China’s AI Office Race Is Turning Productivity Apps Into Goal-to-Result Agents

The emerging AI-office model is not “open a spreadsheet, document, or meeting app.” It is: state an outcome—such as preparing a sales review, identifying exceptions, assigning owners, and initiating approvals—and let an agent assemble authorized context, take steps across tools, and return a result for human review.

That is the strategic logic behind a rapid set of product consolidations by ByteDance, Tencent, Alibaba, and Kingsoft/WPS. Their products differ, but the common ambition is to make an agent the front door to work while conventional office applications become execution tools behind it.

The shift: from artifact creation to outcome execution

Traditional productivity software starts with an artifact: a blank document, workbook, slide deck, calendar event, or meeting. An agent-oriented product starts with a business objective.

To complete that objective, an enterprise agent needs more than a capable model. It needs access—within the user’s authorized scope—to current documents, messages, meetings, calendars, workflows, and business systems. It also needs a way to turn an answer into an action: create a file, update a record, notify colleagues, route an approval, or assign a task.

This reframes the competitive unit. The contest is no longer simply document editor versus document editor, or chat app versus chat app. It is a stack of model capability, agent runtime, enterprise context, identity and permissions, integrations, and controlled execution.

ByteDance: Doubao becomes the work entry point; Feishu supplies the context

ByteDance made the most explicit organizational move toward this architecture. On July 30, it merged Feishu’s product team with the Doubao product organization, while moving Feishu’s marketing, sales, and customer-service operations into Volcano Engine’s enterprise go-to-market organization. 3738

The split is revealing: Doubao is positioned closer to product intelligence and agent experience, while Volcano Engine carries more of the enterprise delivery and commercial motion.

On August 25, ByteDance launched Doubao Work, a standalone AI productivity product built around deep Feishu integration. Signing in with a Feishu account can give the agent access, within the user’s permissions, to work context including chats, documents, meeting notes, and calendars—and allow it to write results back into the workplace system. 1726

ByteDance subsequently folded the TRAE coding and Coze agent-building teams into Doubao, further concentrating work-oriented AI capabilities under the Doubao brand. 2931

Some reports also describe strong Feishu commercial growth, including reported 2025 revenue above RMB 3 billion and more than 100% year-on-year growth in the second quarter of 2026. Those figures are media-reported rather than company-audited disclosures, so they should be treated as directional rather than definitive. 13

Tencent: consolidating desktop agency around WorkBuddy

Tencent’s July 20 reorganization moved QClaw’s business and part of its team into the cloud division that houses WorkBuddy, while QClaw continued operating as a product. 410

The likely product rationale is to bring together a broad workplace agent and computer-use or desktop-control capabilities under one management system. Rather than asking users to choose between disconnected agent products, Tencent can present WorkBuddy as an AI workstation and use QClaw-related capabilities as part of the execution layer.

The products already had notable reported PC reach before the consolidation. QuestMobile data cited in subsequent coverage put July 2026 monthly active users at 6.582 million for WorkBuddy and 2.259 million for QClaw. These are third-party estimates, not Tencent disclosures. 6

For Tencent, the open question is how deeply agent execution can connect to enterprise systems and collaboration surfaces, including WeCom, rather than remaining primarily a powerful desktop assistant.

Alibaba: Qianwen Office unifies desktop, cloud, and collaboration agents

Alibaba’s Qianwen Office, also referred to as QwenWork, consolidated three previously separate products on August 3: QoderWork for desktop control, MuleRun for cloud-side execution, and Wukong for enterprise collaboration agents. It entered public beta at the same time. 420

The combination makes the intended architecture unusually clear:

• a model supplies reasoning and generation;
• an agent layer coordinates work across desktop and cloud environments; and
• enterprise systems provide the data, organizational context, permissions, and workflow endpoints.

