"In the past few months, every company has been working hard to prove that AI can do more and more work. But to be honest, just because one person can do the work, it does not mean that a group of people can collaborate."

An industry insider who has been paying attention to AI office for a long time found that after trying out the mainstream AI office products on the market, there are still some "trouble" that cannot be avoided. The most typical one is that some tasks have to be done manually.

"Agent can complete tasks independently, but it is difficult to handle the flow of tasks between people." Sometimes, after AI delivers the results, handing over between people takes more time and effort than the Agent doing the step itself.

If AI office wants to fight To B, it must face up to this dilemma.

When it comes to AI, companies really care about two things: one is personal productivity, whether AI can do more work for employees; the other is organizational productivity, whether AI can reduce information loss and duplication of work in team collaboration.

In the past few months, AI office has overcome many obstacles in personal productivity, but there is still a long way to go in terms of organizational productivity.

On November 9, Qianwen Office released version 2.0, launching the industry's first "collaboration" function that supports multi-person collaboration with Agents, and announced on its official blog that it is moving into the Team Agent era. It wants to respond to the above-mentioned problems: How can the enterprise's organizational productivity keep up with the improvement of individual productivity?

The domestic AI office craze has lasted for nearly half a year. Product competition in the C-end market is becoming increasingly fierce, while the challenges in the B-end market are far beyond simply selling products to companies.

From a product perspective, with the emergence of overseas Personal Agent products such as Muse, Cue, and Dots, the capabilities of personal AI are rapidly being broadened, and enterprise-level AI is also being pushed to higher expectations.

AI office manufacturers are facing a more severe test: when AI can become a personal work partner, can it go one step further and become part of team collaboration?

From the customer side, companies also have very realistic demands: How can the purchased AI office seats truly become organizational productivity?

Leifeng.com observed that the previous seat system was more like an access and budget tool, solving the problem of "who can use it and how much". When AI gradually becomes a digital employee, the value of the seat system also needs to be redefined: from configuring AI to employees to connecting employees to a human-machine collaboration system.

A relatively intuitive feature of Qianwen Office 2.0 is that it supports collaboration between multiple people and the same Agent.

Leifeng.com used this function to simulate a process in which reporters, senior editors and Agent (Qianwen Xiao Q) discuss FDE topic selection together.

After receiving the instruction to "plan FDE-related topics based on existing interview materials", Qianwen Xiao Q automatically read the four interview materials in the space data area, listed 6 industry observation draft topics and 1 alternative character topic, and also gave risk warnings.

Later, during the discussion between reporters and senior editors, Qianwen Xiao Q will refer to the opinions of both people at the same time to continuously improve the topic selection framework.

After experiencing it, the most intuitive change is that people participating in the discussion no longer need to repeatedly add background to the Agent. It can not only read the information in the space, but also keep up with the changes in information in multi-person discussions. Information that was originally scattered between people and between people and AI can now be brought together in the same collaboration space.

However, media discussion topic selection is a relatively lightweight scenario. In the survey of Intime Department Store, Leifeng.com saw business collaboration scenarios involving multiple people, multiple groups, and multiple links:

Intime Department Store’s group-buying coupon business is particularly complex—more than 62 stores, more than 200 collaborators, and about 50 pieces of information must be checked before each group-buying coupon is put on the shelves. This information was originally scattered in more than 60 DingTalk groups, and staff had to repeatedly read chat records and manually summarize the progress.

After using the "collaboration" function of Qianwen Office 2.0 in internal testing, the team associated these group chats and the card and coupon AI forms to the same collaboration space. Qianwen XiaoQ can not only summarize progress and update forms, but also regularly check to-do items and remind the specific person in charge.

Leifeng.com has observed that Team Agent changes not only the collaborative relationship between humans and AI, but also the way tasks are initiated and completed.

Most AI offices follow a relatively fixed process: people make demands, AI executes them, and people check the results. Even if the Agent can perform some tasks autonomously, the work rhythm is still mainly driven by humans.

But this paradigm is changing.

Previously, Anthropic launched Claude Tag and introduced Claude to Slack, allowing AI to participate in the daily work of the channel as a team member. It can read the context of the channel in which it is located, call the tools authorized by the organization, and can also proactively synchronize progress when tasks are completed, and follow up on stalled discussions and unresolved tasks.

This represents a product exploration direction: using the team collaboration channel as the entrance, making AI a working member that multiple people can call together and continuously assign tasks.

Qianwen Office 2.0 relies on DingTalk's enterprise collaboration ecosystem to connect DingTalk group chat, data space, AI forms and other capabilities, allowing agents to participate in discussions, summarize progress, and follow up on to-dos around specific business data and task nodes.

Looking at the two together, it is not difficult to find that AI office is showing a new trend: Agent is changing from a tool that passively responds to instructions to a collaborator that can actively undertake tasks and participate in the work process.

How does AI dance with "people" and reshape work processes? In this experience of Qianwen Office 2.0, Leifeng.com (public account: Leifeng.com) made two discoveries:

The first is human-machine co-creation: from delivering results to completing work together

In the past, the collaboration between humans and AI was somewhat like a relay race: employees handed over their requirements to AI, waited for it to generate results, then copied the results and put them into documents or other tools for further modifications. AI is responsible for generation, and humans are responsible for judgment, adjustment and integration.

