在大模型在拼命刷 Benchmark 卷编程的时候,「不能说话」的 Jev 却意外成了 AI 圈的新顶流,究其原因,大模型应用的价值正在从「生成什么」,进一步延伸到「能否理解复杂信息,并据此做出可靠判断」。

放到企业办公情形里,这个难题尤为显著。
员工的一句「帮我整理报价」,背后可能同时牵涉客户历史、库存状态、折扣规则、审批权限和部门协作。AI 如果只会生成文档,却不了解这些业务背景,输出再完整,也很难真正进入公司的工作流。
企业真正需要的,是一个可以理解业务上下文、知道该和谁协作、也清楚自己权限边界的 AI。

就在刚刚,千问办公在杭州云栖大会发布企业上下文(Enterprise Context)、应用、QwenNote、安全中心等能力和机型。核心目的就是要为 Agent 进入企业业务流程给出一套完整的兼容体系。
阿里巴巴集团副总裁、千问办公 CEO 陈宇森表示:「千问办公是业内首个企业级 Agent 机型,Agent 要在企业真实情形落地,既要懂业务,也要像人一样在企业和组织内协同,且每一次任务执行要有边界、能追溯。」

「懂业务、能协同、信得过」,三个要求对应着企业引入 AI 后必须回答的难题,涵盖它了解多少业务、如何与同事配合,以及可以被授予多大权限。
企业 AI 的第一课,是理解上下文
撰写报告、整理表格已是办公 Agent 的常见能力,但企业还有自己的业务规则。
同样是「给客户报价」,不同企业的机型体系、折扣权限和审批流程各不相同。即使在同一家公司,三个月前的会议结论也可能已经失效,群聊里的临时讨论未必代表最终决定。
一个 Agent 在不同公司能做成什么,很大程度上取决于它对业务上下文的了解。
过去的数字化建设,让企业把大量工作记录保存在平台里;Agent 进入工作流,则要求企业进一步整理记录之间的关系,涵盖谁负责项目、哪个版本已经生效,以及任务开展到了哪一步。
此前,千问办公团队已推出个人上下文项目 MyContext,它可以把海量异构的个人工作数据加工成 Agent 能直接使用的上下文,让 Agent 更懂工作也更懂使用者,提高 Agent 在工作情形下的表现。
此次发布的企业上下文管理机型企业上下文( Enterprise Context),进一步把这一能力延伸到企业级情形。

它可以将分散在群聊、文档、知识库等位置的信息整理成结构化的企业上下文,按任务给出相关内容,并随着业务变化持续更新。按任务组织上下文,有助于减少无关信息干扰,也让积累多年的资料在需要时被正确调用。
古茗的门店运营就是一个很贴切的案例。新品频繁上线,意味着制作要求、物料规范和培训安排也在不断变化,一线员工需要从多个渠道找到与当前工作有关的最新说明。

现在,Enterprise Context 将古茗飞书文档、知识库、答疑群和培训日程中的内容整理为「门店运营知识空间」。店员询问海报张贴规范时,Agent 调取相应信息,并随标准更新维护知识空间,减少员工翻找资料和反复询问的时间。
据介绍,基于 Enterprise Context 和千问办公的 Agent 托管能力,千问办公还可以协助企业构建懂业务的数字员工,让他们在钉钉等各类 IM 上和人类员工开展协同,重塑业务流程,大幅提高组织效率。

企业还有大量信息产生在线下,会议讨论和客户拜访中的要求,同样会影响 Agent 的任务执行准确率。
此次发布的 Agent 硬件 QwenNote 及首款机型 A2,兼容语音转写和语音调用千问办公,把线下交流转化为可供后续工作使用的记录;其转写内容仅保留文字和纪要,不保留原始音频。

