ByteDance's major AI consolidation: Coze Enterprise Edition is withdrawn, and the era of internal horse racing draws to an end.

At the end of October, Coze, the AI application development platform under ByteDance, officially released the "Coze Enterprise Edition Offline Announcement". The announcement states that the Enterprise Standard Edition and Enterprise Flagship Edition will stop accepting new customer purchases from November 30, 2026, and all existing packages will be fully offline no later than December 31, 2027, with automatic refunds for unused portions beyond the validity period.

The news caused a stir across the industry. According to 2025 market statistics, Coze once ranked first in China's public cloud market for agent development platforms with a 19.3% market share. However, in ByteDance's huge business empire, the title of top market share cannot guarantee an independent right to survive for a product line.

The exit of Coze Enterprise Edition is not an unexpected accident, but a top-down strategic decision. It marks that ByteDance's AI strategy is fully shifting to "centralized management", and also demonstrates the underlying logic of ByteDance's internal resource allocation: under the "thick main trunk" strategy, any edge product that fails to form an explosive commercial closed loop and overlaps with the main business may face downgrading, integration or even shutdown at any time.

Organizational Consolidation First

The fate of Coze Enterprise Edition was already determined in the summer of 2026.

Looking back at the end of July 2026, ByteDance carried out an extremely low-key but far-reaching organizational restructuring. The Feishu product team and the Doubao product team were officially integrated, with the head of Doubao taking overall responsibility; at the same time, the Feishu GTM (Go-to-Market) team was transferred to Volcano Engine. Shortly after, on August 24, the TRAE and Coze teams were fully merged into the Doubao system, with office-related capabilities deeply integrated with Doubao, and TRAE IDE and CLI were transformed into programming product lines under the Doubao brand. Immediately after that, ByteDance officially launched the independent AI office product "Doubao Work" on August 25.

This series of intensive actions is the prelude to the consolidation of ByteDance's AI strategy. At the subsequent all-hands meeting, Liang Rubo, CEO of ByteDance, clearly put forward the business principle of "high priority, thick main trunk, long-term optimization", and hopes that Doubao can become a main business as influential as Douyin.

Under the logic of the "thick main trunk", ByteDance's resources are tilting towards a very small number of core goals. In the past, ByteDance was good at the "horse racing mechanism", allowing various teams to freely explore in segmented tracks, and the outstanding performers would obtain full access to resources. But in the AI era, the training cost of underlying large models, computing power consumption and delivery cost of enterprise-level services are extremely high. As the AI competition enters the deep water zone, large tech companies must concentrate superior resources to achieve decisive victories, and can no longer afford high trial-and-error costs on multiple fronts at the same time.

Since its launch in 2024, Coze has accumulated a huge number of developers by virtue of its advantages of building Bots with zero code and one-click publishing to ecosystems such as Feishu, WeChat and Douyin. In the second half of 2025, Coze tried to enter the enterprise office efficiency improvement track, and launched the Enterprise Standard Edition and Flagship Edition.

But this exactly touched ByteDance's current strategic red line. The senior management of ByteDance has a very clear judgment: the battle for office Agents must be fought by the main brand "Doubao Work" and Volcano Engine's privatized infrastructure. As an independent office enterprise SaaS package, Coze not only has increasingly blurred functional boundaries with "Doubao Work", but even forms internal homogeneous competition. In ByteDance's business map, the future of office scenarios belongs to Doubao, and the future of underlying computing power and enterprise-level delivery belongs to Volcano Engine. Coze Enterprise Edition, sandwiched in between, has lost the organizational soil for independent existence.

Awkward Position Even for the Market Leader

What made Coze Enterprise Edition lose its qualification for independent existence is not only the organizational will, but also the cold commercial data.

Although Coze captured 19.3% of the public cloud market share of China's agent development platforms in 2025 and ranked first, the same measurement shows that the total size of the public cloud market for agent development platforms that year was only about 250 million yuan. Based on this calculation, Coze's platform product revenue is about 48 million yuan, less than 50 million yuan (this figure is an estimate based on the market size, excluding model calls and cloud resources, and is not ByteDance's official financial data).

This is a very striking data. In the evaluation system of large Internet companies, once a platform product with annual revenue of less than 50 million yuan and requiring continuous investment of a large amount of R&D and computing power resources loses its independent growth momentum, it will quickly be downgraded from a "strategic asset" to a "marginal consumable".

In the second half of 2025, the enterprise-level AI Agent market ushered in a fierce scramble for market share. Tencent's WorkBuddy, relying on WeCom, the B-end organizational connector with the highest penetration rate in China, connected identities, address books, chats and documents, and quickly covered half of China's top 500 enterprises; Alibaba continued to put pressure relying on the DingTalk ecosystem, and later integrated the office capabilities of Tongyi Qianwen, pressing forward step by step on the B-end front.

In such a fierce B-end battle among large tech companies, Coze Enterprise Edition did not exert enough firepower. It neither had a B-end connector with absolute dominance like WeCom as a moat, nor was it as solid as Volcano Engine's infrastructure in the hard core capabilities of privatized deployment, data compliance and enterprise-level governance. For ByteDance, using an independent enterprise SaaS package to fight a war of attrition with Tencent and Alibaba is extremely cost-ineffective.

Therefore, ByteDance chose the solution that best serves its own interests: no longer let Coze conduct trial and error in the form of independent SaaS packages, but dismantle its capabilities. The office-related capabilities are integrated into "Doubao Work", and the development-related capabilities return to a pure platform for developers and vertical scenarios. By concentrating superior resources, the main products are used to fight the head-on war between large tech companies.

