Baidu Dazi reaches 9.77 million MAU as the AI office flanking battle faces a phased major test

By Tang Chen

Who says Baidu Dazi can't fight?

On October 20, QuestMobile released its Q3 2026 AI application market development report. The data shows that AI office monthly active users have reached 142 million, up 50.6% year-on-year.

Among them, Baidu Dazi's performance surprised many. Its combined monthly active users reached 9.77 million, with 83.81 million total usage sessions across platforms, up 5,810% from June.

This scale places it second on the AI office Agent ranking. Ahead of it is Tencent WorkBuddy with 32.353 million, while third and fifth are Qianwen Office with 7.905 million and Doubao Work with 2.049 million, respectively.

By contrast, WorkBuddy, Qianwen Office, and Doubao Work respectively rely on IM (instant messaging) collaboration gateways such as WeChat (WeCom), DingTalk, and Feishu, naturally inheriting existing office traffic.

Baidu Dazi's underlying identity is that of a cloud vendor, and its main pitch is "usable without switching collaboration software." It was able to reach this position by using a "Tencent-style" playbook to fight a flanking battle.

Three leaps in half a year

Reviewing Baidu Dazi's growth trajectory, I found that over the past six months it has completed three leaps.

The first leap came in March this year. Amid the nationwide "raising shrimp" (OpenClaw) wave, Baidu entered with the "desktop shrimp DuMate" form. Subsequently, the "first domestic enterprise-grade full-blooded OpenClaw" was fully launched, positioned as "truly getting work done."

By this time, users' needs were no longer satisfied with chatbots. Official data shows that 60% of tasks require at least 3 steps to complete, and 86% of tasks must ultimately deliver a result.

At Create2026, Robin Li gave it a positioning: "DuMate is a unified entry point, a general-purpose agent that strings together Baidu's search AI API, Miaoda, Famo, and other capabilities."

I once argued that the "One Baidu" concept should be placed on Robin Li's desk and given attention and thought at a higher level. It is similar to "One Ford," "One Microsoft," or "One Geely"—a strategic refocusing of a leading large company amid a dramatically changing competitive landscape.

Next came the "growing up" phase. During this period, the industry focused on WorkBuddy and Qianwen Office. Baidu Dazi was also iterating at a pace of about 3 versions per week.

For example, at Baidu AI Day on August 27, Baidu proposed "delivery that stuns" and launched 96 professional skills at once, packaged into 15 industry suites covering four major functional domains: finance and accounting, business and legal, product R&D, and operations and HR.

Baidu Group Executive Vice President Shen Dou disclosed that they iterated 150 times within 5 months of launch, user growth in the past month was nearly 9 times, and average daily questions increased 60-fold.

Since September, Baidu Dazi began entering enterprises, moving from individuals to organizations. During this period, the enterprise edition launched, and the self-media suite, design suite, finance suite (connected to Tonghuashun data sources), and bidding suite were rolled out one after another, with more than 30 expert suites cumulatively.

At the same time, Baidu Dazi was integrated into Xiaodu hardware, which Xiaodu Technology CEO Li Ying called "a unified agent capability foundation."

More critically, Baidu systematically packed its existing AI assets into one product, including search's low-hallucination retrieval, Wangpan's data foundation, intelligent cloud's computing power, Xiaodu hardware, and the Xinchuang channel, all converging into "one entry point, one general-purpose agent."

While raising the industry's competitive dimension to "model + context + delivery," this also filled Baidu's shortfall at the application-layer entry point. Baidu Dazi became the "ONE" entry point at the application layer for Baidu's full-stack AI of "chips, cloud, models, and agents," and sibling products such as Agent Infra, Wenxin model invocation, Miaoda, Famo, and Shengsuan all followed it into enterprise procurement lists.

Shen Dou's statement about "letting large models turn from technology into economic force" now has a vehicle through which it can be delivered.

