全局协调、理解综合指令并统揽各大业务域的主控 Agent；而一旦深入到底层特定业务垂直领域，例如专门处理采购业务的“采购 Agent（Procurement Agent）”，该 Agent 就必须依赖运行成本极低、响应迅速的高效小模型。通过分工协作，小型模型能够在各自专注的狭窄工作流中输出更高精度与稳定性，同时将整体计算开销降至最低。 ```mermaid flowchart TD accTitle: Multi-Agent Hierarchical Architecture accDescription: 企业级多智能体分层架构示意图，展示了顶层总控与底层细分领域专精小模型的分工与成本优化机制。 subgraph Orchestration["协调调度层"] TopAgent["顶层主控 Agent<br/>（跨领域统筹与路由调度）"] end subgraph DomainAgents["垂直领域专属 Agent（微调小模型）"] Procurement["采购 Agent<br/>（微调 Granite / 开源小模型）"] HRAgent["人力资源 Agent<br/>（领域专精小模型）"] ITAgent["IT 运维 Agent<br/>（高性价比小模型）"] end TopAgent -->|任务分发| Procurement TopAgent -->|任务分发| HRAgent TopAgent -->|任务分发| ITAgent subgraph Benefits["落地收益"] Cost["运行成本降低 ~95%"] Acc["准确率提升 5% - 10%"] end DomainAgents -.-> Benefits ``` - **微调小模型的关键 ROI 收益**：相较于直接调用超大通用模型，专精于特定场景的微调小模型可实现“准确率提升约 5% 至 10%，运行成本降低高达 95%”。 - **架构分工原则**：顶层 Agent 负责全局跨领域调度，底层领域 Agent（如采购 Agent）采用轻量化高效模型，严格专精于特定业务流程。 --- ### [19:58](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=1198s) 从 Agent 构建到全生命周期管理与圆满结语 除了 Agent 的创建与构建（Agent Authoring）之外，演讲者进一步强调了另一个在企业落地过程中愈发关键的维度——Agent 的运营与治理管理（Agent Management）。随着企业内部部署的 Agent 数量不断增加，如何持续管理、监控、调度这些自主智能体并确保其合规高效运行，正在成为构建现代化企业 AI 体系不可或缺的核心议题。演讲在双方对 IBM 与 Oracle 紧密合作伙伴关系的肯定与致谢中圆满结束，双方表示这种深入的行业探索与技术协同将继续助力更多企业解锁生产力飞跃。 - **核心治理关注点**：企业 AI 的演进不仅关乎“如何构建 Agent（Agent Authoring）”，更在于“如何长效运维与管理 Agent（Agent Management）”。 - **总结致谢**：双方高度肯定了本次交流对企业智能化转型的启发价值，并对长期的生态伙伴合作表达了由衷感谢。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=4s) 选择正确的 AI（Doing the right AI）与以正确的方式做 AI（Doing AI right）。 — Shobhit Varshney > [01:07](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=67s) 自左向右路径：首先从核心战略出发，确立市场制胜的北极星指标。 — Shobhit Varshney > [02:35](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=155s) IBM 将自身定位为各项技术方案的“零号客户”（Client Zero），信奉“喝自己的香槟”（Drink our own champagne）。 — Shobhit Varshney > [02:35](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=155s) 两年前，IBM 内部自动化与 AI 的比重约为 50/50；如今，纯自动化占比已降至 25%，而 75% 的降本增效均由 AI 深度驱动。 — Shobhit Varshney > [04:17](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=257s) 打破各部门各自为政的局部优化，围绕客户从市场、销售、定价到签约的“横向端到端旅程”组建跨职能敏捷团队。 — Shobhit Varshney > [07:28](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=448s) 78% 的企业高管高度看好 Agentic AI 对企业业务的重塑能力，76% 的高管清醒地意识到仍需在数据安全、隐私保护以及底层流程重构上进行深度转型。 — Shobhit Varshney > [09:20](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=560s) 这如同组建一支高水平运动队，根据比赛战术灵活引入特定技能的四分卫，实现工作流的敏捷组装与持续替换。 — Shobhit Varshney > [12:52](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=772s) 遵循“AI 追随数据重力”（AI follows data gravity）的原则，在企业防火墙后安全处理核心资产。 — Shobhit Varshney > [12:52](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=772s) 不再将 AI Agent 仅仅视作“削减成本”的手段，而是作为“赋能员工创造更多增量价值”的引擎。 — Federico Torreti > [16:36](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=996s) 最佳的实践心智模型是“让数据保留在源头的记录体系（System of Record）中”，通过联邦查询将计算和推理分发到数据所在之处。 — Federico Torreti > [18:54](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=1134s) 从通用的超大模型实验阶段过渡到专精于特定任务的小模型，任务准确率提升了约 5% 到 10%，而推断成本更是骤降了近 95%。 — Shobhit Varshney > [19:58](https://www.youtube.com/watch?v=PY9SGOi9vzM&t=1198s) 企业 AI 的演进不仅关乎“如何构建 Agent（Agent Authoring）”，更在于“如何长效运维与管理 Agent（Agent Management）”。 — Shobhit Varshney ## Mentioned IBM, Oracle, Shobhit Varshney, Federico Torreti, Arvind Krishna, Harris Poll, IBM Consulting, IBM 商业价值研究院 (IBV), watsonx, Watsonx Orchestrate, IBM Granite, OCI Data Science, Oracle Fusion HCM, Outfront Media, Wood, Slack, RPA, Agentic AI ## Video description Shobhit Varshney, Head of Data & AI, IBM Federico Torreti, Senior Director AI/ML, Oracle Abstracts IBM, like everyone, is excited about AI’s potential. Where we differ, though, is that we’re not experimenting with AI – we’re doing. By determining the areas where AI could have the most beneficial impact on our people and our business, we unlocked $3.5B in productivity savings over two years. Now, we will showcase how we did it and how you can also overcome siloed approaches and drive impactful change across HR, Sales, IT, Procurement, and more with AI Agents. In this session, discover: - Our big learnings with AI-driven productivity enhancements. - Practical use cases on how to augment your workforce with domain-specific AI Agents. - How you can build a team of agents in hours and enable multi-agent collaboration in one interface. Don't get left behind, learn how to leverage the power of AI and get ahead with IBM. [Source](https://www.youtube.com/watch?v=PY9SGOi9vzM)

