Why did ByteDance restructure Feishu?
On July 30, ByteDance made major adjustments to its AI business, reorganizing the organizational capabilities of Feishu, its collaborative office software, to be absorbed by different teams respectively.
According to the adjustment plan released by the media, the Feishu product team will merge with Doubao to form a new Doubao product team led by Zhao Qi, with Xie Xin, head of Feishu, reporting to Zhao Qi; the GTM (Go-To-Market) team of Feishu covering marketing, sales and customer service will be incorporated into Volcano Engine to form a ToB service organization named "Creativity Service Platform", managed by Tan Dai, head of Volcano Engine, with Lin Chan, head of Feishu sales, and Shi Zhijun, head of Feishu marketing, reporting to Tan Dai.
After the reorganization, Feishu's brand name and existing products and services will remain unchanged. In addition, the "Doubao Enterprise Edition", which was jointly developed by the Feishu team, has launched internal tests for some customers.
On the surface, this reorganization means Feishu no longer retains the status of an independent business unit, but in essence it marks a shift in ByteDance's ToB strategic path. It breaks the boundary of the traditional SaaS business unit, further raises the priority of the AI to B strategy, and directly injects the enterprise scenarios and customer resources accumulated by Feishu into the "large model + cloud infrastructure" system.
Public data shows that in the second quarter of 2026, more than 90% of Feishu's new customers purchased Feishu AI products simultaneously. It is understood that according to the average consumption caliber in July, the ARR (Annual Recurring Revenue) of ByteDance's large model business reached 4 billion US dollars, exceeding the total ARR of other domestic model companies.
It is undeniable that after this round of adjustment, ByteDance's enterprise-level AI capabilities have been further enhanced, and the industry competition with peers has evolved from a competition between products to a comprehensive contest across models, applications, cloud computing and commercialization. However, we still need to ask a question: Why did ByteDance reorganize Feishu and inject its two most important capabilities into Doubao and Volcano Engine respectively?
Large Models Have Rewritten the Logic of Enterprise Service Software
Back in 2016, Feishu was not built for the industry, but to meet ByteDance's internal organizational needs amid rapid expansion. After growing rapidly within ByteDance in 2019, Feishu began to open to the outside world, exporting ByteDance's organizational culture standards through an office software. Products well-known to the public, such as multidimensional documents, OKR and Feishu Meetings, all carry strong ByteDance characteristics.
Since then, Feishu has operated as an independent combat unit, and previous adjustments were only limited to the vertical reporting relationship between the group and Feishu. This adjustment is the most important organizational restructuring of Feishu since its launch in 2016.
This move has aroused widespread industry attention not because Feishu is in decline. On the contrary, in terms of product capabilities and industry position, Feishu still firmly ranks in the first echelon of domestic enterprise software, and has even left its competitors far behind. Up to now, Feishu has completed the layout of multiple dimensions including instant messaging, online documents, video conferencing, knowledge base, multidimensional tables and open platform, covering most scenarios of enterprise collaborative office.
Global enterprise software has long evolved along the direction of efficiency improvement. Microsoft expanded its Office product matrix, Google launched Workspace, Salesforce improved its CRM platform, Slack reshaped enterprise communication, and Notion promoted knowledge collaboration; in China, in addition to Feishu, DingTalk and WeCom have also continuously expanded their product boundaries, integrating documents, meetings, approval, knowledge base and process management into a unified platform.
The core of enterprise software competition has always centered on "organizational efficiency". Therefore, the smoothness of information flow and the online degree of business processes constitute the core criteria for enterprise procurement decisions.
We can also clearly see that the core responsibility of all major enterprise software lies in "connection" and "carrying": employees obtain materials, analyze data, write reports, initiate approvals and advance projects, and the software is responsible for storing information, connecting processes and reducing communication costs. Documents carry knowledge, meetings form decisions, knowledge bases precipitate experience, and multidimensional tables organize data, all serving the goal of improving organizational operation efficiency.
The situation was almost all moving in a positive direction. IDC data also shows that China's collaborative office software market maintains growth, enterprise digitalization continues to advance, and the collaborative office track has always been a stable part of enterprise IT investment.
The market has not seen recession, and products have not stagnated, so the organizational adjustment appears particularly special.
The variable occurred after large models entered enterprise application scenarios, which profoundly rewrote the technical base and demand logic of enterprise software. Microsoft embedded Copilot into Microsoft 365, covering Word, Excel, PowerPoint, Outlook and Teams; Google integrated Gemini into Workspace; Salesforce launched Agentforce to promote agents to participate in sales, customer service and operation processes; in China, Alibaba, Tencent and ByteDance have successively released enterprise large model products and agent platforms, and even Alibaba and ByteDance have launched hardware products for information collection - AI recording devices.
What follows is the change in demand, and the focus of enterprise software procurement decisions gradually tilts to model capabilities and task execution effects. The working methods in the workplace have been restructured accordingly - sorting out meeting minutes after meetings, generating customer analysis after sales visits, summarizing market research, assisting in R&D code generation, and standard customer service replies are all deeply participated by AI models.
