Baidu’s Dazi grows 1,063.79% to 6.74 million users

At Baidu’s August 27 product release, one number put the office-AI race into perspective: Dazi had reached 6.74 million monthly active desktop users in July, growing 1,063.79% in a month. The AICPB ranking placed it second behind Tencent’s WorkBuddy, at 11.15 million users. Baidu said Dazi’s user base had grown nearly ninefold over the previous month.

The speed is visible in the product itself. Since launching in March, Dazi has gone through roughly 150 iterations, sometimes shipping almost one release a day. Its enterprise edition now bundles 15 suites and 96 skills covering finance, legal, product and research and development, operations, and human resources. Daily questions have grown 60-fold since launch, while website visits rose 845% month over month in July.

Baidu’s argument is that a useful agent needs more than a stronger model. Executive vice president Shen Dou said it must understand the user, assemble the right tools, and apply professional methods. Dazi draws on Baidu’s AI cloud and its ability to customize delivery, while Tencent’s WorkBuddy is oriented toward communication and documents, and Alibaba’s Qwen toward team collaboration.

That product focus fits chairman Robin Li’s broader measure of progress: DAA, or daily active agents, which counts agents completing tasks each day rather than tokens consumed. Baidu is applying the idea to ERNIE through products including AI search, digital humans, Miaoda, Famou and Dazi. The proposed loop is straightforward: applications expose valuable tasks, the model improves at them, and usage supplies data for the next iteration.

So what changes in practice? Enterprises can access one Dazi edition across a wider set of office functions, with customization built into Baidu’s pitch rather than treated as an afterthought. For Baidu, the early evidence is reach and rapid iteration; it is not yet proof of durable retention or task quality. The user and growth figures come from the AICPB ranking and Baidu’s own disclosure, and the company’s larger bet is still that repeated task completion—not raw model usage—will decide which office agent lasts.