§ blog · AI & ML10/18/2026
← All articles

AI virtual KOLs livestreaming 24/7: how far China has gone, and what bringing it to Vietnam takes

In China, AI livestream hosts already sell real goods: Luo Yonghao's AI avatar brought in 55 million yuan of orders in a single session, and more than 70,000 sellers on JD use virtual hosts. But the rules have tightened too: AI content must be labelled, a real person must supervise, and the operator answers for every word the virtual KOL says. This article sums up what China has done, Vietnam's legal framework in 2026, and how to run a virtual KOL within the law — a service KonexForge is ready to deploy for businesses in Vietnam.

Virtual KOLLivestreamDigital humansAIE-commerce13 min read
By KonexForge Engineering Team
KOL ẢO LIVESTREAM · NHỮNG GÌ NẰM SAU GƯƠNG MẶTKHO TRI THỨCtài liệu sản phẩmđã kiểm chứngMÔ HÌNH + RAGchỉ trả lời từkho tri thứcBỘ LỌCchặn câu cấmchuyển câu nhạy cảmKOL ẢOgương mặt + giọngcó quyền sử dụngLIVESTREAMnhãn AI suốt buổibáo là quảng cáobình luận người xem quay lại mô hìnhNGƯỜI GIÁM SÁTxem bình luận và câu trả lời của vài phiêncó nút dừng · tắt sóng ngay khi có sự cốCÂU NHẠY CẢM → NGƯỜI THẬTkhiếu nại · đổi trả · sức khoẻ · giá riêngkể cả lúc nửa đêm · hẹn giờ phản hồi“24/7” · CHIA CA THEO VAI TRÒ (VÍ DỤ)KOL ảo dẫn · có nhãn AI · có người giám sátngười thật dẫn0h6h12h18h24h19–22hBỐN RỦI RO CẦN TÍNH TRƯỚCBỊA CÔNG DỤNGnói trước hàng nghìn ngườilưu lại ít nhất 1 nămMẠO DANHhình, giọng người thậtcần hợp đồng đồng ýCHÍNH SÁCH SÀNđổi theo từng quýthử ở kênh phụ trướcNGƯỜI XEM CHÁNtò mò qua rất nhanhđổi kịch bản theo nhật kýVIỆT NAM 2026 · BA LUẬT ĐÃ CÓ HIỆU LỰCLUẬT TRÍ TUỆ NHÂN TẠO · 1/3gắn nhãn AI · báo khi khách nói với AILUẬT TMĐT · 1/7định danh người bán · lưu buổi phát 1 nămLUẬT QUẢNG CÁO SỬA ĐỔI · 1/1kiểm tài liệu · báo rõ là quảng cáotham khảo: Baidu 6/2025 · JD 3/2026 · luật Việt Nam 2026konexforge.com

Imagine a cosmetics shop in Can Tho that livestreams every evening from 8 to 10 pm. The host is the owner, presenting products, answering comments and closing orders. Those two hours sell. But customers still visit the page at 11 in the morning and at midnight — and by then the channel is off. Hiring hosts for the other shifts would cost more than the extra sales.

In China, a great many sellers have answered this with a virtual KOL — a KOL (Key Opinion Leader, an influencer) who is not a person but a digital human: an AI-built avatar that speaks with a synthetic voice, reads a script and answers comments using a language model. It does not tire, does not change shifts, and can stay on air for hours.

This article covers three things: what China has achieved (with figures published by the companies themselves), how the rules have changed in China and Vietnam, and how to run a virtual KOL that is effective without exposing the business to legal risk.

How far China has gone

Luo Yonghao's livestream on Baidu (15 June 2025). Luo Yonghao — one of China's best-known livestream sellers — and his co-host put AI avatars of themselves on air on Baidu's Youxuan platform for more than six hours. According to Baidu, the session reached 55 million yuan in GMV (Gross Merchandise Value — the total value of orders), about $7.65 million, with more than 13 million views. The avatars were trained on five years of the pair's own livestream footage to mimic their voices and style; the system drew on a knowledge base of about 13,000 product items and generated about 97,000 words of product presentation during the session.

JD and virtual hosts for small sellers (March 2026). On 25 March 2026, JD said more than 70,000 sellers were using its virtual hosts to livestream, after the tool was made free to all sellers from December 2025. JD said the cost of a virtual host had fallen to about one tenth of a human host, and that GMV from virtual hosts reached tens of billions of yuan in 2025.