[0:04] Shobhit Varshney: 选择正确的 AI（Doing the right AI）与以正确的方式做 AI（Doing AI right）。 [1:07] Shobhit Varshney: 自左向右路径：首先从核心战略出发，确立市场制胜的北极星指标。 [2:35] Shobhit Varshney: IBM 将自身定位为各项技术方案的“零号客户”（Client Zero），信奉“喝自己的香槟”（Drink our own champagne）。 [2:35] Shobhit Varshney: 两年前，IBM 内部自动化与 AI 的比重约为 50/50；如今，纯自动化占比已降至 25%，而 75% 的降本增效均由 AI 深度驱动。 [4:17] Shobhit Varshney: 打破各部门各自为政的局部优化，围绕客户从市场、销售、定价到签约的“横向端到端旅程”组建跨职能敏捷团队。 [7:28] Shobhit Varshney: 78% 的企业高管高度看好 Agentic AI 对企业业务的重塑能力，76% 的高管清醒地意识到仍需在数据安全、隐私保护以及底层流程重构上进行深度转型。 [9:20] Shobhit Varshney: 这如同组建一支高水平运动队，根据比赛战术灵活引入特定技能的四分卫，实现工作流的敏捷组装与持续替换。 [12:52] Shobhit Varshney: 遵循“AI 追随数据重力”（AI follows data gravity）的原则，在企业防火墙后安全处理核心资产。 [12:52] Federico Torreti: 不再将 AI Agent 仅仅视作“削减成本”的手段，而是作为“赋能员工创造更多增量价值”的引擎。 [16:36] Federico Torreti: 最佳的实践心智模型是“让数据保留在源头的记录体系（System of Record）中”，通过联邦查询将计算和推理分发到数据所在之处。 [18:54] Shobhit Varshney: 从通用的超大模型实验阶段过渡到专精于特定任务的小模型，任务准确率提升了约 5% 到 10%，而推断成本更是骤降了近 95%。 [19:58] Shobhit Varshney: 企业 AI 的演进不仅关乎“如何构建 Agent（Agent Authoring）”，更在于“如何长效运维与管理 Agent（Agent Management）”。

[0:04] Shobhit Varshney: Right-to-left is a recipe for AI experimentation; left-to-right figures out your true north strategy first. [11:22] Shobhit Varshney: Your team of athletes is going to keep changing over time... you're essentially assembling a team of agents that's helping you transform your workflows, just like we would go out and bring in a quarterback with specific skill sets to join the team.

[0:04] I'm Shbat Washnne. I'm the head of data [0:06] and AI for IBM consulting. Um I spend my [0:09] day applying data and AI across a [0:11] variety of different use cases for [0:12] enterprises across each industry. Today [0:15] we're going to do a little bit of deep [0:16] dive into how IBM as a company has been [0:19] leveraging AI and AI agents to transform [0:21] ourselves. [0:23] Within IBM, we have a huge focus on [0:26] doing the right AI and doing AI right. [0:29] is it's a yin-yang on the left hand [0:31] side. We want to not only figure out [0:33] what are the right use cases but also [0:35] things around AI literacy. How do you [0:37] drive AI adoption? How do you ensure the [0:40] right uh set of use cases are delivering [0:42] value quarter after quarter on the right [0:44] hand side you want to ensure that you're [0:46] doing AI right with the right set of [0:48] governance. You want to make sure that [0:49] trust is baked in from the very start [0:52] and the way the organization thinks [0:53] about AI should be fairly aligned to [0:55] your own core values. So this yin-yang [0:58] of unlocking ROI from your AI [1:00] investments and making sure that you're [1:02] taking a responsible approach for [1:04] efficient AI deployed for each use case. [1:07] I recently uh commissioned a study with [1:09] Harris pole on about 2,000 seuitees that [1:12] we've worked with asking them about who [1:15] can attribute actual business value [1:18] created to the atscale AI deployments [1:21] that they have done. About 15% of the [1:24] seauite feel confident and saying [1:26] they're actually delivering insane [1:27] amount of value with their investments. [1:29] We call them the AI leaders. The other [1:32] 85% are what we would call AI learners. [1:35] And there's a very stark difference [1:36] between the way they approach the [1:38] enterprise AI. At the bottom, you're [1:40] looking at the strategy itself. What how [1:43] are you going to win in this [1:44] marketplace? Say you are a mortgage [1:46] company and you're going to do mortgage [1:48] takes 3 weeks to approve. If you do this [1:50] in 24 hours, 48 hours, you should [1:52] meaningfully move the needle for your [1:54] business. If you agree on that true [1:56] north principle, then you start figuring [1:58] out what's the value orchestration [1:59] around it. Do I have enough data exhaust [2:01] that tells me where the value lies? Am I [2:03] able to capture that value versus my [2:04] competitors that impacts workflows and [2:06] then human talent? And then you [2:08] eventually get to the tech and data [2:09] governance. Now, here's the kicker. The [2:12] 15% [2:14] leaders go left to right. But the 85% AI [2:17] learners start with free uh pilots and [2:20] then change the way humans do work and [2:21] then build a business case around it. So [2:23] right to left is recipe for AI [2:25] experimentation. Left to right figuring [2:27] out your true north strategy first and [2:29] then bringing in AI and agents [2:31] surgically to re adjust those workflows. [2:35] I'm going to go deeper into IBM as a [2:38] client example of one of the AI leaders [2:39] who have been able to unlock about $3.5 [2:42] billion of productivity on an annual run [2:45] rate. We have about a $15 billion uh [2:48] expense envelope. This represents three [2:50] and a half billion represents the amount [2:51] we've taken out and all those values are [2:54] shared publicly in our earnings calls [2:55] and so forth. They're backed with actual [2:57] data behind them. So, we pride ourselves [2:59] by calling it client zero. So, we drink [3:02] our own champagne. It's a flywheel that [3:04] we have created of delivering a lot of [3:06] productivity that unlocks investments [3:08] that we can plow back in and then that [3:10] drives further growth for the business [3:12] and ultimately we are unlocking free [3:14] cash flow for our shareholders and [3:16] that's very well represented in the [3:18] growth that our IBM stock has had versus [3:20] our competitors in the last few years. [3:22] So I'll give you a little bit more [3:23] detail around how we're getting to this. [3:26] Irvind set a very audacious goal for us [3:28] to get to two billion dollars of [3:30] productivity as a run rate exiting 2024 [3:33] and we smashed that and we're at about [3:35] three and a half billion plus at this [3:36] point about two billion of three and a [3:38] half billion comes from actual [3:40] productivity through automation and AI [3:42] about 1.5 billion comes from simplifying [3:44] the processes consolidation and things [3:46] of that nature [3:49] a lot has gone into getting us to this [3:50] three and a half billion there are [3:52] things around automating workflows for [3:54] sure but there's also things around how [3:55] do to transform your own talent and the [3:57] culture of figuring out where do we [3:59] apply AI. I'll get into a little bit [4:01] more details into the different areas [4:02] we've deployed this at. [4:06] Two years back we are at about 50/50 [4:08] split of where we are with AI and [4:10] automation. Today we are at about 75% AI [4:13] and 25% is just automating regular [4:16] processes. [4:17] At the bottom are some examples of how [4:19] we've gone domain by domain and looked [4:21] at where we can find value that can be [4:23] unlocked. The first one around here is [4:26] our financial performance management [4:28] system. We have a single pane of glass [4:29] that we use to manage our entire [4:31] business. That means that every decision [4:33] around financials all rolls up into one [4:35] pane of glass that has unlocked $200 [4:38] million of business value. If you start [4:40] going across the board, customer [4:42] support, we're an insanely large [4:43] organization globally. IT modernization [4:47] looking at digital labor for automating [4:49] a lot of the processes