A livestream room in Hangzhou never goes dark. At 3 a.m., a host is still describing a blender's six speeds, answering a buyer's question about warranty, and switching camera angles — except the host is not human.
Across China, thousands of these rooms now run on AI digital humans (数字人) that talk, react to comments, and sell around the clock.
The question is no longer whether the technology works, but who is liable when it misfires.
From demo reel to checkout counter
The "virtual host" was a conference curiosity three years ago. In 2025 it became a line item. JD Cloud (京东云) says its AI digital humans served more than 9,000 brands and generated roughly 14 billion RMB (about US$1.96 billion / HK$15.3 billion) in incremental GMV — sales it attributes to AI-run broadcasts rather than human-led ones.
That scale did not come from one platform. Alibaba, Douyin, Kuaishou, JD and Baidu all ship digital-human tooling, and during the 2025 mid-year shopping festival alone, JD reported some 17,000 merchants turning on AI hosts. In the most standardized categories — consumer electronics and daily goods — industry trackers put AI-hosted livestream penetration above 70%.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the share of standardized-category livestreams running at least partly on AI hosts crossed 70% in 2025, a threshold that marks commercial, not experimental, adoption.
The "digital employee" with a filing
The clearest signal that this is now a real business came from Nanjing. Silicon Intelligence (硅基智能), a local developer of "silicon-based labor" (硅基劳动力), filed for a Hong Kong IPO in late 2025. Its prospectus, citing market researcher CIC (灼识咨询), claims a 32.2% share of China's digital-human agent market by 2024 revenue — the largest domestically, and second globally.
The numbers behind the filing:
- Cumulative delivery of 80,000+ digital employees to clients in telecom, finance, government and e-commerce.
- 2024 revenue above 600 million RMB (about US$84 million / HK$656 million).
- A cloning platform that reproduces a person's voice, appearance and "cognitive pattern," then wires it to an industry knowledge base.
The pitch is simple: a human host caps out at 6–8 live hours a day; a digital one runs 7×24 and can span time zones without jet lag.
Why merchants actually say yes
Strip away the sci-fi framing and the adoption driver is cost and coverage, not spectacle.
- Coverage: a digital host fills the dead 2 a.m.–8 a.m. window when human teams log off but shoppers are still browsing.
- Cost: AI-run rooms routinely cut operating cost by 70–90% versus a human crew once the avatar and script are built.
- Consistency: no off-script moments, no fatigue, no reputational slip from an exhausted presenter.
For small brands that could never afford a rotating team of anchors, the math is what unlocked the long tail. Industry estimates put China's core virtual-digital-human market at about 48 billion RMB (US$6.7 billion / HK$52.5 billion) in 2025, up from single digits a few years earlier — though those figures are analyst estimates, not audited totals. That spread is why the category is now treated as infrastructure for commerce, not a marketing experiment.
The rules are catching up
Rapid deployment created a regulatory headache. In late 2025, China released its first national standard for customer-service virtual digital humans — Technical Requirements for Customer-Service Virtual Digital Humans (客服型虚拟数字人) — setting baseline expectations for identity labeling and content control. A separate draft from the cyberspace regulator in 2026 pushes a blunt rule: label the thing as a digital human, and don't ship "virtual romance" or deepfake-impersonation features aimed at minors.
The line the industry is learning to walk: AI hosts are allowed and even encouraged, but undisclosed impersonation of a real celebrity to sell goods is not.
Honest limitations
The 70–90% cost savings are vendor-cited averages, not guarantees for every SKU — conversion still depends on script quality and product fit. Emotion and intent recognition in live chat remains imperfect, so escalations still need a human. Labeling rules exist, but enforcement across millions of rooms is uneven, and the deepfake-impersonation risk is real until platforms close detection gaps. None of this is a substitute for the human judgment a real sales or support team provides on complex or high-value purchases.
What readers can do now
- If you run a store on a Chinese platform, pilot one AI-hosted slot for your lowest-traffic hours before committing budget to a full-day deployment.
- When buying from a livestream, check whether the host is labeled as AI; treat claims the same as you would a human pitch and verify warranty and return terms yourself.
- For founders: the defensible opportunity is vertical knowledge (a trained avatar that actually knows your category), not generic face-and-voice cloning.