The first eighteen months of enterprise AI were spent proving that one agent can finish one task. The harder problem sits one level up: an agent can complete a job, but it cannot catch a job as it passes between people.
That gap is where most of the lost time hides. When a colleague hands work to another colleague after the AI is done, the re-briefing can take longer than the task the AI just performed. Qwen Office, Alibaba’s workplace product, shipped version 2.0 on 9 November with a feature aimed squarely at that handover: shared workspaces where several people and one agent work on the same material at once. The company calls the direction Team Agent.
Two kinds of productivity, one of them neglected
Companies evaluating AI ask two questions. Can it take work off an individual, and can it cut the information loss and duplicated effort inside a team. The first question has been answered often enough this year. The second has not.
The commercial logic of seat-based pricing is being renegotiated along with it. A seat was originally an access and budget control: who may use the tool and how much. Once the agent becomes a digital colleague, the seat stops meaning that and starts meaning how many people are wired into one human-agent working system.
In a trial run, reporters and a senior editor used the shared workspace with the agent, called Xiao Q, to plan a feature. Given the instruction to develop story angles from existing interview material, the agent read four source documents, produced six industry-observation pitches plus one profile alternative and flagged the risks. As the two humans argued, it tracked both sets of comments and revised the framework. Nobody had to re-explain the background.
Sixty-two stores, sixty group chats
The lighter case is a newsroom. The heavier case is retail. In a pilot at Intime Department Store, the group-buy voucher business involves more than 62 stores and over 200 people, and each voucher requires about 50 checks before it goes on sale. That information lived in more than 60 DingTalk group chats, and staff rebuilt progress by scrolling and copying.
Linking the chats and the voucher spreadsheet into one workspace changed the shape of the job. The agent summarises progress, updates the table, runs scheduled checks on outstanding items and reminds the specific owner. The work stopped being a relay race through chat history.
The comparison is Slack, not a chatbot
Anthropic’s Claude Tag put an assistant into Slack as a team member: reading channel context, calling authorised tools, posting progress and chasing stalled threads. Qwen Office 2.0 does the equivalent inside DingTalk, connecting group chats, document spaces and AI tables so the agent works against real business material and real deadlines.
Two behaviours matter more than the feature list. The first is co-editing: instead of generating a document and handing it over, the agent takes comments in place. In one test it reviewed a draft, the human marked specific typos and structural problems in the file, and the agent applied all three instructions directly to the original document in one pass. The copy-paste loop between tools disappears.
The second is task initiation. A Google Research survey this year noted that agents can reason, plan and call tools but still wait for a human to notice that work needs doing. Qwen Office 2.0 adds trigger-based tasks, where a sentence of natural language defines the condition and the handling rule. In testing, the agent watched a DingTalk group for new interview notes, summarised them and posted the summary back without being asked.
From entry point to operating layer
Look past the product and the strategic shift is visible. The last year of Chinese workplace AI was horizontal: consolidate agents, pool resources, win the entry point on the desktop. That contest is not settled, but the weight is moving vertical, to whoever can get an agent actually scheduled inside a company’s processes.
One pilot customer described the result as an agent operating system. The comparison is loose but useful: an OS schedules applications and compute, and this layer schedules tools, data and agent capabilities against a task. The user hands over the job, the digital colleague interprets it, dispatches the capability and pushes it forward.
What that changes is who does the assembling. An assistant helps an individual write, analyse or summarise. An agent that holds business context and can call across tools can absorb work that used to require several roles in sequence, which is where the cross-department friction and the rework actually live.
The ceiling is still the integration. IM and Office are now connected inside the product, and more DingTalk primitives, SaaS tools and professional software are being wired in. But an agent that operates real business process needs the data and the rules behind those tools, not just the interface on top, and that is a slower negotiation than any feature release. As one enterprise software veteran of thirty years put it when the product launched its e-commerce edition, AI office software only becomes alive once it stops being only about the office.
The entry point decides who gets seen. Whether the agent gets inside the process decides who stays.
Editor’s note: This is an adapted translation of the original Leiphone report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.leiphone.com/category/ai/Zt5mFmrAgQP43K9x.html.