and eventually HR [4:51] transformation on the extreme right hand [4:52] side. We are close to 300,000 people [4:55] globally. I'll give you some examples of [4:57] how we've leveraged AI in those [4:58] processes too. [5:01] In order to deliver this at scale, you [5:04] need to have a very disciplined approach [5:06] to ensuring that you have a structure [5:08] that can support a cultural change. [5:11] Transformation using AI requires us to [5:14] think differently. So the way we have [5:16] dealt with this internally, we also have [5:19] a multi-tered oct way of looking at how [5:21] do we support our team members. If you [5:24] look at the very bottom, we started our [5:26] journey by doing external benchmarking [5:28] for each of our domains. What does [5:30] procurement look like for a company of [5:31] our size with a global footprint? What [5:33] does it support? What's the ratio of an [5:35] engineer to support tickets? What should [5:37] be the rate at which we're delivering [5:38] all of this? So we start off by some [5:40] external benchmarking so we know what [5:42] we're heading towards and then we [5:44] started to create these cross functional [5:45] teams. So we got to a point where every [5:48] department financial team was optimizing [5:51] for what they have visibility over but [5:53] our clients engage with us across the [5:55] horizontal from marketing to getting to [5:58] account to actual pricing sending them a [6:01] contract through and through. So that's [6:02] a horizontal. So we had to pivot our [6:04] organization's thinking to go horizontal [6:07] and cut across the cross functional [6:08] teams so we can create the right KPIs [6:10] that we should be tracking. Then we got [6:12] to very quick twoe sprint cycles. We'll [6:14] go try something out if we can validate [6:16] if it's working or not. So this started [6:18] off by defining the playbook by each [6:20] group ensuring that we can demonstrate [6:23] where the opportunity lies. We want to [6:25] make sure that we establishing the why [6:28] we should be doing this and once we [6:29] agree build a team and go go. And we've [6:31] been doing this repetitively over and [6:33] over for each line of business. The [6:35] various layers here you see the the [6:37] structure of our organization from the [6:39] steering committee all the way down. [6:41] Productivity discovery team would be say [6:43] a set of team leaders who understand the [6:45] business and say hey here are the five [6:47] use cases in my team that I think will [6:48] deliver value. That would have gotten us [6:51] to about a billion dollar of unlock [6:52] value. The last layer here is the [6:55] workforce itself. So we had to empower [6:57] our team members to figure out where [6:59] they can apply AI and automation and [7:01] generative AI to automate the workflows [7:03] they're held accountable for. That [7:05] started to unlock a grassroot level [7:08] movement within IBM. We just recently [7:10] concluded our Watson X challenge for the [7:12] year. 170,000 people across the globe [7:15] participating in figuring out where they [7:17] can find value unlock in their [7:19] day-to-day workflows. the that culture [7:21] grassroot level up supported with a [7:23] really good champion backbone is what is [7:25] getting us here. [7:28] All of us have been talking about AI [7:30] agents at scale. In fact, they could [7:31] have probably renamed this conference as [7:33] AI agent for as well. [7:37] There's no doubt that we are seeing a [7:38] lot of value that we are getting through [7:40] applying agents that can transform the [7:42] whole workflows end to end. And I'm [7:44] going to give you a framework of how [7:45] we're thinking about this both [7:46] internally and with our clients as well. [7:49] There's all kinds of studies if you look [7:50] at our own inst institute of business [7:52] value there about 78% of our