智东西（公众号：zhidxcom）
作者 | 江宇
编辑 | 漠影
在华润信托，一笔债券交易的起点，可能只是一条聊天群里的消息。
交易员需要在上千个交易群中接收指令，识别交易要素，再录入交易系统。客户询问持仓、交易进度时，他们还要打开系统查询，把结果发回群里。
这项工作既繁琐，又容不得差错。
如今，华润信托开始让一个智能体接手其中部分工作。它直接进入交易群，识别并拆解交易指令、录入系统，也能查询持仓与交易进度，自动回复群里的咨询。
▲华润信托苏南
据腾讯云披露，这个基于智能体开发平台ADP打造的债券助手，拆单准确率已经从最初的80%以上提高到98%。
但在关键节点，华润信托仍然保留人工确认。
这背后其实是行业的一个非共识：大模型不断展示其可以完成多少复杂任务，但模型能正确理解一条指令，与企业愿意让它执行这条指令，是两回事。
有人认为，模型能力的持续进步，终将抹平企业AI落地的种种技术门槛。
现实未必如此。2026年8月，麦肯锡发布全球AI调查：80%的受访者称AI提高了个人工作效率，但只有37%认为AI对所在企业利润产生了正向贡献。9月，麦肯锡进一步提到，企业大量时间消耗在不同部门和系统之间的任务传递、协调与等待上。
可见，企业AI难以突破的是，模型之外的整个企业生产体系。
这也是不少专注模型或单点应用的中小To B AI公司面临的挑战。模型能力和产品功能可以快速迭代，但要进入企业核心系统，获得客户信任，并建立覆盖不同行业的交付体系，需要更长时间的投入与积累。
10月30日，在腾讯全球数字生态大会上，腾讯云升级智能体开发平台ADP，发布AgentPortal 2.0，并宣布与WorkBuddy双向打通。
此次升级，腾讯云还进一步明确了ADP的Service Agent定位，聚焦企业数据和业务场景，推动Agent深入业务流程，创造实际经营价值。
从企业业务系统接入，到智能体统一管理，再到行业场景与交付生态，腾讯云正在尝试提前占据企业Agent规模化落地的战略要地。
这些探索能否在模型之外，为腾讯云建立起一道更难被后来者快速复制的企业AI护城河？
一、AI敢不敢进入核心业务？企业Agent先要跨过管理关
从理解指令到执行交易，企业Agent首要解决的，是权限与责任问题。
华润信托债券助手仍保留人工确认与操作记录的原因，其实不难理解。一个金融智能体可以查询客户持仓，不意味着它能够查看所有客户的数据；一个负责办理业务的Agent能够调用系统接口，也不意味着它有权完成所有操作。
企业管理Agent，首先要做的就是划清业务权限。
对应这一需求，腾讯云ADP此次升级了安全私网连接、租户及权限隔离等能力，支持不同Agent按照权限访问数据库、ERP等企业内部资源，并将企业系统连接能力统一提供给沙箱、工作流、MCP、连接器等不同模块。
再来看，在云南白药也是如此：云南白药的员工日常需要查询内部制度、药品信息和医学资料，但这些知识的使用对象和权限要求各不相同。
为此，云南白药依托ADP与乐享，将权限管理细化到员工ID。员工通过“白药钉”一个入口提问，智能体自动判断问题类型、路由至对应知识库，再根据员工的实际权限返回内容。
▲云南白药党亚斌
对员工来说，就是问一个问题、得到一个答案；但在背后，是多个知识库和一套权限规则共同发挥作用。
除了明确权限以外，企业同样不能一味追求全自动，不同业务对Agent自主程度的要求也不尽相同。
例如，标准审批、工单处理等有固定规则的业务，往往需要明确的执行顺序和可预测的结果；跨部门分析、经营决策辅助等工作，则可能需要多个智能体分工协作。
为此，腾讯云ADP提供Claw、标准、多智能体、工作流四种模式，让企业能够根据实际业务需求，选择不同的智能体构建与执行方式。
华润信托就采用了这样的组合：通过Claw处理自主规划任务，利用Workflow执行规则明确的业务流程，而关键交易环节始终由交易员确认。
这也与行业对Agent治理的判断相呼应。
2026年5月，全球权威研究机构Gartner提醒，如果企业对所有Agent采用统一的治理方式，可能造成两种结果：简单任务被过度限制，而高风险任务却获得了过多权限。
而当Agent真正进入生产环境，管理也不能止步于上线。
传统软件通常按照预先确定的程序规则运行，但模型输出并不具有同样程度的确定性。一次任务执行成功，也不意味着Agent能够长期稳定地完成相似任务。
尤其是企业开始让Agent持续运行后，如何追踪每次操作、监控成本、定位故障，以及出现问题后如何及时纠正，都成为新的管理需求。
