China’s AI battle moves from chatbots to agents
China’s tech giants Tencent, Alibaba and ByteDance are racing to turn AI agents into the next essential workplace tool, moving beyond chatbots to software that can independently carry out complex tasks. But as they vie to reshape how businesses operate, questions remain over adoption, profitability and whether these agents can deliver measurable value.
16 Oct 2026
Technology
(By Caixin journalists Guan Cong, Yang Zurui, Qu Yunxu and Kelsey Cheng)
Upload a set of spreadsheets to Tencent’s WorkBuddy and an agent can analyse the data and build an interactive dashboard. Ask ByteDance’s Doubao Work to go through product comments, and it can identify what buyers like, recommend products to promote, and prepare advertising images and copy in several languages.
ByteDance Ltd, Tencent Holdings Ltd and Alibaba Group Holding Ltd are reorganising products and teams around AI agents that can carry out office tasks across apps, betting that the software workers use to get things done will become a new gateway to customers.
These tools aim to connect work that once required separate business software. Instead of opening each application and completing one step at a time, an employee can give an agent a task and let it choose the tools and steps needed to finish it.
A person at ByteDance’s workplace-software unit Feishu described the shift as “moving some decision-making from people to AI”. In many office tasks, Chinese general-purpose models are already capable enough; companies now need to redesign their standard operating procedures so agents can act on them, the person said.
Advances in AI coding have supported the transition. Programming tools such as Anthropic Inc.’s Claude Code and OpenAI’s Codex showed how models could handle complex, multistep work. But office agents target a much larger potential audience. While China has an estimated 3 million to 5 million programmers, it has a much larger white-collar workforce, an executive at an internet company developing office agents told Caixin.
But office work also requires specialised knowledge. Finance, law and healthcare tasks, for example, depend on data and expertise that general-purpose models may not have. That gives technology companies a reason to connect agents with established workplace software, where company information and business processes are stored.
The contest has become a new priority for China’s internet giants following their race to build chatbots. A report from Citic Securities Co Ltd estimated the global office-agent market at between US$108 billion and US$1.56 trillion based on white-collar worker adoption, or between US$300 billion and US$1.35 trillion when measured against the value of labour that agents might replace.
AI office products in China had a combined 102 million users in July, up 8.2% from a year earlier, while total usage more than doubled, according to research firm QuestMobile.
Three different strategies
Tencent launched WorkBuddy in March and released an enterprise version in June. QuestMobile counted 6.58 million monthly active users in July, ahead of ByteDance’s and Alibaba’s comparable coding products at the time.
Tencent has since consolidated its agent products. In September, it said QClaw would be discontinued within three months, with the company directing users toward WorkBuddy. WeChat and WeCom have enabled users to transfer chat records to WorkBuddy, which can generate WeChat mini programmes from prompts without requiring users to write code.
A Tencent employee told Caixin that WorkBuddy had an early advantage because it was built as a consumer product with a simple setup, while many competing agents were modified from the viral, open-source AI assistant OpenClaw and were harder to configure. WorkBuddy was also opened to Tencent employees early, which helped it connect with the company’s tools and teams, the employee said.
Tencent is also directing more computing resources toward the product. The company formed a cloud unit in April dedicated to WorkBuddy and CodeBuddy, led by Tencent Cloud Chief Technology Officer Wang Huixing. In the second quarter, Tencent shifted priorities away from its chatbot Yuanbao and toward WorkBuddy.
Tencent Chief Strategy Officer James Mitchell said WorkBuddy subscriptions were already beginning to generate returns, with revenue flowing to Tencent Cloud. He said Tencent sees the product as having greater long-term value than renting out computing capacity or selling model access alone.
WorkBuddy’s subscription revenue has gross margins comparable to Tencent Cloud’s, Mitchell said, though subsidies to acquire users are weighing on the product’s overall margin. Tencent has set aside enough GPUs to expand its cloud business, he added.
The investment comes as profits are being squeezed. Tencent’s second-quarter net profit attributable to shareholders grew 9% year-on-year, breaking a three-year streak of double-digit expansion, according to its earnings report. Capital expenditure, primarily for purchasing computing power, surged 65% from the previous quarter to 59.3 billion RMB (US$8.9 billion).
Meanwhile, Alibaba is emphasising business customers and collaboration through DingTalk, its workplace platform. In July, it combined QoderWork, Wukong and MuleRun under the QwenWork platform, bringing together a coding agent, an agent that works with DingTalk, and a marketplace for agent skills. Alibaba said QwenWork had more than 30 million users by 4 September, about a month after its launch. It has connected with DingTalk, Feishu and WeCom.
The consolidation followed management changes. DingTalk founder Chen Hang left in June, and Chen Yusen, who had led MuleRun, took over. A week later, he announced the merger of MuleRun and Wukong. The changes followed internal criticism of DingTalk’s product organisation, which drew attention to difficulties coordinating the company’s AI products.
