BAT can no longer afford to wait in the battle for the AI office sector.

In the past two months, China's three major tech giants collectively referred to as "BAT" have completed the exact same move: merging, streamlining and unifying the caliber of product lines that previously operated under internal horse-racing mechanisms.

On July 20, Tencent adjusted the QClaw business and part of its team to the department where WorkBuddy belongs, wrapping up the internal intelligent office agent horse-racing process; 10 days later on July 30, Liang Rubo, CEO of ByteDance, sent an internal letter to announce that the Feishu product team was merged into Doubao, establishing a "Creativity Service Platform"; on August 3, Alibaba integrated three intelligent agents QoderWork, MuleRun and Wukong into "Qianwen Office", which is overseen by Chen Yusen, the new CEO of DingTalk, and officially launched public beta testing.

On August 14, at its AI Day, Baidu officially renamed its general intelligent agent GenFlow to "Kuku AI", launching an independent product matrix covering PC clients, web pages, mini-programs and enterprise editions, formally upgrading from functional modules embedded in Baidu Wenku and Baidu Netdisk to standalone products.

The intensive organizational adjustments and product upgrades of the three BAT giants all point to the same industry judgment: the AI office track is shifting from multi-line horse racing to centralized operations, and the competition dimension is also changing from function stacking to Agent-based workflow reconstruction.

To put it more plainly, a trend has become very clear: Agents are becoming the new office entry point, while traditional office suites such as Word, Excel and PPT are being reduced to the execution layer of Agents.

When the entry point changes hands, the seating order of players at the old table will be rearranged. Baidu laid out its plans three years in advance, but Tencent, coming from behind, leads in monthly active users, while Alibaba has bet on the DingTalk ecosystem and arrived late to the game.

The three BAT giants seem to have different calculations, but the real problems they face are highly consistent.

01 Each of BAT Has Its Own Calculation

First, we need to clarify the differences in product layout and progress among the three BAT players. This is not a simple comparison of function checklists, and "Business Show" believes that it represents a divergence of three underlying logics.

Let's start with Baidu, which has the earliest layout of AI office business. Since 2023, both Baidu Wenku and Baidu Netdisk have completed large model reconstruction, taking the lead in launching functions such as intelligent PPT and in-depth AI research reports; by 2024, the two platforms were fully integrated, connecting 1.8 billion professional documents on Baidu Wenku, 700 million academic materials and the private domain knowledge base of Netdisk, and this set of content data assets is the unique source of barriers for Baidu. In April 2025, GenFlow 1.0 was launched, and it was iterated to version 4.0 within one year.

At the AI Day on August 14, Kuku AI released the latest data: in April 2026, the total monthly active users of the entire platform exceeded 100 million, of which the MAU of AI office exceeded 25 million.

In two rounds of authoritative evaluations by the National Industrial Information Security Development Research Center, the 2025 "Large Model Empowered Smart Office Evaluation Report - PPT Generation" and the 2026 "Office Agent Workflow Evaluation Report" both showed that Baidu Wenku ranked first, occupying the first echelon alone.

Baidu's playbook logic is very clear: use the content data flywheel to build first-mover advantages, and then adopt a multi-agent collaborative architecture (MoE Mixture-of-Experts architecture + cloud long-term memory + end-cloud collaboration) to realize parallel invocation of PPT, Excel and Word Agents with one instruction to complete complex tasks. The newly launched independent terminal has also been deeply optimized for the financial investment research scenario, with built-in databases of listed company quotes, financial reports and brokerage research reports, which can generate financial models such as SOTP, LBO and M&A with one click.

Looking at Alibaba next, its pace is obviously a beat slower. Qianwen Office did not start public beta testing until August 3, integrating three previously independent internal products, which are based on the desktop AI intelligent agent tool QoderWork, deeply integrating Wukong, the enterprise collaborative office agent incubated by DingTalk, and MuleRun, the Agent execution engine of Alibaba Cloud's internal startup, all overseen by Chen Yusen, who took over as CEO of DingTalk on June 11.

According to the "2026 Q2 China Office Intelligent Agent Platform Market Insight Report" released by Analysys, as of June 2026, the monthly access volume of QoderWork before integration was 7.88 million, ranking third among 17 mainstream desktop AI-native office intelligent agents, which is only more than one third of Tencent WorkBuddy (20.97 million times).

The differentiated positioning of Qianwen Office lies in the enterprise level. According to media reports, it is also the first product in the industry that supports desktop Agent, cloud Agent and enterprise collaborative Agent at the same time, and has initially connected DingTalk IM, supporting 25 enterprise-level capabilities such as message group chat reading, attendance approval, meeting schedule, and document and form generation. At the model level, it is equipped with Alibaba's latest flagship Qwen3.8-Max-preview by default, with a total parameter of 2.4 trillion, and the maximum context length of advanced and cutting-edge gears supports up to 1 million Tokens.