腾讯云ADP此次强化了运行记录、成本与状态监测、故障定位和版本回滚等能力。同时，ADP还可以利用AI分析真实业务反馈，提出智能体优化建议，再由人工审核后生效，让Agent在运行过程中持续受到评估和调整。
企业管理Agent，要让它能够在明确的业务规则下持续执行任务。企业级工程与治理能力，正在成为模型之外的第一层技术壁垒。腾讯云ADP此次升级，正是围绕这些生产环境中的管理需求展开。
二、员工都用上AI了，为什么企业还没同样“快提效”？
企业不是一群员工的简单相加。每个人都节省一个小时，并不意味着公司就能少花同样多的时间完成一笔业务。
企业AI的第一个问题，是把员工的工作拆成了一个个彼此独立的任务。
过去两年，AI最先进入的是文档撰写、数据分析等个人办公场景。这些工作有一个共同特点：任务边界相对清楚，员工自己就能够判断结果好坏。
但企业业务的运行方式并非如此。一项任务往往需要经过相关负责人、区域团队和一线人员等多个环节，在不同部门、人员和业务系统之间流转。
如果只是某个环节完成得更快，AI带来的效率优势就可能被原有流程消耗掉。
由此，延伸至企业AI的第二个问题：交接与等待，拖慢了整个业务流程。
2026年9月，麦肯锡在一项研究中分析了一家财富500强工业企业的生产流程，发现整个业务流程所花费的时间，大部分是在等待任务进入下一环节。
麦肯锡将其称为“协调接口成本”（Coordination Interface Tax）。
对于大型企业而言，单独优化一个步骤，未必能解决不同团队和系统之间的协作问题。这也解释了为什么员工使用AI的频率越来越高，但企业整体的效率和经营收益未必能够同步增长。
企业AI需要解决的，已经不局限于让每个员工更快完成自己的工作，还包括让AI真正进入业务流程，把不同环节连接起来。
腾讯云与伊利的合作探索提供了另一案例。伊利将区域销售、导购表现和市场数据汇聚起来，由智能体辅助分析经营情况、形成策略，再将相关能力延伸到一线导购场景。
在这个过程中，WorkBuddy承担了业务人员使用AI、开展经营分析的部分工作；ADP构建的导购智能体，则进一步进入面向消费者的销售场景，根据客户需求生成商品推荐和沟通话术。
据腾讯云披露，这套组合让导购运营团队效率提升超过50%。同时，导购智能体在A/B测试中带来26%的销售订单提升。前者体现内部效率，后者直接触及经营结果。
AI的价值不再只体现为员工少做了多少工作，也可能表现为企业完成了更多业务。
但从个人工作流走向企业级服务，还需要让AI能力能够跨系统、跨人员复用。
这也是Personal Agent与Service Agent的核心区别。腾讯云副总裁吴永坚告诉我们，Personal Agent更多从个人使用出发，而Service Agent需要站在企业角度，考虑不同角色的权限、安全，以及构建好的能力能够分享给谁、如何管理。
▲腾讯云副总裁吴永坚
例如，员工可以通过个人办公入口分析销售数据，但不能仅凭一段AI对话，就让经营策略自动进入全公司的销售流程。企业还需要完成权限验证、能力分发、任务执行和效果反馈。
腾讯云此次将ADP与WorkBuddy双向打通，正是试图缩短个人AI使用与企业业务执行之间的距离。
一方面，企业通过ADP构建的Agent和Skill，可以进入WorkBuddy，供员工在日常办公中直接调用；另一方面，员工通过WorkBuddy探索和验证的个人工作流，也可以转化为ADP智能体，进入企业统一管理与分发体系。
这意味着，个人积累的有效工作方法，有机会进一步成为企业可以复用的AI能力，而不必长期停留在某个员工的电脑里。
个人Agent解决员工如何使用AI的问题，服务Agent则进一步把AI带入企业业务流程。企业AI真正可能产生更大收益的地方，可能恰好在于重新组织这些步骤之间的关系。
三、Agent造得越来越容易，企业也为“养Agent”发愁
智能体开发门槛降低，可能让企业重新遇到软件时代的老问题：Agent太多了。
不同部门有自己的知识库、流程和业务需求，也可能各自构建智能体。每个Agent单独看都有价值，但当数量不断增加，企业可能不知道究竟有哪些Agent正在运行、谁负责维护、哪些功能重复，以及它们各自拥有多少权限。
2026年4月，Gartner将这种现象称为Agent Sprawl（智能体蔓延）。该机构当时的调查显示，只有13%的受访组织认为自身具备适当的Agent治理机制。
企业可能还没来得及充分利用AI，就先需要面对新的管理负担。
对此，腾讯云采取了一套“两手并进”的方案：ADP降低Agent开发门槛；AgentPortal 2.0解决智能体的统一使用和管理问题。
