IDC's Jim Spare Says 87.5% of Enterprise AI Pilots Fail to Yield ROI

IDC 的 Jim Spare 表示 87.5% 的企业 AI 试点未能产生投资回报 现成的 AI 模型无法创造长期价值，企业需构建嵌入式且经人类验证的智能体工作流，以将专有数据转化为具备复利效应的竞争壁垒。 Jim Spare 探讨了广泛的 AI 实验与企业实际实现投资回报（ROI）之间的脱节，指出尽管 75% 的企业在使用 AI，但仅有 12.5% 实现了规模化的 ROI。他将通过调用通用 LLM API 的“租用智能”，与通过嵌入式自主智能体及专有人机协同（human-in-the-loop）反馈闭环构建的“自建智能”进行了对比。结合 IDC、医疗、消费品和土木工程等领域的案例，Spare 阐释了包含“连接数据、呈现洞察、验证与学习、执行行动”的四阶段框架，旨在大幅加速工作流并打造持久的竞争护城河。 尽管 75% 的企业已开始使用 AI，但仅有 25% 成功进入生产阶段，最终只有 12.5% 能够实现规模化的投资回报率（ROI）。 (0:04) 直接调用开箱即用的通用大模型本质上属于“租用智能”，由于无法沉淀专属能力，企业难以建立竞争壁垒，而 AI 原生企业正通过将自主智能体嵌入业务流程来“构建智能”。 (0:38) IDC 与 A.Team 合作在 8 周内完成原型开发，将 60 年的市场数据资产整合进采购流程，使软件选型周期从 6 个月缩短至 6 周，评估阶段从 5 周缩短至数小时。 (5:27) 嵌入式智能体架构依靠连接分散数据、主动挖掘洞察、结合专家与用户的 RLHF 人机协作验证，并在业务工作流中直接触发自动化执行动作。 (10:05) 多场景落地案例显示，定制化嵌入智能体已助力医疗机构提升护士 40% 的生产力，并将工程咨询公司的环境评估报告耗时减少 70%，每年直接节省 600 万美元成本。 (13:13)

Jim Spare argues that most enterprise AI initiatives stall because organizations merely 'rent intelligence' through generic APIs rather than embedding domain-specific agents into core operational workflows to compound proprietary data. The AI Adoption Gap and Enterprise Blockers (0:04) While interest in artificial intelligence remains high, Spare cites industry benchmarks from McKinsey and BCG showing a steep drop-off between initial experimentation and realized business value: 75% of companies report using AI, but only 25% transition projects into production, and merely 12.5% demonstrably prove ROI at scale. According to Spare, enterprise progression stalls against four primary 'value blockers': a persistent shortage of internal AI expertise, difficulties in evaluating hallucinations to build trust, fragmented and siloed legacy data foundations, and an accelerating pace of model updates that outpaces standard corporate decision cycles. Renting Versus Building Organizational Intelligence (0:38) Spare contrasts the prevailing enterprise reliance on off-the-shelf LLM APIs—termed 'renting intelligence'—with 'building intelligence.' He asserts that querying generic foundation models generates disposable outputs where compounding learning accrues exclusively to the model vendor, failing to produce a proprietary moat for the enterprise. In contrast, an AI-native approach embeds autonomous agents directly into operational processes using a four-stage framework: connecting fragmented data, proactively surfacing insights, collecting human validation, and executing downstream workflow actions to ensure continuous organizational learning. Human-in-the-Loop Validation in High-Stakes Workflows (5:27) Drawing on a case study from market research firm IDC, Spare illustrates how unlocking 60 years of unstructured proprietary data through an eight-week prototype led to an AI-native IT procurement platform. To maintain the precision needed for high-stakes enterprise decisions, the platform integrated a human-in-the-loop feedback architecture, using evaluations from analysts and end users to continuously refine system performance. Rather than presenting a blank chat interface, the application mirrored the end-to-end procurement workflow. This structured approach compressed overall software selection cycles from six months down to six weeks and reduced the data-intensive evaluation phase from five weeks to a few hours. Quantifiable Returns Across Embedded Domain Workflows (13:13) Spare points to cross-industry deployments demonstrating that integrating agents into existing operational tasks delivers significant productivity gains. In healthcare, embedding compliance agents into nurse workflows to automate insurance claim documentation generated a 40% productivity increase following a four-month MVP deployment. Similar implementations yielded major time and cost efficiencies across other sectors: a global consumer goods firm reduced regional sales analysis turnaround from days to hours, while a civil engineering firm lowered Environmental Site Assessment drafting time by 70%, capturing $6 million in direct annual cost savings. Spare concludes that sustainable AI performance requires organizations to treat AI as a core capability to own rather than a tool to rent. To establish clear alignment, he advises enterprise leaders to identify the precise human process being augmented, pinpoint the proprietary data advantage, and define measurable outcomes within a 90-day window.