clients [7:55] that are very positive and uh on the [7:58] fact that agentic AI is going to [8:00] transform their businesses [8:03] 76% still think that they need to max to [8:05] achieve maximum value they still need to [8:07] do a lot more of transformation inside [8:08] of their companies themselves there's a [8:10] lot of issues around making sure that [8:12] data security and privacy and stuff like [8:13] that is is maintained so this is the [8:15] current state of what we found with [8:17] about 800 suites that we discussed the [8:20] specifically about agentic AI the the [8:22] latest wave of AI [8:25] I'm going to make a statement around how [8:26] do we get to incremental value in the [8:29] processes that we've been working with [8:31] there are four levels at which we have [8:33] these conversations uh around processes [8:35] if you look at a simple process like [8:37] submitting a claim we've been doing this [8:39] for a decade plus if you go to an [8:41] insurance provider for auto insurance [8:43] and say I want to automate your process [8:45] of doing claims you can ask a subject [8:48] matter are expert they'll draw you on a [8:49] whiteboard how this is done you take [8:52] that and say there's a smarter way of [8:53] doing it I'm going to codify it I'm [8:54] going to automate it whatever you can [8:56] define on a whiteboard we can obviously [8:58] automate this is the robotic process [9:00] automation era we went through a [9:02] structured flow somebody is defining the [9:04] decision tree saying if this then go do [9:06] this do this all of this is very [9:09] inexpensive very fast it delivers [9:12] consistent responses because a human is [9:14] defining every step along the way then [9:16] we looked at opportunities ities here at [9:18] step number two where we are reading a [9:20] document. We said aha we have large [9:22] language models that can help us. So we [9:24] did one step one level up and said for a [9:27] particular automation flow. Step number [9:29] two I can switch it out to a large [9:31] language model that can visually read a [9:33] particular document and can go do some [9:36] very quick analysis with it. [9:39] From here we got to level three which is [9:41] the inflection point for AI agents. And [9:43] here AI agents are taking a small task [9:46] and independently they're executing on [9:48] it. So it's just like getting a new hire [9:51] or or new intern. And if I ask that [9:53] person, can you go compare our earnings [9:55] with a competitors? She's going to [9:57] independently open up a browser, look at [9:59] my competitors, download them, create a [10:01] table, look at how I did it last time, [10:02] and come back to me when she's satisfied [10:04] that she has an answer. That's [10:06] essentially what an AI agent is doing [10:07] today. But Shut is still doing the ask [10:10] to task in my head. So I get the ask [10:12] from my CEO. I break it down into small [10:14] tasks and I delegate those tasks. When I [10:16] get a response back, I verify and I move [10:18] on. So the ask to start still being done [10:20] by a human, but the small token of work [10:22] task is being done by an AI agent. Once [10:25] you start to create a repository of [10:26] these asks to task, you can then start [10:28] training a large model that takes the [10:31] overall ask and can go delegate it down [10:33] into smaller task. And now you're [10:34] starting to orchestrate agents amongst [10:36] themselves. So now you can start to [10:38] answer complex questions around hey why [10:40] is my claim different from last time [10:42] right that requires multiple things [10:43] multiple systems to get into and now [10:45] these multi- aent frameworks are going [10:47] to answer this in reality we need to [10:51] understand which workflow which specific [10:54] task should go towards a deterministic [10:56] one or really go dial this up and