Meta’s personal artificial intelligence assistant, Muse, recorded an estimated 10.9 million downloads within six weeks of its September launch, according to Sensor Tower. Its early popularity has provided a reference point for a growing number of Chinese startups developing similar products. But attracting users is only the first hurdle. These companies must also secure access to the services people use, manage the computing costs of proactive assistance, and convince customers to pay.
Qi Junyuan, the former head of Doubao’s desktop product, left ByteDance in 2025 and founded a startup earlier this year. Its product, Today AI, emphasizes persistent memory and proactive assistance.
Manus, known for its general-purpose AI agent, introduced Cue in September, equipping its personal agents with tools including dedicated email accounts and virtual computers.
Other startups are taking different approaches. Some offer standalone software for phones and computers, while others integrate their assistants with hardware. Their software strategies also vary, from learning about users over time to providing AI agents with their own computers and accounts, or allowing them to operate existing devices through multiple interfaces. These approaches are not mutually exclusive.
In China, all these teams face competition from free products offered by much larger companies. Doubao and Qwen already have substantial user bases, and their respective developers, ByteDance and Alibaba, have been adding task execution capabilities. Independent assistants must offer enough convenience to retain users and persuade some to pay despite the availability of free alternatives.
Even paying customers cannot guarantee that an assistant will complete a task. Shopping, travel, and payments are spread across platforms that can restrict automated access. After ByteDance introduced a protocol allowing developers to opt out of automated actions by its second-generation Doubao mobile assistant, apps including WeChat, Taobao, Alipay, and Meituan appeared on the opt-out list.
Interviews conducted by 36Kr suggest that startups are responding by narrowing their initial ambitions. Lidan is starting with notes and schedules before expanding into other aspects of users’ lives. WishSeen is concentrating on a single everyday function, while UniCone, which emphasizes proactive assistance, is initially targeting creators and people developing personal brands.
Some are beginning with a particular function, others with a defined customer group. Their shared goal is to establish recurring usage and a willingness to pay before expanding what their assistants can do.
“Personal agents are a good story, but if users don’t keep using them and remain willing to pay, it simply won’t work,” said WishSeen co-founder Cao Fangyongzheng.
China’s user habits and service ecosystem differ from those Muse encounters. Whether a Chinese counterpart can follow its path, and whether Muse itself can establish a sustainable business, remain open questions.
Give users a reason to choose an assistant
AI agents have expanded from coding into workplace tasks and, increasingly, everyday life. As their applications become more specialized, each product needs a reason for users to choose it over existing alternatives.
Zha Yujie’s Shenzhen-based team began with a ChatGPT-based chatbot for WeChat and developed it into CowAgent, an open-source assistant capable of calling software tools and operating a computer. Its enterprise platform, LinkAI, helps customers build agents, knowledge bases, and automated workflows.
“Personal agents aren’t really a new category or technology. They are more a use case,” Zha, LinkAI’s founder, told 36Kr.
Ordinary users already have Doubao and Qwen, while professionals have workplace agents.
“If you develop another similar product, how to get individual users to pay is a real question.”
Menlo Ventures’ 2026 consumer AI report offers one possible answer. The share of US AI users who use both general-purpose assistants and specialized AI tools rose from 60% in 2025 to 72% in 2026. The venture capital firm found that consumers who already use ChatGPT or other general assistants may adopt additional products if those tools address particular needs more effectively.
Personal agents could therefore complement general AI tools, provided they offer tangible everyday convenience.
Lidan, for example, turns users’ requests into tasks displayed on its main screen rather than leaving them buried in chat histories. Its founder, A Tang, describes the approach as task-centered. The product currently helps users manage notes and schedules, with plans to support hotel and flight bookings later.
The company’s social media profile reads: “In Chengdu, five people building China’s version of Muse.”
Tang said the team is already considering international expansion.
“Our next focus is overseas, but we want to refine the product in the Chinese market first.”
An assistant designed to serve users continuously must also know when to intervene. Shenzhen-based UniCone is developing a proactive assistant that identifies tasks and breaks them into steps using information users authorize it to access, including screen content and chat histories.
“It has to be proactive, which is different from simply helping you complete something,” its CEO said.
Achieving that requires an ongoing understanding of a user’s circumstances, supported by background computing that determines when assistance might be useful.
Personal information, however, changes over time. TechRadar editor Lance Ulanoff described an example from his trial of Muse: the assistant found an old plan for a trip to Portugal that conflicted with a more recent Italy itinerary in his email. Rather than automatically choosing a destination, it asked him to confirm before making the correction.
