In two weeks this summer, Alibaba, Tencent and ByteDance each reorganized their desktop office AI agent businesses, converging on a market that drew more than 60 million monthly visits in June.
In two weeks this summer, Alibaba, Tencent and ByteDance each reorganized their desktop office AI agent businesses, converging on a market that drew more than 60 million monthly visits in June.

Alibaba, Tencent and ByteDance reorganized their desktop office AI agent units within two weeks this summer, converging on a market where Tencent's WorkBuddy drew 20.97 million monthly visits in June, more than its next two rivals combined.
"The primary user of software is changing from humans to Agents," Akai, an industry insider, said.
ByteDance on July 30 folded its Feishu product team into Doubao and merged Feishu's sales arm with Volcano Engine to form a new To B unit called the "Creativity Service Platform." Alibaba on July 23 combined QoderWork, Wukong and MuleRun into Qianwen Office under DingTalk. Tencent on July 13 moved its QClaw product center into the WorkBuddy team.
The prize is more than market share. Desktop agents that operate files, browsers and spreadsheets capture a new data type — the full trajectory of how a task is executed — that model makers are paying top dollar to obtain. The shift is already hitting legacy office software: Goldman Sachs on July 21 downgraded Kingsoft Office to "sell" from "neutral," citing WPS AI's slow adoption against better-priced rivals.
WorkBuddy, launched in February, drew more than 8 million visits in its first month and exceeded 13 million daily active users by July, according to Analysys data. Its June desktop traffic of 20.97 million topped the combined total of the second- and third-ranked products, making it the reference point that forced rivals to consolidate. Tencent has defined WorkBuddy as a "super project," with Ma Huateng personally overseeing progress.
The three companies are taking different routes. Alibaba's Qianwen Office runs as a standalone desktop client that connects to DingTalk and, with authorization, reads documents on external platforms including Feishu — though it still relies mainly on the Qwen model. ByteDance keeps two lines: Doubao Enterprise Edition, built on Feishu's documents, meetings and organization charts, and TRAE Work, which grew out of a coding tool and offers multiple model choices. Tencent's WorkBuddy aggregates several models and can call WPS and Tencent apps, while a separate in-testing agent called Dayuan grows inside WeCom.
The deeper value lies in what these agents record. When an agent searches, reads a PDF, drafts a document and revises it with a user, it logs a complete execution trajectory — not just a question-and-answer pair. A Springer review of agent evaluation this year concluded that assessing an agent requires recording the full trajectory, including tool calls, intermediate reasoning and error correction, because the final answer alone is insufficient.
That data is becoming a scarce commodity. "There are model companies that approached us hoping to cooperate, to obtain industry-specific long-horizon task data through us. They are all paying high prices to buy this kind of scarce dataset," the head of a Shanghai AI platform company said.
The appetite extends beyond China. Anthropic in May formed an enterprise AI services company with Blackstone, Hellman & Friedman and Goldman Sachs, sending engineers into client firms to find where Claude fits. A week later, OpenAI launched OpenAI Deployment Company with more than $4 billion in initial investment and plans to acquire consulting firm Tomoro and its roughly 150 field engineers. A Shanghai AI company executive said the deeper goal is to learn real workflows and absorb the methodology, since most models today deliver outputs but lack understanding of how businesses actually run.
The economics are lopsided. Doubao, with 200 million daily active users in the first half of 2026, earns less than 1 million yuan a day while consuming tens of millions of yuan in computing power daily, according to LatePost. ByteDance plans to raise 2026 capital expenditure above 200 billion yuan, about 60 percent of its 2025 profit. Feishu, with about 4,000 employees and annual labor costs of 3.2 billion to 4 billion yuan, posted ARR above 3 billion yuan in 2025 and grew second-quarter 2026 ARR more than 100 percent year on year — yet still runs a profit gap after computing and marketing costs.
iiMedia Research estimates China's AI agent market reached 80.4 billion yuan in 2025 and will grow to 696.8 billion yuan by 2030. The three-way race is far from settled, and each company still runs competing internal projects. But the direction is clear: office software is becoming a tool that agents call, not a destination users visit. The winner will be the company whose agents complete the most work — and capture the most workflow data to train the next generation of models.
This article is for informational purposes only and does not constitute investment advice.