下一轮企业数字化,需要把公司整理成 Agent 看得懂的公司。 过去依赖老员工记忆、跨部门询问才能解释清楚的业务关系,需要逐步变成可以维护、查询和授权使用的组织知识。
给 Agent 一张组织里的工牌
理解业务之后,Agent 还要找到自己在组织里的位置。
一份报价从销售流转到财务、采购和法务,会经过多个岗位,每个环节都有自己的输入和职责。企业效率的提高,很少来自某一个人突然快了十倍,更多来自任务在不同岗位之间可以连续流转。
企业可以基于千问办公和钉钉构建数字员工,相当于给 Agent 发了一张组织里的工牌。
它有名称、所属部门、负责人和岗位职责,也有明确的授权范围与生命周期,可以被员工识别、被平台授权,也可以被管理员停用,所有执行记录都能追溯到对应身份。
钉钉等协作平台决定数字员工可以进入哪些群聊、访问哪些平台,Enterprise Context 为其给出完成工作所需的业务信息,Managed Agents 则负责管理它的工作方式。
企业可以在平台中设定 Agent 的目标和边界,规格模型、指令与工具,并授权运行环境和凭证,让不同岗位的数字员工按照各自职责持续工作。
仓储物流包装公司富立卡已经将这种思路用于销售流程。其销售团队每人每天需要处理 300 多条客户询盘,同时与六个部门协作。

千问办公为其规格了三个数字员工,借助 Enterprise Context 共享同一笔业务当前需要的信息,让询盘、报价和合同沿着同一条流程推进。Agent 在这里承担的已经超出一个孤立任务,逐步成为组织分工中的一个岗位。
数字员工进入公司以后,协作关系也会扩展到人与 Agent、Agent 与 Agent 之间。
千问办公推出「协作」能力,让同事与数字员工在同一空间内,围绕项目资料讨论、分工,并共同保留进度和产出。人与 Agent、Agent 与 Agent 之间可以共享必要背景,后续参与者也能沿着已有记录继续工作。

当然,任务最终还要回到业务平台,因而新推出的「应用」能力,允许企业和 SaaS 厂商在保留原有页面与数据的基础上引入 Agent。

它可以了解使用者当前的操作及进度,协助完成任务;以投研为例,研究员可以在应用内完成分析、核对和修改,再将结果保存到业务平台。
旧平台不用推倒重来,新能力见缝插针——这种「不折腾」,本身就是企业级机型最被低估的美德。
把这些能力放在一起,才能看清企业级 Agent 与通用办公 Agent 的分水岭。过去半年,行业竞争大多围绕写文档、做表格、生成 PPT 和操作电脑展开,随着主流机型陆续补齐能力,这些能力正在迅速成为标配。
真正进入企业的 Agent 需要读懂公司的上下文,带着明确身份与人协作,接入原有平台,并把结果交给下一个岗位。企业需要的也随之从一件通用办公工具,变成一套可以承接真实业务流程的组合思路。
当然,能力越深入业务流程,Agent 取得的权限越大,可能产生的风险也越接近企业核心。
Agent 时代新的组织语言
当 Agent 可以帮企业查询客户、生成合同,甚至直接调用业务平台,企业衡量它的标准也会发生变化。内容浮现偏差还可以修改,数据越权访问或错误操作已经执行,则可能直接影响真实业务。
陈宇森表示,企业不能只在平台外围增加防线,而要从架构外观之初,就把权限越界、数据泄露和误操作等 Agent 风险考虑进去。
飞行前的检查单就是一个类似的例子。
1935 年,波音 299 型轰炸机原型机试飞时坠毁,调查结论让人哭笑不得:飞机没有任何故障,是飞行员忘了解锁升降舵和方向舵的控制机构。后来的应对思路土得掉渣,一张飞行前检查单。
这张纸片后来救了无数人的命,也让复杂机器的操作权,一步步交到了普通人手里。双发飞机今天能跨大洋飞行,靠的也是几十年里每一趟航班的可靠记录,一点点攒出来的信任。
此次发布的千问办公安全中心,也是同样的道理。
它统一管理数据访问、工具调用和任务执行:企业可以借助沙箱限定运行环境,对高风险操作设置审批和拦截,并在异常发生后依据操作记录与恢复机制处理难题。
结合企业权限体系,管理员还可以按岗位授予必要权限,识别敏感信息并审计执行过程,追溯任务由谁发起、使用了什么数据、开展了哪些操作。数字员工的岗位职责,由此对应到可检查的访问和执行记录。
业务负责人可以检查执行依据、在必要时接管任务,才更有条件把涉及真实客户和内部平台的工作交给 Agent,信任需要借助每一次可追溯的执行逐步建立。
微软《2026 年工作趋势指数》里面有个细节值得注意,在其调查的 AI 使用者中,66% 表示 AI 让自己能投入更多时间从事高价值工作,58% 表示可以完成一年前无法完成的工作;报告同时指出,许多组织的工作方式尚未跟上员工能力的变化。
行业数据也在佐证同一个拐点。Gartner 预测,到 2026 年底,40% 的企业应用将嵌入特定任务的 AI Agent,而 2025 年这个数字还不足 5%。
无论国内还是海外,头部模型厂商近来也纷纷把 Agent 装进自己的办公生态。风向已经很清楚了:竞争的主线正在从模型能力的比拼,转向企业组织生产力的提高。
千问办公此次升级关注的,正是个人生产力提高之后,怎样继续传导到整个组织。据团队介绍,相关机型当下已向古茗、富立卡、长安汽车等企业开放邀请测试。
钉钉和 QwenNote 硬件入口负责沉淀企业每天产生的沟通、会议和文档, Enterprise Context 负责把这些信息整理成可调用的组织上下文,千问办公则在此基础上执行任务,并把结果回流到协作情形。
由此,千问办公也协助企业把协作入口、上下文处理、Agent 执行, 并与多种设备连接成一套完整的企业 AI 体系。

