By | Bishan

Source | Bowang Finance

Four months ago, Xin Xie stood on the stage at the Peking University Entrepreneurship and Innovation Center and delivered a talk titled, 'How Did I Connect the Dots?'

As the CEO of Feishu, he has led this business for a full decade. In that speech, he reflected on the 'dots' he had personally connected throughout his career: from CTO of Kuxun to ByteDance's first HR leader, and finally to the creator of an office productivity tool. Those once-scattered dots eventually formed a line leading to Feishu. However, he did not elaborate on where the other end of that line would lead.

At 9:30 AM on July 30th, the answer arrived in the form of a company-wide email. The Feishu product team merged with the Doubao product team to form a new Doubao product team, led by Zhao Qi, Doubao's head, with Xie Xin now reporting to Zhao Qi. Meanwhile, Feishu's marketing, sales, and customer service teams were integrated into Volcano Engine, forming a new organization called the 'Creativity Service Platform,' led by Tan Dai, Volcano Engine's head. Feishu's existing products and services remain unchanged.

Brief words, significant changes. This once-independent BU, which stood alongside Douyin and Volcano Engine, has now been split in two: product capabilities now belong to Doubao, while commercialization functions have been transferred to Volcano. Xie Xin's reporting line has shifted from ByteDance CEO Liang Rubo to Zhao Qi. For the first time in a decade, he is no longer independently responsible for a business unit.

It should be clarified that only Feishu has been dismantled. Volcano Engine, in this email, serves as the recipient, taking over Feishu's entire commercialization front.

01

The First Phone Call

The story began with a phone call twelve years ago.

On September 2, 2014, Zhang Yiming called Xie Xin and asked if he would consider changing jobs to join ByteDance as HR leader. Xie Xin declined on the spot. At the time, he was CTO of Kuxun Travel, with a bachelor's and master's degree in computer science from Peking University. He had served as a project manager at Microsoft Research Asia, a core engineer for Baidu's web search, and even appeared on Tianjin TV's 'Only You' to recruit for his company—a true technical manager. 'I'm currently the CTO; having me join such a small company as an HR leader feels like a significant step down,' he later recalled of his initial reaction.

Zhang Yiming did not give up. He explained why he insisted on finding a CTO to handle HR: 'Strategy is important, but the team is equally crucial—I've learned this firsthand. Kuxun and Qunar were in direct competition, with clear strategic directions, yet the gap between them widened.' Three months later, Xie Xin joined ByteDance. At the time, the company had around 300 employees, and he became its first HR leader. His first task was to map out the company's organizational structure using tools—a task no one had truly accomplished for the 300-person company before.

As HR leader, Xie Xin quickly turned his attention to tools. In 2014, he stated at the QCon conference: 'When you use advanced tools, you're not just using the tools themselves; you're unconsciously adopting advanced work methods.' Over the next few years, ByteDance's internal office tools changed annually, shifting from Skype to WeChat Enterprise, then to Slack, and finally to DingTalk—Feishu later officially claimed that ByteDance had once been DingTalk's largest client. The breakup occurred around 2016: ByteDance proposed customization requests to DingTalk that were not met, and high-level leadership disagreed with DingTalk's manager-oriented product philosophy. They believed that good efficiency tools should serve employees.

In 2016, Xie Xin sent an email to Zhang Yiming, proposing that the company, now with 1,200 employees, develop its own office productivity tool. By the end of that year, a ten-person team initiated Project Lark, with Xie Xin personally writing the product specification—using GTD principles for instant messaging, treating communication content as documents, and designing workflows that could be marked as complete like read emails. This product later received a Chinese name: Feishu.

02

The Sharpest Years

Feishu grew out of ByteDance's internal needs, answering a fundamental question of the digital age: How can information flow more freely within an organization?

It solidified ByteDance's proven organizational practices—transparent information, document-driven workflows, distributed decision-making, and OKR alignment—into a product later summarized as 'Context, not Control.' In 2019, Feishu opened to external users and launched an international version. By 2021, as ByteDance restructured into six independent BUs, Feishu stood alongside Douyin, TikTok, and Volcano Engine, with Xie Xin reporting directly to Liang Rubo. That marked Feishu's peak within ByteDance's ecosystem.

