企业AI Agent落地指南：数字化三阶段与API、RPA协同实践需要先看清适用场景和关键步骤，避免只记结论却忽略实际限制。

{"type":"doc","content":[{"type":"paragraph","attrs":{"id":"ff841b1a-f2c0-4c07-a205-df9e9a7d9612","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"中国信通院和三六零今年6月底联合发布的一份报告提到，"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"企业级智能体正式走出实验室概念，进入规模化商用周期。"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"报告测算，2026年国内企业级智能体市场规模有望突破800亿元，年复合增长率超120%。Gartner的预测认为，到今年年末，全球40%的企业软件将内置智能体能力，这个数字在去年还不到5%。"}]},{"type":"paragraph","attrs":{"id":"96d1a09f-4546-4bea-943e-bd10f5060cd4","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"数字增长很快，落地效果却参差不齐。同一套智能体产品放进不同企业，结局可能完全不同。有企业把它嵌进核心流程，越用越深，也有企业做了几个像样的演示，汇报一结束就没了下文，还有企业买了平台，员工偶尔拿它写材料，跟普通聊天工具没什么区别。"}]},{"type":"paragraph","attrs":{"id":"a1120b28-6616-43f1-987a-ade444520ced","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"模型能力不是决定因素，真正拉开差距的，是企业自己走到了数字化转型的哪一步。"}]},{"type":"paragraph","attrs":{"id":"e9676ac1-4b49-4d28-8135-8185b4acc2f4","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"企业数字化大体分三个阶段："},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"先把业务搬进系统，再把分散的系统连起来，最后用智能能力重新组织流程。"},{"type":"text","text":"AI智能体最合适的进场时机，通常落在第二阶段。"}]},{"type":"heading","attrs":{"id":"2ebfa2bc-05f7-4a0f-a295-470e543eb50e","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","text":"阶段一：系统建好了，但数据不流动"}]},{"type":"paragraph","attrs":{"id":"b4c361d8-7d71-40ff-97f4-90fa6028aef6","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"企业数字化的第一步是上系统。财务上ERP，销售上CRM，生产上MES，审批搬进OA，这一步解决的是"有没有"的问题：有没有统一入口，有没有线上数据，有没有可以照着走的流程。"}]},{"type":"paragraph","attrs":{"id":"19259fe8-2847-4028-aa44-c6d76fb5678e","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"问题是，系统建起来之后，数据未必真正流动。很多企业表面上有几十套系统，日常工作仍然靠人工搬运，从邮件下载附件、登进系统查数据、结果再复制进Excel，审批走完OA，还要人工把结果录进业务系统，客户信息散在CRM、工商数据、内部交易记录里，最后靠客户经理自己拼起来。"}]},{"type":"paragraph","attrs":{"id":"1734f745-f6a2-4137-8436-7f7ec01c2cc6","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"这个阶段上智能体，最容易做成的是知识问答、材料总结、邮件生成。"},{"type":"text","text":"智能体能看懂一封邮件，未必有权限进入后续系统，能判断材料缺了什么，未必知道该触发哪条审批。它缺的不是更聪明的入口，而是一条走得通的路。"}]},{"type":"paragraph","attrs":{"id":"ec0e20ee-9887-4a43-a729-ceb96a49d936","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"所以"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"第一阶段适合做边界清楚、风险较低的场景，"},{"type":"text","text":"比如制度问答、文档提取、会议纪要。跨部门、涉及资金和审批的核心流程，先别急着交给自主智能体。"}]},{"type":"image","attrs":{"id":"ea7bbab4-0841-43fa-b332-5bafd2fd3d99","src":"https://developer.qcloudimg.com/http-save/audit-11837377/1c374bd3378d4076a0f03d42866b6fe2.jpg","extension":"jpg","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":1100,"aspectRatio":"1.776833","status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"heading","attrs":{"id":"9a7f195e-fa74-4fa5-b914-c06cf70cd62e","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","text":"阶段二：系统开始互联，人力被消耗在衔接上"}]},{"type":"paragraph","attrs":{"id":"2de099a5-3647-4941-a5d0-6366606453d3","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"企业进入第二阶段之后，管理者常见的困惑是，系统明明比以前多了，效率为什么没有明显提升？"}]},{"type":"paragraph","attrs":{"id":"e449f292-f332-4ef4-92ef-603954a14e2b","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"答案往往在系统之间。一个动作在系统A完成，不代表流程结束，后面可能还要进系统B核验、系统C审批、系统D更新状态，真正消耗人力的，是这些环节之间反复的下载、录入、核对和催办。"}]},{"type":"paragraph","attrs":{"id":"14b99f80-ec3c-49de-97e2-2ace47972133","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"也是在这个阶段，API、RPA、低代码、数据集成平台开始大规模出现。企业先啃最规则化的工作，例如自动下载报表、批量录入、核对账户、生成凭证。