ByteDance’s AI Strategy Now Runs Through Doubao

ByteDance’s August 6, 2026 all-hands pointed to a clear change in AI priorities: consumer products still matter, but enterprise productivity is becoming the main strategic front. Doubao is no longer being treated as one AI product among many. It is being positioned as the group’s central model, agent, and distribution layer, with workplace software and cloud services organized around it. 1410

The shift is notable because it comes with an unusually candid admission from CEO Liang Rubo: ByteDance’s large language models have fallen further behind leading overseas competitors. At the same time, Liang described the consumer-facing Doubao app as competitive and Seedance as remaining at the technological frontier, while committing ByteDance to continued in-house research rather than a short-term shortcut. 124

The reorganization puts Doubao at the control point

The organizational changes make the strategy more concrete than the all-hands messaging alone. ByteDance merged the product team of Lark—known domestically as Feishu—with Doubao’s product organization. The expanded team is led by Doubao chief Zhao Qi, while former Lark leader Xie Xin reports to Zhao. Lark’s sales, marketing, and customer-service functions were moved toward Volcano Engine, ByteDance’s cloud and enterprise-services arm. 8912

This separates two responsibilities that previously sat more closely together:

• Doubao owns the model and agent product experience.
• Volcano Engine becomes the main enterprise go-to-market and infrastructure channel.
• Lark remains important for workplace workflows, collaboration context, and customer relationships, but is no longer the primary product center.

That is more than a reporting-line adjustment. It suggests ByteDance wants customers to encounter its AI through an agent first, with collaboration tools, business data, and cloud services supporting that experience rather than defining it.

Why enterprise agents changed the calculation

Traditional workplace software is organized around documents, chat, calendars, and workflows. AI-native office agents reverse that order: a user starts with an instruction, and the system searches information, creates content, coordinates tasks, or invokes tools across the workplace.

The competitive pressure is visible in China’s desktop-agent market. Tencent’s WorkBuddy recorded 20.97 million monthly PC-side visits in June 2026, more than the second- and third-ranked products combined in the cited market ranking. 3236 But the figure needs to be read carefully: the underlying measure is web-visit volume across desktop platforms, not enterprise revenue, paid seats, daily active users, or mobile usage. 34

Even with that limitation, the ranking is strategically relevant. It indicates that the entry point to workplace AI may be shifting quickly toward dedicated agents. A collaboration product built around chat and documents could lose influence if another company owns the assistant that interprets requests, retrieves knowledge, and orchestrates work across those systems.

For ByteDance, folding Lark’s product team into Doubao is a way to put the agent above the workflow rather than leaving it beside the workflow. Lark can provide enterprise context and collaboration surfaces, but Doubao becomes the place where the user asks for work to be done.

Doubao brings consumer scale to an enterprise bet

ByteDance’s strongest argument for making Doubao the center of this strategy is distribution. Reports citing QuestMobile data put Doubao at about 345 million monthly active users in the first quarter of 2026. 17 Volcano Engine-related figures reported in June also put daily model usage at 180 trillion tokens, roughly 1,500 times the volume recorded around the model’s 2024 launch. 1828

Those numbers do not prove that Doubao will win enterprise AI. Consumer activity and corporate adoption have different requirements: enterprises care about reliability, access controls, data handling, administration, and measurable productivity. But the scale gives ByteDance a large testing environment for models, interfaces, and high-compute features before those capabilities are packaged for organizations.

The company also began testing direct consumer monetization. Doubao Pro was reported with monthly tiers of 68 yuan, 200 yuan, and 500 yuan, while basic access remained available for free. The paid offering was aimed at more complex, compute-intensive tasks such as document work, data analysis, and video production. 172324

That pricing experiment matters strategically. It indicates that ByteDance is testing whether a large assistant user base can support premium usage, not merely serve as a free acquisition channel. It also creates a bridge between consumer demand and the economics of enterprise AI, where intensive model use must ultimately be paid for by customers or sponsors.

A three-layer platform is emerging

Taken together, the restructuring points toward a three-layer enterprise stack:

1. Volcano Engine provides infrastructure and enterprise distribution. Its role connects cloud deployment, model services, and customer-facing sales.
2. Doubao owns the intelligence layer. Models, agents, and the primary user experience sit under the organization led by Zhao Qi.
3. Lark supplies workplace context. Collaboration, permissions, workflows, and existing customer relationships give the agent a setting in which to operate.

This three-layer description is an interpretation of the organizational changes, not a formally documented product architecture. The underlying restructuring—Lark product moving into Doubao and its go-to-market functions moving toward Volcano Engine—is established in company and media reports. 81012

The logic is straightforward: infrastructure without a compelling agent is difficult to differentiate; an agent without secure enterprise context is difficult to deploy; and workplace software without an intelligent interface risks becoming a back-end system rather than the place where work begins.

“High priority, thick trunk, long horizon” is a resource-allocation strategy

Liang’s reported emphasis on a high-priority, “thick trunk” and long horizon can be understood as a commitment to concentration rather than another round of loosely connected AI experiments. In practical terms, ByteDance is concentrating model development, agent products, product leadership, enterprise distribution, and computing resources around Doubao.

The long-horizon element is especially important because ByteDance is not claiming an uncontested lead in foundation models. Liang acknowledged that the gap with overseas leaders is widening, while arguing for continued in-house research and stronger underlying capabilities. 12

That creates a deliberately asymmetric strategy:

• Defend the consumer foothold through Doubao and Seedance.
• Build the enterprise interface around Doubao’s models and agents.
• Use Volcano Engine and Lark to make that interface deployable in corporate environments.
• Keep investing in foundation models even if near-term benchmark leadership remains out of reach.

The organization is therefore being optimized around where ByteDance believes value will accrue, not necessarily around where it is strongest today.

The bet is powerful—but still unproven

If AI agents become the main interface for search, creation, automation, transactions, and workplace tools, Doubao could become a major routing layer for ByteDance’s broader businesses. Its position could eventually connect AI with content distribution, e-commerce, monetization, search, and collaboration—the kind of cross-business leverage that made Douyin strategically important to the group. 71015

That outcome is not guaranteed. ByteDance must still demonstrate that its models can compete on quality, that agents can complete tasks reliably, and that enterprises will trust the company with sensitive data and permissions. The WorkBuddy traffic figures show the speed of the competitive race, but they do not establish who has won enterprise adoption. 34

The clearest conclusion is organizational rather than financial: ByteDance has chosen to make Doubao the trunk of its AI strategy. Lark is being repositioned as enterprise context and workflow infrastructure, Volcano Engine as the commercial and deployment channel, and Doubao as the model-and-agent surface where the company hopes the next generation of work will begin.