其中，ADP此次升级有两个重要变化：一是从对话式搭建智能体，走向AI原生贯穿整个开发流程；二是从搭建智能体，走向构建与部署完整的业务应用。ADP正从4.0版本的“智能体构建专家”，升级为“智能体全生命周期专家”。
过去，业务人员可能需要向技术团队提交需求，再由开发者分别完成智能体配置、系统集成和应用界面搭建。现在，部分构建工作可以在同一工作区中完成，配置、预览、调试和评测也能衔接起来。
AgentPortal 2.0则进一步解决智能体统一使用和管理的问题，让企业可以将不同部门构建的Agent纳入统一入口，同时支持对ADP及第三方智能体的集中管理。
员工可以直接提出业务需求，由系统自动选择适合的智能体处理；企业也可以按照部门、职能分发Agent，统一管理上下架和使用权限。
对于需要长期协作的复杂项目，多个智能体可以组成团队，由项目经理智能体协调任务；对于直接服务消费者的业务，多个专业Agent也可以汇聚在统一服务入口背后。
由此，企业不必让员工分别寻找和调用不同的AI工具，而是可以通过统一入口完成任务分派与多智能体协作。
与此同时，一套区别于过去软件时代的AI建设新打法，也正在逐渐成形。
过去，企业软件通常由IT部门采购、部署，再要求业务部门学习和使用。但个人Agent普及后，一线员工可能先发现某种AI工作方法有效，再推动企业将其变成正式应用。
ADP与WorkBuddy的双向打通，让这种路径在技术上更容易实现。
员工可以先通过WorkBuddy探索和验证工作方法，再将优质工作流转化为ADP智能体，纳入企业统一管理；企业已有的Skill，也能够回到员工的日常办公入口继续使用。
AI能力开始从个人经验沉淀为组织资产。由此，企业Agent平台的竞争，还延伸到企业内部AI能力的组织与运营。
四、从1个项目到1000个项目，腾讯云能否抢下企业AI的生态位？
如果说前面三个问题，分别对应企业Agent如何进入业务、产生经营价值和实现统一管理，那么接下来还有一个更艰巨的挑战：如何把一个成功项目复制到更多企业？
企业Agent越深入真实业务，越需要专业工程团队进入客户现场。这也是FDE（前沿部署工程师）模式重新受到关注的原因。
FDE是海外企业软件行业已有多年的工作方式。工程师直接进入客户的业务环境，与业务团队一起确定问题、接入系统并开发解决方案。
AI兴起后，FDE在企业Agent交付中开始扮演更重要的角色。
2026年10月，澳大利亚软件龙头Atlassian披露，其FDE团队已经与超过100家企业客户合作，推动80多个AI Agent进入生产环境。
FDE受到重视，恰恰是因为真实企业往往无法仅凭产品说明书就完成Agent部署。
业务数据可能分布在不同系统，权限规则与实际工作流程也未必一致，很多问题只有工程师真正进入业务现场后才会暴露出来。
但定制化交付也存在一个悖论：越依赖专家，越难降低交付成本。
2026年9月，Gartner提出一项预测：到2028年，70%的企业可能放弃由供应商FDE团队构建的Agent，主要障碍包括成本增加、独立维护困难。
这项预测也提醒行业，一个项目如果只能依靠最初的实施团队维持运行，就很难形成可持续的商业模式。
企业Agent需要的不只是完成一个项目，还要将项目中的经验沉淀下来，让后续客户能够以更低成本复用。
吴永坚告诉我们，腾讯云并不希望把企业Agent交付变成单纯的项目或人力外包生意。他打了一个比方：过去做企业Agent项目，有点像经营一家私房菜馆，每个客户来了，都需要厨师按照具体需求定制。但如果希望把生意做大，就需要根据过去服务大量客户的经验，逐渐沉淀出可以复制的菜品。
“只有复制效率提升了，我觉得这个生意才能够跑得转。”吴永坚说。
为此，腾讯云正在尝试以“FDE＋ADP＋场景包”突破定制化交付的困局。
在此次大会上，腾讯云推出FDE认证，并进一步升级ADP开放接口及ADP CLI，推动企业、开发者和生态伙伴参与智能体建设与交付。
除此以外，腾讯云还将过去几年积累的项目经验，进一步沉淀为一套可以交给合作伙伴使用的交付方法论。
这套FDE体系覆盖五个阶段：接入期，打通系统和数据，明确权限与安全边界；建设期，组合Claw Agent与Workflow，形成业务方案；验证期，通过真实场景检验效果、把控风险；推广期，推动Agent从试点走向规模化应用；治理期，持续管理知识、效果、成本与安全。
每个阶段都配有相应的任务、交付物和验收标准，以及模板、评测用例和操作手册，让合作伙伴能够按照相对统一的流程完成交付。
在此基础上，腾讯云还将过往项目经验归纳为覆盖零售、制造、教育、金融、医疗五大行业的25个服务场景包。
