China's AI Assistant Startups Face a Monetisation Problem

A crowded race with uncertain economics

Within weeks of Meta releasing its Muse assistant in September, the product accumulated more than ten million installs globally, data from Sensor Tower shows. That early traction has drawn attention in China, where a cohort of new ventures is building proactive AI helpers designed to remember user preferences, anticipate needs, and execute tasks without prompting. Yet attracting downloads represents only the opening challenge. These startups must also persuade platforms to permit automated actions, manage the expense of running inference continuously in the background, and ultimately convince people to subscribe when ByteDance and Alibaba distribute comparable tools at no charge.

Qi Junyuan, who previously led the desktop edition of ByteDance's Doubao assistant, departed the company in 2025 and launched Today AI earlier this year. The product prioritises persistent memory and unprompted intervention. Manus, a startup known for its general-purpose agent framework, unveiled Cue in September, furnishing its assistants with dedicated email addresses and sandboxed virtual environments. Other teams are pursuing divergent architectures: some ship standalone mobile and desktop applications, others bundle their software with consumer hardware. Implementation strategies vary from gradual learning of user habits to provisioning agents with isolated computing resources or granting them control over existing devices through multi-modal interfaces. These technical choices are not mutually exclusive, but every team confronts the same structural obstacle in China - entrenched free offerings from companies with vast user bases and the financial capacity to subsidise compute indefinitely.

Doubao and Qwen already count tens of millions of active users, and both ByteDance and Alibaba have begun layering task-execution capabilities into their platforms. Independent assistants must deliver enough incremental convenience to retain users and convert a meaningful fraction into paying subscribers, even as free alternatives improve. Paying customers, however, cannot guarantee task completion. Shopping, travel booking, and payments flow through siloed platforms that can block automated access at will. After ByteDance introduced the Screen Automation Execution Protocol (SAEP) in advance of its second-generation mobile assistant, developers of WeChat, Taobao, Alipay, and Meituan placed their apps on the opt-out registry. Interviews with startup founders suggest a tactical retreat: narrow the initial scope, establish recurring usage in one domain, then expand once retention and payment behaviour are proven.

Specialisation as survival strategy

AI agents have moved from code generation into workplace automation and, more recently, personal productivity. As applications proliferate, each new entrant needs a defensible reason for adoption. Zha Yujie, founder of Shenzhen-based LinkAI, which helps enterprises build agents and automated workflows, argues that personal assistants represent a use case rather than a novel technology category. Consumers already have access to Doubao and Qwen, while professionals deploy workplace-specific agents. Convincing individuals to pay for another variant demands a clear value proposition.

Research from Menlo Ventures offers one pathway. The share of US consumers using both general-purpose chatbots and specialised AI tools climbed from 60 per cent in 2025 to 72 per cent in 2026, the firm reported. Users who rely on ChatGPT or similar platforms will adopt additional products if those tools address distinct needs more effectively. Personal assistants could therefore coexist with general AI interfaces, provided they deliver tangible, repeated convenience.

Lidan, a Chengdu-based startup, renders user requests as discrete tasks on its home screen rather than burying them in conversation threads. Its founder, A Tang, describes the design as task-centric. The product currently handles notes and calendar management, with hotel and flight booking slated for later releases. The team's social media bio declares it is building China's answer to Muse, and Tang confirmed the company is already mapping overseas expansion. Refining the product in the domestic market comes first, he said.

An assistant intended to operate continuously must also determine when to act. UniCone, based in Shenzhen, is developing a proactive system that identifies tasks and decomposes them into steps using information users authorise it to access, including screen captures and messaging histories. The company's chief executive emphasised that true assistance requires initiative, not merely reactive execution. Achieving that demands ongoing awareness of a user's context, supported by background inference that evaluates when intervention might prove useful.

Personal circumstances, however, evolve. TechRadar editor Lance Ulanoff recounted an instance during his trial of Muse in which the assistant surfaced an outdated plan for a trip to Portugal that clashed with a more recent Italy booking in his inbox. Rather than resolving the conflict autonomously, it asked for confirmation before updating its understanding. Muse can pull current information from email and other linked services. In China, much of that context resides in superapps such as WeChat, Alipay, and Dianping. Without hooks into these platforms, delivering useful proactive assistance becomes significantly harder.

