Source | Benyuan Finance

Author | Li Youshan

The AI sector continues to be a hotbed of rapid change. Leading companies have pivoted from early-year strategies focused on acquiring consumer-end (C-end) users through various entry points, parameter benchmarks, and rankings, to now emphasizing business-end (B-end) productivity solutions, offering office assistants tailored for enterprises and their workforce.

Tencent's WorkBuddy exemplifies a scenario-driven model, leading the pack in office intelligence agents. In July, Tencent merged its QClaw product center with WorkBuddy. Following Alibaba's move to integrate AI into DingTalk, Chen Yusen, former head of MuleRun, took the helm at DingTalk. In August, Alibaba unified several office AI products, including QoderWork, MuleRun, and Wukong, under the "Qianwen Office" brand, thus eliminating internal product rivalry. Baidu also recently consolidated its office intelligence agents, merging the team and resources of its internal office product, dodo, into Baidu Dazi, centralizing its office agent operations.

ByteDance, renowned for its prowess in "chat" and emotional engagement, has made the most significant strategic shifts among the four, demonstrating a pragmatic approach.

In late July, ByteDance merged the Feishu product team with Doubao, while integrating sales, marketing, and customer service teams into Volcano Engine to bolster computational and data capabilities. According to The Information, during the Seed all-hands meeting, ByteDance founder Zhang Yiming expressed his stance on model distillation, stating unequivocally: "Closed-source and open-weight models cannot be distilled. 'Seed' is prepared to face temporary setbacks but will not resort to distilling competitors' models to overcome them."

With major companies struggling to find commercial success through ordinary consumers and existing products, they have abandoned the traditional internet traffic mindset and are collectively returning to the fundamentals of business. The question remains: who has truly crossed the productivity threshold?

Breaking Free from the Chat-Centric Mold

Contrary to the traditional internet product paradigm where increased user numbers lead to lower marginal costs, the growth in large model users and usage directly drives exponential increases in computational costs. This results in a mismatch between user scale and monetization efficiency, trapping companies in a commercialization quandary.

The initial "AI craze" at the beginning of the year, where AI evolved from passive responders to proactive executors, attracted a surge of high-net-worth office workers and showcased the industry's potential for "AI to break free from the chat-centric mold."

Office software has entered an era of intelligent collaboration, beginning to handle complete workflows.

Anthropic introduced the Claude for Microsoft 365 plugin, embedding AI assistants into office scenarios. Microsoft Windows 365 added enterprise-grade AI agents and GPU cloud computers, releasing the Copilot Cowork agent to integrate AI capabilities into enterprise office and development workflows. Google upgraded Gemini Notebook, integrating it into the Gemini App, Search, and Workspace office ecosystem.

These international examples also confirm that the B-end is the primary arena for profit.

While OpenAI gained consumer recognition with ChatGPT, it incurred operational losses exceeding $20 billion in 2025. In contrast, Anthropic, focusing on government and enterprise clients, has surpassed OpenAI in B-end revenue, proving that the enterprise market is the key to unlocking the commercial value of large models.

The domestic market is closely following this trend, with major and mid-sized companies shifting from pursuing general-purpose chatbots to betting on office AI.

According to Analysys' "Q2 2026 China Office Intelligence Agent Platform Market Insights" report, in June 2026, the 17 mainstream desktop-end AI-native office intelligence agents included in the statistics had a combined monthly page view count exceeding 60 million, tripling from March.

A report by Kezhi Consulting indicates that the Chinese enterprise-grade AI agent market was approximately 21.2 billion yuan in 2025, expected to grow to 44.9 billion yuan in 2026 and surpass 332 billion yuan by 2029.

In the latter half of 2026, Tencent, Alibaba, and Baidu concentrated their efforts on releasing integrated AI office products, marking the enterprise AI market's transition from cultivation to rapid implementation, with customer needs gradually maturing. Against this backdrop, ByteDance's thorough integration does not seem so abrupt.

Specifically, the Feishu product team now reports to Doubao, while the marketing, sales, and customer service (GTM) systems are integrated with Volcano Engine to form a new ToB organization, the "Creativity Service Platform." Zhao Qi oversees the product side, Tan Dai handles commercialization, and Wu Yonghui leads basic model R&D.

The declining trend of Feishu and the rising trajectory of Doubao became evident in the first half of the year.

In the previous round of office application competition, DingTalk and WeChat Work took the lead, while Feishu, a latecomer, was designed for efficient collaboration, providing solutions for enterprises with the brand mindset of "advanced enterprises use Feishu." By 2023, its Annual Recurring Revenue (ARR) had exceeded $200 million. Without the large model wave, Feishu would likely have continued to deepen and thicken its business scenarios.

However, the AI era has disrupted old logic, with over 90% of Feishu's new clients simultaneously purchasing Feishu AI products. As instant messaging's role in offices diminishes, AI agents can directly disassemble tasks, generate solutions, and perform initial executions. The challenges faced by these advanced enterprises have also evolved, requiring Feishu to reprove its value.

