The chatbot in the corner of the screen is gone. In its place is something that opens the spreadsheet itself. Across China's biggest tech platforms, the "assistant" grew hands this year. The question is no longer what AI can say. It is what it is allowed to do.
The launch window
The summer of 2026 was a near-simultaneous release of office agents (智能体):
- ByteDance launched Doubao Work (豆包工作) on 25 August 2026 — an agent product integrated with Feishu (飞书) that can operate a browser and PC, take long-running cloud tasks, and accept mobile remote commands (Guangming, Xinhua Finance).
- Alibaba's Qwen Office (千问办公) targets enterprises and plugs into DingTalk (钉钉), with 25 business modules and over 1,000 commands.
- Tencent WorkBuddy shipped an enterprise edition and an Agent Suite connecting Hunyuan, DeepSeek, Zhipu GLM and Kimi; on 2 September its open platform went live with 100+ ecosystem partners and 9 co-branded devices.
- Baidu's Kuku AI (库库AI) reported more than 25 million monthly active users.
- Kingsoft WPS pushed two agents: WPS Comate, used at Peking University First Hospital, and Lingxi, a personal assistant.
The race is no longer about who has the best large model (大模型). It is about who controls the layer where the model actually touches work.
A hospital, not a keynote
The most concrete proof is mundane. At Peking University First Hospital, WPS Comate cut the data-processing cycle for quality-supervision reports from two to three weeks down to two to three days. WPS also topped an international spreadsheet-agent benchmark, SpreadsheetBench V2, at 50.86% — ahead of Claude Opus 4.6 (34.89%), GPT-5.2 (26.79%) and Gemini 3.1 Pro (23.68%) as reported by the榜单.
That is the kind of before/after number procurement teams actually trust: not "AI is smart," but "this report used to take 15 days and now takes 3."
Why now
Two signals explain the timing. Analysys (易观) reported that in Q2 2026, monthly visits to 17 mainstream desktop AI office agents rose from 20 million in March to 60 million in June (China Economic Net). And analysts at Gartner and IDC forecast China's enterprise AI Agent (智能体) market at RMB 44.9 billion (≈ US$6.3 billion / HK$49 billion) in 2026, up 112% year on year (21st Century Business Herald).
The users are no longer playing. They are processing documents, meetings and data.
The pattern: context is the moat
Every vendor chose the same strategic move — embed the agent inside the collaboration tool people already use, rather than ship a new app:
- Doubao + Feishu: the agent inherits the organization's chats, docs and calendars.
- Qwen + DingTalk: AI is injected into approvals, forms and meetings as workflow nodes.
- WorkBuddy + WeChat Work: an ecosystem play bridging hardware, apps and developers.
The shared design is "AI processes information, a human makes the key decisions." Agents trigger inside existing business flows — invoice pre-checks, contract review, weekly reports — rather than replacing the system of record.
The readiness gap
Not every team is set up to absorb an agent. Companies with clean, structured document warehouses and clear approval chains see value in weeks; teams still running ad-hoc file shares and undocumented workflows mostly get a smarter search box. The differentiator is rarely the model itself — it is whether the surrounding process is legible enough for an agent to act inside it without a human firefighting every exception. Procurement teams should budget as much for workflow cleanup as for the agent license.
The worries
Scaling agents inside real companies surfaces three hard problems. Data security: an agent that can read files and send email is a real action, not a suggestion, so 360 built "native security" into its Nanowork product from the factory. Token cost at scale: long-running multi-agent tasks rack up usage fast. Core-system integration: deep customization of ERP and CRM still needs connectors, APIs and zero-code orchestration.
Honest limitations
- Launch dates and product claims come from company announcements and Chinese state and business media (China Economic Net, Guangming, Xinhua Finance, 21st Century Business Herald); we did not test the products ourselves.
- The 50.86% benchmark is WPS's result as reported on SpreadsheetBench V2; leaderboard positions shift monthly.
- The RMB 44.9 billion market forecast is an analyst figure (Gartner/IDC via media), not our estimate, and "agent (智能体) market" definitions vary.
- Baidu's 25 million MAU is vendor-reported; we have no independent usage data.
- FX approximated at RMB/USD ≈ 7.1, RMB/HKD ≈ 1.10.
What readers can do now
- Pilot one agent inside your existing suite — Feishu, DingTalk or WPS — on a single repetitive workflow before buying any standalone tool.
- Put human-approval gates on any agent that can write or send. Treat it like a junior employee with access, not a chatbot.
- Use the RMB 44.9 billion forecast as a barometer of enterprise pull — but verify against your own procurement and renewal data before betting the budget on it.