The End of Diversified Deployment

The offline of Coze Enterprise Edition is a microcosm of ByteDance's overall AI strategy shifting from "scattered operations" to "main force decisive battle".

Sorting out ByteDance's layout in the AI field, three main lines can be clearly seen: the Doubao App for C-end users, Feishu and the upcoming "Doubao Work" for B-end office scenarios, and Volcano Engine for underlying infrastructure and ToB delivery. Coze was originally expected to become a "universal glue" connecting developers, enterprises and ByteDance's ecosystem.

But "universal" often means "not specialized". In the early stage of AI development, the industry needed low-threshold construction platforms to educate users and prosper the ecosystem. But when the industry enters the commercial implementation stage, enterprise customers no longer need a simple platform that can casually assemble plug-ins, but digital employees that are deeply embedded in business processes and have extremely high security and observability.

This is exactly what the combination of "Doubao Work" and Volcano Engine can provide. Feishu team's years of B-end product experience, Doubao's continuous iteration of model capabilities, plus Volcano Engine's infrastructure in cloud-native and privatized deployment, form the complete puzzle of ByteDance's ToB strategy.

In this new puzzle, the position of Coze Enterprise Edition becomes very awkward. If the independent SaaS package is retained, it will not only cause repeated investment of internal resources, but also bring cognitive confusion to customers — should customers buy Coze Enterprise Edition, or wait for "Doubao Work"? For ByteDance, which is famous for its "go all out to achieve miracles" and "extreme efficiency", such internal friction and consumption are unacceptable.

Therefore, the exit of Coze Enterprise Edition is essentially an internal resource restructuring. Through organizational means, ByteDance has straightened out its product lines, ensuring that all forces converge on the two most core fulcrums: "Doubao" and "Volcano Engine".

Products as Consumables

The fate of Coze Enterprise Edition is also a typical endgame of the internal horse racing mechanism in large tech companies.

In the tech circle, there is an increasingly obvious trend: the service life of AI products of large tech companies is even shorter than that of products from startups. In 2025 alone, a large number of AI tools around the world stopped services; OpenAI itself also shut down multiple AI tools within less than a year of their independent service, for example, the ChatGPT Atlas browser and the visual Agent building tool Agent Builder were announced to be offline shortly after their launch.

The essence behind this is not that large tech companies cannot make good AI products, but that the logic of large tech companies treating AI products has changed: products are no longer just independent entities for making money, but also "consumables" for verifying strategic directions.

The short life cycle of products from large tech companies often comes from "capabilities being absorbed". Whether it is the office capabilities of Coze Enterprise Edition integrated into "Doubao Work", or various small OpenAI tools dismantled into the main ChatGPT client. From the perspective of product brands, they have exited; but from the perspective of capabilities, they continue to exist in a more powerful and more core container. For large tech companies, this is a necessary cost for resource reallocation, and an inevitable choice to maintain organizational combat effectiveness.

In the machine of large tech companies, no specific product is irreplaceable, only the company's strategic goals are unshakable. When a product cannot independently support a hundred-billion-level market imagination, or when its existence hinders the advancement of the main business, being incorporated or shut down is its inevitable fate. ByteDance's decisive cut of Coze Enterprise Edition demonstrates the giant's ruthlessness and efficiency in strategic focus.

Ecological Niche Rearrangement

The exit of Coze Enterprise Edition has left profound enlightenment for the entire AI Agent track.

For product teams inside large tech companies, to judge whether a product can survive independently for a long time, we should not only look at the growth of external users, but also see whether it is on the "main trunk" of the company. Once it overlaps or conflicts with the core strategy, even the highest market share may be withdrawn overnight.

For startups and small and medium-sized manufacturers, the strategic consolidation of large tech companies is exactly a huge opportunity. The strategic consolidation of large tech companies means that they have given up those segmented markets that are "tough and tiring", need to be highly adapted to specific industry processes, and cannot form platform-level monopolies.

Take the open source AI platform Dify as an example. It completed a $30 million financing in 2026, and chose a tough path that large tech companies are unwilling to take: give control to customers, and earn subscription fees paid by enterprises for "data staying on premises" and "governable systems". Dify does not compete for a huge number of individual developers, but directly targets large and medium-sized enterprises that are extremely sensitive to data security, allowing customers to deploy the system in their own computer rooms.

Avoiding the "main trunk" channel of large tech companies, doing the tough and messy work that large tech companies are unwilling to do, and serving the long-tail demands that large tech companies' standardized products cannot cover, can instead build a moat that large tech companies cannot easily pull out. In the era when large tech companies pursue "thick main trunk", verticalization and customization are the best barriers for startups.

The Agent track is far from reaching its end. When the billing method of AI Agent shifts from "software subscription" to "task delivery", the hundred-billion-level market space has just opened.

ByteDance's consolidation and the exit of Coze Enterprise Edition are just the first milestone in this long marathon. It reminds all practitioners: in the business map of giants, products always serve the strategy and may be pressed the stop button at any time; while in an independent closed loop, products must serve customers, and only by solving specific pain points can they master their own right to survive. In the cruel knockout round of AI, recognizing your own ecological niche is more important than blindly pursuing large and comprehensive coverage.

This article is from the WeChat official account "Tech Business", written by Tech Business, and published with authorization by 36Kr.