A.Team执行顾问兼IDC前首席产品官Jim Spare指出，多数企业受困于人才与数据壁垒，导致AI实验难以转化为可规模化的投资回报；企业若要突围，必须从调用通用API的“租用智能”转向把自主智能体深度融入业务流的“构建智能”，通过私有数据与人机协同反馈构筑复合竞争壁垒。 AI落地的价值转化断层与核心阻碍 (0:04) 尽管企业普遍开启了AI技术探索，但从实验走向实际业务价值的过程中存在显著断层。Spare援引麦肯锡与波士顿咨询（BCG）的市场数据指出，约有75%的企业报告以某种形式使用AI，然而仅有25%（约占使用者的三分之一）能从原型测试推进至生产环境，最终仅有12.5%（约占进入生产阶段企业的一半）能够在大规模部署中切实证明可衡量的商业投资回报率（ROI）。 这种价值耗损源于企业普遍面临的四大核心阻碍：首先是内部极度缺乏能够设计并实施专业AI架构的复合型人才；其次是对模型准确性评估困难以及模型幻觉问题，使得建立正式的投资论证与信任机制充满挑战；第三是企业内部四处分散的数据孤岛阻碍了动态AI工作流所需的数据整合；最后则是AI技术的超快迭代速度与企业漫长的决策周期之间存在剧烈脱节。 “租用智能”与“构建智能”的战略差异 (0:38) 面对落地阻碍，许多企业倾向于依赖现成的通用工具，但Spare指出这往往只能获得平庸的结果。他将直接调用通用大语言模型（LLM）API的做法比作“租用智能”。在这种模式下，企业花费在API调用上的资金仅能获得短暂且一次性的回答，系统性价值无法在企业内部沉淀，反而是模型提供商在交互中不断积累复合学习，企业自身未能建立任何技术护城河。 与之相对，AI原生企业主张“构建智能”，其核心是将自主智能体（Autonomous Agents）直接嵌入到企业的实际业务工作流中。通过这种方式，企业的日常业务操作能够转化为可持续积累的专有反馈闭环，推动内部数据资产与工作流知识在本地持续增值，进而建立起难以被复制的长期竞争壁垒。 IDC转型实践与人机协同闭环 (5:27) 在IDC的转型案例中，该市场研究机构拥有60多年积累在PDF、分析师笔记和数十个分散数据库中的专有资产。面对四大阻碍，IDC与A.Team合作在8周内构建了一个可运行的原型，使用户能够基于IDC专属数据获得精准回答。基于此，IDC开辟了新产品开发、核心研究报告流程加速与内部职能工作流优化三大战略赛道，并重点推出了一款AI原生SaaS产品，端到端辅助企业IT高管的软件采购流程。该产品将软件选型周期从6个月缩短至6周，并将数据密集型评估阶段从5周压缩至数小时。 为了保障ERP选型等高风险决策的准确性，IDC部署了“人机协同（Human-in-the-Loop）”反馈架构。分析师首先根据专家结论评估智能体生成的建议，随后系统在工作流内直接收集终端用户的交互反馈。这套机制通过人类反馈强化学习（RLHF），使企业的专有智能层在每一次实际交互中持续演进优化。 嵌入式智能体框架与跨行业落地效益 (13:13) A.Team总结的嵌入式智能体系统涵盖四阶段框架：打通分散的数据孤岛（Connect Data）、驱动智能体主动发现模式（Surface Insights）、依托人工验证持续强化系统学习（Validate & Learn），并在标准运营流程内触发下游业务动作（Execute Action）。Spare强调，企业推进AI原生化需遵循四项原则：视专有数据为根本护城河、通过无缝工作流集成驱动ROI、以周为单位快速原型迭代，以及重视实战部署而非停留在演示幻灯片。 这一模式在多个行业取得了具体效益：在医疗保健领域，定制智能体协助护士在HIPAA合规约束下自动化生成保险理赔记录，MVP开发仅耗时4个月，将护理生产力提升了40%；在跨国消费品企业中，智能体嵌入财务与营销流实时分析区域销售下滑原因，将报表周期从数天缩短至数小时；在土木工程公司中，智能体协助解析环境法规与地块档案并起草环境评估报告草案，单份报告耗时减少70%，每年直接节省成本600万美元。 Spare总结建议，企业决策者在启动AI项目前必须明确三项核心问题：该项目具体增强了哪项人类工作流程、企业拥有哪些赋予自身优势的专有数据，以及如何在90天内明确衡量成功。领先企业必须将AI视为自身必须构建并掌控的核心组织能力，而非向外部租赁的通用外购软件。

## Background 本视频由 Ai4 于 2025 年 8 月 29 日在 YouTube 平台发布。视频内容为前 IDC 首席产品官、现 A.Team 执行顾问 Jim Spare 的主题演讲，主要探讨企业如何告别单纯“租用智能”的工具调用模式，通过将自主 Agent 深度嵌入业务工作流来构建具备自我进化能力的专有智能系统，从而跨越 AI 原型到高 ROI 生产环境的落地鸿沟。 ## TL;DR Jim Spare 剖析了约 75% 的企业 AI 项目停留在原型阶段或无法实现可衡量 ROI 的核心原因，指出仅调用通用大模型属于将数据与学习价值流失给外部厂商的“租用智能”。他结合在 IDC 将 60 年沉淀数据打造为软件选型 Agent 产品的亲历经验，以及在医疗、快消、土木工程等领域的实战案例，拆解了结合 Human-in-the-Loop 机制的“从洞察到行动”（Insight-to-Action）飞轮模型。演讲强调，AI 原生企业的成功关键在于以专有数据为壁垒、与核心工作流紧密集成、以周为单位快速迭代，并最终建立起随着业务行动持续复利的组织专有智能资产。 ## Atmosphere & style 整场演讲风格沉稳务实、直击痛点，开场的现场互动调研迅速打破了行业对 AI 繁荣的表面幻象，营造出极具说服力的求真氛围。作为拥有丰富企业管理与创业经验的老将，Jim 的叙述条理清晰且极具实战导向，既有高屋建瓴的战略模型提炼，又有细致到具体行业收益的量化数据支持。现场节奏紧凑明快，充满了一线构建者的敏锐洞察与务实笃定，能够让身处 AI 转型迷茫期的业务与技术决策者产生强烈的共鸣与行动紧迫感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=2utRWU6AtBI&t=4s) 行业基准调研显示，约 75% 的企业正在尝试使用 AI，但仅有 1/3 能将项目推进至生产环境，最终仅有约半数项目能证明明确的商业 ROI。 - [02:33](https://www.youtube.com/watch?v=2utRWU6AtBI&t=153s) 企业实现 AI 价值面临四大核心阻碍：AI 专业人才匮乏、输出信任与量化评估困难、底层数据基础设施分散孤立，以及决策流程难以跟上技术极速迭代。 - [02:33](https://www.youtube.com/watch?v=2utRWU6AtBI&t=153s) 采购现成通用工具调用 API 属于“租用智能”，数据与学习复利全部回流给模型厂商；真正高 ROI 的企业专注于“构建智能”，将专有智能沉淀为自身竞争壁垒。 - [05:27](https://www.youtube.com/watch?v=2utRWU6AtBI&t=327s) IDC 实战案例表明，将数十年专有数据与具体采购工作流结合构建 Agent 系统，可将长达 6 个月的重大选型流程压缩至 6 周，供应商深度评估耗时从 5 周缩短至数小时。 - [10:05](https://www.youtube.com/watch?v=2utRWU6AtBI&t=605s) 在低容错商业决策中，人机协同（Human-in-the-Loop）架构不可或缺，结合领域专家与终端用户的持续反馈，能借助强化学习机制形成自我进化的学习闭环。 - [10:05](https://www.youtube.com/watch?v=2utRWU6AtBI&t=605s) “洞察到行动”（Insight-to-Action）框架推动系统从被动的单次问答交互，转变为主动整合全域数据、提炼商业洞察并直接驱动业务执行的智能伙伴。 - [13:13](https://www.youtube.com/watch?v=2utRWU6AtBI&t=793s) AI 原生企业的四大核心支柱：独家数据是核心差异化壁垒、深度集成业务流程驱动 ROI、敏捷迭代以周为单位快速推进、崇尚实战构建而非方案纸上谈兵。 - [13:13](https://www.youtube.com/watch?v=2utRWU6AtBI&t=793s) 企业在 AI 项目立项初期应明确回答三问：正在增强哪项人类业务流程、依托什么专有数据建立优势，以及如何在 90 天内衡量项目成功。 - [15:00](https://www.youtube.com/watch?v=2utRWU6AtBI&t=900s) 跨行业实战用例验证了 Agentic AI 的价值：医疗领域提升护士文档效率 40%，快消领域将数据分析从数天缩短至数小时，土木工程领域缩短 70% 报告生成时间并年省 600 万美元。 - [17:33](https://www.youtube.com/watch?v=2utRWU6AtBI&t=1053s) 企业在 AI 时代制胜的关键在于认知转变：AI 不仅是一个拿来使用的外部工具，而是企业必须主动构建并实现复合增长的核心能力。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=2utRWU6AtBI&t=4s) 讲者背景与现场调研：企业应用 AI 的现状与真实 ROI 困境 大会开场，Jim Spare 进行了简短的自我介绍。他在职业生涯中一直致力于推动颠覆性技术业务的增长，经历涵盖大型成熟企业与独立风投支持的初创企业。他曾先后任职于微软（Microsoft）、通用仪器（General Instrument）以及全球知名的市场研究机构 IDC，并担任过 IDC 的首席产品官（Chief Product Officer）。在他创立和参与的风险投资公司中，有成功也有失败，其中一家被微软收购，另一家则与沃尔玛（Walmart）合并成立了大型合资企业。目前，Jim 担任 A.Team 的执行顾问。A.Team 是一家专注于为企业将 Agentic AI（智能体系统）嵌入实际业务工作流的 AI 原生系统集成商。基于在第一线观察到的大量成功经验与落地痛点，Jim 希望在此分享关于企业如何真正实现 AI 价值的洞见。 为了解现场参会者的真实进展，Jim 向观众提出了三个层层递进的问题。首先，现场绝大多数人都举手表示所在企业目前正以某种形式使用 AI；然而，当问到有多少人将实验推进到了“超越原型、进入生产环境”阶段时，举手的人数明显减少；最后，当问及有多少人认为自己已经在大规模生产中“成功证明了投资回报率（ROI）”时，现场只剩下寥寥几人。 Jim 指出，现场的举手情况与麦肯锡（McKinsey）和波士顿咨询（BCG）等机构的行业调研基准高度吻合。行业数据显示，约有四分之三的企业正在探索或使用 AI，但这其中仅有三分之一能将原型推进至实际生产环境；而在进入生产环境的项目中，又只有半数能够真正实现可衡量的业务 ROI。这说明绝大多数企业仍处于 AI 转型的早期阶段，普遍面临着难以获取实质商业价值的挑战。 - 行业基准现状：约 75% 的企业在使用 AI，仅 1/3 推进至生产环境，其中仅 50% 实现了清晰的业务 ROI。 - 演讲核心主旨：剖析 AI 原生企业的核心差异，讲解如何构建能够“随着每次行动不断自我进化”的智能体架构，并将 AI 沉淀为组织的核心能力。 --- ### [02:33](https://www.youtube.com/watch?v=2utRWU6AtBI&t=153s) 企业的四大价值阻碍与“租用智能”的认知陷阱 企业在尝试构建 AI 系统时，通常会遭遇四大核心价值阻碍（Value Blockers）。第一是 AI 专业人才的匮乏（Elusive Expertise），多数企业缺乏实现其宏大 AI 目标的技术储备；第二是 AI 的信任与评估难题（Trust & Evaluation），除了众所周知的模型幻觉问题外，企业很难对 AI 系统的输出进行科学量化评估，无法评估就极难建立起充分的商业投资论证；第三是底层数据基础设施薄弱（Data Foundations），绝大部分企业的数据资产处于分散、孤立状态，尚未做好接入 AI 的准备；第四则是技术迭代节奏极快（Rapid Pace），正如业内常调侃的“距离上一次 AI 重磅发布才过去零天”，而传统的企业决策流程根本无法跟上这种演进速度。 受制于这些阻碍，绝大多数企业的应对方式是直接采购市面上现成的通用工具，并试图将它们硬塞进现有的业务流程中。然而，这种做法的直接后果是“通用的工具只能产生通用的平庸结果”。Jim 强调，这种模式在本质上属于“租用智能”（Renting Intelligence）而非“构建智能”（Building Intelligence）。 当企业仅仅通过 API 调用基础大模型时，实质上只是在租用模型厂商的算力与智力。企业投入的每一笔费用，其产生的学习反馈和数据复利全部回流给了模型供应商，自身并没有沉淀下任何随时间复合增长的专有能力。相比之下，真正实现高 ROI 的领先企业从一开始就专注于“构建智能”：将自主 Agent 深度嵌入到具体的工作流中，让系统在自身专有业务语境中不断学习，使得 AI 随着每次业务操作变得更加聪明，最终将专有智能沉淀为坚不可摧的竞争护城河。 ```mermaid flowchart TD accTitle: Renting Intelligence vs Building Intelligence subgraph Renting["租用智能 (Renting Intelligence)"] R1["单次 API 调用 / 通用工具"] --> R2["生成通用结果"] R2 --> R3["数据与学习价值流失给外部厂商"] end subgraph Building["构建智能 (Building Intelligence)"] B1["专有数据与工作流"] --> B2["嵌入式 Agent 智能体系统"] B2 --> B3["业务行动与人机反馈"] B3 --> B4["专有智能持续复利，形成竞争壁垒"] B4 --> B2 end ``` - 阻碍价值变现的四大难题：人才短缺、信任与评估困难、数据底座薄弱、技术演进过快。 - 租用智能的弊端：企业付出的成本无法形成学习闭环，技术资产与经验积累全部归属于基础模型供应商。 - 智能体系统的转变路径：从被动的提示词问答转向主动检测业务信号、记住行为模式、深度融入现有流程并实现持续改进。 --- ### [05:27](https://www.youtube.com/watch?v=2utRWU6AtBI&t=327s) IDC 实操案例：将 60 年沉淀数据转化为全新 AI 产品 为了直观呈现这一转变，Jim 分享了自己在 IDC 担任首席产品官期间亲历的实战案例。作为一家拥有 60 年历史的全球市场情报巨头，IDC 积累了海量关于科技市场、产品、服务以及客户落地案例的专有数据。然而，当管理团队希望拥抱 AI 时，他们同样面临着上述全部四大阻碍：缺乏 AI 专长、必须快速行动、数据高度分散在几十个内部数据库、PDF 文档和分析师笔记中。 IDC 选择与 A.Team 合作，在短短 8 周内开发出了一个工作原型（Prototype）。该原型接入了 IDC 的多源异构数据，使用户（包括内部员工和外部客户）能够直接对该专有数据库进行自然语言查询，并获得基于独家洞察的高质量智能回复。这个原型的成功不仅验证了可行性，更为团队指明了后续的三条业务发展跑道（Lanes）： 1. **开发全新产品**：打造全新的 AI 原生产品并直接对外商业化销售。 2. **赋能核心业务**：加速分析师团队撰写市场调研报告与交付数据的全流程。 3. **优化内部职能**：在财务、人力资源、销售等职能部门中落地 AI 自动化。 IDC 决定首先聚焦于最具商业价值的第一条跑道——打造一款全新的数字化产品，解决 IT 高管在软件采购中的决策痛点。过去，企业选型一套核心业务软件（如 ERP）是一项重度依赖人工咨询、流程极度漫长的业务，涉及定义目标、了解品类供应商、明确技术需求、筛选候选名单、最终定选与商务谈判等标准环节。IDC 没有做一个空洞的聊天框界面，而是严格按照该采购工作流构建了一个 Agent 辅助系统，大幅降低了用户的学习门槛并最大化了系统采纳率。 产品上线后获得了早期客户的极高评价。客户反馈显示，原先耗时可能长达 6 个月的重大软件采购选型流程，通过该平台有望缩短至 6 周；其中最耗费人力与数据分析时间的“供应商深度评估”阶段，直接从原先的 5 周大幅压缩至数小时以内。 - 原型开发周期：借助 A.Team 的工程能力，仅用 8 周时间即完成多源数据接入原型的构建。 - 业务效率提升：软件采购选型全周期从 6 个月缩短至 6 周，核心评估分析阶段由 5 周压缩至数小时。 - 成功关键：将 AI 与已有的成熟业务流程精密结合，依托独家数据解决高价值业务痛点。 --- ### [10:05](https://www.youtube.com/watch?v=2utRWU6AtBI&t=605s) Human-in-the-Loop 架构与“洞察到行动”框架 Jim 进一步拆解了该系统的架构设计逻辑。在最高层面上，系统的运作分为两步：第一，理解并结构化处理海量的非结构化与半结构化数据；第二，提炼出可操作的商业洞察，驱动符合业务目标的实际行动。然而，像更换企业 ERP 系统这种耗资数百万美元且影响未来多年的重大决策，客户对准确性的容错率极低，系统必须具备绝对的公信力与品牌信任度。 为了保证输出质量，系统在架构上深度引入了“人机协同”（Human-in-the-Loop, HITL）机制。最初，系统生成的推荐由 IDC 资深分析师进行审核，确认“基于这些数据，我是否也会给出相同的建议”；随后，这种反馈机制延伸到了实际使用的终端用户，用户在工作流的各个关键节点与最终结果上给予即时反馈。这种机制将强化学习（RL）深度融入系统运行中，不仅保障了严谨度，还构建起了一个不断学习的反馈闭环，持续积累专有智能。 基于此类实践，A.Team 总结出了一套普适性的系统构建框架： ```mermaid flowchart LR accTitle: Insight-to-Action Framework subgraph DataLayer["1. 数据连接"] D1["结构化数据"] D2["半/非结构化数据"] D3["行业专属语料"] end subgraph AgentLayer["2. 智能提炼"] A1["Agent 整合与分析"] A2["生成业务洞察"] end subgraph HumanLayer["3. 人机闭环"] H1["人类专家校验"] H2["用户行为反馈"] end subgraph ActionLayer["4. 业务行动与复利"] ACT["执行业务动作"] COMP["专有智能持续复利"] end DataLayer --> AgentLayer AgentLayer --> HumanLayer HumanLayer -->|"优化反馈"| AgentLayer HumanLayer --> ActionLayer ActionLayer -->|"强化学习循环"| AgentLayer ``` - **全域数据打通**：将结构化、半结构化与非结构化数据清洗并统一暴露给智能体。 - **自主提炼洞察**：AI 持续分析全量信息并主动提炼出商业洞察。 - **专家参与校验**：结合人类专家与终端用户的反馈，校准并强化模型表现。 - **驱动业务落地**：系统或业务人员依据洞察采取行动，随着行动次数的增加，专有智能持续复利。 - **主动型工作流协作**：摒弃“提问-等待单次回复”的被动交互模式，让 Agent 成为工作流中主动提供决策支持与内容生成的业务伙伴。 --- ### [13:13](https://www.youtube.com/watch?v=2utRWU6AtBI&t=793s) AI 原生企业的四大核心支柱与启动三问 要在激烈的技术变革中真正成为一家“AI 原生”企业，组织需要牢牢确立四大核心支柱： 1. **数据是核心差异化壁垒（Data is your differentiator）**：通用的基础模型人人可用，唯有企业沉淀的独家专有数据与独特工作流逻辑，才是不可复制的竞争护城河。 2. **深度集成驱动真正 ROI（Integration drives ROI）**：AI 智能体系统与现有核心业务工作流整合得越紧密，产生实际投资回报的速度就越快。 3. **敏捷迭代以速度致胜（Speed wins）**：在 AI 时代，最小可行性产品（MVP）与功能迭代的周期应当以“周”为单位计算，而不是以“月”为单位。 4. **重视实战构建而非纸上谈兵（Builders over slides）**：真正成功的公司不会停留在方案汇报和 PPT 上，而是迅速将可用系统交到一线用户手中，在真实互动中摸清用户痛点、系统边界与产品能力。 为了帮助企业清晰地开启这段转型旅程，Jim 建议业务领导者在立项初期必须明确回答三个关键问题： - **我们正在增强哪项人类业务流程？**（What human process are we augmenting?） - **我们正在利用什么专有数据来建立竞争优势？**（What proprietary data gives us an advantage in doing this?） - **我们打算如何在 90 天内衡量该项目的成功？**（How do we want to measure success in 90 days?） --- ### [15:00](https://www.youtube.com/watch?v=2utRWU6AtBI&t=900s) 跨行业落地的 Agentic AI 实战用例 除了 IDC 的软件选型案例外，Jim 还分享了 A.Team 在医疗、消费品和土木工程等不同行业落地的三个具体案例： - **大型医疗系统服务商（Healthcare）**：该机构面临的巨大痛点是护士需要耗费海量时间撰写临床文档，尤其是涉及复杂合规与保险索赔的文档，同时必须严格满足 HIPAA 隐私保护标准。团队构建了一套 Agentic 解决方案实现文档撰写自动化，在 4 个月内交付了 MVP，最终使护士的工作效率提升了 40%。 - **大型快速消费品公司（CPG）**：该企业在某关键区域的季度销售额持续下滑却找不到根因。