Qianwen Office spent 41 minutes verifying a single report

Qianwen Office (千问办公) spent 41 minutes generating a single one-week embodied AI industry report, prioritizing deep search over rapid generation. In a benchmark reviewed by woshipm, TRAE Work finished in 16 minutes, and WorkBuddy finished in 7 minutes. That speed sacrificed accuracy. WorkBuddy claimed no major public developments emerged for Leju Robot (乐聚机器人) or Tashi Zhihang (它石智航). Yet the Shenzhen Stock Exchange accepted an IPO application from Leju Robot, and a Tashi Zhihang co-founder made the MIT TR35 list during that period.

Speed without verification simply transfers labor back to human reviewers. When a model overlooks material events, an operator must audit every missing line.

Woshipm observed similar failure modes in TRAE Work, which marked six companies as having no verifiable developments while relying on aggregation platforms like Toutiao. Fixing this gap requires broader workspace integration. According to ifanr, Qianwen Office launched a Collaboration feature where human workers and digital employees share project context and divide labor in a unified environment. The decisive metric is no longer draft generation time, but how many manual handoffs remain before the work is done.

The 41-minute audit against the seven-minute draft

When tasked with drafting a one-week embodied AI industry report, WorkBuddy finished in 7 minutes, while Qianwen Office (千问办公) took 41 minutes according to woshipm. Speed concealed omissions. WorkBuddy reported that no major public developments were retrieved for Leju Robot (乐聚机器人) or Tashi Zhihang (它石智航). During that test week, the Shenzhen Stock Exchange accepted Leju Robot's IPO application, and an MIT TR35 honor went to a Tashi Zhihang co-founder.

Speed does not settle the workload. A fast draft missing core public filings forces human analysts to redo the search manually.

By contrast, Qianwen Office spent 41 minutes verifying citations. When citing US Federal Communications Commission import restrictions on Chinese humanoid robots, it flagged the claim as a single source and recommended further verification. According to ifanr, Qianwen Office manages these tasks through a Collaboration feature where human staff and digital employees share context in a unified space. That extra time caught what others skipped. TRAE Work finished in 16 minutes, yet its report pulled links from Toutiao aggregators and marked six firms-including Qianxun Intelligence (千寻智能) and LimX Dynamics (逐际动力)-as having no verifiable developments.

In June, Tencent disclosed that WorkBuddy can collaborate inside Tencent Docs, while Tencent Meeting opens context through Model Context Protocol (MCP) and Skills interfaces. Yet document embedding cannot fix missing facts. In a Jefferies benchmark evaluating eight agents across five office tasks, Qianwen Office ranked first with 95 points, while WorkBuddy tied Gemini Spark for last place with 66 points. Qianwen Office's foundation model, Qwen 3.8 Max, scored 56 on raw intelligence, but engineering harness mechanisms lifted its total score to 95. Human cleanup remains the decider. A woshipm author argues the benchmark has shifted to the manual steps remaining after an AI answers.

Structuring live project context across chats and recordings

At the Apsara Conference in Hangzhou, Qianwen Office (千问办公) introduced Enterprise Context alongside QwenNote, Applications, and a Security Center. The system grew out of an earlier personal project named MyContext. Instead of relying on static prompts, it organizes information from group chats and documents around concrete tasks, updating automatically as business routines shift. According to ifanr, beverage chain Guming applies this setup by aggregating Feishu documents, knowledge bases, Q&A channels, and training schedules into a store operations knowledge space.

Live context requires clean input. Raw transcripts frequently fail when meeting jargon collides with generic speech engines.

Surrounding text resolves these speech errors. In a test reading DingTalk AI hardware A1 transcriptions, woshipm observed Qianwen Office correcting acoustic mistakes like 'VOA' to 'VLA' and 'cpr' to 'CVPR' through contextual analysis. The companion QwenNote A2 hardware assists by retaining text summaries rather than hoarding raw audio recordings.

Context also extends to peripheral hardware. Asus announced plans for Xiaoshuo Zhidao 2.0 (小硕知道2.0) to write long-term memory directly to personal storage drives during local model updates. In wearable gear, ifanr noted that Guangfan AI OS on Shokz earphones filters incoming notifications across WeChat and Feishu, summarizing critical alerts while dropping casual chatter.