四十年前,微软用 Office 定义了 PC 时代的办公,文档、表格和幻灯片由此成为全球通用的生产力语言。
今天,当数字员工开始领工牌、进群组、处理任务和参与流程,「办公」的对象也从文件延伸到了人类与 Agent 协作的组织本身。
在 APPSO 看来,企业 AI 的长期价值,要看它做过的工作,能为下一次工作留下什么。
文件是 PC 时代的工作语言,在线协作是移动互联网时代的工作语言,而千问办公正在押注的企业上下文,也将会成为 Agent 时代新的组织语言。
让 AI 读懂世界,考验的是模型具备多少知识;让它读懂一家公司,考验的则是它能否理解那些藏在群聊、流程和人情世故里的言外之意。千问办公瞄准的,恰好是后一个更接近真实工作的命题。
本文来源:https://www.ifanr.com/1681363
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
While large models are frantically brushing benchmarks and competing in programming, the "mute" Jev unexpectedly became the new top of the AI circle. The reason is that the value of large model applications is extending from "what is generated" to "whether complex information can be understood and reliable judgments can be made based on it".
In the context of enterprise office, this problem is particularly significant.
An employee's sentence "Help me organize the quotation" may involve customer history, inventory status, discount rules, approval authority, and department collaboration at the same time. If AI only generates documents but does not understand these business backgrounds, no matter how complete the output is, it is difficult to truly enter the company's workflow.
What enterprises really need is an AI that can understand the business context, knows who to collaborate with, and is also aware of its own authority boundaries.
Just now, Qianwen Office released enterprise context (Enterprise Context), applications, QwenNote, security center and other capabilities and models at the Hangzhou Yunqi Conference. The core purpose is to provide a complete compatible system for the agent to enter the enterprise business process.
Chen Yusen, Vice President of Alibaba Group and CEO of Qianwen Office, said: "Qianwen Office is the first enterprise-level agent model in the industry. To land in real enterprise scenarios, the agent must not only understand the business, but also collaborate within the enterprise and organization like a human, and every task execution must have boundaries and be traceable."
"Understand the business, be able to collaborate, be trustworthy", the three requirements correspond to the problems that enterprises must answer after introducing AI, covering how much it knows about the business, how to cooperate with colleagues, and how much authority it can be granted.
The first lesson for enterprise AI is to understand the context.
Writing reports and organizing tables are common capabilities of office agents, but enterprises also have their own business rules.
For the same "send a quote to the customer", different companies have different model systems, discount permissions, and approval processes. Even within the same company, the conclusions of a meeting three months ago may have expired, and temporary discussions in the group chat may not represent the final decision.
The extent to which an agent can succeed in different companies depends largely on its understanding of the business context.
Previous digitization efforts have allowed companies to store a large amount of work records in the platform; agents entering the workflow require companies to further organize the relationships between these records, including who is responsible for the project, which version is effective, and what stage the task has reached.
Previously, the Qianwen Office team had launched the personal context project MyContext, which can process a large amount of heterogeneous personal work data into context that agents can directly use, making agents more intelligent in work scenarios and more familiar with users, thereby improving the performance of agents in work scenarios.
This time, the enterprise context management model enterprise context (enterprise context) was released, which further extends this capability to the enterprise level.
It can organize information scattered in group chats, documents, and knowledge bases into a structured enterprise context, providing relevant content according to tasks and continuously updating as the business changes. Organizing context by task helps reduce interference from irrelevant information, and also makes many years of accumulated materials correctly called when needed.
Gumeng's store operation is a very pertinent example. Frequent launch of new products means that production requirements, material specifications and training arrangements are also constantly changing. Front-line employees need to find the latest instructions related to their current work from multiple channels.
Now, Enterprise Context organizes the content in Gumeng Feishu's documents, knowledge base, Q&A group, and training schedule into a "store operations knowledge space". When employees inquire about poster posting specifications, the agent retrieves the corresponding information and maintains the knowledge space as the standards are updated, reducing the time employees spend looking up materials and asking questions repeatedly.
According to reports, based on Enterprise Context and Qianwen Office's agent hosting capabilities, Qianwen Office can also help companies build business-savvy digital employees, allowing them to collaborate with human employees on various instant messaging platforms such as DingTalk, reshaping business processes and significantly improving organizational efficiency.