The macro AI structure is changing, and the micro behavior changes are more drastic. The value forms of documents, knowledge bases, organizational structures, permission systems and business processes have also been migrated. Documents become the input source for models to understand enterprise knowledge, knowledge bases form the cornerstone for models to obtain context, organizational structures limit the data call boundaries of models, permission systems standardize the task execution permissions of models, and business processes guide models to complete specific work. Collaborative office has evolved into the underlying supporting capability for model operation.
In this context, the competition dimension of enterprise software has shifted from collaborative experience and function coverage to model capabilities, enterprise knowledge precipitation, agent ecology and enterprise-level deployment capabilities. Gartner predicts that by 2028, more than one-third of enterprise software will have built-in Agent capabilities, and at least 15% of daily business decisions will be completed with the participation of Agent. A single office product cannot independently support the complete link of model, computing power, data, security, deployment and commercial delivery.
Against this background, ByteDance's organizational reclassification has clear industrial logic. Feishu has precipitated enterprise knowledge, organizational relationships, permission systems and business processes, Doubao has large model and agent capabilities, and Volcano Engine has cloud computing, model deployment and enterprise service infrastructure. The three capabilities were originally scattered in different business systems. With the entry of enterprise AI into the implementation stage, maintaining three independent systems will push up collaboration costs and affect product evolution and commercialization efficiency.
The product team is incorporated into Doubao to promote the software to evolve around model capabilities; the marketing, sales and customer service teams are merged into Volcano Engine to realize unified undertaking of enterprise customers. Feishu retains its product form, and the redefinition of organizational boundaries reflects the transformation of the competition direction in the enterprise software track.
What Feishu faces is not homogeneous competition with DingTalk and WeCom, but the change of the form of enterprise AI infrastructure. Once the positioning of enterprise software shifts from assisting employees to complete work to directly participating in business execution, collaborative office can no longer define the next generation of enterprise software alone. Therefore, Feishu has become the entry point for ByteDance to restructure its enterprise-level AI system.
The Basic Unit of Enterprise AI Competition Has Changed
Faced with similar industry challenges, ByteDance's choice is decisive and firm. It did not choose to continue to expand the Feishu business unit. On the contrary, it reconstructed the organizational system behind the brand, adapting to new industry changes by adjusting organizational relationships, which is completely different from the logic of other large manufacturers.
Feishu has long maintained a relatively complete business closed loop. From product R&D, marketing promotion to customer sales and after-sales service, all are concentrated in Feishu, and organizational goals are fully centered on Feishu.
We can summarize Feishu's original organizational form as a "vertically developed business unit". This is also the organizational paradigm that software enterprises have used for a long time. Microsoft Office, Salesforce, DingTalk and WeCom have all established similar systems - the product leader has high overall planning authority over R&D, sales and customer operation, so the business unit has become the basic unit of enterprise software development, and it is easy to form its own small independent ecosystem.
But the bottom line for this model to work is actually very high: the product must be able to operate independently and make profits independently. To put it plainly, it is to run a product as a small company - R&D, sales and promotion all revolve around it, and the final account calculation clearly shows whether it is profitable or not. It is not only a place to make products, but also a place to calculate business performance.
However, with the rapid development of AI technology, the survival soil of software has changed, and the ceiling above Feishu and the battlefield under its feet have also hit a wall.
In the early stage, Feishu served Internet companies and growing enterprises, with its advantages concentrated in collaborative office and organizational efficiency. With the deepening of large model commercialization, traditional industries such as finance, manufacturing, energy, retail and government affairs have become the core demand side of enterprise AI. Such customers have long procurement cycles, high deployment requirements, and put forward strict requirements for cloud resources, security capabilities, localized delivery and industry solutions. Relying solely on office software cannot undertake complex projects. The enterprise service capabilities accumulated by Volcano Engine have become the key support for ByteDance to expand the enterprise AI market.
We observe that when enterprises buy AI now, it is far more than just buying a piece of software. Models, platforms and applications have long been deeply bound together, and it is difficult to deliver satisfactory results to customers by relying on only one link.
Microsoft's strategic adjustment shows the characteristics of this trend. Microsoft 365 Copilot, Azure AI, GitHub Copilot and Azure OpenAI Service are fully connected at the underlying level, the sales system shifts to take the overall enterprise AI solution as the core, and the procurement process usually covers models, cloud platforms, security systems and office software. Google also closely integrates Gemini, Workspace and Google Cloud.
In China, the evolution path of several giants is very similar to that of the US market. Alibaba deeply binds the Tongyi large model, Alibaba Cloud and DingTalk; Tencent promotes the Hunyuan large model to run through WeCom, Tencent Cloud and Tencent Meeting. Alibaba and Tencent have gone through the multi-line horse racing stage, gradually converged their fronts, reduced frequent horse racing projects, unified the projects that have achieved initial results, and entered an era dominated by "integration".
Therefore, ByteDance's reorganization of Feishu conforms to this trend. More importantly, ByteDance's judgment on the basic unit of enterprise AI competition has undergone essential changes.