Qianwen Office was placed under the ATH organization led by DingTalk CEO Chen Yusen, tying the agent strategy closely to DingTalk’s enterprise environment. 320 Coverage also reported initial DingTalk connections for functions including group-chat reading, attendance and approval handling, meeting scheduling, and document or form generation. 25

That matters because DingTalk is not merely a distribution channel. It can be the governed environment in which an agent retrieves information and carries work into operational processes.

Kingsoft/WPS: placing an agent layer in front of the document suite

Kingsoft Office represents the incumbent-suite version of the same shift. It released WPS Comate for organizational customers in mid-July, positioning it as an AI office entry point spanning tasks, experts, teams, skills, knowledge resources, applications, and automation. 3

Reports also describe Lingxi Professional as an independent agent rather than only an in-app WPS feature, able to call WPS through plugins and command-line interfaces. 13 The reporting available here offers less independently detailed information about Lingxi Professional’s packaging and organizational structure than it does for the ByteDance, Tencent, and Alibaba changes.

Still, the direction is clear: WPS’s durable advantages in document compatibility, editing, and enterprise deployment are being recast as capabilities that an agent can invoke. The suite is no longer just the place where a user manually creates an artifact; it can become the system that completes part of an agent-directed workflow.

Why Feishu, DingTalk, and WeCom still matter

A general-purpose model can draft a memo. It cannot safely act on behalf of an organization without a governed operating layer. This is why collaboration platforms may become more—not less—important in an agent-first market.

Identity, authorization, and least privilege

Enterprise platforms hold directories, user accounts, roles, and permission inheritance. These determine what an agent may read, modify, send, or approve on a particular employee’s behalf. Without those controls, a highly capable agent is not enterprise-ready; it is an unmanaged automation risk.

Live work context

The facts needed to complete work are usually scattered across shared files, chats, meeting notes, calendars, project records, and approvals. Doubao Work’s account-level Feishu integration is a practical example: the agent can use workplace information available within a user’s permission boundary rather than relying solely on a generic prompt. 1726

Workflow and execution surfaces

The final step is often the hard part. Organizations need agents to create records, route approvals, send notifications, update tasks, and preserve an audit trail—not merely recommend what a worker should do next. DingTalk, Feishu, and similar systems can provide the workflow primitives and integration points that make this possible.

Distribution and trust

Feishu, DingTalk, and WeCom already occupy daily coordination surfaces. One QuestMobile-based report estimated that the three together reached 92% coverage of China’s collaborative-office market in 2026; that figure is third-party market research and should not be read as audited platform data. 21

The implication is that a collaboration platform does not need to own the strongest foundation model to remain indispensable. It can be the permissioned system of record and action plane that makes agents useful in real organizational work.

What the market may look like next

The July–August 2026 consolidations suggest several likely changes.

Competition moves from apps to integrated stacks. Vendors will increasingly compete on the combination of model quality, agent reliability, enterprise data access, permissions, connectors, and workflow completion—not on isolated features in documents or meetings.

The interface may become secondary to the task. Workers may begin with requests such as “prepare the regional review and start follow-up approvals,” then inspect, edit, and approve the agent’s output. Documents, spreadsheets, and chat threads remain important, but increasingly as places where work is executed and reviewed.

Enterprise infrastructure becomes a stronger moat. An agent connected to a company’s identity graph, permission model, knowledge base, workflow history, and internal integrations is harder to replace than a standalone drafting tool. That favors platforms with existing enterprise footprints.

Standalone tools face a sharper test. A point application can still win through specialized workflows, proprietary data, compliance capabilities, or superior execution in a domain. But generic productivity functions risk becoming services called by an agent rather than destinations users visit directly.

Governance will decide adoption. Impressive demos are not enough for high-value workflows. Reliability, delegated authority, traceability, human approvals, data controls, and the ability to correct or reverse actions will determine whether these products graduate from personal assistance to enterprise operations.

The core bet is therefore simple: the winning AI-office product may not be the best place to make a document. It may be the default place to state a goal—connected to the enterprise systems that can understand, authorize, and complete the work.