At 4 a.m., shopping livestreams on Taobao can still be busy. Some of the hosts pitching products, however, are digital replicas built from short videos of real people. They can speak from scripts, gesture toward products, and keep a sales channel open while human hosts are away.
For brands, the appeal is continuity at lower cost. For workers and platforms, the shift raises questions about jobs, disclosure, and how much human participation a livestream needs.
From a short video to a working sales host
Chinese startups and major technology companies began offering AI avatars for e-commerce livestreaming in 2022. A basic clone can be made from a brief sample video, then animated to match scripted audio. Staff enter product details, review the generated script, and start the stream.
The setup process has become faster. Silicon Intelligence, a Nanjing-based startup founded in 2017, once needed 30 minutes of training video to create a digital host. The requirement fell to 10 minutes, then three, and eventually one minute. The company’s basic avatar costs about 8,000 RMB ($1,100), including a year of maintenance. More capable versions can cost several thousands of dollars.
Scriptwriting can also be automated. Companies are using large language models to draft sales pitches, while more advanced systems can respond to comments by matching them with answers in a database. Sima Huapeng, Silicon Intelligence’s founder and CEO, says the technology can adjust its marketing approach based on the number of viewers.
Gestures help make the pitch feel familiar
These avatars draw on the repeated patterns of online selling. Xiaoice, a Chinese AI company, uses a database of nearly a hundred pre-designed movements. Its virtual hosts can point upward when asking viewers to follow a channel or down toward the shopping cart when introducing a product.
That coordination matters because a spoken sales pitch and a mismatched gesture can make an avatar look unnatural. Xiaoice’s Huang Wei, who directs the company’s virtual influencer livestreaming business, says the aim is to align body language with speech.
Xiaoice focuses on avatars that can show emotion, reflecting the difference between a static product listing and a host speaking to an audience. After trial projects with clients, it launched a service this year offering digital clones for under $1,000. Clients can pay more for detailed versions, such as the clone of sports announcer Liu Jianhong that read match results and related news on Douyin during the 2022 FIFA World Cup.
Lower costs come with trade-offs
AI hosts are not expected to displace the most successful livestream stars. Huang says they are better suited to replacing mid-tier presenters. Yet that still puts pressure on human workers: the source reports that livestream hosts in China found it harder to get jobs this year, and their average salary fell 20% compared with 2022, according to iiMedia Research.
Brands may see the greatest value in keeping streams active when audiences are smaller. AI-generated hosts already appear during post-midnight hours on platforms including Taobao and JD. Chen Dan, CEO of Quantum Planet AI, argues that companies can use virtual hosts for lower-return roles while retaining their strongest performers.
The replicas also struggle with demonstrations that depend on a real body. Chen says many of his clients sell furniture, but an AI host cannot sit on a sofa or lie in a bed to show how it feels to use. That leaves a gap between describing a product and physically testing it for viewers.
Platforms and regulators are still working out the boundaries
The market extends beyond smaller specialist companies. Alibaba, Tencent, Baidu, and JD have all launched versions of services that let brands create AI livestreamers. Marketing agencies have also begun working with avatar providers, including Foshan Yowant Technology’s announced collaboration with Xiaoice.
Douyin has emphasized transparency. In a May document, the platform said AI-generated videos should be clearly labeled and virtual influencers should be operated by real humans. It has also banned recorded video used as livestream content, leaving AI streams that contain little real-time human input in an uncertain position under that rule.
Chinese government laws covering synthetic media and generative AI apply to e-commerce streaming, but the effects of platform and government rules remain unclear while the technology is new and enforcement has not yet been serious.
Silicon Intelligence says it wants to add emotional responses to its avatars, such as sadness in response to abusive comments or happiness when products sell well. Sima says the company has generated 400,000 virtual streamers and has a longer-term goal of creating “100,000,000 silicon-based laborers” by 2025. For now, AI hosts can extend a sales pitch across more hours, but their limits and the rules around their use are still taking shape.