与此同时，腾讯云还展示了智能客服、视觉智能体等垂类应用的进展，尝试将更多成熟能力沉淀在ADP平台上。
目前，ADP已经覆盖千余个公有云应用场景、百余个私有云应用场景，并在零售、金融、医疗、工业等行业沉淀出可以复制的场景包。
这些场景背后有不少共性需求。例如，不同企业都需要安全连接内部系统、管理Agent权限、监控运行情况，以及根据业务反馈调整应用。
如果能够通过平台统一提供这些能力，工程师便不必在每一个项目中重新解决相同的问题。但行业场景包并不意味着所有企业都可以直接复制。所以，FDE和企业内部团队仍然积极参与适配。
在一些业务流程相似、标准化程度较高的场景中，这种模式已经展现出规模化部署的潜力。
例如，基于ADP打造的酒店智能体管家“华小AI”，通过38条工作流覆盖152个客服子场景，已经在超过10000家门店落地，自动处理超过70%的高频客服问询。酒店门店之间存在大量相似的客服需求，智能体可以复用共性工作流，再根据不同门店的具体需求进行调整。
这类可以在大量门店重复部署的场景，更容易获得规模化优势。
因此，腾讯云此次推进FDE生态建设，也是在尝试利用ADP和行业场景包，让工程师不必在每个项目中重复解决相同的问题。
FDE解决进入现场的问题，平台化降低重复开发成本，行业场景与生态伙伴则支撑进一步复制。这也是企业AI护城河可能逐渐形成的地方。
对于中小To B AI公司而言，开发相似的模型应用或产品功能，未必是无法跨越的技术难题。但进入企业核心业务系统、获得客户信任、积累跨行业经验，并建立覆盖开发、交付和持续运营的伙伴网络，都需要更长时间的投入。
吴永坚也提到了类似的判断。在他看来，模型以及围绕模型的基础技术能力固然重要，但企业Agent落地还有大量“最后一公里”的问题。如何把这些问题逐一解决，并将解决方案沉淀进产品，再通过合作伙伴规模化交付出去，难度并不低。
他将这一过程比作从做好一道菜，到能够经营上万家门店。前者考验单个产品的能力，后者则需要一整套能够长期稳定运行、持续复制的体系。
企业AI的护城河在于持续积累的客户连接、行业经验与交付生态。
腾讯云利用已有的企业生态和行业落地基础，正在争取先发位置。这条路径能否形成规模优势，将决定腾讯云在企业Agent市场中建立的护城河究竟有多深。
结语：模型之外，企业AI的护城河才刚刚开始
过去，企业购买软件，是为了让员工更高效地工作。如今，Agent开始直接连接业务系统、执行任务，人的角色逐渐转向设定目标、审核决策和处理异常。
但企业可以把工作交给AI，却不能把责任一并交出去。
腾讯云此次ADP升级，正是围绕这些问题展开：
- 向下，通过安全连接、权限隔离、运行监控和反馈优化，让Agent能够进入业务系统并持续运行；
- 向上，通过AgentPortal 2.0统一管理和分发智能体，打通WorkBuddy，让个人工作流能够沉淀为企业可复用的服务；
- 向外，则借助行业场景包与FDE生态，尝试降低不同企业重复开发和交付的成本。
这也反映出企业Agent行业正在发生的一种变化：竞争的重点正从让Agent具备更多能力，转向如何将这些能力组织起来，形成能够长期运行的企业服务。
腾讯云此次围绕ADP展开的一系列探索，正试图在模型能力之外，为企业AI的规模化落地筑起第一道护城河。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Intellectual (WeChat Official Account:zhidxcom)
I'm sorry.
Edit Queen
At the Wigand Trust, the starting point for a bond transaction may be just a chat.
Dealers are required to receive instructions in thousands of trade groups, identify the elements of the transaction and re-enter the system. When clients ask about holdouts and the progress of their transactions, they also open a system query and send the results back to the group.
This work is cumbersome and leaves no room for error.