Alibaba has also created a business unit, Alibaba Token Hub, to bring models and applications closer together. QwenWork’s pitch draws on Alibaba Cloud’s business customer base and sales network.
A person at Alibaba told Caixin that adapting mature software for agents is more difficult than launching an AI-native product because agents must be able to use years of company data and understand business context. But once the work is done, the person said, it can create an important advantage.

Get the ThinkChina Weekly Newsletter
Insights on China, right in your mailbox. Sign up now.
By subscribing, I agree to SPH Media's Terms and Conditions and Privacy Policy.
At the same time, ByteDance has been reshaping its products and organisation around Doubao Work, while combining it with Feishu. In July, it split Feishu’s product and sales functions, moving the product team under its AI product unit and the sales team to its cloud business. In August, it brought its coding tool Trae and agent platform Coze into the AI product unit. In September, ByteDance Chief Executive Liang Rubo made a rare public appearance to introduce Doubao Work and pledged more resources for the enterprise market.
QuestMobile estimated that Doubao had 310 million users who used it for work in July, compared with 126 million for Alibaba’s Qwen app. ByteDance’s strategy relies on combining Doubao’s consumer reach with Feishu’s access to company information and collaboration tools.
Feishu Chief Executive Xie Xin said the latest version was redesigned to treat agents as a core part of the workplace. Users can add agents to group chats, where several agents can work together. Feishu’s annual recurring revenue grew 2.5 times in the first half of 2026 from a year earlier, its fastest growth since the product was founded, Xie said.
Adoption, economics remain unsettled
Office platforms and software-as-a-service products are expected to continue handling more structured tasks, such as approvals, compliance and workflows, a data-analysis software executive told Caixin. Agents, meanwhile, can take on flexible work using the data stored in those systems.
Businesses are still testing the products. Budgets for software and digital transformation are usually set in advance, while agent spending can fluctuate with token use. A Baidu Inc executive told Caixin that companies were trying several vendors and comparing their products, rather than deciding purchases based on measured returns.
A September survey by research firm IDC found that enterprise agent use remains concentrated in tasks such as customer service, knowledge management and enterprise search. More than half of surveyed companies said measuring return on investment was a key hurdle to broader deployment.
Pricing is unsettled, too. Many products charge both a seat, or per user, fee and model usage fees. WorkBuddy’s personal subscriptions cost between 70 RMB and 700 RMB a month, while enterprise seats cost between 198 RMB and 316 RMB.
The commercial model for agents is still taking shape, a software executive told Caixin. To deliver useful results, agents need a deeper understanding of the customer’s business — a capability that depends on expertise and data accumulated by software providers. Those companies could become partners if they make their tools and knowledge accessible to agents, the executive said.
Companies are also looking beyond subscriptions. A Tencent employee said agents could eventually charge users for specialist skills or expert services.
The data-analysis software executive told Caixin that its specialist agents, available through WorkBuddy, have generated tens of thousands of calls. The company sees an opportunity to charge for results. For instance, it could use an agent to identify irregularities in a retailer’s sales rebates and charge for that result, rather than selling access to software alone.
From office tools to platforms
The companies’ longer-term ambition is to build platforms that connect workplace software, outside services and specialist agents. Doubao Work’s desktop app, for example, can connect to services including NetEase Mail, WeCom, business-data provider Tianyancha and legal database Pkulaw.
Traditional software companies are adapting, too. US enterprise software provider Salesforce Inc. has broken parts of its customer-management software into APIs and tools that agents can call. SAP SE has said it is shifting toward becoming a “business AI company” and has introduced agents for tasks such as entering financial records and reconciling accounts.
Microsoft says its Microsoft 365 Copilot had more than 30 million paid seats as of the end of July, after adding 10 million in one quarter. The company has announced agents that can run in the background of Microsoft Office applications and act with their own identities and permissions.
In China, Beijing Kingsoft Office Software Inc. is promoting its WPS Comate agent, drawing on its established software and government and corporate customers. Kingsoft Vice President Wang Shaokang told Caixin that users still need documents and spreadsheets to review and edit work, even as agents make it easier to retrieve data or generate reports.
Agents are also expanding beyond office work into everyday tasks such as restaurant booking and travel planning. Meta Platforms Inc introduced its Muse personal-agent app in September, while OpenAI and Google have also launched personal-agent products. ByteDance is developing a similar project, and Tencent introduced LightVela, a personal agent accessible through WeChat.
Gu Zhaowei and Bao Yunhong also contributed to this story.
This article was first published by Caixin Global as “Cover Story: China’s AI Battle Moves From Chatbots to Agents”. Caixin Global is one of the most respected sources for macroeconomic, financial and business news and information about China.
Related: Why Chinese celebrate AI but Americans hesitate | China’s second Belt and Road may be built on AI
Popular This Month

Get the ThinkChina Weekly Newsletter
Insights on China, right in your mailbox. Sign up now.
By subscribing, I agree to SPH Media's Terms and Conditions and Privacy Policy.