I gave Tencent’s WeChat AI agent control for 24 hours: where it excelled – and stumbled

Tencent Holdings' WeChat, a dominant super-app in China with over a billion users, is now incorporating an AI agent called Xiaowei. Tencent discussed Xiaowei for the first time in its latest earnings report, emphasizing its focus on user privacy and inference efficiency. I had the opportunity to test Xiaowei for nearly 24 hours, examining how much of my daily routine an AI could manage.

The results were a mix of impressive capabilities and frustrating limitations. Xiaowei excels at tasks that involve action, such as making restaurant reservations and placing food orders through mini-programs. It can send messages, initiate audio calls, set reminders, and retrieve information from WeChat. Additionally, it offers an AI-assisted writing function for WeChat Moments, though this feature has faced criticism for potentially diminishing personal connections.

However, Xiaowei falls short in some areas. Navigating mini-program menus proved slow, making routine tasks like ordering a favorite meal faster when done manually. Voice calls to friends with similar names also resulted in failed attempts, as Xiaowei struggled to identify the correct contact. The assistant relies on WeLM, a family of AI models developed by WeChat's internal team, rather than Tencent's flagship Hy3 model.

While Xiaowei is not yet a fully autonomous assistant, Tencent's advantage lies in its ecosystem where social networking, payments, and daily services are deeply integrated.