Today, the Wigand Trust began to get a AI agent to take over some of these tasks. It goes directly into the transactional cluster, identifies and dismantles the transactional instructions, enters the system, and can also search for hold and transactional progress and respond automatically to queries in the group.
Zhu Hua Jun Trust Sunan
According to Tencent clouds, the bond assistant based on the AI agent development platform ADP has increased the accuracy of the breaking order from over 80 per cent to 98 per cent.
However, at the key nodes, the human confirmation remains in the Wigand Trust.
Behind this is actually a non-consensus of the industry: LLM constantly shows how many complex tasks it can accomplish, but the model is different from the one that companies are willing to implement.
The view was expressed that continued progress in modelling capacity would eventually level the technological thresholds for the landing of corporate AI.
This is not necessarily the case. In August 2026, the McKinsey publish Global AI Survey: 80 per cent of respondents stated that AI had improved individual work efficiency, but only 37 per cent believed that AI had made a positive contribution to the profits of the enterprise in which it was located. In September, McKenzie further mentioned that enterprises consume a significant amount of time for task transfer, coordination and waiting between different sectors and systems.
It is therefore difficult for BAI to break through the entire enterprise production system outside the model.
This is also a challenge for a number of small and medium-sized To B AI companies with focused models or single-point applications. Model capabilities and product functions can evolve quickly, but entering the enterprise ' s core system, gaining customer trust and establishing delivery systems that cover different industries requires longer periods of input and accumulation.
On 30 October, at the Tencent Global Digital Ecology Congress, Tencent clouds upgraded the AI agent development platform ADP, publishAgentPortal 2.0 and announced a two-way link with WorkBuddy.
The upgrade also resulted in the Tencent cloud further clarifying the location of ADP ServiceAgent, focusing on business data and business scenes, and promoting Agent ' s in-depth business processes to create real business value.
From business system access to AI agent integrated management, to industry scenes and delivery ecology, Tencent clouds are trying to pre-empt the strategic location of Agent ' s size.
Can these explorations, outside the model, create a corporate moat for Tencent clouds that is more difficult to quickly replicate by latecomers?
I. Is AI going into core business? The company Agent has to cross management.
From the understanding of the directive to the execution of the transaction, the first solution for the enterprise Agent was permission and liability.
It is not difficult to understand the reasons why the human confirmation and the operating records are still being kept by the Wigand Trust Bond Assistant. A financial AI agent can search for customer hold, not to imply that it has access to the data of all clients; or that Agent, who handles business, has access to the system interface, or that it has the right to complete all operations.
Enterprise management Agent, the first thing to do is to sort out business permission.
In response to this need, the Tencent cloudADP has upgraded its capacity to secure private network connectivity, tenants and permission isolation, supporting different Agents to access databases according to permission, in-house resources such as ERP, and to provide integrated enterprise system connectivity to different modules such as sandboxes, workflow, MCP, connectors, etc.
This is also true in Yunnan, where employees are required to consult internal systems, pharmaceutical information and medical information on a day-to-day basis, but the users of this knowledge differ from permission requirements.
To this end, Yunnan White Medicine relies on ADP and pleasure to fine-tune permission management to employee ID. The employee asked questions through a “white nail” entry point, and AI agent automatically determined the type of problem, route to the corresponding knowledge base and returned to the content according to the actual permission of the employee.
Yabin, Yunnan White Medicine Party
For employees, it is a question and an answer; but behind it is a multiple knowledge base and a set of permission rules that work together.
With the exception of permission, it is equally impossible for enterprises to pursue full automaticity, and different business requirements for Agent ' s autonomy vary.
For example, standard approval, worksheet processing, etc. with fixed rules, often requires a clear sequence of implementation and predictable results; cross-sectoral analysis, business decision support, etc., may require multiple AI agent division of labour collaboration.
To this end, Tencent CloudADP offers four models of Claw, Standards, Multi AI agent, and workflow, enabling enterprises to choose different AI agent construction and implementation modalities according to actual business needs.
The Wigand Trust uses a combination of Claw to handle autonomous planning tasks, using Workflow to implement well-defined business processes, while key trading links are always identified by traders.
This also echoes industry's judgement on the governance of Agent.
In May 2026, Gartner, a global authority research institute, warned that a single approach to business governance for all Agent could result in two outcomes: simple tasks being over-restricted, while high-risk tasks received too much permission.
And when Agent actually enters the production environment, management does not stop on the line.
Traditional software usually operates according to predetermined procedural rules, but model outputs do not have the same degree of certainty. The success of a mission does not mean that Agent will be able to perform similar tasks in the long term and in a stable manner.
