Most people’s first impression of AI customer service is poor. The parcel never arrived, yet the bot loops ‘delivered.’ The question is half-asked before the system jumps topics and answers the wrong thing. Reach a human at last, and you repeat the whole story from scratch. That fatigue is universal – and it has been quietly eroding trust. Yet enterprises cannot walk away: customer service is the most direct bridge between a company and its users, covering pre-sale, after-sale and complaints, and the work is repetitive, fragmented and cheap to automate.
The frustration is measurable. China’s market regulator disclosed that in 2024, e-commerce after-sales complaints mentioning ‘smart customer service’ reached 6,969 cases, up 56.3% year on year, with users citing off-topic answers, unreachable humans and low efficiency. So the bots are both heavily deployed and heavily disliked.
The technology has evolved in three waves. Early systems matched keywords to fixed scripts – the ‘artificial idiocy’ era. Next came semantic recognition, intent classification and history-based training, less rigid but expensive to retrain and still stumped by unseen phrasing. The large-model era changed the calculus: with knowledge bases, retrieval-augmented generation and prompt design, bots can finally ‘learn on the fly’ and handle expressions they were never explicitly trained on.
The market is heating up. IDC data shows the top five smart-customer-service vendors held 35% share in 2024 – Alibaba Cloud at 11.4%, Baidu AI Cloud at 10.4%, alongside Ronglian Qimo, Zhongguancun Kejin and Zhichi. The giants bundle the capability into cloud, model and agent platforms, but the sector is too fragmented – every industry has different rules, knowledge bases and escalation standards – for them to fully control it, leaving room for vertical players and AI-voice startups.
The defining shift is agentisation. OpenAI launched Presence for enterprise voice support in July. Sierra, the bellwether US startup, raised $950 million in May at a valuation above $15 billion and acquired a long-horizon-agent company, signalling that AI customer service is no longer a Q&A bot but a front-end agent linking users, orders, tickets, knowledge bases and human agents. The yardstick has moved from ‘can it answer’ to ‘can it resolve.’
In standard scenarios – order status, logistics, account info, booking changes, policy explanation – the improvement is clear: identify intent, call the right data, close the ticket. China’s leading systems now deliver millisecond real-time voice. China Telecom’s TeleAI says its GOAT-SLM model catches not just the words but the subtext – sensing anxiety in a caller’s tone and responding with matching empathy. In complex after-sales – refunds, complaints, liability, fund safety, medical faults – AI still cannot judge independently, and regulators such as Zhejiang’s consumer council now advise routing routine queries to bots and escalating the messy, urgent ones to people.
The louder complaint is often not that the AI is dumb but that the human is unreachable. Engineering helps: NetEase’s Qiyu and WoFun auto-extract key facts, summarise and pre-fill tickets, so users no longer repeat themselves to a human. The next competitive frontier is entering those complex service flows, not answering more standard questions.
Customer service is a decades-old, unfashionable market – which is precisely its value. Demand does not vanish with the economic cycle, making it real, verifiable and unconstrained by education cost. Most Chinese vendors call off-the-shelf models – Baidu ERNIE, Alibaba Tongyi, ByteDance Doubao, DeepSeek – and invest in the application layer rather than training their own. The business models are still settling: SaaS subscriptions versus project fees, with Sierra even experimenting with ‘pay for results,’ charging only when an issue is actually resolved. No one knows the endpoint, but the road itself is already reshaping how people work.
Source (Chinese original): Sohu IT
Translated and adapted from Sohu IT (it.sohu.com).