尽管拥有海量数据，但传统的 IT 架构与静态分析流程极为迟缓，出具一份多源分析报告往往需要数天。团队通过在财务与营销工作流中嵌入定制 AI Agent，自动拉取多源异构数据，将原本静态的滞后分析转变为近乎实时的洞察输出，将分析周期由数天直接压缩至数小时。 - **大型土木工程咨询公司（Civil Engineering）**：任何大型工程或土地交易前，都必须进行耗时极长的前置环节——环境场地评估（Environmental Site Assessment, ESA）。该环节涉及跨联邦与各州的繁琐法规，且现场环境数据极其复杂，过去完全依赖领域专家（SME）耗费数周手工查阅、比对并撰写报告。团队构建了一套能够解析各项法规并关联特定场地数据的 Agentic 系统，自动起草评估报告供专家审核确认。这套系统将报告生成时间缩减了 70%，并立即为该企业每年节省了 600 万美元的运营成本。 --- ### [17:33](https://www.youtube.com/watch?v=2utRWU6AtBI&t=1053s) 打造 AI 原生飞轮：从“租用 AI”迈向“拥有 AI” 在演讲的尾声，Jim 总结了 AI 原生飞轮（AI-Native Flywheel）的运转机理：通过 Agent 连接碎片化系统中的多维数据，将智能洞察实时推送到业务第一线；一线团队基于这些洞察迅速采取行动——无论是自动生成合规内容、辅助重大采购决策，还是自动化业务审批流；而随着每一个行动的发生，系统通过反馈闭环变得更加聪明。这种机制最终在企业内部构建起一个具备未来扩展性的专有智能层，持续拓宽竞争护城河。 Jim 总结道，在这场技术变革中能够脱颖而出的企业，往往既不是资金最雄厚的巨头，也不是理论最高深的公司，而是那些在认知层面上发生根本转变的组织。他们深刻意识到： > "AI is not a tool that you use. AI is a capability that you build."（AI 不仅仅是一个你拿来使用的工具，而是你需要主动构建的核心能力。） 领先的企业正在停止“租用 AI”，并坚定地走向“拥有 AI 与专有智能”。演讲在现场热烈的掌声中结束，Jim 邀请有兴趣深入探讨的企业在会后及展台进一步交流。 ## Notable quotes > [02:33](https://www.youtube.com/watch?v=2utRWU6AtBI&t=153s) 通用的工具只能产生通用的平庸结果 — Jim Spare > [02:33](https://www.youtube.com/watch?v=2utRWU6AtBI&t=153s) 距离上一次 AI 重磅发布才过去零天 — Jim Spare > [10:05](https://www.youtube.com/watch?v=2utRWU6AtBI&t=605s) 基于这些数据，我是否也会给出相同的建议 — Jim Spare > [13:13](https://www.youtube.com/watch?v=2utRWU6AtBI&t=793s) Data is your differentiator（数据是核心差异化壁垒） — Jim Spare > [13:13](https://www.youtube.com/watch?v=2utRWU6AtBI&t=793s) Integration drives ROI（深度集成驱动真正 ROI） — Jim Spare > [13:13](https://www.youtube.com/watch?v=2utRWU6AtBI&t=793s) Speed wins（敏捷迭代以速度致胜） — Jim Spare > [13:13](https://www.youtube.com/watch?v=2utRWU6AtBI&t=793s) Builders over slides（重视实战构建而非纸上谈兵） — Jim Spare > [13:13](https://www.youtube.com/watch?v=2utRWU6AtBI&t=793s) What human process are we augmenting?（我们正在增强哪项人类业务流程？） — Jim Spare > [13:13](https://www.youtube.com/watch?v=2utRWU6AtBI&t=793s) What proprietary data gives us an advantage in doing this?（我们正在利用什么专有数据来建立竞争优势？） — Jim Spare > [13:13](https://www.youtube.com/watch?v=2utRWU6AtBI&t=793s) How do we want to measure success in 90 days?（我们打算如何在 90 天内衡量该项目的成功？） — Jim Spare > [17:33](https://www.youtube.com/watch?v=2utRWU6AtBI&t=1053s) AI is not a tool that you use. AI is a capability that you build.（AI 不仅仅是一个你拿来使用的工具，而是你需要主动构建的核心能力。） — Jim Spare ## Mentioned Jim Spare, IDC, A.Team, 微软 (Microsoft), 通用仪器 (General Instrument), 沃尔玛 (Walmart), 麦肯锡 (McKinsey), 波士顿咨询 (BCG), HIPAA, ERP ## Video description Jim Spare, CxO, f. Chief Product Officer at IDC, A.Team Abstract Most enterprises are stuck renting intelligence, using AI tools without building lasting value. Learn why 75% of AI pilots fail and how the winning 25% build learning systems that improve with every interaction. Discover how top companies are turning decades of disparate data into real-time intelligence, plus real-world use cases across healthcare, finance, and manufacturing that showcase the shift from disconnected tools to embedded learning systems. Stop experimenting. Start compounding. See how AI-Native companies win. [Source](https://www.youtube.com/watch?v=2utRWU6AtBI)

[2:33] Jim Spare: 通用的工具只能产生通用的平庸结果 [2:33] Jim Spare: 距离上一次 AI 重磅发布才过去零天 [10:05] Jim Spare: 基于这些数据，我是否也会给出相同的建议 [13:13] Jim Spare: Data is your differentiator（数据是核心差异化壁垒） [13:13] Jim Spare: Integration drives ROI（深度集成驱动真正 ROI） [13:13] Jim Spare: Speed wins（敏捷迭代以速度致胜） [13:13] Jim Spare: Builders over slides（重视实战构建而非纸上谈兵） [13:13] Jim Spare: What human process are we augmenting?（我们正在增强哪项人类业务流程？） [13:13] Jim Spare: What proprietary data gives us an advantage in doing this?（我们正在利用什么专有数据来建立竞争优势？） [13:13] Jim Spare: How do we want to measure success in 90 days?（我们打算如何在 90 天内衡量该项目的成功？） [17:33] Jim Spare: AI is not a tool that you use. AI is a capability that you build.（AI 不仅仅是一个你拿来使用的工具，而是你需要主动构建的核心能力。）

[0:38] Jim Spare: zero days since the last major announcement [0:38] Jim Spare: renting intelligence [0:38] Jim Spare: building intelligence [5:27] Jim Spare: New Product Development [5:27] Jim Spare: Operational Acceleration [5:27] Jim Spare: Departmental Workflows [10:05] Jim Spare: Connect Data [10:05] Jim Spare: Surface Insights [10:05] Jim Spare: Validate & Learn [10:05] Jim Spare: Execute Action [13:13] Jim Spare: Data as Differentiator [13:13] Jim Spare: Workflow Integration Drives ROI [13:13] Jim Spare: Speed Over Perfection [13:13] Jim Spare: Builders Over Slides [13:13] Jim Spare: What human process are we augmenting?

