AI Agents are moving beyond chat windows and stepping into the organizational fabric of enterprises, with Alibaba's Qwen Office filling out the product chain from "understanding business" to "executing tasks."

On September 22, Qwen Office unveiled a series of enterprise-grade AI Agent products at the 2026 Hangzhou Apsara Conference, introducing features such as Enterprise Context, Digital Employees, Collaboration Spaces, and a Security Center. This rounds out the Agent product system spanning business understanding through to execution and collaboration. On the hardware front, Qwen Office simultaneously launched the AI Agent personal assistant, the QwenNote A2.

Among the announcements, Enterprise Context integrates enterprise data to help Agents retrieve task-relevant information, laying the groundwork for building Digital Employees that genuinely "understand the business."

Chen Yusen, Vice President of Alibaba Group and CEO of Qwen Office, stated on stage, "Qwen Office is the industry's first enterprise-grade Agent product." He emphasized that for Agents to truly land in the enterprise world, they must not only grasp the business but also collaborate like humans within companies and organizations. Moreover, each task execution must have clear boundaries and be fully traceable.

Regarding hardware, Chen Yusen revealed that the previous generation, DingTalk A1, has achieved sales of "several hundred thousand units" domestically in China. He set an ambitious sales target of "ten million units" for the new QwenNote A2, underscoring Qwen Office's further strategic push into the AI hardware market.


Enterprise Context: Enabling Agents to Truly Decipher Business Data

One of the core products launched this time is Enterprise Context, positioned as a context data management product designed for businesses.

The underlying logic is that while enterprises often accumulate vast amounts of data, Agents executing specific tasks don't need all the information available—they need the most relevant content for the immediate task at hand. Enterprise Context uses built-in proprietary models to compress and structure enterprise data at a low token cost, updating in real-time as business conditions change. This allows Agents to make judgments based on the latest business information.

This product directly tackles the data silos and information overload issues prevalent in enterprise AI deployment. Leveraging Enterprise Context and Qwen Office's Agent hosting capabilities, businesses can construct Digital Employees that "understand the business" and deploy them on instant messaging platforms like DingTalk, enabling them to collaborate with human employees to complete tasks.


Digital Employees and Collaboration Spaces: Agents Move Toward Organizational Operation

At the Digital Employee level, Qwen Office endows Agents with a complete "organizational identity": they have a name, a department, a manager, and defined job responsibilities. They also have authorization scopes and lifecycle management. Agents can be recognized by employees, authorized by the system, and deactivated by administrators. All execution records can be traced back to this corresponding identity.

Simultaneously, Qwen Office introduced a "Collaboration" feature that allows enterprises to create collaborative spaces. These spaces unify information scattered across group chats, documents, and knowledge bases, and allow team members and Digital Employees to be added as needed. Within these spaces, employees can directly discuss with colleagues or Agents, and Agents can communicate, divide labor, and collaborate with one another. Task outputs and progress are consolidated within the space, reducing the friction of information being passed back and forth across different groups.

For work scenarios that aren't suitable for a chat interface, Qwen Office also launched an "Applications" feature. This supports enterprises and SaaS vendors in integrating their business systems as applications, preserving the original interface and data while layering Agent capabilities on top. Users can directly modify and save AI-generated results within their business systems, eliminating the need to switch between multiple tools.


Security Center: Managing Agent Risks at the Architectural Level

As AI transitions from "answering questions" to "executing actions," security boundaries become a critical consideration for enterprise Agent adoption. Chen Yusen noted that companies cannot merely add defenses at the system's perimeter; they must incorporate considerations for permission breaches, data leaks, and operational errors from the very start of architecture design.

The newly launched Security Center for Qwen Office can centrally manage Agents' data access, tool calls, and task execution. The specific mechanisms include: setting execution boundaries through a sandbox before execution, approving and intercepting high-risk operations during execution, and handling anomalies post-incident through operation logs and recovery mechanisms.

Combined with the DingTalk enterprise security system, this product also supports least-privilege authorization, sensitive information detection, and comprehensive auditing. This allows enterprises to trace the task initiator, the data used by the Agent, and the specific actions executed.


QwenNote A2: AI Hardware Differentiated by Privacy Protection

On the hardware front, Qwen Office introduced the AI Agent personal assistant, the QwenNote A2, which is promoted with the tagline "True AI, No Recording," addressing privacy concerns associated with traditional AI recording devices.

In terms of privacy protection, the QwenNote A2 features a "single transcription, burn-after-transcription" design: the device does not record by default. After a conversation ends, the audio is transcribed into text once in the cloud, after which the original recording is permanently deleted. While the device is working, a discreet green light illuminates to signal to the other party that only notes are being taken, not recorded. Weighing 67 grams and equipped with a six-microphone array, the device supports independent network connectivity and can invoke Agents without needing a paired smartphone.

In terms of use cases, the QwenNote A2 supports one-click conversion of offline conversations into cloud-based tasks. For example, after a salesperson visits a client, they can use voice commands to sync the conversation content to Qwen Office, automatically generating PPTs or extracting key information.

Chen Yusen disclosed that the previous generation product, DingTalk A1, has already sold "several hundred thousand units" in China. He believes that AI recording represents only a small niche; products like AI note-taking and AI-powered meeting intelligence have much larger application potential and could realistically achieve sales at the "ten million unit" level in the future.


Qwen Intelligence: Expanding the Qwen LLM's Presence in Personal Intelligence

In a related move, the Qwen large language model also released a full-stack AI smartphone solution called Qwen Intelligence. Aimed at phone manufacturers, it provides models, platforms, and scenario-based solutions to push phone-embedded Agents from simple Q&A toward complex task execution.

Addressing the core needs of mobile Agents, Qwen Intelligence debuted three solution suites: the Mobile Planner Agent, which handles understanding requirements, decomposing tasks, and dynamic planning; the Mobile-Use Agent, which handles actual phone operations; and the Mobile Creative Agent, aimed at imaging and creation, transforming a user's one-sentence request into a complete creative workflow. Official data shows that the Mobile-Use Agent achieves a 90% end-to-end task success rate in real-world testing, with single-task completion times reduced from 83.6 seconds to 59.5 seconds.

Currently, Qwen Intelligence has already entered practical application. Alibaba has partnered with Honor, using Qwen Intelligence and Honor Magic OS as the foundation, to jointly build vertical domain models and solutions for next-generation AI smartphones. Official data indicates that this solution achieves an overall task accuracy rate of 91.8%, can execute complex, long-horizon tasks involving over 100 steps, and reaches a 90% end-to-end service closure rate.