The fight over who will control the entry point for China's enterprise AI office agents has officially begun this summer.
On July 30, ByteDance initiated a restructuring of its AI operations: its Feishu product team merged with the Doubao product team to form a new Doubao product unit; the Feishu GTM team merged with the Volcano Engine team to create a "Creativity Service Platform," which will handle the sales and service for both MaaS and SaaS. In simple terms, ByteDance is embedding Feishu into Doubao and moving Feishu's sales force into Volcano Engine. A beta version of Doubao Enterprise is already being tested among some Feishu clients, integrating Feishu's documents, spreadsheets, meetings, and group chats, so users won't need to deploy anything extra.
With this move, ByteDance has finally made the same decision as Alibaba. Just a week earlier, Alibaba announced it would consolidate its three competing products—QoderWork, Wukong, and MuleRun—into a single offering called Qianwen Office, to be managed by DingTalk's new CEO, Chen Yusen. Ten days before that, TENCENT integrated its QClaw product center into the cloud products department where WorkBuddy resides. All three companies, with their respective internal races, have arrived at the same strategic choice this summer: concentrate resources and unify the user entry point.
Why Racing to Experiment, Then Merging to Compete
After OpenClaw's "Lobster" was open-sourced in early 2026, China's major internet firms nearly simultaneously began an experiment: having multiple teams develop their own AI office products to see which would prevail. TENCENT internally spawned three products: WorkBuddy, QClaw, and Marvis. ByteDance had four exploration lines: Trae, Kouzi, Doubao Office, and Feishu Aily. Alibaba ran three parallel tracks: QoderWork, Wukong, and MuleRun. This "horse racing" mechanism was logical in the industry's early days; with no clear idea of what form the AI Agent would ultimately take, laying down multiple paths increased the chances of finding the right one. The core value of the OpenClaw open-source AI agent framework is to "give large models hands and feet"—enabling them to directly manipulate users' file systems, browsers, terminal commands, and various APIs. The products built around it by the tech giants all aim to do the same thing: package the open-source capabilities for enterprise use, lower deployment barriers, add security and access controls, and connect them to their own ecosystems and user bases.
However, the cost of this "horse racing" was significant: duplicated effort, internal talent wars, and user confusion. When Alibaba's Wukong launched, it was touted as "the world's first enterprise-grade AI-native work platform," but it remained in a semi-finished state for so long that even Alibaba employees admitted they mostly just used QoderWork. ByteDance's Trae was positioned as a programming tool, Kouzi followed the Agent-building route, and Doubao gradually added office features—leaving users unsure which product to actually use.
In July, a report from Analysys became the final straw that ended the internal races. The data showed that in June, monthly visits to China's desktop AI office agents surpassed 60 million. TENCENT's WorkBuddy topped the chart with 20.97 million visits, exceeding the combined total of second-place ByteDance's Trae (12.79 million) and third-place Alibaba's QoderWork (7.88 million). This figure revealed a harsh reality: the market had already made its choice. Continuing to experiment would mean ceding the primary battlefield to competitors. The purpose of dividing into teams was to experiment. Merging them is done to fight the real war.
Tencent: The Dark Horse That Won the Most Impressive Victory
WorkBuddy's climb to the top was unexpected just a few months ago. In early 2026, TENCENT's position in AI was precarious. It had large models, cloud computing, and social networks, but it lacked a widely recognized AI product. "Tencent is falling behind in AI" was a common industry view. WorkBuddy started as a small internal tool within Tencent Cloud, developed by a team of about 10 people. The turning point came in March, when TENCENT decided to launch it commercially. The key decision was not to position it as a "second coding assistant," but as an "AI-native workbench." The former helps programmers write code faster, while the latter allows anyone sitting in front of a computer to complete tasks using natural language. This fundamental difference in positioning meant WorkBuddy's user base would defy all expectations.
TENCENT's strategy was to rapidly package the OpenClaw ecosystem for the enterprise: WorkBuddy directly integrated with OpenClaw's skill ecosystem, allowing users to get started immediately without configuring from scratch. This wasn't about reinventing the wheel; it was about adding enterprise-grade packaging, lowering deployment barriers, and adding security and access controls on top of a mature community framework, while connecting to its own ecosystem. After the public beta, the fastest-growing user segments were not programmers but HR professionals, administrative staff, operations personnel, finance teams, and production line engineers. The users who drove WorkBuddy's monthly active users (MAU) to 20 million were "people who never write a line of code."
A July report from Citigroup revealed key figures: WorkBuddy had a cross-platform, de-duplicated MAU of 20 million, daily active users (DAU) exceeding 13 million, and a DAU/MAU ratio of 65%-75%. In enterprise SaaS, this represents the highest level of user stickiness. Citigroup's conclusion was that "the paid conversion rate for high-value users could be higher than typical consumer subscription services like music, video, or even gaming." TENCENT's ecosystem advantages are difficult to replicate—WorkBuddy integrates with Tencent Docs, Tencent Meeting, WeCom, and QQ Mail, and supports task assignment remotely via WeChat and WeCom. Users don't need to install any extra software; they can command an Agent to perform tasks directly from a chat box. However, questions remain for TENCENT. Retention and payment rates are calculated by combining data from CodeBuddy and WorkBuddy, so it will take time to verify if WorkBuddy's standalone business model is viable. Furthermore, TENCENT's real killer feature might not yet be fully revealed—an AI Agent internally tested for WeCom, called "Big Circle," which is embedded directly into every interface of the enterprise app and can be activated by swiping left. According to Sina Technology, Pony Ma's attention to this project is so intense that he "rarely misses any related meetings."