Sandboxing digital employees before they sign or quote

A woshipm author argues that in the agent era, the smallest unit of product design is shifting from software to tasks. When those assignments involve executing commercial steps rather than generating passive summaries, technical safety controls become the decisive constraint. According to ifanr, Qianwen Office (千问办公) introduced an Application feature that allows enterprises and SaaS vendors to integrate agents without altering existing user interfaces or data stores.

When software takes commercial actions, passive drafting gives way to containment. Execution demands hard boundaries before an autonomous agent ever commits to a binding quotation.

Qianwen Office opened invitation-only testing of its new products to enterprises including Guming, Fulika, and Changan Automobile. According to ifanr, packaging company Fulika deployed three digital employees across six departments. These agents handle customer inquiries and quotations, alongside contracts across those units. To protect operations, the Qianwen Office Security Center manages data access and tool invocation during task execution. The architecture confines agent actions to sandboxed environments. It halts high-risk operations until humans issue approvals, while continuously recording audit logs.

Full-workflow execution remains rare across industry tiers. A March report by the China Academy of Information and Communications Technology showed that while over half of more than 6,000 surveyed small and medium-sized enterprises used AI, only 4% achieved full-workflow embedding. Structural scale drives this divide. According to woshipm, the AI adoption rate among large industrial enterprises in China was 19% in the second quarter, compared to 3.4% among small industrial enterprises. Reaching full deployment depends on sandboxes that trap errors before business commitments become binding.

Selecting Enterprise AI-Office Platforms and Edge Deployments

• Deep industry research or competitive intelligence where omissions and unverified citations carry business or legal risks Deploy Qianwen Office over speed-optimized research tools. In head-to-head benchmarking for an embodied AI report, WorkBuddy took 7 minutes but reported no developments for Leju Robot (which had an SSE IPO application accepted) and Tashi Zhihang (whose co-founder joined MIT TR35), while TRAE Work finished in 16 minutes relying heavily on Toutiao aggregators and missed milestones for six companies. Qianwen Office required 41 minutes, but it flagged single-source regulatory claims (such as US FCC restrictions) for verification and corrected ASR transcription errors ('VOA' to 'VLA', 'cpr' to 'CVPR').
• Multi-department operational workflows requiring digital workers to access scattered company knowledge without re-engineering existing enterprise software Use Qianwen Office's Enterprise Context and Application features. This architecture lets organizations aggregate disparate knowledge-such as Guming unifying Feishu documents, Q&A groups, and training schedules-and deploy digital agents across departments (as seen in Fulika's deployment of three digital workers handling customer inquiries, quotations, and contracts across six departments) while governing access via Security Center sandboxes, approvals, and audit logs.
• SMB offices or individual creators needing local, confidential voice transcription and long-term memory without round-tripping data to the public cloud Adopt hybrid edge-cloud hardware configurations like the Asus Poxiao Air running the Xiaoshuo Zhidao & WorkBuddy Exclusive Edition. Powered by an Intel Core 7 processor with integrated NPU, this setup runs Intel local ASR Skills and customized WorkBuddy Skills entirely offline without internet connectivity, storing long-term memory on local personal storage drives.
• Real-time collaborative drafting and meeting context integration within existing team office suites Leverage Tencent WorkBuddy for native document co-authoring inside Tencent Docs and meeting reading via Model Context Protocol (MCP) and Skills in Tencent Meeting, or evaluate Qianwen Office's planned integration into DingTalk, which connects agents directly into an ecosystem of 20 million to 26 million enterprise organizations.

Flat subscriptions break under long-running token consumption

According to ifanr, Gartner projected that 40% of enterprise applications will incorporate task-specific AI Agents by the end of 2026, up from less than 5% in 2025. Capturing that demand requires deeper corporate infrastructure. ByteDance merged its Feishu (Lark) product team into Doubao, while Alibaba established its ATH Business Group in March alongside Alibaba Cloud and e-commerce under Group CEO Wu Yongming. Alibaba focused the group's strategy on creating tokens and transporting them, alongside applying tokens in business software. Distribution alone does not guarantee recurring revenue when agents run indefinitely.

woshipm author Wang Zhiyuan argues that flat subscription models will become financially unviable for enterprise AI deployments because continuous 24-hour operations and risk control lead to unpredictable token consumption. When agents monitor communications or run audits, compute volume spikes. At public API pricing assuming an 80% input and 20% output distribution, 3 billion tokens equates to approximately 76,000 to 84,000 yuan.