In the past, enterprise software relied on "single-point products" to fight independently, and Feishu could open up the market with its ultimate user experience. However, with the implementation of large models entering the deep water zone, the position of a single office software in the entire delivery chain has been redefined. It is no longer the end point of commercialization, but has become a high-frequency interactive entry point for large model capabilities to land. The "product vs product" individual combat model is unsustainable, and the full-stack collaboration of "system vs system" has become the foundation for winning large orders.
The cruelty of this change lies in the migration of customer demand structure. When large and medium-sized enterprises purchase AI today, the core pain point is no longer "lack of a collaborative tool", but complex engineering problems: underlying cloud computing power scheduling, security cleaning and permission isolation of private data, fine-tuning of industry-specific models, and finally how to seamlessly embed these capabilities into existing business processes.
Feishu's advantage lies in providing high-frequency working scenarios and user interfaces, but it is essentially only an upper-layer application, and cannot independently undertake heavy tasks such as computing power base, data security and complex engineering delivery. If we continue to advance along the old track of "centering on Feishu", it will not only fail to meet the integrated demands of large customers for "computing power + data + applications", but also build an invisible department wall between Feishu, Doubao and Volcano Engine, greatly increasing the cost of communication and interest friction across business units. Breaking the original commercial closed loop of Feishu and allowing models, clouds and applications to re-network is an inevitable move for ByteDance.
The Underlying Change of Business Logic
The organizational reorganization of Feishu reveals the common pain and embarrassment faced by the domestic enterprise software industry. After large models reshape everything, the traditional SaaS business unit architecture can no longer support the delivery and commercialization logic of the AI era.
Similar problems are also placed in front of other giants, but everyone has chosen different coping methods.
Alibaba chose to continue to strengthen the "cloud and DingTalk integration", continuously inject the underlying capabilities of the Tongyi large model into DingTalk, trying to maintain DingTalk's independent position as a ToB super entry; Tencent maintains the lightweight collaborative attributes of WeCom and Tencent Meeting, with CSIG (Cloud and Smart Industries Group) providing Hunyuan large model and cloud base support, taking a robust route of "platform embracing applications"; in Silicon Valley, while retaining the Microsoft 365 business unit, Microsoft formed an AI department that horizontally runs through all business lines, striving to promote Copilot without breaking the original SaaS interest pattern.
In contrast, ByteDance's approach is more radical and more thorough. It did not try to "patch up" within the Feishu system, but directly broke the organizational wall of the business unit, decoupled and repackaged the front-end scenarios, model capabilities and cloud infrastructure. Behind this separation lies the underlying change of business logic.
For a long time, the SaaS industry has adhered to the "per head per year" subscription system. However, the introduction of generative AI and agents has triggered the core contradiction of the business model: the higher the automation degree of agents, the fewer employee accounts required by enterprises, and the commercial growth of SaaS charged by head count will easily hit the ceiling. The core indicator of future enterprise AI procurement is shifting from "how many accounts have been purchased" to "how many Tokens have been consumed, how many Agent tasks have been completed, and how much cloud computing power and deployment resources have been called".
This change is extremely obvious in industry data. An IDC data shows that global AI spending in 2024 was about 235 billion US dollars, and it is expected to exceed 631 billion US dollars by 2028, with a five-year compound growth rate of nearly 30%. IDC also predicts that China will remain one of the fastest growing AI markets in the world, and AI investment in the next few years will maintain a growth rate higher than the global average.
What is more critical is the migration of investment structure. IDC data shows that enterprise AI investment is continuously concentrated in AI infrastructure, AI platforms and intelligent software, and infrastructure capabilities have become an important investment direction for enterprises to deploy generative AI. Gartner predicts that by 2028, about one-third of enterprise software will have built-in Agent capabilities, and more and more enterprises will build intelligent applications around a unified data and AI platform, instead of deploying isolated AI tools.
In this context, if we continue to maintain the independent business unit architecture of Feishu, the front-end sales sell SaaS software accounts, and the back-end accounting counts software subscription revenue, which will inevitably cause serious internal consumption and customer acquisition overlap with Volcano Engine's cloud computing and Doubao's model calls. Incorporating the GTM (marketing, sales and service) into Volcano Engine uniformly and the product R&D into Doubao is essentially to completely switch the business billing and delivery mode from the traditional SaaS software one-time sales system to the track of MaaS (Model as a Service) and cloud computing.
This also indicates a major shift in the domestic enterprise software track. In the past, manufacturers competed for "who can create office software with more ultimate experience", but now pure collaborative tools are losing their independent commercial value, and have degraded into scenario interfaces for models to obtain enterprise knowledge and business calls.
Feishu is the first sample in China to actively break the SaaS organizational paradigm. This skeletal reorganization triggered by large models will eventually spread to the entire Chinese enterprise service track - the future competition is no longer the confrontation between one software and another, but who can take the lead in completing the full-stack commercial closed loop of "large model + cloud infrastructure + scenario entry".
This article is from the WeChat official account "Xiutai", author: Shi Can, published with authorization from 36Kr.