Using Tencent's speed to fight its own flanking battle

At the industry level, it validated an easily overlooked proposition: AI office does not necessarily have to depend on IM relationship chains, and the industry's competitive focus has shifted to two variables: delivery capability and private data context.

I believe this is because Baidu found a self-consistent playbook: learn from Tencent's low-price push for incremental growth, use rapid iteration to trade for maturity, and choose a battlefield that maximizes strengths and avoids weaknesses.

For example, in addition to iteration speed, on pricing, Baidu Dazi's personal edition is 59 yuan for a continuous monthly subscription, versus WorkBuddy at 70 yuan, Qianwen Office at 78 yuan, and Doubao Work at 68 yuan—it is the lowest; the enterprise edition is limited-time 132 yuan per seat, with a minimum purchase of 1 seat, and also offers a 9.9 yuan trial price.

But this playbook cannot be copied wholesale. WorkBuddy has WeChat and WeCom behind it, while Baidu has obvious shortcomings in models and collaboration entry points. If it fully replicated the "IM + office" model and competed head-on with Tencent, Alibaba, and ByteDance, it would be using its own weaknesses to hit others' strengths.

"Marketing Warfare" divides competition into four forms: defensive, offensive, flanking, and guerrilla. Among them, the core of flanking warfare is: avoid the opponent's dominant main battlefield, open battle in unclaimed territory where its defenses are weak or even absent, and expand results through tactical surprise and pursuit of victory.

Baidu Dazi chose four "no-man's lands" and fought a flanking battle.

First is the delivery standard. Baidu proposed "delivery that stuns," and created DuMateBench, the industry's first open leaderboard evaluating agents' delivery capabilities. It incorporates more than 200 real tasks, pulling the competitive yardstick away from "how smart the model is" toward "whether the delivery can be used directly."

This is also the business-level concretization of the DAA (daily active agents) metric proposed by Robin Li.

Second is the data context. This is also Baidu's deepest moat: search gives Dazi low-hallucination retrieval capability, and Baidu Netdisk gives it more than 1 billion users and over 100 billion GB of stored materials. After obtaining authorization, Dazi can read the files, history, and enterprise knowledge across the user's entire data repository.

An industry consensus is: for an AI office Agent to do work, the model plus the execution framework determines whether it can finish the job, and the context determines how far it can go. Baidu is betting on context, while Tencent, Alibaba, and ByteDance are all racing to catch up.

Deeper still are the expert suites. Baidu Dazi has already launched more than 30 expert suites. For example, the finance suite was built with trading operators and connects to Tonghuashun data sources, and the self-media suite was co-created with KOLs and MCNs. You should know that a single suite takes up to a month to polish, and must pass professional authoritative validation before launch. This is the encapsulation of industry know-how that a general chat box cannot replace.

There is also a channel with an extremely high threshold: Xinchuang. Baidu Dazi teamed up with Lenovo Kaitian, Kylin Software, and UnionTech Software to jointly build a full-stack domestically produced general-purpose agent that supports private and hybrid cloud deployment; there is also a "Dazi Workstation" preloaded with models and a runtime environment.

This is an "all-domestic" lineup spanning hardware, operating system, models, and runtime environment, fully able to meet the procurement requirements of central state-owned enterprises and key industries.

Overall, these four no-man's lands share one thing in common: the defenses of the IM giants are all weak.

When Baidu Dazi takes on Tencent, Alibaba, and ByteDance, it does not need to fight over IM either.

If Baidu redefines the winning standard for AI office as: whose delivered results can be used directly, and who can understand an enterprise's data repository. Then the importance of the IM entry point will drop sharply, and Dazi will have found its comfort zone.

In fact, Baidu management's understanding of the flanking battle has always been clear.

At Create2026, Robin Li divided Agent entry points into two generations. Chatbots solve information acquisition, and stop at that step once the answer is found; general-purpose agents solve task completion, and it only counts once the work is done.