[0:04] Uh thank you everyone for coming. Uh [0:05] we'll get right into it. Uh I'd like to [0:07] start with a just a a quick introduction [0:09] so that you know who I am. Uh I've spent [0:12] my career growing disruptive technology [0:14] businesses both inside large companies [0:16] and also as independent uh venture [0:18] funded businesses. Uh I've worked at [0:20] large companies like Microsoft and [0:22] General Instrument and IDC the global [0:24] market research firm and some of my uh [0:28] small venture funded companies uh have [0:30] been successful others haven't. Uh one [0:32] was acquired by Microsoft and another [0:34] merged into a giant joint venture with [0:35] Walmart. Uh I'm now an executive adviser [0:38] at A Team and A Team is an AI native [0:41] systems integrator for enterprises who [0:43] want to build agentic AI uh uh systems [0:47] into their workflows. And so we have a [0:50] lot of experience seeing what is [0:52] successful and what some of the [0:53] challenges are. And so I'd like to share [0:56] those experiences with you today. Uh [0:59] first I'd like to start with a few [1:01] questions to just get a sense of the [1:02] audience here. Um who here in your [1:05] company is using AI in some form or [1:07] fashion today? [1:10] Okay, that's pretty good. Um who has [1:14] moved these experiments past the [1:15] prototype stage and into production? [1:19] Okay. A few few less. And who uh feels [1:23] like you've been able to prove ROI at [1:26] scale? [1:29] All right. A few. Yeah. A few fewer. [1:32] Well, this is actually pretty similar to [1:35] uh industry benchmarks. In fact, [1:36] McKenzie and BCG uh report that about [1:40] threequarters of companies are uh using [1:42] AI today in some form or fashion. But [1:45] only a third of those have taken uh uh [1:48] these experiments past uh prototype [1:50] stage and into production. And once [1:52] they're into production, um only half of [1:55] those have gotten to the point of [1:56] delivering ROI. So um we're at early [2:00] stages um in this whole journey. Um, and [2:03] many organizations still see challenges [2:06] in getting value from AI. So, what I'd [2:10] like to cover today is what separates [2:13] those AI native companies um from those [2:16] that aren't. Um, how to build and [2:19] architect AI agentic systems that learn [2:23] over time and get better uh with every [2:25] action that is taken and how to make AI [2:28] a core capability of your organization. [2:33] So we see that as enterprises approach [2:36] uh building out AI systems, there are [2:38] typically four value blockers. Uh first, [2:41] AI expertise is elusive. Uh many [2:44] companies just don't have the talent [2:46] around to be able to execute on their AI [2:49] goals. Uh second, it's hard to trust AI, [2:53] right? We've all experienced that [2:55] ourselves as users and um hallucinations [2:58] have been talked about uh much um in [3:00] this in this conference. But it's also [3:02] hard to evaluate AI and if you can't [3:04] evaluate something, it's really hard to [3:06] build an investment case against it. Um [3:09] and then third is most organizations do [3:11] not have the data foundations to be [3:14] prepared to execute and build an AI [3:17] system. And then lastly, everything is [3:19] moving so fast, right? It's been like [3:21] zero days since the last major [3:24] announcement in AI and we don't uh make [3:27] decisions in our companies typically [3:29] that fast. So, uh these are some of the [3:32] challenges that we've seen and and um as [3:35] a result, most companies are just using [3:39] generic tools that are available and [3:41] trying to pull those generic tools into [3:44] their workflows. Um, and what we see is [3:46] those generic tools end up delivering [3:49] generic results. And one way we think [3:52] about it is that most organizations are [3:54] renting intelligence and and not [3:56] building it. [3:58] Now, the uncomfortable truth when you uh [4:02] make a query or uh do an API call to an [4:06] an LLM, you're essentially uh renting [4:10] their intelligence. And this is not uh a [4:14] path for widstream success. Every uh [4:17] dollar you spend with them um is is lost [4:20] and you don't have any benefit from the [4:22] compounding of learning of the results. [4:25] They get all of that benefit. However, [4:28] those that are more successful in [4:30] getting to ROI are looking at this as [4:33] building intelligence. They have agents [4:35] embedded into their workflows. They've [4:38] architected the system so that the AI [4:40] gets smarter in the context of their [4:42] business with every action that's taken. [4:44] And thus this proprietary intelligence [4:47] compounds and really forms a basis of a [4:49] competitive mode. [4:52] And so in order to really get uh move [4:56] from these reactive appro from the [4:58] reactive approach to autonomous agentic [5:01] systems, you've got to change the way [5:03] you think about it. Um, you've got to, [5:05] you know, think about how we can detect [5:08] and act on business signals. We've