Muse can draw on current information from email and other connected services. In China, much of that context is distributed across superapps such as WeChat, Alipay, and Dianping. Without access to these services, delivering useful proactive assistance becomes more difficult.
Many Today AI users have authorized access to Lark (also known as Feishu), health information, and calendars, helping the assistant develop a better understanding of their activities and preferences, Suki, a product manager at Today AI, told 36Kr.
Not every startup intends to manage every aspect of a user’s life from the outset. WishSeen plans to retain memory and proactive assistance while concentrating on one everyday function in its mobile assistant.
Cao said developing a comprehensive product would be too costly and difficult to differentiate from competitors.
“I would rather make one function good enough that users want to use it every day, and even pay for it.”
Reaching users is only the first hurdle
Most of these startups have tested their ideas with relatively small groups. A few hundred early users may help validate a concept, but scaling up means competing with established technology companies that have far greater distribution networks and resources.
In January, Alibaba demonstrated how users could order bubble tea through Qwen using integrations with services such as Taobao and Alipay. Wu Jia, president of Qwen’s consumer business group, said he hoped the Qwen app would eventually be sufficient for most everyday tasks.
Alibaba introduced Qwen Intelligence for smartphone manufacturers in September. The following month, it opened applications to test Qwen Go, which supports tasks such as insurance claims, price protection claims, and appointment bookings. ByteDance has also been adding travel and household bill-payment functions to Doubao since late September.
Large companies can distribute these capabilities through existing products while investing in integrations with external services. Independent teams, meanwhile, must acquire users from scratch.
“Big companies have deep pockets. They can offer free products and spend money advertising them. We can’t match that,” Tang said.
He hopes Lidan’s interface will provide a point of differentiation.
“If the products are exactly the same, we definitely can’t compete with them.”
“Reminders and to-dos shouldn’t be buried in the conversation stream. They should be pulled out onto the main page,” he said. “Once the valuable things have been extracted, the chat history isn’t so important.”
Zha doubts that interface differences can provide a lasting advantage.
“There are differences, but the time they last will definitely get shorter.”
Open-source frameworks and increasingly capable development tools have made agents easier to build and replicate. Rapid improvements in underlying AI models have also made some investors cautious.
“Models are iterating so quickly now. The software application you invest in today might be wiped out by a model tomorrow,” an investor at an industry investment fund told 36Kr.
Still, startups see value in establishing even a temporary lead.
“We may be among the earliest teams to try adding memory to agents, and among the more aggressive teams in pursuing proactivity,” Suki said.
She argues that larger companies’ financial resources do not necessarily guarantee sustained commitment to a product.
“A month after launch, if the numbers are bad, or it can’t make hundreds of millions, the boss can decide on a whim to cut the project.”
A startup’s product advantage may not last indefinitely, but “even if it doesn’t become a defensible advantage, that period still matters.”
Beyond distribution and product differentiation, assistants face another obstacle: gaining permission to operate across platforms.
Alibaba and ByteDance can integrate their own services first, but users’ daily activities extend well beyond any single company’s ecosystem.
“If everyone shares their permissions with Doubao, every app becomes just a backend. That is something none of the app makers want to accept,” said Chen Mo, a pseudonym for an agent product manager at a major technology company.
Ahead of the release of its second-generation mobile assistant, ByteDance introduced the Screen Automation Execution Protocol (SAEP). The protocol allows app developers to declare whether AI assistants can operate their services and specify which pages and actions are accessible.
As of October 14, 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 shopping on behalf of users. According to media reports, the retailer argued that third-party applications purchasing goods from other businesses should “operate openly and transparently, and respect service providers’ decisions on whether to participate.”
Zha expects platforms to become more accessible eventually, but believes the process will be slow:
“Over the long term, I think they will open up, but it may take years. In the short term, definitely not.”
Consumer platforms have their own data and commercial interests to protect, while smaller AI developers have limited leverage to negotiate access individually.
UniCone’s CEO sees a potential competitive incentive for opening up:
“If one big company doesn’t open up, but other platforms do, traffic will move to them.”
As AI agents become another way for users to discover and access services, he expects platforms to make their interfaces more accessible.
Chen, who has worked on several agent projects at large technology companies, is less optimistic. He believes the challenge of cross-platform integration “won’t be solved within at least five years.”