Those years were also Feishu's sharpest phase. 'Advanced teams' like Li Auto, Xiaomi, and NIO became its clients, choosing Feishu not just for its product but also to gain insights into 'how ByteDance works.' At early Future Infinity conferences, the main speakers were clients from various industries discussing organization, collaboration, and reducing meetings—seemingly introducing software but actually spreading a belief in how companies should operate.

However, after reaching its peak, the trajectory mostly pointed downward. In late 2022, during company-wide layoffs, Feishu cut over 1,000 employees. At ByteDance's 11th-anniversary event in March 2023, Liang Rubo made a comment later frequently cited: Feishu and Volcano Engine's R&D investment was 'no less than Douyin or TikTok,' but their 'ROI wasn't very cost-effective.' He added a second part—'Doing enterprise services well is crucial for the company to become a great tech company.' The first part discussed finances; the second, value. Years later, it's clear that finances outweighed value.

In March 2024, Xie Xin sent a company-wide letter, cutting about 20% of Feishu's 5,000-person team to 'return to a startup's Day 1 mindset.' Three months later, Feishu President Zhang Nan resigned, and the president role was abolished. In May 2025, Feishu People business was spun off. By September 2025, the real foreshadowing emerged: Zhao Qi, Doubao's head, joined, and Feishu product leader Tong Yao began reporting to him in a dotted-line capacity. Ten months later, the dotted line became solid.

03

Why Now, of All Times

The confusion lies here: Feishu's restructuring occurred precisely when its performance was at its best.

A July report from *Caijing* provided numbers: Feishu's 2025 revenue exceeded 3 billion yuan, up ~60% year-over-year; in Q2 2026, revenue grew over 100% YoY. Under the ARR metric, it reached $100 million in 2022 and doubled to over $200 million in 2023. More critically, its customer mix shifted: over 90% of new clients also purchased Feishu's AI products—enterprise customers had voted with their feet, seeking not just office software but AI solutions.

Growth was real, but so was pressure. With a ~4,000-person team, industry estimates put Feishu's annual labor costs between 3.2 and 4 billion yuan. Xie Xin once extended profitability targets to 2030. Relying solely on collaboration licensing fees for documents, spreadsheets, and meetings, even with rising ARR, would struggle to cover escalating labor, R&D, and new AI computing costs. Meanwhile, competitors multiplied: DingTalk and WeCom squeezed the existing market, while startups like Moonshot AI directly entered the office intelligence space. Feishu's window to break through growth ceilings via standalone SaaS was narrowing.

ByteDance, however, had repositioned Feishu within its AI strategy. At the January 2026 all-hands meeting, Liang Rubo judged AI as 'at least a PC + Web-scale peak.' At the June 23 FORCE conference, he clarified: 'Scaling the AI peak' was ByteDance's top priority, with Volcano Engine's MaaS becoming a foundational business. 36Kr summarized it broadly: 'Everything but AI is contracting.'

Contraction coexisted with investment. Public reports indicate ByteDance increased AI new business investment in 2025; capital expenditures rose to over 200 billion yuan in 2026, with ~85 billion yuan earmarked for AI chip procurement, even discussing further increases. Such investment demanded a sufficiently large monetization outlet.

04

What Doubao Lacked, Feishu Had

The preferred outlet was initially Doubao, but its financials were unimpressive.

By user scale, Doubao dominated: MAUs grew from 345 million in Q1 2026 to 382 million in June, with DAUs exceeding 200 million and daily token usage surpassing 180 trillion. Revenues told a different story—according to *LatePost*, Doubao's daily revenue was under 1 million yuan, primarily from in-app e-commerce commissions. Based on Volcano Engine's public API pricing, Doubao's daily inference computing costs likely ran in the tens of millions of yuan. Its massive user base brought not commercial value but sustained bleeding.

In May, rumors of paid messaging emerged, and Doubao's MAUs dropped by 6.1 million that month—its first significant decline since launching in 2023. In June, despite controversy, Doubao launched paid subscriptions. More critical was user demographics: Citigroup research noted that students and elderly users dominated Doubao's 300+ million MAUs, using it mainly for casual chatting and information queries, with average daily usage under one hour. Professionals and developers with high-frequency productivity needs found it inadequate, preferring ChatGPT, Claude, or domestic alternatives like Kimi.

