The internal horse-racing mechanism has been suspended, and Tencent, Alibaba and ByteDance have consolidated their AI office product resources to compete head-to-head.

AI Agent (Intelligent Agent) office use is becoming a new battlefield where leading manufacturers compete head-on. Over the past six months, multiple large tech firms have run internal parallel projects and conducted multi-line trial and error, but from late July to early August, Tencent, Alibaba and ByteDance have successively rolled out their respective integrated plans:

Tencent has adjusted the relevant business and part of the team of the QClaw Product Center to the 6th Cloud Product Division, bringing it under the same management system as WorkBuddy, another AI office agent. Alibaba has integrated three office products — QoderWork, Wukong and MuleRun — to launch Qwen Office. ByteDance has fully merged the Feishu product team into the Doubao team, with the two products deeply bundled in office scenarios.

This fierce head-to-head competition is closely related to market data. According to a report by Analysys, the total monthly visits of mainstream domestic desktop AI-native office agent platforms exceeded 60 million in June, while three months ago, this figure was only just over 20 million.

Major tech firms no longer hide their ambitions, as they strive to seize the entry point for next-generation productivity.

WorkBuddy breaks through to the mass market, Tencent fights for a leading position

Multiple Tencent employees disclosed that WorkBuddy is the most popular product inside the company recently, with obvious resource tilting.

A Tencent employee close to the WorkBuddy team said that the product is regarded as another national-level product of Tencent following QQ and WeChat, shouldering the important task of making up for the gap Tencent fell behind in the large model competition. This high level of emphasis can be seen from its update frequency: WorkBuddy has released 27 versions in total from June to July, with updates almost every two days.

A human resources staff member at a manufacturing company told *Economic Observer* that before using WorkBuddy, her understanding of agents was limited to chat-based products. She only decided to download and try it out after seeing WorkBuddy's advertisements on WeChat and several video platforms one after another, and recommended it to her colleagues. She mainly uses it to handle attendance statistics, talent inventory and other work. "The threshold for use is very low, no extra learning is needed, just throw the tasks in and you can get the results quickly."

A group of people who were once excluded from the Agent wave are getting closer to this revolution through WorkBuddy. When you search for "workbench" on platforms such as Xiaohongshu and Douyin, various outputs from WorkBuddy come into view: middle school students build learning progress dashboards, financial staff digitalize project ledgers, and teachers upload student scores and connect to large models for academic performance analysis.

"Combining one's own position and work content, using AI to customize products and handle transactional demands is what major tech firms have been doing internally for more than half a year," said a product operations specialist at an internet giant, "WorkBuddy brings this capability outside the big tech firms."

On the traffic side, Tencent is the current winner. Data from Analysys shows that Tencent WorkBuddy ranks first with 20.97 million monthly visits, exceeding the sum of the second place (the domestic version of Byte TRAE IDE) and the third place (Alibaba QoderWork).

The product operations specialist from the internet giant also added that the actual number of WorkBuddy users, how much of the above user practices come from marketing activities, and whether the product solves real demands remain to be further observed. Monthly visits are a relatively rough indicator, which is prone to inflation when the product is just launched or during marketing nodes such as large promotions. "To evaluate the market performance of a product, we still need to look at DAU (Daily Active Users) and retention rate."

Alibaba makes a tough three-in-one integration to move closer to DingTalk

Compared with Tencent's aggressive advancement, Alibaba's moves appear more restrained.

On July 27, Alibaba launched Qwen Office to the market, which integrates three existing products: QoderWork derived from the AI programming tool Qoder, Wukong, the AI workbench launched by the DingTalk team for the B-end market, and MuleRun, the AI Agent platform running on cloud hosts.

Before the integration, QoderWork attracted the highest market attention. Wu Yongming, CEO of Alibaba, once mentioned at an internal meeting in June that the daily active users, weekly active users, and Token consumption of QoderWork rank first among all AI tools of the Alibaba Group.

In the 2025 China AI Programming Market Share Report released by IDC (International Data Corporation) in July, Alibaba Qoder ranked first with a 47.6% share, far exceeding other similar products such as Zhipu CodeGeeX, SenseTime Raccoon, Tencent CodeBuddy and so on.

With three products with converging functions in hand, Alibaba's integration move seems reasonable; but by discarding the already well-known product name QoderWork and launching Qwen Office which is strongly bundled with its own model side, an Alibaba R&D staff disclosed that there is no lack of "confused" voices inside the group about this decision.

From a broader perspective, Alibaba's strategic consideration becomes clear. In the view of the Alibaba R&D staff, Tencent WorkBuddy is growing rapidly with the ecological bonus of products like WeChat, and has obvious advantages on the C-end; this adjustment by Alibaba may aim to highlight model capabilities and rely on DingTalk's existing customers to break through the B-end market. "Each has its own path, you can't use your own shortcomings to hit others' strengths."

As of press time, QoderWork is still described as an independent product on the official Qoder website and supports download and use. Does this mean that Alibaba plans to use two products to target C-end and B-end users respectively? A relevant person from Alibaba Cloud disclosed that "they will all be merged later."

Apart from Qwen Office, Alibaba is still exploring other possibilities for AI office. According to *Economic Observer*, Alibaba Cloud is currently conducting internal testing of a B-end oriented multi-Agent collaboration platform called "Wanyou Wujie". Users can input specific project scenarios, and the platform will automatically build a collaboration team composed of multiple Agents, which will jointly promote the project and deliver results collaboratively. The product is currently in the internal testing stage, with some pre-configured general Agents built in, and supports users to create required Agents by customization.