Finally, let's look at Tencent, whose path is the most special. The underlying layer of WorkBuddy is not an office product built from scratch, but reuses the Coding Agent kernel polished by CodeBuddy, the AI programming tool that Tencent Cloud started developing three years ago. According to Liu Yi, Vice President of Tencent Cloud and Head of CodeBuddy and WorkBuddy, this gives WorkBuddy a solid foundation from the first day of its launch, rather than being encapsulated based on the open source project OpenClaw like many peers, whose performance is unsatisfactory.

Data shows that WorkBuddy has grown rapidly since its release in March 2026. According to data from Analysys, the number of visits in June alone reached 20.97 million, exceeding the sum of ByteDance's TRAE IDE (12.79 million) in second place and Alibaba's QoderWork (7.88 million) in third place.

Tencent's playbook is built on three capabilities: scenario connection capability (high-frequency touchpoints such as WeCom, WeChat, QQ Browser, and Tencent Docs), engineering capability (Harness system, including context engineering, Memory system, Action system, SubAgent and AgentTeams), and collaborative design of model and product. On July 20, after integrating QClaw into the WorkBuddy system, Tencent completed the unified scheduling of internal personal entry points and enterprise entry points.

02 Unified Entry, or Admitting Defeat in Horse Racing?

A noteworthy issue arises: these integrations ostensibly focus resources on achieving major goals, but from another perspective, they also expose an awkward fact: the previous multi-line internal horse racing of each company has at least partially failed.

Take Alibaba as an example. The three products QoderWork, MuleRun and Wukong integrated by Qianwen Office originally belonged to different teams, with overlapping and different positioning. Before the integration, the monthly access volume of QoderWork was only 7.88 million, ranking third in the market. The three products were scattered in investment and fought their own battles, but none of them became the leading top runner.

According to the research report of Guosheng Securities, after Chen Yusen took over as CEO of DingTalk in June, he sent an internal letter on June 18 to integrate the R&D teams of Wukong and MuleRun, announced on July 2 to integrate capabilities based on QoderWork, launched internal beta on July 22, and started public beta on August 3.

This fast pace is not so much a well-considered strategic upgrade as an emergency streamlining forced by the market window.

Tencent's situation is slightly different. QClaw comes from the Tencent PC Manager team, positioned as an AI remote control for the WeChat ecosystem, encapsulated based on the open source framework OpenClaw. WorkBuddy comes from the Tencent Cloud CodeBuddy team, positioned as a full-scenario workplace AI desktop workstation, reusing the Coding Agent kernel polished by CodeBuddy.

They come from different teams and their positioning can be described as complementary. The logic of streamlining lies in avoiding repeated investment in internal products and unifying back-end resource scheduling.

But it is worth asking: since the two products do the same thing and have complementary positioning, why not unify the team from the very beginning? Instead of developing separately for more than a year before merging, who will account for the consumed R&D resources and missed market windows in the process?

ByteDance's playbook seems more extreme. According to 36Kr reports, ByteDance directly merged the Feishu product team into Doubao, and the Feishu GTM team into Volcano Engine. This means that Feishu's strategic status as an independent office suite has been substantially downgraded, and it has been incorporated into the creativity service platform centered on AI models.

According to 36Kr, some industry analysts put it bluntly: "The window period for desktop-level office products is closing rapidly on a monthly basis." This sentence also explains why the three companies took intensive actions within two weeks: it is not a well-considered strategic synergy, but an emergency stop-loss forced by the window period.

But the deeper problem is, after streamlining, can the integrated products truly achieve "1+1>2"? Organizational merger is easy, but product capability integration is difficult. The original technical architecture, Agent framework, model invocation strategy and interaction logic of the three products need to be unified into one set of product experience in a short time, and the engineering difficulty is no less than redeveloping the products from scratch.

Alibaba still needs to give away points (2000 points for new user registration, 500-2000 points per day in the first week) during the public beta period to attract users, which shows that the natural attractiveness of the product itself is still under further verification.

The narrative of unifying entry points sounds attractive, but the cost of admitting defeat in horse racing is silent. The window period will not wait for any company to complete internal integration.

03 With 25 Million Monthly Active Users, Why Is It Still Not Profitable?

From the perspective of commercialization, the most interesting one is Baidu's "Kuku AI", whose AI office MAU exceeds 25 million, while Tencent WorkBuddy has more than 20 million monthly active users and over 13 million daily active users.

In the context of traditional Internet products, this is an enviable set of data. However, AI office has a fatal problem that traditional SaaS does not have: the more users there are, the more losses the company suffers.

This is not an exaggeration, but a "structural rupture" of the cost structure.

The marginal cost of traditional SaaS tends to be zero. When a set of systems is delivered to 10 or 100 customers, the incremental computing power and R&D investment hardly increase. But every service of AI office corresponds to real and rigid computing power consumption.

According to the Securities Times, when the old Internet logic of exchanging free services for scale hits the new AI cost structure of metered billing, a scale trap where more users bring more losses is emerging. The more successful the product is, the more serious the loss is, which is a widespread paradox of diseconomies of scale in the downstream application side of the AI industry chain.