Doubao wasn't short on users but lacked enterprise scenarios and high-paying demographics. Feishu had both: its enterprise intelligence platform Aily, multidimensional tables with over 10 million MAUs, and years of accumulated enterprise workflows and mature paying clients. The Doubao Enterprise Edition, co-developed with Feishu's team, began internal testing among select Feishu clients in late June—employees could summon Doubao directly from Feishu's knowledge Q&A, documents, group chats, or meetings. By mid-July, Feishu's 'knowledge Q&A' feature quietly switched to Doubao.

Thus, this merger wasn't Feishu being abandoned but a mutual need: Doubao gained enterprise access and scenarios, while Feishu tapped into large model tech and Volcano Engine's sales channels. A person close to ByteDance put it bluntly: 'A business with tens of billions in ARR can't serve as an entrance alone; only by merging with Doubao and Volcano can it unleash greater potential.'

05

The Entrance Shifted

Zooming out, Feishu's restructuring is part of a broader industry trend this summer.

On July 20, Tencent integrated parts of the QClaw product center into WorkBuddy, already designated a 'super project' overseen personally by Pony Ma. On July 27, Alibaba consolidated multiple Agent products into 'Qianwen Office,' handed to DingTalk's new CEO with pricing directly targeting WorkBuddy. Huawei took a bolder approach: Xiaoyi, its AI assistant since inception, is becoming the unified AI entrance across all scenarios.

Within ten days, major tech firms had consolidated their AI office strategies. Behind this lay a shared data point: EasyView reported that in June, 17 mainstream desktop AI office agents in China recorded over 60 million total visits, with Tencent WorkBuddy leading at 20.97 million—exceeding the sum of second and third place. ByteDance's Trae ranked second with 12.79 million visits. These visits occurred on PCs, far more valuable than mobile fragmented time—users activated Agents for writing long documents, handling complex spreadsheets, and managing corporate emails, indicating clear productivity needs. Enterprise users weren't chatting; they were working.

A clear judgment emerges: the first work software employees open won't be collaboration tools but AI assistants, with office functions as extensions of AI capabilities. Since late March, DingTalk, Feishu, and WeCom have sequentially open-sourced their CLI interfaces—the logic is clear: when Agents become the first window opened at work, traditional office software must retreat to the backend as execution engines. Microsoft 365 Copilot's paid enterprise seats have surpassed 20 million, adding 5 million in a single quarter. IDC predicts China's enterprise-grade token consumption will grow ~20x from 2025 to 2026.

The entrance has shifted—this is the true context of Feishu's restructuring. During office software's golden decade, Feishu was the most advanced; but when 'advanced' is redefined by AI, clinging to an independent entrance becomes the riskiest choice. Maintaining separate teams would only drain resources, raise client cognitive costs, and even cause internal competition. Unifying products, sales, and services is essential for enterprise AI to deliver real value.

06

What Those Dots Connected To

Returning to Xie Xin's speech title.

When Feishu launched in 2016, it aimed to solve how information flows within organizations. A decade later, enterprises' most pressing question became: Can AI read corporate knowledge, understand organizational permissions, invoke business tools, and ultimately get work done? Feishu's decade-long accumulations—documents, messages, organizational relationships, permissions, and business data—are prerequisites for Agents to enter organizations. However, their value now manifests not as a standalone office software but as ByteDance's AI foundation.

For ByteDance, this is arithmetic: facing 200 billion yuan in AI infrastructure investment, Feishu's 3 billion yuan revenue could no longer independently dictate its fate. For the industry, it signals a shift: the first half focused on concepts and parameters; the second half hinges on which solutions truly solve problems, with AI's center of gravity moving from 'internal competition' to 'entrance contention.' For Xie Xin personally, over a decade, he grew a ten-person team into a 4,000-person BU, only to willingly place it on a larger platform.

The dots he connected—from HR to tools, from tools to Feishu, from Feishu to AI—now form not Feishu's empire but a cornerstone of ByteDance's AI landscape.

This isn't a tale of failure. Feishu remains a strong product with loyal clients and doubling growth rates. But in the AI era, the definition of 'a strong product' is being rewritten—advanced tools may not own their era, but they can choose to remain part of it.