Anthropic's enterprise edition reflects this reality. It charges a flat monthly fee of $20 per seat for platform and administrative management, while billing all model usage separately at API rates without an upper quota cap.

Even with lower domestic rates-the API price of Qwen 3.8 Max is one-fourth the price of GPT-5.6 Sol-autonomous background tasks drain budgets. Meanwhile, converting corporate seats remains slow. Microsoft reported 15 million paid Copilot seats out of 450 million Microsoft 365 users, representing a 3.3% enterprise penetration rate.

Enterprise AI Office Ecosystem Comparison: Alibaba (Qianwen Office) vs. Tencent (WorkBuddy)

According to ifanr, AI hardware competition is shifting away from single-point functional features toward integrated system-level capabilities. Specialized hardware makers deploy external software like Guangfan AI OS across accessories, avoiding internal development costs for model scheduling or maintenance. The interaction remains hands-free. Shokz earphone wearers call on "Xiao Fan" (晓帆) to arrange appointments or set memos. When calendar conflicts appear, the system detects the clash and verbally prompts the user to reschedule. It also screens notifications from WeChat and Feishu (Lark), reading important alerts while ignoring casual chatter across three summarization speed modes.

Reliability in bounded, reversible handoffs decides agent adoption. When an assistant fails to screen message noise or arbitrate a schedule, it cannot be trusted with critical business systems.

For readers outside China

• Availability: The enterprise solutions discussed are focused primarily on mainland China's enterprise software ecosystem. Qianwen Office launched invitation-only testing for domestic enterprise clients, including beverage chain Guming, packaging company Fulika, and Changan Automobile, with ongoing integration into DingTalk. Tencent WorkBuddy has achieved broad domestic desktop adoption (recording 20.97 million visits, 20 million monthly active users, and 13 million daily active users across platforms in June). International general availability, Western cloud marketplace listings, and cross-border enterprise licensing are not disclosed in the sources.
• Pricing: Platform-specific seat pricing for Qianwen Office and WorkBuddy is not disclosed in the sources. At the model layer, Qwen 3.8 Max API pricing is established at one-fourth the price of GPT-5.6 Sol. For high-volume agent usage, public API consumption assuming an 80% input and 20% output distribution costs roughly 76,000 to 84,000 yuan per 3 billion tokens. By comparison, the sources note that Anthropic's enterprise edition bills a flat $20 per seat monthly fee for administrative tools while passing model consumption through at uncapped API rates, and Kingsoft's WPS 365 generated 497 million yuan in first-half revenue under its enterprise subscription model.
• Closest Western equivalents: Microsoft 365 Copilot; Claude Cowork; OpenAI Codex; Gemini Spark
• Data residency: Data protection is approached through architectural boundaries and edge isolation rather than Western regulatory compliance certifications (e.g., SOC 2 or GDPR, which are not disclosed in the sources). Qianwen Office operates an enterprise Security Center featuring sandboxed execution environments, mandatory approval gates for high-risk operations, and audit logs; its companion QwenNote A2 hardware stores text summaries while discarding raw audio recordings. For sensitive local processing, the Asus and Intel collaboration on the Poxiao commercial PC series deploys on-device ASR and customized WorkBuddy Skills that process data locally without internet connectivity and commit long-term memory to private personal drives.

Sources

• 36kr 科氪 | 破晓先行，智领未来！华硕商用破晓家族拥抱AI Agent时代 https://36kr.com/p/3993338830486272
• ifanr 千问办公押注的企业上下文，是 Agent 时代的组织语言 https://ifanr.com/1681363
• woshipm 耗时41分钟，千问办公押注了怎样的Agent未来？ https://woshipm.com/ai/6445266.html
• ifanr 光帆携手韶音，以 AI OS 破局单点功能内卷 https://ifanr.com/1681475
• woshipm 阿里、腾讯重新打办公这一仗：烧掉30亿Token后，我选阿里 https://woshipm.com/it/6442725.html
• woshipm AI 办公的产品仗已经打完了 https://woshipm.com/ai/6457926.html