Tencent's WeChat is the undisputed super-app of China — over a billion people use it for everything from splitting dinner bills to hailing taxis. Now, Tencent is embedding an AI agent named Xiaowei directly into the platform, adding a hands-free automated assistant to its vast digital ecosystem. After gaining early access to the trial feature, one SCMP journalist spent nearly 24 hours letting Xiaowei run their digital routine. The verdict? A fascinating, promising glimpse into the future of apps — and an occasional exercise in frustration.
Beyond Chatting: An Agent That Actually Works
Unlike standard chatbots that simply answer questions, an AI agent is designed to take action. Where a chatbot recommends a hotpot restaurant, Xiaowei is meant to make the reservation and place your order, too. This is where Tencent holds a major home-field advantage. While rivals like Alibaba Group's Qwen AI app must build connections with external services from scratch, Xiaowei sits directly inside an ecosystem where messaging, payments and millions of mini-programs have already coexisted for years.
In testing, Xiaowei handled some tasks smoothly. It successfully helped send messages to specific recipients, initiated audio calls, set up reminders and pulled information from across WeChat. When asked to order food through mini-programs, Xiaowei navigated the multi-step process independently, leaving the user with only the final tap to confirm and pay.
Slow Speeds and Misheard Names
Despite those bright spots, Xiaowei is not quite ready to replace thumbs. Navigating through mini-program menus took long enough that, for routine tasks like ordering a favourite meal, doing it manually was still faster. Trying to place a WeChat voice call to a friend with a phonetically similar name turned into a loop of failed attempts, as Xiaowei struggled to pick the right contact from the list despite multiple nudges.
Xiaowei also offers an "AI-assisted writing" function for WeChat Moments mini-blog, but automated social posts have already drawn backlash from early users who worry that generative text dilutes real human connection. The tension between convenience and authenticity is playing out in real time.
Powered by WeLM, Not Hy3
For the record, Xiaowei is supported by WeLM, an in-house AI model developed by the WeChat team, rather than Tencent's flagship Hy3 foundational model. According to a social-media post by the WeChat team that unveiled WeLM, it is a family of AI models with two different parameter sizes: 80 billion and 617 billion. This separate model lineage suggests Tencent is treating super-app AI as a distinct discipline from general-purpose LLMs — one optimized for task execution within a closed ecosystem rather than open-ended reasoning.
The Bottom Line
Xiaowei is far from a fully autonomous assistant. It remains constrained by the familiar hurdles facing consumer AI agents: sluggish execution, errors in precision and the need for human supervision. Yet, Tencent's real edge may not be Xiaowei's current intelligence.
Very few AI developers control a digital space where social networking, payments and daily services are already so deeply intertwined. Xiaowei may be clunky today, but it is standing on the ultimate launch pad. While Western companies build AI agents that need to integrate with a fragmented landscape of apps and services, Tencent already owns the landscape. The question isn't whether Xiaowei will improve — it's how quickly it will, and what that means for the rest of the world's super-app ambitions.
🔥 Hot Takes
1. Tencent is playing 4D chess while OpenAI is playing checkers. Everyone's obsessed with making AI smarter, but Tencent's play is different: make AI useful by embedding it where people already live digitally. WeLM doesn't need to pass the Turing test — it just needs to order your bubble tea without screwing up. That's a far more defensible moat than raw intelligence.
2. The "AI dilutes human connection" backlash is the canary in the coal mine. WeChat Moments users already pushing back on AI-generated posts signals a cultural fault line. In China, where guanxi (relationships) is everything, automating social interaction isn't just inconvenient — it's suspect. Tencent will have to walk a tightrope between utility and authenticity, and the West will be watching closely.
3. Xiaowei's failures are Tencent's roadmap. Slow navigation, misheard names, the need for human supervision — these aren't bugs, they're a feature list. Each stumble maps directly to a engineering sprint. The real story isn't what Xiaowei can do today; it's that Tencent has a billion users feeding it interaction data every day, and every error is training signal. By the time the West figures out how to build useful agents, Tencent will have billions of real-world corrections baked in.