make [10:58] it multi- aent multi- aent will help you [11:00] do a lot more complex things but it [11:02] comes with its own set of uh things [11:04] around hey how do I know that the answer [11:06] is correct Correct. There's a lot more [11:08] governance that's needed. It takes it's [11:10] more expensive. It takes more latency. [11:11] The extreme left hand side I have to [11:12] define every step around how to get to [11:14] the answer but it's very deterministic. [11:16] So in reality it's a combination of [11:18] these fixed and agetic workflows that [11:20] deliver value for us. The way we're [11:22] doing this internally within IBM we have [11:25] a product called Watsonx orchestrate [11:27] which is the orchestration of [11:28] intelligence across all the agent take [11:30] and fixed workflows. And these could [11:32] come from a variety of different [11:34] partners. It could be our own AI. It [11:36] could be our our ecosystem partners AI. [11:38] But we are orchestrating all these [11:40] agents with the right set of tools. [11:43] One level down if you look at the actual [11:45] details around it. You can have multiple [11:47] personas of people coming in. Each one [11:49] of them could come through a different [11:50] channel. So you could be on Slack, you [11:52] could be calling in, but eventually the [11:54] orchestrator agent needs to figure out [11:56] how do I delegate this particular work [11:58] to which sort of a AI agent? Either [12:00] that's from IBM or from one of our [12:02] ecosystem partners. [12:04] So this is an example of how it would be [12:06] if you ask a question around can I help [12:09] onboard this new company. You take that [12:12] ask you understand and you determine [12:14] that here's the particular path I need [12:16] to do. There are three steps here. I [12:18] need to make sure I can go profile this [12:20] company that I'm onboarding. I need to [12:22] have an onboarding agent that could come [12:24] from one of our ecosystem partners. You [12:26] could have another legal adviser that's [12:28] from an IBM agent AI. But eventually you [12:30] get to a point where it is very modeler [12:33] and your best your team of athletes is [12:35] going to keep changing over time and [12:36] each athlete will have a specific uh [12:38] role that they are playing and you're [12:40] essentially assembling a team of agents [12:42] that's helping you transform your [12:43] workflows just like we would go out and [12:45] bring in a quarterback with specific [12:47] skill sets to join the team to deliver [12:49] that particular work really well. [12:52] Great segue to talk about our our [12:54] ecosystem partners like IBM as a [12:56] company. We've we've taken this this [12:57] mission to ensure that we are leveraging [13:00] a large ecosystem of our partners to [13:02] build together with them. So the clients [13:04] have an opportunity to go swap things in [13:06] and out. They can keep the data where [13:08] where they have confidence and behind [13:10] the firewalls. AI would would follow the [13:12] data gravity. One such of our partners [13:15] is is here with us today. Oracle. I [13:17] would love to get Fetty on stage to come [13:19] join me. [13:23] Thank you so thanks so much for for [13:25] joining us. So with IBM and uh and [13:27] Oracle both on the software side as well [13:29] as consulting side we've had a really [13:31] deep partnership. Uh all of our software [13:33] uh in AI like Watsonx orchestrate.ai and [13:36] LLMs are available natively on Oracle as [13:39] well as our IBM consulting team. We have [13:40] an insanely large partnership that goes [13:42] back decades. We've done a lot of really [13:44] good work. A lot of assets built with [13:46] Oracle. We also happen to have some [13:48] incredible AI talent in our team. seems [13:49] we've been working really really well. [13:51] So Freddy, can we just start there and [13:52] just give a give a couple examples of [13:53] how we're partnering where we are doing [13:56] well and it'll really help for the [13:57] audience to hear some client examples of [13:59] where we are doing this well. [14:00] >> Yeah, of course. Of course. Um the I [14:04] think you covered very well during the [14:06] the first part of the of the [14:08] presentation. A key component of [14:10] actually delivering ROI into enterprise [14:13] engagement is really this combination [14:15] between the IBM's consulting expertise [14:18] with Oracle's infrastructure and and [14:20] that has been extremely important as as [14:23] we look at 2024 and into 2025 as many [14:26] companies migrated away from pilot stage [14:29] into um looking to move some of these [14:33] experiments into those 15% leading AI [14:36] companies that are trying to move into [14:38] production And a large part of that has [14:40] really built on decades like you [14:42] mentioned of experience together. We've [14:43] seen not just tech companies embrace AI [14:46] agents. We've seen a very wide range of [14:48] organizations from traditional brick and [14:51] mortar mort and construction uh in the [14:54] construction business. They've looked at [14:56] Oracle's technologies to um understand [14:59] how can we use automations and AI um [15:03] agents for um accelerating the time to [15:07] bid for their processes. Uh Outfront [15:10] Media is another company that comes to [15:11] mind and Outfront Media has been [15:13] particularly uh effective in looking at [15:16] their sales representative um uh [15:18] processes. So it's less about an [15:20] efficiency plane. It's more about [15:21] enhancing customer engagement uh [15:24] specifically around how do we provide [15:26] the richer information at point of views [15:29] to the sales representative that is [15:31] engaging with uh with an with with a [15:33] prospect. And then we have a services [15:36] business um would um which is an [15:39] engineering consulting business. they [15:41] deployed our fusion uh HCM agents uh um [15:44] to accelerate the um the IR to or the [15:48] source to hire process and so they're [15:50] about 56% productivity gains coming from [15:54] looking at inefficiencies in the process [15:56] and one final thing I would say to build [15:58] on what you mentioned earlier is that [16:02] when you look at the segmentation of the [16:04] 15 and 85% of companies uh we're seeing [16:07] companies that are moving away of [16:09] looking at AI agents as a cost [16:11] efficiency play and rather looking at [16:15] how can they enable their workforce to [16:19] do more and that's that's a complete [16:22] shift in mindset as you approach problem [16:24] because you're no longer looking at how [16:26] do I take cost out of a process and [16:28] instead you're like trying to think okay [16:30] what else could I do what are the [16:32] problems that I haven't been able to [16:34] tackle up until now [16:36] >> so let me ask you a question around uh [16:38] where the data resides you made a really [16:39] good point about the about companies who [16:42] are not tech forward quite yet. A lot of [16:44] them have data on prem, they have data [16:46] all over the place. How do you think [16:48] about where do we build intelligence and [16:50] uh if you were to believe that [16:51] intelligence is your IP? Where would you [16:53] want the teams or any advice to [16:55] companies on thinking about hey my data [16:56] is still on prem do I really need to [16:57] move things out or how do I do this [16:59] balance of uh of intelligence? Uh I [17:02] think this is where um what's on x [17:04] orchestrated comes uh uh very effective [17:07] because u the mental model is that the [17:11] you you want to try and avoid and [17:13] minimize data movement. [17:15] >> Um and the way I would approach it and [17:19] the way our customers are