In particular, following the start of the business ' s continuous operation of Agent, tracking each operation, monitoring costs, positioning failures, and the timely correction of problems becomes a new management requirement.
This time Tencent CloudADP enhanced its operational records, cost and status monitoring, trouble positioning and version roll-back. At the same time, ADP could use AI to analyse real business feedback and make recommendations for AI agent optimization, which would be validated manually and would allow Agent to be continuously evaluated and adjusted during its operations.
Enterprise management Agent should enable it to carry out its mandate on an ongoing basis under clear operational rules. Enterprise-level engineering and governance capabilities are becoming the first layer of technical barriers beyond the model. This upgrade of Tencent CloudADP revolves around the management needs of these production environments.
ii. Why is it that businesses are not doing the same "fast-out" work?
Business is not a simple addition to a group of employees. The fact that everyone saves one hour does not mean that the company spends as much time on a business as it does on a business.
The first problem with Enterprise AI is to break down the work of employees into a separate task.
In the past two years, AI has been the first to enter individual workspaces such as document writing, data analysis, etc. One common feature of these efforts is that the mission boundaries are relatively clear, and the staff themselves can judge the results to be good or bad.
However, the business of the enterprise does not operate in the same way. A mission often involves a cross-section of responsible persons, regional teams and front-line personnel, moving between different departments, personnel and operational systems.
If only one link is completed faster, the efficiency advantage of AI may be consumed by the original process.
As a result, the second issue that has been extended to Enterprise AI is that of handover and waiting, slowing down the entire business process.
In September 2026, McKenzie analysed the production process of a wealth 500 industrial enterprise in a study and found that most of the time spent on the entire business process was at the next stage of waiting for a mission.
McKinsey calls it “Coordination Interface Tax”.
For large enterprises, the optimization of a single step does not necessarily address the issue of collaboration between different teams and systems. This explains why employees are using AI more and more frequently, but the overall efficiency of the enterprise and business returns are not necessarily growing simultaneously.
The enterprise AI solution is no longer limited to enabling each employee to do his or her job more quickly, but also includes real entry into business processes and linking different links.
Another case is the exploration of Tencent clouds in cooperation with Ely. Ely brings together regional sales, lead-purchase performance and market data, supported by AI agent to analyse business performance, develop strategies and extend related capabilities to the first-line lead-purchase scene.
In the process, WorkBuddy undertook part of the work of business staff using AI and conducting business analysis; the purchase of AI agent, built by ADP, went further into the consumer-oriented sales scene and generated product referral and communication skills based on client demand.
According to the Tencent cloud, this portfolio has resulted in efficiency improvements of more than 50 per cent for the lead purchase team. At the same time, the lead purchase of AI agent resulted in a 26 per cent increase in sales orders in the A/B test. The former represents internal efficiency, while the latter directly touches on business results.
The value of AI is no longer only reflected in the low level of work done by the staff, but may also be reflected in the fact that the enterprise has completed more business.
However, moving from an individual workflow to an enterprise-level service also requires that AI competencies be reused across systems and people.
This is also the core difference between Personal Agent and Service Agent. Vice President Tencent Yun U Yongjin told us that Personal Agent is more personal, and Service Agent needs to be on the corporate side, considering permission for different players, security, and how to build the capacity that can be shared and managed.
Vice President Wu Yongjin, Tencent Yun
For example, staff can analyse sales data through their own office entrances, but business strategies cannot be automatically entered into the company-wide sales process by a single AI dialogue. Enterprises also need to complete permission validation, capacity distribution, task execution and impact feedback.
The Tencent cloud, which connects ADP and WorldBuddy in two directions this time, is an attempt to reduce the distance between individual AI use and business performance.
On the one hand, Agent and Skill, which are built through ADP, have access to WorkBuddy for direct calls by employees in their day-to-day office; on the other hand, workflow, which is explored and validated by employees through ADP, can be transformed into ADPAI agent, which is integrated into a corporate management and distribution system.
This means that the effective working methods accumulated by individuals have the opportunity to further become the AI capacity that businesses can reuse without having to stay in the computer of an employee for long periods.
Individual Agent addresses the use of AI by employees, while service Agent further brings AI into business processes. Where there is real potential for greater gains for BAI, it may happen that the relationship between these steps is reorganized.
iii. Agent's making it easier, and the company's worried about `Agent'.
The lower development threshold of AI agent may bring businesses back to the age of software: there are too many Agents.
Different departments have their own knowledge base, processes and business needs and may build AI agent separately. Each of the Agents is worth looking at alone, but when the number increases, the enterprises may not know which of the Agents is running, who is responsible for maintaining, which functions are duplicated, and how many permissions they each have.