Two things to read carefully before turning these numbers into a plan:

  • The figures are company-reported. We found no independent verification of the 55 million yuan or of JD's sales. GMV is also the value of orders placed, before cancellations and returns.
  • Luo Yonghao's session had advantages that are hard to repeat. It was the first event of its kind by an extremely famous seller, heavily promoted. The avatars borrowed a reputation the real people had built over years. A brand-new virtual character launching on a small shop's channel has none of that.

The genuinely notable number is JD's: most of those 70,000 sellers are ordinary shops using virtual hosts not to create a sensation but to fill the hours when nobody is available. That is also the use that fits small and medium businesses in Vietnam best.

China has tightened the rules

As virtual hosts spread across platforms, regulation followed quickly:

  • AI content labelling (in force from 1 September 2025): the rules on identifying AI-generated content require both a visible label for viewers and an identifier in the data (metadata).
  • Livestream e-commerce supervision measures (in force from 1 February 2026), issued by the State Administration for Market Regulation and the Cyberspace Administration of China: for the first time they bring AI-generated virtual hosts under supervision, and they prohibit using AI to fabricate misleading commercial information or impersonate others.
  • Platform rules: according to reports on Douyin's rules, owners of virtual characters must register under their real identity, and livestreams using virtual characters must be supervised by a real person. A lawyer quoted by Sixth Tone said the operator of an avatar is legally liable for any false advertising during a broadcast.

The common message: AI may be the host, but it may not be the one held responsible. Behind every virtual KOL there has to be an identified person or business, and the AI label has to be visible to viewers.

In Vietnam, October 2026

Vietnam has no regulation written specifically for virtual KOLs, but three new laws that took effect in 2026 are enough to shape the approach:

  • The Law on Artificial Intelligence (Law No. 134/2025/QH15, in force from 1 March 2026): AI-generated or AI-edited audio, images and video that simulate real people must be labelled so they can be told apart from real content, and users must be told when they are interacting with an AI system rather than a person. Under the implementing decree, the disclosure must be clear, visible, and present before or while the viewer accesses the content; notices designed to obscure it do not count.
  • The E-commerce Law (Law No. 122/2025/QH15, in force from 1 July 2026): livestream sellers must complete electronic identity verification, and platforms must keep the video and audio of each livestream for at least one year. In other words, every virtual-KOL session is still tied to an identified person or business, and everything it says can be traced.
  • The amended Advertising Law (Law No. 75/2025/QH15, in force from 1 January 2026): anyone who directly presents products online must check the product documentation and state clearly that the content is advertising immediately before and during it; false advertising can be fined and require compensation.

The Advertising Law was written for people — a virtual KOL cannot "have tried" a product. Our cautious reading, while specific guidance is awaited: that responsibility passes to the business operating the virtual KOL, and the way to meet it is to let the virtual KOL say only what is in verified product documentation. This is our reading, not legal advice; for sensitive categories such as supplements or cosmetics, ask a lawyer as well.

Platform policy moves faster than law. Since mid-2026, TikTok Shop's US seller policy has banned AI-generated voices and pre-recorded audio from shopping livestreams — speech must come from a real person, live. We have not seen an equivalent rule published for TikTok Shop Vietnam, but the direction is clear. Before going live on any marketplace, re-read the livestream policy in that marketplace's seller centre.

What a virtual KOL is made of

Viewers see only a face and a voice. What decides whether a virtual KOL is safe to leave on air for hours sits behind it:

  • Face and voice. A new character, or a digitised version of a real host with a contract permitting it. For Vietnamese, the synthetic voice needs testing with real customers from each region — a voice that gets tones wrong or breaks sentences oddly drives viewers away very quickly.
  • A product knowledge base: specifications, ingredients, prices, promotions, return policy, product registration records. The language model may only answer from this base — the same RAG technique (retrieve documents first, then generate the answer) we described in RAG pipelines in production.
  • A filter before speaking. Every answer is checked before it is read out: does it contain a forbidden claim, does it quote a price not in the price list, is it a question that should go to a human.
  • A supervisor (human-in-the-loop): one person watching the comments and answers of one or several sessions at once, receiving the questions handed over, and holding a stop button to end the stream the moment something goes wrong.
  • Labels and logs. The AI label shows for the whole session; every question, answer and filter decision is recorded — both to find errors and to check against when a complaint comes in.

