AI Agents Are Learning to Predict What Users Want – Before They Ask

Researchers at Shanghai Jiao Tong University and Chinese technology giant Tencent are said to have carried out the study I have an assistant which uses the quiet time between conversations to predict what users will ask next – and prepare answers before they ask.

The system, called ProActit works differently than most AI assistants that wait for users to ask a question before answering. Instead, ProAct uses the pause between messages to review previous conversations and stored user information, then prepares useful information in the background before the next query.

“Although AI agents show remarkable abilities in reasoning and using tools, they remain passive: They calculate responses only when they are clearly informed,” the researchers wrote. “This paradigm ignores a critical opportunity: Idle time between conversations is wasted, which leaves agents less able to plan for future needs.”

This system works in several stages. The first, called Future-State Prediction, predicts future questions by analyzing past conversations, preferences, and missing items.

The second phase, called Idle-Time Acquisition, selects which predictions to investigate based on the importance, timing, and usefulness of the new information.

A separate system decides whether to deliver scheduled data, store it for later, or store it until needed, creating a “loop-loop” system designed to anticipate and respond to user needs.

“After each interaction, the agent updates its memory, predicts what will happen in the future, assigns people who need to take action, and decides how those preparations should be carried out,” he wrote. “This design integrates prediction, discovery, and delivery into a single point, instead of taking up idle time like an endless search.”

According to the researchers, ProAct was tested in 200 simulations in 40 areas, including financial planning, software release management, and cybersecurity. According to the paper, the system reduced conversations by 14.8% and cut follow-up requests by 11.7%. When compared using a benchmark called ProActEval, ProAct anticipated 703 user identification needs versus 32 for the older system. The researchers reported a 28.1% decrease in mood.

The research comes as independent AI agents spread across the tech industry, with projects like OpenClaw and Helper of Hermes Providing continuous AI assistants that can perform long-term, autonomous tasks—such as copying, editing, searching, and automated task management—without direct human input.

The research also comes as a separate researcher earlier this month warned that AI assistants could be completed dangerous work without understanding the consequences.

“Like Mr. Magoo, these agents move toward a goal without fully understanding the consequences of their actions,” lead author Erfan Shayegani, a UC Riverside student, said in a statement. “These agents can be very helpful, but we need protection because sometimes they can prioritize achieving their goal and understanding the big picture.”

The researchers admitted that the ProAct study had several limitations, including that in 3% of cases, the system made the answers worse by bringing in useless information. The paper also said that any kind of real world would need privacy protection, because the system constantly analyzes the conversations and stores the user’s data.

“Our budget analysis also shows that large amounts of Idle-Time Acquisition raise fixed costs and lower returns,” they wrote, “thus accounting for idle-time acquisition rather than leverage.”