ByteDance: After Spending the Most Money, It Finally Decides to Consolidate
The organizational restructuring announced today is ByteDance's clearest statement of intent in the AI office domain. The restructuring path mirrors Alibaba's almost exactly: merging the product teams of Feishu and Doubao, and merging the GTM teams of Feishu and Volcano Engine. The Doubao Enterprise edition, as the first integrated product, is already in beta testing with some clients. Why is ByteDance doing this? Because its AI business has a critical awkwardness. Looking solely at Doubao's data, ByteDance appears to be China's most powerful AI company, with 382 million monthly active users—ranking first among domestic AI applications, more than second-place Qianwen (167 million) and third-place DeepSeek (130 million) combined. The problem, however, is how to convert these 382 million users into revenue from the enterprise office space. ByteDance is projected to spend 200 billion yuan on AI infrastructure in 2026, yet according to LatePost, Doubao generates less than 1 million yuan in daily revenue—amounting to just 365 million yuan for the entire year.
In the battle for the office entry point, ByteDance's product lines are also the most fragmented. Trae is geared towards programming, Kouzi is an Agent-building platform, Feishu has its own Aily intelligent agent, and Doubao is gradually opening up office capabilities. These four lines operate independently without a unified entry point. However, Feishu has shown a significant trend over the past year. Its revenue in 2025 exceeded 3 billion yuan, and in Q2 2026, it grew 100% year-over-year, narrowing the revenue gap with DingTalk to a historic low. More importantly, the composition has changed: since 2026, over 90% of Feishu's new clients have simultaneously purchased AI features. Its client base is expanding from internet and new energy sectors to retail, manufacturing, and energy—with new clients including Luckin Coffee, Haitian Flavouring, Deli, and Nanjing Iron & Steel. Feishu has proven that enterprise clients are willing to pay for AI embedded in their workflows. But it lacks Doubao's consumer-side influence. This integration essentially combines Feishu's enterprise B-side capabilities (corporate workflows, client payment systems) with Doubao's massive C-side influence (382 million MAU). If successful, ByteDance's competitiveness for the office entry point could leap to a new level.
Alibaba: The First to Integrate, but Also the Longest to Suffer the Pain
Alibaba was the first of the three to initiate integration, but it has also experienced the most pain. In June, DingTalk's founder, Chen Hang, stepped down as CEO, succeeded by the post-90s Chen Yusen. His first major task was to merge the three products—QoderWork, Wukong, and MuleRun—into Qianwen Office. Each product has a different origin: QoderWork came from Alibaba Cloud, Wukong from DingTalk, and MuleRun from a startup team. Their similar positioning and overlapping functions created a de facto internal competition, with limited computing and R&D resources spread across multiple projects. An insider familiar with DingTalk told Caijing, "It's like choosing between one's own children. Integrating products is easy; integrating people is much harder." After integration, Qianwen Office uses QoderWork as its technical foundation, incorporating Wukong's DingTalk enterprise scenarios and MuleRun's overseas growth. Alibaba Cloud has also found its role in the OpenClaw ecosystem by offering a one-click deployment solution, allowing users to run OpenClaw on a lightweight application server for as little as 56 yuan per month, with its 52 pre-installed official Skills averaging 120 million daily calls. But Alibaba still needs to solve two problems: first, how Qianwen Office, known for its reputation in the tech community, can integrate with DingTalk's large enterprise client base; and second, how to convert DingTalk's huge client base in traditional industries like manufacturing, government, and education into AI business revenue. DingTalk's revenue in 2025 exceeded 4 billion yuan, making it the largest of the three in terms of size. However, the digital maturity and AI acceptance of these clients are significantly behind the user groups in the internet-native Agent space. Convincing them to pay extra for AI "is something that requires a slow, patient approach."
The Winner of the Entry Point Determines the Winner of the Era
The entry point for the PC era was the browser and the search box, won by Google and Baidu. The entry point for the mobile era was the super app, won by WeChat and Douyin. Generative AI is turning the entry point into a spoken command to an AI, and the office desk is the most natural starting point for that command—a person spends half their waking hours at their desk, with every interaction generating the most valuable work data. WorkBuddy's data reveals a fact overlooked by many: the first heavy users of AI office tools are not programmers, but those who "most need AI to help them with their work"—HR, admin, operations, finance, and production line workers. An HR professional who can't code uses WorkBuddy to automatically screen resumes and generate interview reports; she needs an AI colleague more than a programmer does. This means the competitive logic for AI office products is completely different from programming tools. Programming tools compete on model performance and capability ceilings; office products compete on how quickly an average, non-technical white-collar worker can learn to use them—within three minutes—and how indispensable they become within three weeks. The previous round of entry-point wars—from group buying to food delivery to ride-hailing—took a decade to resolve. DingTalk CEO Chen Yusen offered a different timeline at the May Alibaba Cloud Summit: "In the next 18 months, we will see very strong productivity transformations across all industries." Now that ByteDance has completed its restructuring, China's major internet companies are all entering a phase of resource consolidation. The next step is where the real product showdown begins: whose Agent can truly integrate into the daily lives of white-collar workers will have the best chance of becoming the core of the next-generation enterprise AI ecosystem.
Comments