Researchers from Shanghai Jiao Tong University and Tencent have developed ProAct, an AI agent that uses idle time between conversations to predict user needs and prepare answers in advance. Unlike traditional reactive AI agents that wait for explicit prompts, ProAct analyzes past conversations, user preferences, and missing information to anticipate follow-up questions. It operates in stages: Future-State Prediction forecasts likely queries, Idle-Time Acquisition determines which predictions are worth researching, and a delivery system decides how to present the prepared information. In simulations across 40 domains including finance, software management, and cybersecurity, ProAct reduced conversation turns by 14.8%, cut follow-up requests by 11.7%, and decreased hallucinations by 28.1%. On the ProActEval benchmark, it anticipated 703 predictable needs versus 32 for a prior system. However, the system had limitations: in 3% of cases, it introduced irrelevant information, and privacy concerns arise as the system constantly analyzes conversations. The research highlights the potential of proactive AI agents but notes trade-offs between computational cost and performance.

Key facts

• ProAct uses idle time to predict user needs and prepare answers proactively.
• Reduced conversation turns by 14.8% and follow-up requests by 11.7% in simulations.
• Anticipated 703 predictable needs vs. 32 for earlier system on ProActEval benchmark.
• In 3% of cases, ProAct made responses worse by introducing irrelevant information.
• Privacy and cost trade-offs noted; real-world deployment requires protections.

KeyAudit data perspective