规则越是清楚的地方，传统自动化跑得越稳。"}]},{"type":"paragraph","attrs":{"id":"d7676d0a-361d-4390-a50a-b6f8add27f42","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"卡点在于流程里总有不那么规则的部分，合同格式会变，邮件写法不统一，客户材料常有缺漏，系统告警需要结合上下文判断，传统自动化只能执行，不会理解，遇到模糊情况就停下来等人。"}]},{"type":"paragraph","attrs":{"id":"6506b8b1-ee7e-4de2-9a5e-9a8102e0e497","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"这正是AI智能体最适合进场的位置。"},{"type":"text","text":"企业已经有系统、接口和自动化脚本，智能体不用从零开始，它负责读懂任务、识别材料、判断异常，把明确的操作交给已有工具执行。一份客户材料进来，智能体先提取信息、判断是否齐全，符合条件的自动进入审批，不符合的生成补充清单，涉及敏感事项再转人工。运维告警触发后，智能体结合知识库判断故障类型，能自动处理的交给执行程序，无法确认的升级给工程师。"}]},{"type":"paragraph","attrs":{"id":"7be1efd2-6c33-4fed-af46-c3062b3130ba","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"理解和执行，第一次有机会被放进同一条链路。"}]},{"type":"paragraph","attrs":{"id":"36dab995-8fd8-4a81-b25a-83a52145f75f","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"对多数企业来说，这一步的投入容易算清楚，效果也容易验证，数字化底座已经不算空白，流程里又有大量人工接力可以被替代。这比在第一阶段做一个孤立的AI入口，或者等所有系统全部改造完再启动，都更现实。"}]},{"type":"image","attrs":{"id":"7a90029b-e0ca-4435-98f7-19961c183e11","src":"https://developer.qcloudimg.com/http-save/audit-11837377/b6a10cbe517b9ad741ef5bb351ee43c6.jpg","extension":"jpg","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":1100,"aspectRatio":"1.776833","status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"heading","attrs":{"id":"68e7760a-e740-4b69-86c5-09150ea8e6ff","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","text":"阶段三：流程打通之后，风险也一起被放大"}]},{"type":"paragraph","attrs":{"id":"bb33a3f1-4bb0-4657-a5b6-a841b6377bee","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"到了第三阶段，重复性劳动大多已被自动化吸收，企业开始关心更复杂的问题："},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"能不能提前发现风险，能不能动态调度资源。"},{"type":"text","text":"智能体从流程助手，逐步走向运营角色，采购智能体结合库存和供应商信息生成补货建议，客户运营智能体识别投诉风险并完成核查分派，财务智能体持续监测异常数据。多个智能体之间也开始协同，一个理解需求、一个查找数据、一个执行操作、一个检查结果。"}]},{"type":"paragraph","attrs":{"id":"eb87cc29-61a2-4627-b9aa-a5330b24f738","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"能力越强，容错空间越小。第一阶段，智能体答错一句话，顶多是员工重新修改，第三阶段，它判断错一个业务状态，可能自动触发一连串查询、审批和系统操作。权限有没有隔离，关键动作要不要人工确认，出了问题能不能回退，这些问题直接决定智能体能不能进入核心业务。"}]},{"type":"paragraph","attrs":{"id":"079fa3c3-e4ae-44c9-b031-680f2c51024f","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"这里还有一条硬约束。《人工智能生成合成内容标识办法》从2025年9月1日起施行，要求AI生成合成的内容必须显式或隐式标识，平台要核验、要留存日志。智能体一旦深入到对外内容生成和决策环节，合规就不再是上线前补一份材料那么简单，而要嵌进执行链路本身。"}]},{"type":"paragraph","attrs":{"id":"454ca799-07e3-4cfc-89c3-5b79c49b9eb9","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"第三阶段的智能体价值最大，试错成本也最高，它已经嵌进了企业运营体系，不再是一次性采购的软件。"}]},{"type":"heading","attrs":{"id":"140cf49c-948b-4620-a778-f8a3c3119799","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","text":"什么信号说明企业可以出手了"}]},{"type":"paragraph","attrs":{"id":"36a56fb9-71d4-4a9b-89a4-b04b1436e395","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"不用先给自己贴一个"数字化二阶段"的标签，观察这几个信号就够："}]},{"type":"paragraph","attrs":{"id":"36b88579-7dd0-4949-b1b9-ed900ef5ed96","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"1、员工花在系统之间搬运数据的时间，超过花在业务判断上的时间：说明问题不在缺系统，在系统之间缺协同。"}]},{"type":"paragraph","attrs":{"id":"c00bc6da-eee5-40b3-bbe3-20cd70e4648c","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"2、传统自动化开始频繁撞上非结构化信息：合同、邮件、图片、自然语言要求，总要等人读完才能继续，这通常意味着流程已经需要AI的理解能力。"}]},{"type":"paragraph","attrs":{"id":"03e8967f-78ec-4e86-9db5-52c7e678c779","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"3、企业能说清一条流程的边界：数据从哪来，能调用哪些系统，哪些步骤可以自动执行，哪些必须人工确认，而不是完全依赖某个老员工的经验。"