In April 2026, Gartner referred to the phenomenon as Agent Sprawl (AI agent spread). The Agency ' s survey at that time showed that only 13 per cent of the organizations interviewed felt that they had the appropriate Agent governance mechanisms.
Businesses may have to face new management burdens before they can take full advantage of AI.
In response, the Tencent cloud adopted a “two-pronged” programme: ADP lowers the Agent development threshold; and AgentPortal 2.0 addresses the issue of the uniform use and management of AI agent.
Among them, the ADP upgrade involved two important changes: one from dialogue to AI agent, to AI to the whole development process; and the second from AI agent to the construction and deployment of full operational applications. ADP is upgrading from "AI agent Builder Specialist" version 4.0 to "AI agent Life Cycle Specialist".
In the past, operational staff may need to submit their needs to the technical team and developers will have to complete AI agent configurations, systems integration and application interfaces. Part of the construction can now be done in the same workspace, and configuration, preview, debugging and evaluation can be linked.
AgentPortal 2.0 further addresses the issue of AI agent harmonization of use and management, allowing enterprises to integrate Agent, which is built in different sectors, into the integrated portal while supporting centralized management of ADP and third party Z42Z.
Employees can submit business needs directly, and the system automatically selects the appropriate AI agent; the enterprise can also distribute Agent by department, function, and manage up and down and using permission.
For complex projects requiring long-term collaboration, multiple AI agents can form teams, coordinated by project manager AI agent; for direct consumer service operations, multiple professional Agents can also come together behind a single service portal.
As a result, enterprises do not have to allow their employees to find and access different AI tools, but rather to work with multi-AI agent through a single access point.
At the same time, a new set of AI-building approaches, distinct from the software age of the past, is taking shape.
In the past, enterprise software was usually procured, deployed by the IT sector and required learning and use by the business sector. However, with the spread of the individual Agent, front-line employees may first discover that some of the AI working methods are effective and then promote their formal application.
The two-way connection between ADP and WorkBuddy makes this path technically easier to achieve.
Employees can first explore and validate working methods through WorkBuddy, and then transform high-quality workflow into ADPAI agent into integrated corporate management; the existing Skyll of the enterprise can also continue to be used at the staff ' s daily office entrance.
AI capacity began to sink from personal experience into organizational assets. As a result, competition from the Enterprise Agent platform also extends to the organization and operation of AI capabilities within the enterprise.
IV. Can the Tencent cloud take the ecological position of Enterprise AI from one to 1,000 projects?
If the first three questions are one of how Agent gets into business, generates business value and achieves integrated management, then there is a more daunting challenge: how to replicate a successful project to more enterprises.
The more the enterprise Agent goes into real business, the more professional engineering teams are required to access the client site. This is also why the FDE (Front Deployment Engineer) model has received renewed attention.
FDE has been working in the software industry of overseas enterprises for many years. Engineers enter the client ' s business environment directly, working with business teams to identify problems, access systems and develop solutions.
With the rise of AI, the FDE began to play a more important role in the delivery of the Enterprise Agent.
In October 2026, Australia ' s software champion Atlassian disclosed that its FDE team had worked with over 100 corporate clients to facilitate the entry of more than 80 AI Agent into the production environment.
FDE is valued precisely because real enterprises are often unable to complete Agent deployment on the basis of product specifications alone.
Business data may be distributed across different systems, and the permission rules may not necessarily be consistent with the actual workflow process, and many problems will only be revealed when engineers actually enter the business site.
However, there is also a paradox to customizing delivery: the more reliance on experts, the more difficult it is to reduce delivery costs.
In September 2026, Gartner presented a projection that by 2028 70 per cent of enterprises might abandon Agent, which was built by the supplier FDE team, with major obstacles including increased costs and difficulties in independent maintenance.
This projection also reminds industry that it is difficult to develop a sustainable business model if a project is to operate only on the basis of an initial implementation team.
The business Agent needs to not only complete a project, but also to settle down the experience of the project so that subsequent clients can be reused at a lower cost.
Wu told us that Tencent clouds do not want to turn the business Agent delivery into a simple project or manpower outsourcing business. He gave the example of a former business, Agent, sort of like running a private restaurant, where every customer came and needed a cook to tailor it to his needs. However, if business is to be expanded, it will be necessary to gradually sink replicable vegetables, based on past experience in serving a large number of clients.
"Only if the replica efficiency improves, I think the business will be able to make a difference." Go said.
To this end, Tencent clouds are trying to break down the delivery of customised delivery with the FDE+ADP+ scenario package.