On the other end, Doubao boasts over 300 million monthly active users and 200 million Daily Active Users (DAU), making it the largest AI application in China by user scale, far ahead in the C-end. As of June 2026, Doubao's large model had a daily average token usage exceeding 180 trillion, with soaring computational costs. Unsatisfied with lightweight chatting and basic copywriting generation, Doubao began exploring commercialization, launching a professional subscription service in June to penetrate office scenarios.

Feishu becoming a division of Doubao's office scenarios also avoids internal resource waste, as reflected in the integration of the ToB GTM organization.

Previously, Feishu's Xie Xin oversaw both product and had an independent sales, marketing, and customer service team. After integration, the hierarchy significantly decreased, reporting to Zhao Qi instead of CEO Liang Rubo, aligning with ByteDance's consistent talent strategy of prioritizing business accountability.

Additionally, according to Tianyancha and third-party research institution data, Volcano Engine's share in the domestic public cloud Model as a Service (MaaS) layer is nearly 50%, but its Infrastructure as a Service (IaaS) accumulation is far weaker than Alibaba's and Tencent's.

Commercialization is now unified under Volcano Engine, with all B-end products sharing a single sales team, shortening decision-making chains internally and streamlining delivery processes externally. If they can smoothly integrate, the organizational adjustment's momentum will translate into corresponding market share, addressing ecological shortcomings.

Rejecting Distillation as a Shortcut

Having risen on the dividends of traffic and algorithms, ByteDance is often perceived by the outside world through Douyin, Toutiao, Hongguo, Doubao, and other pan-entertainment scenarios, with its hardcore technical strength rarely becoming the main topic of industry discussion.

With this round of deep AI strategy and organizational restructuring, ByteDance's technical route choices have been pushed into the spotlight. Zhang Yiming's clear rebuttal of large models "taking shortcuts through distillation" has sparked concentrated industry discussion. According to The Information, Zhang Yiming stated at the Seed all-hands meeting, "We are willing to sacrifice short-term gains for long-term goals."

CEO Liang Rubo also recently stated, "I've seen external reports claiming we insist on self-research due to external pressure, which is just guesswork. The real reason is that we hope the team delays gratification and builds a solid technical foundation, crucial for long-term Artificial General Intelligence (AGI) realization."

Using distillation, where small models learn and imitate large models to reduce training time and costs, has become common in the AI industry. Distilling Claude can only infinitely approach but not surpass this ceiling. Seed internally discussed the feasibility of the distillation route multiple times, but now the two executives have set the tone for ByteDance's self-research, prompting different external voices.

Some commentators believe ByteDance fears a repeat of the TikTok review incident, giving others leverage, claiming it's hard to define the line between borrowing and innovation in models. Another voice attributes Doubao's lag behind other companies to refusing to distill foreign large models.

This essentially reflects the contradiction between rapid iterative evolution and breakthroughs in foundational capabilities in the AI industry.

Investment queen Xu Xin revealed on a podcast that while Douyin is a significant business, Zhang Yiming now spends 50% of his time on Seed.

Seed is the technical source and strategic trump card of ByteDance's AI system. The language model Seed 2.0 had a modest response, not ranking first in coding, agents, or office scenarios. Liang Rubo proposed Douyin and Doubao as the two "broad mainlines" at the All Hands meeting while acknowledging the gap with overseas leaders in large language models.

The video generation model Seedance series is an exception, widely recognized as the world's most powerful video model, maintaining State Of The Art (SOTA) status. Volcano Engine set its overall revenue target for 2026 at over 40 billion, highly dependent on the video model structure, validating that ByteDance is not without technical strength and that powerful model capabilities can indeed translate into commercial moats.

ByteDance, lagging in language models, also wants to take a gamble. According to LatePost, ByteDance is discussing training a model with over 5 trillion parameters, led by Seed Foundation head Xiang Liang, with LLM pre-training data head Shen Ke cooperating.

This model surpasses Alibaba's Qwen 3.8-Max (2.4 trillion parameters) and Yuezhi's K3 (2.8 trillion parameters). Whether this gamble succeeds ultimately depends on implementation.

The Final Stretch

Shifting from "selling manpower" to "selling tokens," major companies are now competing across model capabilities, data assets, tool ecosystems, and delivery capabilities until truly solving the "last mile" of enterprise implementation.

Tencent's WorkBuddy leads with its product logic, reaching 20.97 million monthly page views in June, surpassing the combined total of ByteDance's TRAE IDE (domestic version) and Alibaba's QoderWork, with increasing online and offline promotion efforts. Alibaba's Qianwen Office and Kingsoft's WPS AI are also attempting to embed workflows to help users solve problems. With each company integrating its ToB business, they stand at similar starting points. The excitement around AI office solutions has just begun, and ByteDance is not the only one changing.

|References
[1] "ByteDance's AI Strategy Has Changed" by Huang Qingchun Channel
[2] "ByteDance Discusses Training a Model with Over 5 Trillion Parameters; Zhang Yiming: No Distillation, Don't Be Influenced by Short-Term Trends Like Coding" by LatePost
[3] "Feishu-Doubao Integration: A Goal Reshaping for ByteDance" by Geek Park
Operation/Yu Shuya
Design/Yanweier
*All rights reserved, no reproduction without authorization

end