DingTalk's Post-90s CEO Debuts 'Qianwen Office' Desktop Agent

DingTalk was born 12 years ago, and the attendance scheduling, hierarchical permissions, and approval flows of tens of millions of enterprises nationwide are deeply embedded in it. For customers of this scale, deleting or modifying any old component within the system could trigger business shocks; but blindly stacking new AI modules will only make the already heavy architecture more bloated. When the technology generation shifts from mobile internet to Agent intelligence, this "additive inertia" has become DingTalk's greatest constraint.

At this moment, the post-90s new leader Chen Yusen took on this challenge. Previously, he led the incubation of the personal long-range desktop Agent "MuleRun" at Alibaba Cloud, achieving a commercialization result of "34% of users paying over $200 per month," which became a key foreshadowing for his takeover of DingTalk and Alibaba's enterprise AI business. But how the individual-track single-soldier execution capability can penetrate the intricate business flows of government and enterprise remains an unsolved problem.

103 days later, Chen Yusen delivered his first report card at this year's Yunqi Conference. He did not announce rebuilding an "AI version of DingTalk," but instead merged and consolidated the previously scattered product line resources such as MuleRun, Wukong, and QoderWork, creating an independent desktop Agent execution entry point called "Qianwen Office" on top of DingTalk, and bringing the core capability Enterprise Context to the forefront.

Chen Yusen used two sets of vastly different numbers on site to expose the physical bottleneck behind the "efficiency illusion": the average internal data scale of mid-to-large enterprises is about 150PB, while the physical context window limit of current top large models is only around 1 million tokens. The magnitude difference between the two determines that the brute-force approach of directly feeding enterprise data to the model will not work. In actual business, merely uploading a multi-dimensional report or project collection over 50MB is enough to overload and overflow the model's context window.

A McKinsey survey reflects the collective anxiety of enterprise CIOs: corporate investment in AI has surged to nearly 4 times the original average, with 93% of surveyed enterprises experiencing budget overruns. But the more awkward contrast is: although 80% of surveyed employees believe AI significantly improved personal productivity, only 37% of managers believe it truly contributed positively to organizational performance.

The deeper resistance lies in the interaction form itself. If AI is confined to "one-on-one private chats" in instant messaging, it essentially creates new information silos. Employees repeatedly debug prompts in private chat boxes, and after working out results, they still need to manually copy and paste them into DingTalk project groups to coordinate all parties and repeatedly follow up. This not only fails to shorten the collaboration chain but adds an unnecessary manual relay loss.