It goes without saying that every additional question from a user and every additional task run by an Agent is backed by real money spent on Token consumption.

What is more severe is that the computing resource consumption of Agent-type tasks far exceeds that of traditional chatbots. According to 36Kr citing data from Anthropic, the Token consumption of a single Agent to complete a typical task is about 4 times that of ordinary dialogue, while multi-agent collaborative tasks can reach 15 times that amount.

This means that the more intelligent the product is and the closer it is to complex business scenarios, the more non-linear growth the unit Token cost will show.

Tencent's financial report data may partially confirm this dilemma. Tencent's management disclosed that the gross profit margin of WorkBuddy's paying users is comparable to the overall gross profit margin of Tencent Cloud, but the overall gross profit margin is low, mainly because there are a large number of free users, and Tencent exchanges market share and user growth through subsidies. The enterprise version is priced at 198 yuan per user per month at the starting point, but most users are still using it for free. The conversion rate from free to paid and customer acquisition cost are still unknown.

Moreover, Tencent also faces a unique structural problem: WorkBuddy is not uniquely bound to Hunyuan. The management clearly stated at the conference call that WorkBuddy can invoke third-party models (DeepSeek, Zhipu GLM, Kimi, etc.).

This multi-model compatibility strategy is a common practice in the industry, but it means that the user scale and traffic of WorkBuddy cannot be directly equivalent to the commercial success of the Hunyuan model. If the performance and score of Hunyuan in actual tasks are not the first choice of users, there is a risk of fracture in the link from the model to the application.

Baidu's path is no exception. Although "Kuku AI" claims to have 25 million MAU, ranking first in the industry, Baidu has never disclosed the paid conversion rate, ARPU value or the proportion of Token cost of these users. How much of the 25 million monthly active users is supported by free quotas and how much is diverted from the traffic of Baidu Wenku and Netdisk cannot be verified by the outside world.

More critically, Baidu Wenku and Netdisk have long operated with free or low-price strategies, and Kuku AI has inherited this gene. The migration of user mentality from "free tool" to "paid Agent" cannot be completed simply by changing a name.

Alibaba's situation is slightly different, but equally difficult. Qianwen Office is positioned at the enterprise level, which is theoretically the customer group with the strongest willingness to pay. However, enterprise-level AI Agents face the customized cost black hole. Large-scale deployment involves customized development, system docking, and knowledge base fine-tuning. In addition, enterprise business processes, policies and internal data continue to change, so the Agent knowledge base, tool chain and prompt word engineering need to be iterated monthly, and each investment is showing exponential growth.

It can be said that the three mainstream pricing models (account subscription, Token billing, project-based one-time purchase) each have fatal flaws, and the entire industry has not yet formed a standardized result-based pricing system.

The most critical thing is the change in the macro environment. Since 2026, domestic AI large models have entered the era of collective price increases: Tencent's Hunyuan large model API has a maximum price increase of 463%; Zhipu released GLM-5.1 on April 8 and raised its price by 10% for the third time; Alibaba Cloud Bailian suspended the subscription of Lite packages on March 20. Zhipu AI raised prices three times in the first quarter, with a cumulative increase of 83% for core APIs, Alibaba Cloud's AI computing power services increased by 5% to 34%, and MiniMax's core model prices increased by 30% to 50%.

Although leading large manufacturers can develop models independently to reduce their dependence on external APIs, the training cost of self-developed models is also staggering. Tencent's capital expenditure in the second quarter of 2026 reached 52.8 billion yuan, totaling 84.7 billion yuan in the first half of the year, and is expected to exceed 200 billion yuan for the whole year, most of which is used for computing power infrastructure. According to Caijing Magazine, affected by capital expenditure, Tencent recorded negative free cash flow of 13.8 billion yuan in the second quarter of 2026.

Cases of out-of-control Token consumption have already appeared among giants. According to 36Kr citing iResearch reports, Uber's full-year planned Token budget was exhausted in only four months; in a multi-agent experiment of miHoYo, dozens of intelligent agents fell into an infinite loop, consuming 2 million yuan worth of Token resources overnight.

Goldman Sachs predicts that driven by the large-scale invocation of AI intelligent agents, the global Token consumption in 2030 will reach 24 times that of 2026. If the current extensive management model is not changed, this growth rate will completely exceed the payment capacity of enterprises.

25 million monthly active users is not a moat, but an increasingly expanding bill. Before the Token cost structure is fundamentally restructured, all players in the AI office track are exchanging losses for scale, scale for valuation, and valuation for time.

The problem is that time will not be given infinitely.

03 Three Years of First-Mover Advantage, Why Can't It Stop Functional Substitutes?

A thought-provoking topic is that the narrative Baidu repeatedly emphasizes is its three-year layout: entering the market in 2023, completing integration in 2024, launching GenFlow in 2025, and exceeding 100 million monthly active users in 2026.

This narrative