Enterprises also generate a lot of information offline. Requirements during conference discussions and customer visits will also affect the accuracy of the agent's task execution.
The QwenNote agent hardware and its first model A2 released this time support voice transcription and voice-calling Qianwen Office, converting offline communications into records that can be used for subsequent work; the transcribed content retains only the text and minutes, not the original audio.
The next round of enterprise digitization requires the company to be organized into a company that the agent can understand. Business relationships that previously required relying on old employees' memories and inter-departmental inquiries to explain clearly now need to be gradually turned into organizational knowledge that can be maintained, queried, and authorized.
Give the agent an ID card in the organization
After understanding the business, the agent needs to find its place in the organization.
A quote goes through multiple posts from sales to finance, to procurement, and to legal, each with its own inputs and responsibilities. Enterprise efficiency improvements rarely come from a single person suddenly becoming ten times faster, but more often from tasks flowing continuously between different posts.
Enterprises can build digital employees based on Qianwen Office and DingTalk, which is equivalent to issuing an ID card to the agent within the organization.
It has a name, a department, a manager, and a job description, as well as a clear authorization scope and lifecycle. It can be identified by employees, authorized by the platform, and disabled by administrators, and all execution records can be traced back to the corresponding identity.
DingTalk and other collaboration platforms determine which group chats the digital employee can enter and which platforms it can access. Enterprise Context provides the business information required to complete the task, and Managed Agents is responsible for managing its working method.
Enterprises can set the agent's goals and boundaries on the platform, specify the model, instructions, and tools, and authorize the running environment and credentials, allowing digital employees in different positions to work continuously according to their respective responsibilities.
Fulika, a packaging company in the warehousing and logistics industry, has already used this idea in its sales process. Its sales team has to deal with 300 inquiries per day and collaborate with six departments.
Qianwen Office has configured three digital employees for them, sharing the same information required for the current business through Enterprise Context, allowing inquiries, quotations, and contracts to proceed along the same process. The agent here is no longer a single isolated task, but is gradually becoming a position in the organizational division of labor structure.
After the digital employee enters the company, the collaborative relationship will also extend between people and agents, and between agents and agents.
Qianwen Office launched the "Collaboration" function, allowing colleagues and digital employees to discuss project materials, divide labor tasks, and jointly maintain progress and output in the same space. People and agents, as well as agents and agents, can share the necessary background information so that subsequent participants can continue working based on existing records.
Of course, the task ultimately has to return to the business platform, so the newly released "Application" function allows enterprises and SaaS vendors to introduce agents while retaining the original page and data.
It can understand the user's current operation and progress, and assist in completing the task. For example, in the field of securities research, researchers can complete analysis, verification and modification in the application, and then save the results to the business platform.
The old platform does not need to be torn down and rebuilt, and the new function can be seamlessly integrated. This "non-interference" is the most underrated virtue of enterprise-level models.
Putting these capabilities together allows us to clearly see the dividing line between enterprise-level agents and general-purpose office agents. Over the past six months, industry competition has mostly focused on writing documents, making tables, generating PPTs, and operating computers. As mainstream models gradually fill these gaps, these capabilities are rapidly becoming standard.
Agents that truly enter enterprises need to read the company's context, carry out clear identities to collaborate with people, connect to existing platforms, and turn the results over to the next post. The needs of enterprises have also shifted from a general office tool to a combined idea that can support real business processes.