He said, "Chatbots solve the problem of information acquisition and are the first-generation entry point; general-purpose agents solve the problem of task completion and are the second-generation entry point."

In September this year, Shen Dou proposed the "industrial agent operating system" at the Intelligent Economy Forum, with the approach of having three types of agents work in coordination: general-purpose agents, professional agents, and customized agents. Work in enterprises is also cut into three parts: 20% daily operations, 30% professional domains, and 50% core tasks that determine operations, handed to these three types of agents.

In the operating system proposed by Shen Dou, Dazi happens to sit exactly at the general-purpose agent layer, and the competitive dimension of AI office has also shifted to the Agent operating system.

This is also what Tencent is doing.

On September 2, Liu Yi, vice president of Tencent Cloud and head of Tencent CodeBuddy & WorkBuddy, said publicly, "We hope to build WorkBuddy into an operating system for the Agent era, not a stronger tool, but a platform that can carry all tools."

The decisive factor for Baidu Dazi

But the outcome of the flanking battle does not depend on how great Baidu's determination is, but on how Dazi avoids its own soft spots.

The most direct threat comes from AI office giants. Concepts such as delivery and context are being followed by WorkBuddy, Qianwen Office, Doubao Work, and Kingsoft Office, and they also have traffic funneled in from IM collaboration entry points.

Baidu's short-term advantage lies in the data-context combination of "Netdisk + search." If it cannot quickly turn this into a product system, once the window period passes, the field of "context" will fall into a red ocean.

Moreover, Baidu Dazi also has shortcomings in creative scenarios. Jiemian News' hands-on testing shows that Doubao Work has obvious advantages in tasks such as text-to-video and SVG animation, and Dazi's lack of third-party experience endorsement will weaken the persuasiveness of "delivery that stuns."

There is another trouble from within Baidu. How to truly consolidate Baidu AI's entry points into "ONE" still requires greater determination from Robin Li.

The most fundamental test is the business model. In Baidu's Q2 2026 financial report, total AI business revenue was 12.5 billion yuan, accounting for half of general business revenue; of that, AI application revenue was about 2.5 billion yuan, up only 3% year on year, while the real bulk was GPU cloud, up 283%.

First, solidify the DAA measurement yardstick and make delivery and data these two cores into industry standards. DuMateBench is currently still Baidu's own leaderboard, and the next step is to push it to become a general evaluation benchmark. Whoever defines the evaluation system first holds the discourse power.

At Create2026, Robin Li proposed DAA, that is, daily active agents, and the core logic is that Token counts input, not output.

Second, adjust the focus of attack, lead with enterprise knowledge base understanding and deep cultivation of industry suites, and avoid involution over IM traffic. Songxianxian is a typical example: supply chain cost-reduction analysis was compressed from two working days to 3 minutes, with scheduled output of reports, dashboards, and alerts, then pushed to IM. This kind of deep moat of data + delivery is hard for competitors to copy.

Third is sustained pursuit. If Baidu Dazi wants to build an industrial agent operating system, it needs to dig deeper along the industrial dimension. At the same time, beyond growth rate, Baidu also needs to pragmatically consider user retention and renewals.

This time window is not too long. Li Jingqiu's judgment is "from the third quarter of this year to the end of the year."

I further think that Baidu World 2026, which is about to begin on November 12 and is themed "The Next Act, Intelligent Organization," focusing on new progress in enterprise scenarios, may be the phased major exam for this flanking battle.

Baidu Dazi has proved one thing over the past six months: without an IM entry point, it can still become one pole of the AI office landscape. But the endgame of the flanking battle will not be the ranking on some leaderboard; its real meaning is to make "delivery" and "data context" the industry default for AI office.

Next, all the focal points are: among 9.77 million monthly active users, how many will still open it the next day; whether enterprise customers will still pay this seat fee by next year; and besides Baidu itself, who else is willing to use the DuMateBench standard to measure their own products.

These three things are the real decisive factors in this flanking battle.