With an extra handover in the middle, the work can easily be cut into two parts.

Qianwen Office 2.0’s on-screen editing eliminates the step of trying to send results back and forth between humans and AI.

How is the actual experience? This time Leifeng.com asked Qianwen Office to review a first draft of FDE: after Qianwen Office gave its opinions, Leifeng.com marked out the corresponding typos and grammatical errors in the document (Opinion 1), then asked AI to adjust the position of some content (Opinion 2), and divided the first part of the article into two parts (Opinion 3).

AI processed these three revision comments in batches and modified them directly on the original document at once, eliminating the workload of repeatedly copying text, describing requirements, and switching software.

This is different from the past where AI generated a complete document and then had people take over the modifications: people can put forward opinions at any time, the Agent continues to improve based on the opinions, and both parties work together to "sharpen" a work.

The second is automatic tasks: Agents change from waiting for people to dispatch tasks to proactively serving them.

Traditional AI office has an invisible premise: most tasks rely on "people" to initiate them.

This is also a hurdle that many agents have not yet overcome. A review paper released by Google Research this year pointed out that although many agents already have the ability to reason, plan, call tools, and perform complex tasks, they still rely on humans to identify work worth processing and initiate tasks.

In other words, AI can take over execution, but it has not yet truly taken over the initiative to initiate tasks.

This is also room for exploration left to AI office manufacturers. Can AI shift from waiting for employees to dispatch tasks to proactively providing services based on task conditions?

Qianwen Office 2.0 responded to this problem in the feature update. "Automatic tasks" supports triggered tasks. Use a natural language to clearly explain the trigger conditions and processing requirements, and the tasks can run automatically.

Leifeng.com also tried letting Qianwen Office monitor the updated interview minutes in the DingTalk group, proactively summarize the key points and return them to the group, eliminating the trouble of going back and forth between AI tools and IM systems in the past.

Looking beyond the product perspective and from an industry perspective, this update of Qianwen Office 2.0 is also releasing new signals.

In the past few months, various manufacturers have continued to integrate agent products, allocate organizational resources, and "concentrate on big things"—relying on ecological and desktop touch points to compete for users' access to AI.

It is not difficult to see that the previous stage was mainly about closing the front line "horizontally" and seizing the entrance first. But the battle for entrance is not over yet, and the focus of competition has begun to shift "vertically": who can allow Agents to be truly scheduled in enterprise processes.

Judging from the recent iteration direction of domestic AI office products, many manufacturers have realized that they must first deepen their own capabilities. Qianwen Office falls exactly in this direction from 1.0 to 2.0.

"This is a bit like an Agent OS." An enterprise user who participated in the internal beta intuitively felt that Qianwen Office 2.0 not only made a stronger AI office entrance, but also tried to make Agent become the middle layer connecting users, tools and businesses.

Just like traditional operating systems schedule applications and computing resources, Agent OS schedules different tools, data resources, and Agent capabilities. After the user hands over the task to Qianwen Xiao Q, this "digital employee" will understand the needs and scheduling capabilities and push the task forward.

For businesses, this means digital workers are becoming new players in the organizational division of labor.

Previously, AI was more like an assistant at hand for employees, helping to write documents, perform analysis, and organize information. When an agent can understand the business context, call different tools, and continue to advance work among multiple tasks, it has the opportunity to take over some of the work that originally needed to be completed by different positions, helping the enterprise reduce losses in cross-department communication and process connection.

For the domestic AI office industry, this is an attempt worthy of attention: Although it is still far from the real Agent OS, Qianwen Office is already broadening the boundaries of "office".

A To B veteran who has been in the business for more than 30 years told Leifeng.com when Qianwen Office launched its e-commerce version in October this year: "When AI office is no longer limited to office work, it will truly have vitality."

The implication is that traditional general office capabilities such as documents, PPT, and copywriting are difficult to create strong enough user stickiness. The value of Agent is not only to complete a general task, but also to enter specific business scenarios in the digital world, understand business data, rules and workflows, operate different functions like real people, and promote the implementation of tasks.

Behind this, AI office products are needed to further connect the company's existing data and processes.

It is understood that the integration of IM and Office has been implemented in Qianwen Office 2.0. As more DingTalk atomic capabilities, SaaS and professional tools are gradually connected, Qianwen Office is also trying to reduce cross-tool operations and allow Agents to participate more deeply in the real work of the enterprise.

This also means that Agent OS competes not only on how many tools are connected, but also on whether it can organize the business capabilities behind the tools. For AI office products to move from general tasks to real business, it also requires the participation of professional software vendors and deeper data and context connections.

In the final analysis, the entrance determines who is seen, whether they can truly enter the enterprise process, and who is left behind.

Big factory AI To B war, why is the battlefield this time office?

2026-08-24

The "survival record" of the first batch of FDE

2026-10-30

Original articles from Leifeng.com, unauthorized reproduction is prohibited. See the reprint instructions for details.

edit

Resend Email Modify Email