AI Agents Are Learning to Predict What Users Want—Before They Ask for It
Researchers at Shanghai Jiao Tong University and Tencent developed ProAct, an AI agent that uses downtime between conversations to anticipate and prepare answers for future user questions. Unlike typical reactive AI systems, ProAct analyzes past conversations and user data during idle moments, predicts likely follow-up queries, and gathers relevant information in advance. The system consists of three stages: predicting future user needs, prioritizing which predictions deserve further research, and deciding how to store or deliver the prepared information. Testing ProAct across 200 simulations in 40 different domains, researchers found it reduced conversation turns by 14.8%, follow-up requests by 11.7%, and hallucinations by 28.1% compared to prior systems. ProAct anticipated user needs far more frequently than older agents. However, limitations included occasional irrelevant responses and concerns over privacy, since continuous monitoring and data storage are required. The researchers emphasized proactive processing should be balanced, as excessive background activity increases computational cost with diminishing benefits.