The underlying conflict is commercial. Personal assistants promise to spare users from navigating individual apps, while platforms depend on direct engagement to sustain their businesses.
Until agents can reliably perform everyday tasks across these services, their developers may struggle to persuade customers to pay for them.
Who will pay for a personal agent?
A personal agent needs to become useful enough to form part of a user’s routine before many people will consider paying. Yet continuous background operation and proactive assistance incur computing costs from the outset.
Startups therefore need to fund the experience before they have established recurring subscription revenue.
“We don’t just need users. We need revenue. Every additional user means another expense. We can’t afford to support so many users for free,” Tang said.
Lidan may withdraw its free usage allowance to test how many users are willing to pay, then decide whether to invest in advertising and influencer promotions.
“We won’t spend heavily to expand the user base at the beginning.”
Running AI models is a significant cost that distinguishes personal agents from many traditional internet products. Tang’s team has tested several models, finding them either too expensive or too slow. Cheaper, faster alternatives sometimes struggle to understand instructions, use external tools, or complete longer tasks.
He remains optimistic about improvements in model capabilities and efficiency.
“I’m optimistic about improving model capabilities and falling prices. Useful models that help ordinary people get things done will eventually become affordable for everyone.”
WishSeen faces a similar constraint. Its team estimated that a general-purpose assistant meeting its expectations for proactive service and reliable task execution would cost at least RMB 1,000 (USD 150) per user each month.
“No user is going to pay you RMB 1,000 every month,” Cao said.
Large companies can subsidize user acquisition and computing expenses, while smaller teams need to determine early whether revenue from each customer can cover the cost of providing the service.
“The biggest problem isn’t technology. It’s cost.”
Founders are still looking for models that combine sufficient capability, fast responses, and low operating costs.
“Current models and agents weren’t designed for the personal AI use case,” said UniCone’s CEO, who conducts AI research at Tsinghua University.
The company is experimenting with smaller models and additional training of existing ones to reduce the cost of continuous assistance.
Another approach is to charge a narrower audience for a more specialized service. In an essay published last year, a16z partner Anish Acharya argued that higher-priced AI subscriptions often address specific needs, such as coding or video creation. Startups could therefore serve smaller groups of customers willing to pay more for a product that meets their needs, rather than immediately pursuing the mass market.
Muse offers a free tier alongside subscriptions priced at USD 20 and USD 100 per month. Meta has not disclosed specific revenue figures for the service.
Transaction commissions offer another potential revenue stream, but they create questions about whose interests an assistant ultimately serves.
Today AI’s head of R&D, Gao Ce, sees a potential conflict.
“Revenue sharing will inevitably push product design toward higher conversion rates,” he said, which could affect the user experience.
An assistant making decisions on someone’s behalf would need to balance that person’s interests against its own incentive to earn commissions.
The remaining question is whether investors are willing to finance development before customers demonstrate a willingness to pay.
One AI industry investor told 36Kr that he had recently concentrated on computing and AI infrastructure because demand in those areas was easier to measure. Growth among some application developers has since drawn his attention back to the sector.
“Lately, I’ve been spending some time looking at applications again. Some of their growth looks quite decent.”
Cao previously worked at an AI model startup whose investors lost confidence midway through its development. This time, she wants WishSeen to establish product-market feedback and user data before raising capital.
“If the product can’t run well on its own, investors’ money won’t make it run well.”
UniCone is also speaking with investors but wants funding to support expansion rather than sustain its basic operations.
LinkAI has taken a different approach. Its enterprise business supports the broader team: open-source and personal products attract users, while enterprise subscriptions, model and computing services, and private deployments generate revenue.
The company has not raised external funding, and Zha said financial pressure is currently limited. Its structure allows personal AI products to attract users while enterprise customers help fund development.
For now, the eventual form of the personal assistant remains uncertain, as does the identity of the companies best positioned to deliver it.
“What will the final product look like? Will it be hardware or software, and which company will build it?” asked Zha, who left a major technology company to become a founder.
He does not doubt that demand exists.
“Who will ultimately build them, and in what form, is still very uncertain, particularly in China.”
KrASIA features translated and adapted content that was originally published by 36Kr. This article was written by Li Zhaofeng for 36Kr.
Note: RMB figures are converted to USD at rates of RMB 6.69 = USD 1 based on estimates as of November 10, 2026, unless otherwise stated. USD conversions are approximate and, where appropriate, rounded for ease of reference. They may not fully match prevailing exchange rates.