}]},{"type":"paragraph","attrs":{"id":"32bf7266-d816-42a1-877b-a537c3fad4c0","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"4、企业关心的不再是一次性采购价格，而是长期运行成本：模型调用、接口维护、知识更新都要算进账。高频、长链路、需求稳定的流程，更适合率先智能化。"}]},{"type":"image","attrs":{"id":"fbed102c-a9ca-420f-911f-534b5be1cc5b","src":"https://developer.qcloudimg.com/http-save/audit-11837377/193e581d9435741cd9dc82eb7ede8d6b.jpg","extension":"jpg","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":1100,"aspectRatio":"1.776280","status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"heading","attrs":{"id":"4ebeecaa-ac89-4135-8fdb-09f83585473f","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","text":"厂商路线不止一条，执行层才是分水岭"}]},{"type":"paragraph","attrs":{"id":"2fe9a3dc-1f84-4e9a-96d6-da4b62f3be1c","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"企业级智能体的供给侧目前大致分几类：大模型和云厂商提供模型与开发框架；ERP、CRM等业务系统厂商把AI嵌进原有产品；低代码和集成平台解决连接问题；RPA AI阵营则更专注智能体如何操作存量系统，把任务真正执行下去。"}]},{"type":"paragraph","attrs":{"id":"ae1da7ff-0b42-4155-bf16-671a2cb8168b","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"路线没有绝对优劣，取决于企业自身的系统基础。应用高度云化、接口齐全的企业，用模型平台做编排会比较顺；内部还有大量桌面软件、内网系统、没有开放接口的老旧客户端，能不能非侵入式接进现有环境，往往比模型参数更重要。金融、政务、能源行业尤其如此，这些企业通常不缺系统，缺的是如何在不推倒重建的前提下，把智能能力安全接进现有流程。"}]},{"type":"paragraph","attrs":{"id":"96b8d14a-da21-4611-9df4-e062d4686044","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"IDC发布的《中国RPA AI解决方案，2024》报告显示，这个市场规模已达31.5亿元，同比增长27.4%。金智维、艺赛旗、来也科技位列前三，份额分别为10.1%、9.1%、8.4%，实在智能、弘玑等厂商紧随其后，前六家合计占据41.7%。行业格局仍在变化，有的厂商依托早期RPA积累转型，有的从数字员工或大模型侧切入，具体选型还是要看企业自身的系统环境和场景需求。"}]},{"type":"paragraph","attrs":{"id":"43d0c58b-0b20-413c-966c-e9395815712a","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"以金智维为例，它的架构思路是把大模型放在理解和规划层，API、RPA和既有工具放在执行层，敏感节点保留人工确认，过程通过权限、日志和监控管理，先给智能体划定工作区：能读什么数据，能调用哪些工具，能把流程推进到哪一步，遇到什么情况必须停下来找人。目前和金智维合作过的海通证券的数智员工平台已覆盖总部多个部门和上千个自动化流程，湖北社保也把自动经办和统一管理纳入了同一套体系，这类项目的共同点是，智能能力没有绕开原有流程另起炉灶，而是在已有自动化基础上继续往理解和判断延伸。"}]},{"type":"paragraph","attrs":{"id":"b256a337-4e71-48ac-b718-8d3265761af6","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"市场份额说明一条路线积累了足够的实践样本，但选型终究要落回企业自己的系统环境和治理要求。"}]},{"type":"image","attrs":{"id":"3cd2fbd9-e8d8-46e4-a228-70115a9356ad","src":"https://developer.qcloudimg.com/http-save/audit-11837377/ce2eaa9c0c4abf47f4f485467ecf5d36.png","extension":"png","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":738,"aspectRatio":"1.324955","status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"ebce7e6c-27bd-4b11-9063-fb9a490adfe9","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"回到最初的问题：企业数字化走到哪个阶段，引入AI智能体效果最好？多数情况下是第二阶段，系统已经建起来，流程开始互联，但仍有大量跨系统操作和人工判断没被消化。智能体这时候进场，既有资产可以复用，也有明确问题要解决。"}]},{"type":"paragraph","attrs":{"id":"4ac1a54d-cc89-4e26-bda4-05ed474abc30","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"但这不是一张放之四海皆准的门票，有些企业系统不多，流程和数据却足够清楚，可以更早启动，有些集团信息化建了十几年，内部权限和业务边界依然混乱，也未必适合让智能体碰核心流程。"}]},{"type":"paragraph","attrs":{"id":"1831b7ad-47c0-41f7-bcab-9d2a136ba374","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"在讨论"现在要不要上AI"之前，可以先在企业内部找这样一条流程：人工成本足够高，业务边界足够清楚，执行结果能验证。找到这条流程，再谈平台和模型选型。"}]},{"type":"paragraph","attrs":{"id":"9fe38e7a-9b05-4f39-83d3-a763efc96d98","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"太早进场，智能体只能站在业务的外面，太快铺开，一上来接管整个部门，最后被权限和系统协同拖住。合适的节奏，是在企业已经有路、但这条路还需要大量人工接力的时候进场，让AI看懂下一步，让自动化工具把动作做稳，让人在关键节点保留判断权。"}]}]}","createTime":1786502744,"ext":{"closeTextLink":0,"comment_ban":0,"description":"","focusRead":0},"favNum":0,"html":"","isOriginal":0,"likeNum":0,