ByteDance integrates Feishu and Doubao, creating mutual reinforcement inside and outside the ecosystem

On July 30, ByteDance announced the launch of organizational adjustment for its AI business: the Feishu product team and the Doubao product team are integrated to form a new Doubao product team, which is led by Zhao Qi, the former head of Doubao, and Xie Xin, the head of Feishu, reports to Zhao Qi.

From the data perspective, this integration is more like a mutually beneficial resource matching. The QuestMobile report shows that as of June 2026, Doubao has 382 million monthly active users, ranking first among AI-native applications, exceeding the sum of Qwen and DeepSeek. The daily average Token call volume of the Doubao large model has exceeded 180 trillion.

Doubao has a huge C-end user base, but interactions are concentrated in the free chat scenario, so its monetization paths are limited. In contrast, Feishu's market scale is much smaller: QuestMobile data shows that its monthly active users are about 30 million, less than one seventh of DingTalk's, but it shows strong B-end monetization capability. According to a report by *Caijing*, Feishu's revenue in 2025 exceeded 3 billion yuan, maintained rapid growth in the first half of 2026, and its revenue in the second quarter of 2026 increased by more than 100% year on year.

Doubao needs Feishu's monetization capability, while Feishu needs Doubao's underlying traffic, so the in-depth bundling of the two is logical.

At present, the integration of Feishu and Doubao remains at the organizational level, with little impact on the products themselves. Feishu's native AI Agent function "Feishu aily" is still running, and C-end users of Doubao can enable the "office task" mode on the computer side, allowing the Agent to operate local files directly.

ByteDance is not only betting on the Feishu-Doubao integration in the AI office track. TRAE Work, another general office assistant under the company iterated from an AI programming product, is temporarily unaffected by this integration. Before this organizational adjustment, TRAE Work was popular for its ability to directly read Feishu messages, handle approvals and arrange schedules. A user who uses TRAE Work intensively introduced that she once used both WorkBuddy and TRAE Work to handle work, and there was no significant difference in the performance of the two products, and their operation methods were basically the same. Since her enterprise uses Feishu for collaboration, she decided to keep only TRAE Work.

A person close to the TRAE Work team told *Economic Observer* that it is quite common for large tech firms to retain similar competing products; the biggest difference between the Feishu-Doubao system and TRAE Work lies in the degree of openness. The former only supports calling ByteDance's self-developed models, while the latter supports calling mainstream external models and configuring APIs by users themselves.

Screening for genuine retained users from "coupon scalpers"

Major tech firms are fighting fiercely in the cloud, while users at the center of the whirlpool are voting with their feet.

A C-end user who has tried all three products admitted that the three have not yet formed a significant gap in core functions. Taking advantage of the recent marketing campaigns where all manufacturers are giving away large amounts of credit points, he is switching between the three products to collect benefits just like he did during the "food delivery war" in the past.

What determines the final retention of users is not only the intensity of subsidies from each platform, but also whether the platform is willing to abandon model monopoly.

"It doesn't matter which company the product belongs to. Being practical and cost-saving matters more." The user said that besides the credit point giving intensity and credit point consumption speed, another factor directly affects the use cost: whether users can use their own APIs.

The three products differ in openness: WorkBuddy and TRAE Work are open to the outside, clearly supporting users to configure custom models, with external models such as DeepSeek, Zhipu and Kimi built in; Qwen Office and Doubao stick to their own systems, only allowing the use of their own or designated internal model tiers. Among them, Qwen Office provides three model tiers: Advanced, Basic, and Economy, as well as its own Qwen3.8-Max, while only ByteDance's self-developed models can be used in Doubao.

For intensive users, "whether they can use their own APIs" directly determines the use cost. The C-end user said frankly that if the platform does not support calling low-cost external APIs that users have already purchased (such as DeepSeek), users have to pay for platform credits inside the platform to cover computing power consumption.

In terms of payment mechanisms, all these products adopt the structure of "free drainage + tiered subscription". Users can get free credits by registering, signing in or participating in activities. After the credits are exhausted, monthly payment is required, and different tiers correspond to different monthly credit quotas and other rights. The lowest monthly paid tier for personal versions ranges from 45 yuan to 78 yuan.

The aforementioned user introduced that taking WorkBuddy as an example, if you only handle simple work such as writing documents and spreadsheets, subscribing to the lowest paid tier is suitable; but when handling complex tasks such as writing programs and analyzing data, the credits are consumed extremely fast, "a package of around 140 yuan tier is basically enough."

A cloud computing agent who sells AI Agent office rights has encountered many "coupon scalpers": "They get all their relatives and friends to register accounts, so that they can keep using the free credits given away."

The aforementioned product operations specialist believes that the current monthly visit data of various products represents "whether users are willing to try", and retention, payment and renewal are the key to testing whether AI Agent office products can achieve sustainable development. "Doubao and DeepSeek both have huge C-end users, but how many of them are willing to pay?" Tencent, Alibaba, ByteDance and other manufacturers are all thinking about this question.

This article is from the WeChat official account "Economic Observer", author: LIU Sixuan, republished with authorization from 36Kr.