At the conference, the Tencent cloud launched the FDE certification and further upgraded the ADP open interface and the ADP CLI to promote the participation of business, developers and ecological partners in the construction and delivery of AI agent.
In addition, the Tencent cloud further sanctifies the project experience accumulated over the past few years into a delivery methodology that can be entrusted to partners.
The FDE system covers five phases: access, access to systems and data, and identification of permission and the security boundary; construction, a combination of Claw Agent and WorldFlow, to form an operational programme; validation, to test effects and manage risks through real scenes; roll-out, to drive Agent from pilot to scale applications; and governance, to continuously manage knowledge, effectiveness, cost and security.
Each phase is accompanied by the corresponding tasks, deliverables and acceptance criteria, as well as templates, evaluation examples and operational manuals to enable partners to deliver according to a relatively uniform process.
On this basis, the Tencent cloud also summarizes past project experience as 25 service scenarios covering five main sectors: retail trade, manufacturing, education, finance and health.
At the same time, the Tencent cloud also demonstrates the progress of ingenious applications such as smart passenger service, visual AI agent, and attempts to sink more mature capabilities on the ADP platform.
Currently, ADP has covered over 1,000 public cloud applications, over 100 private cloud applications, and has deposited replicable scenarios in retail, financial, medical, industrial, etc.
There are many common demands behind these scenes. For example, different enterprises need to secure access to internal systems, manage Agentpermission, monitor operations and adjust applications based on business feedback.
If these capabilities can be provided uniformly through the platform, engineers will not have to re-establish the same problems in each project. However, the industry scenario does not mean that all enterprises can copy directly. Therefore, FDE and the intra-enterprise team are still actively involved in adaptation.
In some scenarios where business processes are similar and more standardized, this model has demonstrated the potential for scale deployment.
For example, the manager of the ADP-based hotel, AI agent, “Ai Hua, China”, covered 152 passenger service scenes through 38 workflow, and had landed in over 10,000 shops, processing over 70 per cent of high-frequency client queries automatically. There is a large amount of similar demand for guest services between hotel stores, and AI agent can revert to common workflow and then adjust to the specific needs of the different stores.
Such scenarios, which can be replicated in a large number of shops, make it easier to obtain economies of scale.
Therefore, the Tencent cloud this time advanced the FDE ecological construction and was also trying to use the ADP and industry scenarios so that engineers did not have to repeat the same problems in each project.
FDE addresses access to the field, platformization reduces the cost of duplicate development, and industry scenarios and eco-partners support further replication. This is also where the Enterprise AI moat may be formed.
The development of similar model applications or product functions is not necessarily a technical challenge that small and medium-sized To B AI cannot cross. However, access to the core business systems of the enterprise, client trust, cross-practice experience and the establishment of a network of partners covering development, delivery and ongoing operations require longer inputs.
U Yongjin also referred to a similar judgement. In his view, while models and the underlying technological capabilities surrounding them were important, there were still a number of “last-kilometres” problems for the enterprise Agent to land. It is not too difficult to resolve these problems one by one and to sink solutions into products and deliver them through the scaling up of partners.
He compared the process with the preparation of a meal to the ability to run a million stores. While the former test the capacity of individual products, the latter require a set of systems that can operate in a stable and continuous manner over time.
Enterprise AI ' s moat is based on a continuous accumulation of customer connections, industry experience and delivery ecology.
The Tencent cloud is seeking to pre-position itself, taking advantage of the existing ecological and industrial foundations of the enterprise. Whether or not this path has the advantage of size will determine how deep the moat created by the Tencent cloud in the company Agent market is.
Conclusion: outside the model, the corporate AI moat is just beginning.
In the past, companies had purchased software to enable their employees to work more efficiently. Today, Agent has started to connect directly to the operational system and to implement its mandate, and the role of the human being has gradually shifted to targeting, vetting and handling anomalies.
However, businesses can hand over their jobs to AI, but they cannot hand over responsibilities together.
The Tencent cloud has been upgraded to the ADP on these issues:
- Downwards to enable Agent to access and sustain operational systems through secure connectivity, permission isolation, operational monitoring and feedback optimization;
- Upward, the integrated management and distribution of AI agent through AgentPortal 2.0, and access to WorkBuddy to enable the individual workflow to sink into a reusable service;
- Outwards are being used to try to reduce the costs of duplication of development and delivery between different enterprises through the industrial scenarios and the FDE ecology.
This also reflects a change that is taking place in the Agent industry: the focus of competition is shifting from making Agent more capable to organizing these capacities into long-term business services.
The Tencent cloud, a series of explorations around the ADP, is trying to build the first moat for the size of the Enterprise AI beyond model capabilities.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.