Many Today AI users have granted access to Lark (known locally as Feishu), health data, and calendars, helping the assistant build a richer model of their routines and preferences, according to Suki, a product manager at the company. Not every startup aims to manage every dimension of a user's life from the outset. WishSeen intends to retain memory and proactive features while concentrating on a single everyday function in its mobile app. Co-founder Cao Fangyongzheng said building a comprehensive product would prove too expensive and difficult to differentiate. Making one function compelling enough that users return daily and eventually pay represents a more pragmatic path, he argued.

Distribution asymmetry and platform gatekeeping

Most of these startups have validated their concepts with cohorts numbering in the hundreds. A few hundred early adopters can confirm product-market fit, but scaling means competing with established technology groups that command far larger distribution channels and deeper capital reserves. In January, Alibaba demonstrated ordering bubble tea through Qwen via integrations with Taobao and Alipay. Wu Jia, president of Qwen's consumer business unit, said he envisaged the Qwen app becoming sufficient for most routine tasks. Alibaba introduced Qwen Intelligence for smartphone manufacturers in September, then opened applications in October to test Qwen Go, which supports insurance claims, price-protection requests, and appointment scheduling. ByteDance has similarly been adding travel and utility-bill payment functions to Doubao since late September.

Large incumbents can distribute these capabilities through existing products while investing in integrations with external services. Independent teams, by contrast, must acquire users from scratch. Tang acknowledged the asymmetry: deep-pocketed companies can offer free products and fund advertising campaigns that startups cannot match. He hopes Lidan's interface will provide a point of differentiation. Reminders and to-do items should not be buried in conversation streams, he said; they should be surfaced on the main page. Once valuable information is extracted, the chat history becomes less critical.

Zha doubts that interface distinctions can sustain an advantage. Differences exist, but the duration of any lead will shrink, he said. Open-source frameworks and increasingly capable development tools have lowered the barrier to building and replicating agents. Rapid iteration in underlying models has also made some investors cautious. One investor at an industry-focused fund said models are advancing so quickly that a software application funded today might be rendered obsolete by a model update tomorrow.

Startups, however, see value in establishing even a transient lead. Suki argues that larger companies' financial resources do not guarantee sustained commitment. A month after launch, if metrics disappoint or revenue fails to reach hundreds of millions, a project can be terminated on short notice, she said. A startup's product advantage may not endure indefinitely, but that window still matters.

Beyond distribution and differentiation, assistants confront a structural obstacle: securing permission to operate across platforms. Alibaba and ByteDance can integrate their own services first, but users' daily routines span ecosystems controlled by multiple companies. If every user granted full permissions to Doubao, every app would become a backend service, an outcome no platform operator would accept, said Chen Mo, a pseudonymous agent product manager at a major technology firm.

Ahead of its second-generation mobile assistant, ByteDance introduced SAEP, a protocol that lets app developers declare whether AI assistants may operate their services and specify accessible pages and actions. As of mid-October, the published opt-out list contained 155 apps, including WeChat, Taobao, Alipay, and Meituan. The restrictions affect large companies as well as startups. In September, Amazon blocked Muse from making purchases on behalf of users, arguing that third-party applications buying goods from other businesses should operate transparently and respect service providers' decisions on participation, media reports said.

Zha expects platforms to become more accessible eventually, but believes the timeline will stretch across years rather than months. Consumer platforms have proprietary data and commercial interests to protect, while smaller AI developers lack the leverage to negotiate access individually. UniCone's chief executive sees a potential competitive incentive: if one large company refuses to open up but rival platforms do, traffic will migrate. As AI agents become another discovery and access layer, he expects platforms to make their interfaces more accommodating.

Chen, who has worked on several agent initiatives at large technology companies, is less optimistic. He believes the cross-platform integration challenge will not be resolved within at least five years. The underlying tension is commercial: personal assistants promise to spare users from navigating individual apps, while platforms depend on direct engagement to sustain their businesses.

The willingness-to-pay question

Until agents can reliably execute everyday tasks across these services, their developers face a difficult monetisation question. WishSeen's Cao framed the challenge bluntly: personal agents make for a compelling narrative, but if users do not maintain engagement and remain willing to pay, the business model simply will not function.

China's user habits and service ecosystem differ from the environment Muse encounters in other markets. Whether a Chinese counterpart can replicate its trajectory, and whether Muse itself can establish a sustainable revenue model, remain open questions. The startups now navigating platform restrictions, compute costs, and user expectations are testing whether proactive AI assistance can become a category people pay for, or whether it will remain a feature absorbed into the free offerings of the largest technology companies.