got [5:10] to think about, uh, systems that [5:13] remember patterns to guide future [5:14] actions. We think about how to integrate [5:17] the AI systems directly into your [5:19] workflows and [5:22] deliver a system that continuously [5:25] improves with every interaction. [5:27] And so let me talk about a use case uh [5:30] that u I was deeply involved in that uh [5:33] maybe will bring some of this stuff to [5:35] life. So um when I was chief product [5:38] officer at IDC um of course this is a [5:40] large global uh market intelligence firm [5:43] that has deep uh assets in market [5:46] intelligence had built over 60 years of [5:49] proprietary information on technology [5:50] markets and products and uh uh uh [5:53] services and uh uh implementation uh su [5:58] successes. [6:00] And uh we knew that we wanted to uh [6:03] approach uh adoption of AI and get that [6:06] into our business, but we weren't sure [6:07] exactly how. And we literally had all [6:09] four blockers. We didn't have the [6:10] expertise. Uh we we we knew we needed to [6:13] move fast. Uh we our our data systems [6:17] were very desperate. Um stuck in PDF [6:20] documents and analysts notes and dozens [6:22] of different databases. Um, and so we [6:26] weren't quite sure what to do first, but [6:28] we contracted with the A team who had a [6:30] lot of experience in this area. And we [6:32] built a prototype in eight weeks. Um and [6:35] that prototype simply looked at all of [6:37] our existing or many of our existing not [6:40] all but many of our existing data [6:41] sources and enabled our users both [6:44] internal users and external users to [6:47] query that corpus of data and get an [6:50] intelligent response that is based on [6:52] our unique proprietary data. Now once we [6:56] had this in our hands and played with it [6:58] and looked at users playing with it, it [7:00] gave us uh the uh the confidence and the [7:03] insights needed to figure out where to [7:05] go next. And we really defined three [7:09] different lanes from that prototype. Um [7:11] one lane was hey we could actually build [7:14] brand new products that leverage AI that [7:16] we can sell to customers. [7:18] The second lane was hey we can use this [7:21] in our traditional uh operations to [7:24] accelerate the process of creating [7:26] market research reports or uh delivering [7:28] data and then third is okay there's of [7:31] course all kinds of ideas to use AI in [7:34] the departmental operations of the [7:35] company finance HR uh sales and so forth [7:39] and so from that uh uh foundation we [7:43] said all right let's go off and build a [7:46] new product let's take this what we've [7:49] learned so far and and build a product [7:51] that solves a clear business problem and [7:54] in so doing we're going to learn a lot [7:56] and create a new line of business. And [7:58] so the business problem we focused on [8:00] was uh a business problem we knew well [8:03] but delivered as a service not uh [8:06] through a digital uh platform. It was [8:08] helping IT executives pick the right [8:11] software for their business purpose with [8:13] confidence. So a a process that right [8:16] almost every company goes through many [8:17] times a year and uh and and this is a [8:21] process that has a very specific [8:22] workflow uh that is generally used [8:25] across organizations. First you define [8:27] your business goals then you understand [8:30] a little bit about the vendors that are [8:31] in the category that you're interested [8:33] in. Uh then you uh define your [8:36] requirements that you need to that the [8:38] software needs to achieve. You evaluate [8:41] the options and create a short list. [8:42] Then you pick your top vendor. you [8:44] negotiate the deal and you're off to the [8:45] races. Um, and so we built an AI system [8:50] that followed that workflow that wasn't [8:52] just like a blank uh uh canvas of of [8:55] chat, but really supported the user [8:58] through that workflow and that minimized [9:00] uh the training that was needed for new [9:02] users and it also uh maximized adoption. [9:06] So uh as we defined that that product [9:08] and that initiative, we built out the [9:11] rest of the team uh with with uh uh a [9:14] team uh a team uh staff members uh we [9:17] developed the product, we delivered the [9:18] product, we launched the product, we [9:19] signed up first customers and lo and [9:22] behold we found that customers were [9:24] telling us that this new platform now [9:25] takes uh a process of picking a major [9:29] new software package that could take you [9:30] know six months and now they see a a [9:34] possibility to reduce that process. to 6 [9:36] weeks and the evaluation stage which is [9:39] a key datative intensive part of that [9:42] process uh went from like 5 weeks to a [9:45] number of hours. Uh so this is a clear [9:48] case where we built the system around an [9:50] existing workflow. We leveraged our [9:52] proprietary data to be able to recommend [9:55] software uh vendors based on sets of [9:57] requirements and we built a system that [10:00] enabled us to achieve the business goal. [10:05] Now at the highest level uh the the [10:08] process is is you know kind of two [10:10] steps. Uh first make sense of structured [10:13] and unstructured data and then secondly [10:16] uh surface the insights that lead to [10:19] actions that deliver against the [10:21] business goal. In this case it was uh [10:23] picking the right software for your [10:24] business purpose. Now what we found is [10:27] that we were supporting in this product [10:30] big decisions that customers were [10:32] making. you know, if you're picking a [10:34] new ERP system, you don't want to like [10:35] pick the wrong pick the wrong vendor [10:37] because then you're stuck for many many [10:39] years and millions of dollars. And so, [10:40] it was really important that we gave [10:42] that the answers were right. Um, and [10:44] that we trusted uh the system to be able [10:48] to offer it to our customers under our [10:50] brand name. And so, it was it was [10:53] important to us that we have a human in [10:54] the loop. Um, and we architected the [10:56] system to work very very well with AI. [11:00] uh this started out with an analyst [11:02] looking at the results and saying hey [11:03] would I have recommended the same thing [11:05] you know based on uh all the data that [11:07] I'm seeing and then it extended to the [11:10] users themselves uh to give feedback [11:12] along the process and also in the final [11:14] result so that we had reinforcement [11:15] learning uh built in the system and this [11:18] human in the loop enabled us to create [11:21] uh this uh feedback loop that enabled us [11:24] to develop uh proprietary intelligence [11:26] over time. [11:28] So at 18 we've developed a uh a [11:32] framework uh for building these kinds of [11:35] systems and delivering ROI that we find [11:37] works across almost all applications. Um [11:40] so first step connect everything um into [11:43] a gent processing get your unstructured [11:45] unstructured semi-structured and [11:47] structured data um uh visible by agents [11:50] clean and visible by agents. Uh next [11:53] make that visible to the AI and so the [11:56] AI can analyze that information and [11:59] start to surface insights. Uh validate [12:02] that those insights with a human um and [12:05] make sure that human can provide [12:07] feedback back into the AI system so that [12:10] the system can get better. Now, of [12:12] course, as those insights are are [12:14] surfaced, they pro provide the insights [12:17] that are needed so the business can take [12:19] actions or agents can take actions uh [12:22] that achieve the business goal of of the [12:24] the program or the product. And in [12:28] building it out this way, you have [12:30] additional benefits of uh compounding [12:33] intelligence over time. uh the more [12:35] actions you take uh the more intelligent [12:37] the system gets and you continue to to [12:39] build this proprietary advantage. Uh [12:42] second um you have agentic action [12:45] instead of a flat query and wait for [12:48] response uh type of uh uh user [12:51] experience. And so this agentic action [12:54] when built into your workflows [12:56] um uh uh gives you that proactive [12:59] partner that you need instead of a [13:01] reactive approach. And once it's [13:03] integrated into your workflows, uh you [13:06] minimize the training overhead for new [13:08] users and maximize adoption. [13:13] So this enables you as a company to [13:16] build the four pillars that we see that [13:19] are necessary uh uh to be truly AI [13:23] native. Uh first, right, data is your [13:25] differentiator. It's your your [13:27] proprietary data and your proprietary [13:30] workflows that are your competitive [13:32] mode. Build on that. Uh second, um [13:37] pardon that. Um second, integration [13:39] drives ROI. So the uh more that you're [13:43] able to integrate the uh Aentic system [13:46] into existing workflows, uh the faster [13:48] your path to ROI. Of course, speed wins. [13:52] uh you know we measure uh MVPs and [13:55] iterations in weeks not months. And [13:58] lastly, builders over slides. Uh the the [14:02] best AI companies or the best um [14:05] companies the companies that are most [14:06] successful in implementing AI are [14:08] getting systems in the hands of users so [14:11] they can figure out what users really [14:13] want, what they need, what their issues [14:15] are, and also what the systems are [14:17] capable of. And so starting that journey [14:19] as soon as you can uh is incredibly [14:22] important for success. And so to get [14:25] started on this journey, we encourage [14:27] our partners to a ask first, what human [14:31] process are we augmenting? [14:34] Secondly, what proprietary data gives us [14:38] an advantage in being able to do this? [14:40] Are we bringing to the table? And third, [14:43] how do we want to measure success in 90 [14:44] days? [14:48] So uh let me give you a couple more [14:51] examples of uh of some use cases that [14:55] we've seen at a team uh to give you a [14:58] sense of you know how this can work in [14:59] other uh environments. Uh first a large [15:03] health care systems provider uh came to [15:05] us with a challenge that their nurses [15:07] were spending lots of time on [15:09] documentation and were particularly [15:12] challenged around particular kinds of [15:14] documentation around insurance claims. [15:17] And so this is a a a case where uh the [15:22] uh thank you. The [15:24] uh documentation needs to adhere to very [15:27] uh complex sets of regulations and also [15:30] needs to be protected uh uh based on [15:32] HIPPA requirements. And so we built out [15:35] an a uh agentic solution that automated [15:38] these write writeups and re resulted in [15:41] a 40% increase in productivity and [15:43] nurses time. uh in this case the MVP was [15:47] built in four months. Next, a big [15:50] consumer goods company uh was seeing [15:53] quarterly declines in sales and wasn't [15:55] sure in a key region and wasn't sure [15:57] why. And they had all the data in the [15:59] world but weren't able to get the [16:01] analyses from that data with their [16:03] existing systems and their existing IT [16:05] organization. Those analyses took days [16:08] to turn around and so they came to us. [16:10] We embedded custom AI agents into the [16:13] finance and marketing workflows that uh [16:16] pulled data from multiple disparate [16:18] sources and shifted a very static data [16:21] analysis process to real time uh cutting [16:24] their analysis time from days to hours. [16:28] And lastly, working with a big civil [16:30] engineering firm, they came to us and [16:32] shared that one of the big obstacles in [16:35] their business is an upfront step called [16:38] an environmental site assessment. Turns [16:40] out for any civil engineering project [16:42] you need to start with or big uh [16:44] property transaction you need to start [16:45] with an environmental site analysis. And [16:48] this is uh a process or an analysis that [16:51] requires uh uh governance of multiple [16:56] federal and state regulations as well as [17:00] complex information about the site [17:02] itself. And is currently done by subject [17:05] matter experts that spend days and days [17:07] and days and days to create these [17:09] reports. So similarly we created an [17:12] agentic system that looked at all these [17:14] regulations was able to consult the [17:16] information about specific sites and [17:18] create a report for those subject matter [17:20] experts to review and approve and the [17:24] result was a 70% reduction in the time [17:26] that it took to create these reports and [17:29] a $6 million annual savings immediately [17:31] for the company. So the companies that [17:34] are getting this right um end up [17:37] creating uh a native or an AI native [17:40] what we call an AI native flywheel. Um [17:43] the agents you know connect data across [17:45] fragmented systems with insights that [17:48] are pushed to the front lines. Um as [17:50] those insights come in they're turned [17:52] into actions uh whether it be creating [17:55] content whether it be supporting [17:57] decisions whether it be routing [17:58] approvals. Um, and the more uh actions [18:02] you take, the smarter the system [18:04] becomes. And this builds a uh futurep [18:07] proof system or a proprietary uh [18:10] intelligence layer that continues to [18:13] expand your competitive mode. [18:16] And so [18:19] our view is that uh the the companies [18:22] that are best prepared to do this aren't [18:25] the best funded companies or the most [18:27] advanced or or or thinking. Um the [18:31] companies that are most successful are [18:34] the companies that are approaching this [18:36] and saying um AI [18:40] um is not a tool that you use. AI is a [18:43] capability that you build. [18:46] They stop renting AI [18:49] and they start owning it. [18:53] Thank you very much and if you're [18:54] interested in talking more uh please vis [18:56] visit us in the booth or after the uh [18:59] session. Thanks.