Alibaba rolls out multi-person workbench for Qianwen Office as office agent report debuts

According to reports from Leiphone.com and QbitAI, the workbench turns natural-language descriptions into a live web application with role-based permissions, a cloud database, an admin console and one-click online publishing. It targets scenarios including event organizing, home-school collaboration and enterprise teamwork, seeking to address the high cost of standard SaaS subscriptions and the long lead time of custom software development. Users can access the feature from the Qianwen Office homepage.

The tool is designed for multi-party processes rather than personal one-off pages. In a large market event, for instance, it can generate a vendor recruitment workbench where applicants submit materials, organizers review and allocate booths, and finance staff track deposits. In schools, teachers can assign and grade homework, students can submit assignments, and parents can view only their own child's information through role-restricted permissions. For companies, it can build workbenches for influencer partnerships, supplier coordination or multi-city store openings.

Traditional enterprise software purchases, such as ERP, OA or vertical SaaS systems, can cost tens of thousands of yuan per year, while custom development may add hundreds of thousands of yuan and several months of work. Qianwen Office, which the companies behind it describe as the world's first agent product aimed at enterprise scenarios, said it passed 30 million users within its first month, with enterprise users accounting for more than half. The company said the new capability pushes AI from individual efficiency gains to organizational productivity.

Also on September 7, the China Office Agent User Behavior Incomplete Report was released in Beijing, according to an article by QbitAI that cited data from LobsterAI, an open-source desktop office agent developed by NetEase Youdao. The report found Beijing leading all Chinese cities in user numbers, with Shanghai, Hangzhou, Guangzhou and Shenzhen together contributing about 27%. Overseas users accounted for 12.75%, led by the United States with 2.73%.

The report revealed a concentrated usage pattern: the top 20% of users consumed 87.4% of total computing power, and the top 5% consumed 53.5% of tokens. Paying users consumed 6.2 times more tokens, completed 5.2 times more tasks and were active 3.0 times more days per month than free users. Single-task scale grew 3.1 times over five months, with a 53% month-over-month increase in August. Nearly 60% of token consumption happened outside office hours, and 12.2% of model calls were not triggered by a human in real time; each scheduled instruction drove an average of 64 model calls, versus 10.6 for manually triggered tasks.

DeepSeek V4 Flash was the most-called model at 52.2%, or more than 75% when combined with its vision experimental version. New models reached nearly 50% user penetration within three days of release. Programming and debugging dominated task types at 38.5%, followed by document writing at 10.1% and data and spreadsheets at 8.8%. By industry, software/IT contributed 38.9% of tasks, design/media/creation 13.0%, and internet product operations 9.7%. The report described office agents as becoming general-purpose AI productivity infrastructure.