A bank's customer-service window stays open at 3 a.m. The clerk never blinks, never tires, and answers in flawless Mandarin and English. She is not an employee on a night shift — she is a digital human (数字人), and until recently no one could tell you what "good" even meant for one.
That changed on October 5, 2025, when China's first national standard for virtual digital humans took effect. GB/T 46483-2025, "General Technical Requirements for Customer Service-oriented Virtual Digital Humans (信息技术 客服型虚拟数字人通用技术要求)," was approved and released by the standardization administrator and led in drafting by SenseTime (商汤科技). For the first time, a customer-service digital human had measurable bars to clear.
What the standard actually requires
The standard is not poetry; it is a checklist. It sets a lip-sync accuracy of no less than 90%, a gesture-interaction success rate of at least 80%, a voice-interaction success rate of at least 85%, and an emotional-interaction success rate of at least 80%. It also defines reference architecture, visual and voice interaction, image driving and operations-maintenance functions for both 2D and 3D digital humans.
Those numbers matter because, before them, vendors described their avatars with adjectives — "realistic," "natural," "smart" — that meant nothing across companies. A benchmark turns marketing into measurement. SenseTime, whose SenseAvatar product complied early, became the first to pass the standard's conformity test, but the point of the document is industry-wide: every player now knows the floor.
A market that suddenly looks like infrastructure
The timing is no accident. By 2025 the digital human (数字人) had moved out of the novelty section and into daily operations. Research firm iiMedia (艾媒咨询) estimated China's virtual digital human core market at about RMB 48.1 billion (≈US$6.7 billion) in 2025, with a surrounding industry near RMB 640 billion (≈US$89 billion), and projects the core market to roughly double by 2030. A narrower CIC (灼识咨询) definition — revenue from digital-human agent solutions — puts 2024 at RMB 2.0 billion (≈US$278 million), growing from just RMB 0.5 billion in 2021.
The most visible battlefield is e-commerce livestreaming. AI digital human hosts now reach beyond 60% penetration in the sector overall, and above 70% in standardized categories like consumer electronics and daily goods. JD.com reported that digital-human livestreaming across its platform exceeded RMB 2.3 billion (≈US$319 million) in gross merchandise value during the 2025 Double 11 festival, at a service cost the retailer described as roughly one-tenth that of human hosts.
The 32.2% leader goes public
The company that best embodies the shift is Nanjing-based Silicon Intelligence (硅基智能). According to CIC data cited in its Hong Kong IPO prospectus, the firm ranked first among Chinese digital-human agent providers by 2024 revenue, holding a 32.2% market share. By October 24, 2025, it had delivered more than 80,000 "silicon labor (硅基劳动力)" units — digital employees spanning voice, video, livestreaming and interactive agents — to clients in telecom, finance, healthcare, education and public services.
The numbers in the filing show a scaling business with thin margins. Revenue reached RMB 655 million (≈US$91 million) in 2024, up from RMB 530 million in 2023, while gross margin sat in the low-to-mid 30% range and the company posted adjusted profit only in the first half of 2025. Silicon Intelligence filed for a Hong Kong listing on October 31, 2025, with backers including Tencent. The story it tells investors is not "cool avatars" but "rent-a-workforce": a digital human that answers calls, makes sales pitches and hosts livestreams around the clock.
The compliance wave is just starting
A maturing industry attracts rules. Beyond the October 2025 service standard, regulators in early 2026 moved toward mandatory identity labeling for commercial digital humans — a "one identity per avatar" approach intended to stop impersonation and unauthorized use of a real person's likeness. For platforms and brands, that means a digital human (数字人) used in advertising or customer service will increasingly need to be visibly and traceably marked, not slipped into a feed indistinguishable from a human.
This closes a gap the standard did not: GB/T 46483-2025 says how well a digital human must perform, but not how it must be identified. The two together — performance floor plus identity label — are what turn a marketing toy into regulated infrastructure.
Honest limitations
This article draws on the official national-standards platform for the standard's existence and metrics, SenseTime's announcements for conformity testing, and news reports citing Silicon Intelligence's prospectus for market share, delivery counts and financials; I have not reviewed the filings directly. Market-size figures come from iiMedia and CIC, whose definitions differ — iiMedia's "virtual digital human" is far broader than CIC's "digital-human agent revenue" — so the two should not be read as measuring the same thing. Penetration and JD GMV numbers are vendor- or report-cited, not independently audited. The piece does not cover content-labeling (内容标识) enforcement mechanics in detail, nor how small creators will absorb compliance cost, nor the competitive posture of large-model (大模型) clouds such as Baidu, Alibaba and ByteDance in this space.
What readers can do now
- If you deploy a customer-service or sales digital human (数字人) in China, audit it against the GB/T 46483-2025 thresholds — lip-sync, gesture, voice and emotional interaction rates — before your next procurement review.
- Brands using avatars in ads or livestreams should plan now for mandatory identity labeling; design the on-screen mark into the creative rather than bolting it on.
- Buyers should separate "broad virtual human hype" from the narrower, prospectus-backed "digital-human agent" market when evaluating vendors' share claims.