Of course, the deeper the capability penetrates into the business process, the greater the authority the agent acquires, and the closer the possible risks are to the core of the enterprise.
The new organizational language of the agent era
When agents can help enterprises query customers, generate contracts, and even directly call business platforms, the standards for measuring them by enterprises will also change. Content floating deviation can be modified, but data unauthorized access or erroneous operations that have already been executed may directly affect real business.
Chen Yusen said that enterprises cannot simply add a defense line to the periphery of the platform, but must consider agent risks such as permission overflow, data leakage, and operational errors from the beginning of the architecture design.
The pre-flight checklist is a similar example.
In 1935, the Boeing 299 bomber prototype crashed during its test flight, and the investigation conclusion was laughable: there was nothing wrong with the aircraft, it was the pilot who forgot to unlock the control mechanism of the elevator and rudder. The subsequent response was crude, a pre-flight checklist.
This piece of paper later saved countless lives and gradually handed over the control rights of complex machines to ordinary people. The fact that dual-engine aircraft can fly across the ocean today relies on the trustworthy records of every flight accumulated over the past few decades.
The Qianwen Office Security Center released this time has the same logic.
It uniformly manages data access, tool calls, and task execution. Enterprises can use a sandbox to limit the operating environment, set approval and interception for high-risk operations, and handle issues based on operating records and recovery mechanisms when abnormalities occur.
Combined with the enterprise permission system, administrators can grant the necessary permissions according to positions, identify sensitive information, and audit the execution process, tracing who initiated the task, what data was used, and what operations were carried out. The job responsibilities of digital employees are now mapped to auditable access and execution records.
Only when the business leader can check the basis for execution and take over the task when necessary can the work involving real customers and internal platforms be safely handed over to the agent. Trust must be built step by step based on every traceable execution.
A detail in Microsoft's 2026 Work Trends Index is worth noting. Among the AI users surveyed, 66% said that AI allows them to spend more time on high-value work, and 58% said they can complete tasks that were impossible a year ago. However, the report also points out that the way many organizations work has not kept pace with the changes in employee capabilities.
Industry data is also confirming this tipping point. Gartner predicts that by the end of 2026, 40% of enterprise applications will be embedded with specific task AI agents, while in 2025 this figure is less than 5%.
Whether domestic or foreign, leading model vendors are now putting agents into their office ecosystems. The trend is clear: the main line of competition is shifting from the comparison of model capabilities to the improvement of enterprise organizational productivity.
This upgrade of Qianwen Office focuses on how to continue to pass on personal productivity improvements to the entire organization. According to the team, related models have been opened for invitation testing by enterprises such as Gumeng, Fulika, and Changan Automobile.
DingTalk and the QwenNote hardware entrance are responsible for settling the communication, meetings, and documents generated by the enterprise every day. Enterprise Context is responsible for organizing this information into callable organizational context. Qianwen Office executes tasks on this basis and returns the results to the collaboration scenario.
In this way, Qianwen Office also helps enterprises connect the collaboration entrance, context processing, agent execution, and multiple devices into a complete enterprise AI system.
Forty years ago, Microsoft defined the PC era of work with Office, making documents, tables, and slides a universal productivity language worldwide.
Today, as digital employees begin to receive ID cards, join group chats, handle tasks, and participate in processes, the subject of "office" has also expanded from files to the organization where humans and agents collaborate.
In the view of APPSO, the long-term value of enterprise AI depends on what work it has done and what it can leave for the next time.
Files are the working language of the PC era, online collaboration is the working language of the mobile internet era, and the enterprise context targeted by Qianwen Office will be the new organizational language of the agent era.
Teaching AI to read the world is a test of how much knowledge the model has. Teaching it to read a company is a test of whether it can understand the implications hidden in group chats, processes, and human relationships. Qianwen Office happens to be aiming for the latter, which is closer to the real work proposition.