looking to [17:20] approach this problem is how do I retain [17:22] the data the source um and the system of [17:25] records and so having this ability to [17:27] federate out queries is extremely [17:30] important. [17:30] >> Okay. U let me let's get a little bit [17:33] deeper into how we think about AI [17:35] models. There's been a plethora of [17:38] models and every week there's each each [17:40] one of them is out competing each other. [17:41] So the number of choices available to [17:43] our clients are are exponentially [17:44] increasing every day. Can you talk a [17:47] little bit about how important uh is it [17:49] to have a multimodel strategy or where [17:52] how do you how does organization think [17:54] about uh partnerships and building [17:56] intelligence? [17:58] Companies tend to look at this in one or [18:00] two ways. There is one approach which is [18:03] here is the broadest possible selection [18:05] of models and what we found is that that [18:07] typically leads to decision fatigue. [18:09] Totally [18:09] >> agree. [18:10] >> What I think we find that's most [18:11] effective is really working backwards [18:13] from a curated choice of models that's [18:16] optimized for specific use cases. Again, [18:19] this is where uh IBM Granite is is a [18:21] great example of IBM's own models that [18:24] are available on OCI data science and [18:27] and and offers the auditability and the [18:30] compliance that many of our enterprises [18:32] are looking for. Those are factors that [18:35] are much more relevant in an [18:37] enterprisegrade deployment than being at [18:40] the frontier. I would agree with that. [18:42] What we are seeing with our clients as [18:43] we deliver with Watson and Oracle and [18:46] other ecosystem partners, we see a lot [18:48] of experimentation that happens with the [18:50] frontier models first to say that hey is [18:52] this use case even valid but eventually [18:54] when you start to deploy these at scale [18:56] you have this you have other [18:58] considerations like cost envelope [19:00] latency on how quickly I get a response [19:01] back and uh the model if it if the model [19:04] is trying to pull from data it know [19:06] nodes about the entire world or is it [19:08] been fine-tuned for a particular domain [19:10] so we've seen a lot of success with [19:12] taking our granite models and other open [19:14] source models and fine-tuning them for a [19:16] particular use case that reduces the [19:18] scope of what they're doing. But the ROI [19:20] you get you get for our clients we're [19:21] seeing about maybe a 5 10% increase in [19:23] the accuracy but like a 95% decrease in [19:25] the cost to run these if you're moving [19:28] from experimenting with large models and [19:30] eventually getting to smaller models [19:31] that do some things really well. And [19:33] this seems to be accelerating with AI [19:36] agents because if you look at the whole [19:38] hierarchy of the agents that I was [19:39] explaining before, you would have one [19:41] agent on the top that is managing all [19:43] these different domains. But as you get [19:45] down further into a particular domain, [19:47] there's a procurement agent that does [19:49] only procurement and that has to be a [19:51] very costefficient model. So we're [19:54] seeing a lot more smaller models deliver [19:56] higher accuracy with the with a take [19:58] flow. The only one thing I would add to [19:59] that is that agent authoring. I think we [20:02] talked a lot about agent authoring but [20:03] one of the other aspect that is becoming [20:05] more and more prominent is um the [20:08] management of agents and I think again [20:10] this is another important another [20:13] important aspect of the office. This [20:14] >> is extremely uh helpful. Thanks for the [20:16] discussion and thank you so much for the [20:17] partnership as always. [20:18] >> Thank you.