What "24/7" should mean

A virtual KOL running entirely on its own, unattended, streaming non-stop on every marketplace — that is the model Douyin, Taobao and TikTok Shop in the US are all blocking. The model that holds up splits shifts by role:

  • Peak hours: a human host leads. The virtual KOL, if used, plays a supporting part — presenting specifications, reading out promotions, reminding viewers of discount codes.
  • Off-peak hours and nights: the virtual KOL hosts on channels that allow it, labelled as AI, with one supervisor covering several sessions.
  • At all times: complaints, returns, health questions and special prices go to a human, even at midnight — the virtual KOL then replies that staff will respond within a stated window.

On channels the business controls — its own website or app — the virtual KOL can stream for longer, because the business sets the rules within the law. Beyond livestreams, the same character and knowledge base work for short videos and message-based customer advice — two places where platform rules are currently looser.

Four risks to plan for

One — wrong claims about the product. A language model can confidently invent a benefit, an ingredient or a promotion that does not exist. On a livestream, that sentence is said in front of thousands of people and kept for at least a year. This is the biggest risk, and the reason the knowledge base, the filter and the supervisor are not optional.

Two — impersonation. Using a real person's face or voice — even a former employee's — without written consent breaks both platform rules and the law. A contract to digitise a host needs to state the scope of use, the duration and how it ends.

Three — reduced reach or a suspended account. Policies change every quarter. A seller account built over years is not worth risking to try a channel whose rules are unclear. Test on a secondary channel or your own first.

Four — viewers lose interest. Initial curiosity fades fast. A virtual KOL that only recites specifications keeps nobody watching. The script needs to change with the product, the hour and the questions viewers actually ask — and the logs tell you what those are.

How to start

The lowest-risk way to begin is to pick one product group with complete documentation and few sensitive questions (household goods, accessories and fashion are usually easier than cosmetics or supplements), and one time slot your human hosts cannot currently cover. Run it for two to four weeks and measure by the hour: viewers, average watch time, orders, cost per order and — most important — wrong answers per 100 questions. Wrong answers must be reviewed and fed back into the knowledge base or the filter before expanding to another product group.

A good virtual KOL is not the one that answers the most questions, but the one that knows which questions it must not answer on its own.

Conclusion

China shows that virtual hosts sell real goods, and cost far less than human hosts in low-traffic hours. China also shows the limits: AI labels are mandatory, a real person has to stand behind the host, and the operator answers for every word. Vietnam in 2026 is heading the same way with its Law on Artificial Intelligence, E-commerce Law and amended Advertising Law.

So the right question is not "should AI livestream instead of people" but "which hours, which channels and which product groups can be handed to a virtual KOL — and who watches it".

KonexForge is ready to deploy AI virtual KOLs for businesses in Vietnam: character design or digitising a host under a consent contract, a Vietnamese voice, a product knowledge base, an answer filter, and supervision and labelling processes in line with the Law on Artificial Intelligence. If you would like to pilot it on one product group, contact us or see KonexForge's AI & ML capability.

Related articles

AI & ML

KonexForge AI Core: when the AI Orchestrator becomes the central brain of your enterprise ecosystem

Not a chatbot, not an LLM wrapper — AI Core is a unified orchestration layer that connects every AI specialist, enterprise tool, and internal data source into a single automated pipeline. An 8-component architecture, local/cloud routing, and a Critic Engine are the real differentiators.

AI & ML

Legacy websites are falling behind: when SEO isn't enough and GEO doesn't exist

Users are asking ChatGPT, Perplexity, and Google AI Overviews instead of clicking links. If your website isn't cited by AI, you're invisible to a growing share of new searchers — even if you rank #3 on Google.

AI & ML

AI Entry: digitizing image data with Vision-Language technology — replacing 50 data-entry workers in a single deployment

Most businesses still digitize invoices, forms, IDs, and handwritten ledgers with dozens of manual data-entry workers. AI Entry uses Vision-Language (VL) models to read images and emit structured data directly — why VL differs from OCR, the confidence + human-in-the-loop pipeline architecture, and what 'replacing 50 workers' actually means.

Have a similar problem to solve?

Contact the team