Stopping Agent: Learning to Cut Losses in Conversations with Large Language Models

This article introduces an optimal stopping agent based on large language models, aiming to solve the dilemma of "continue following up or give up" in sales dialogues and improve efficiency through real-time judgment. This agent combines classic optimal stopping theory with modern large language models, which can correct human cognitive biases and is applicable to various decision-making scenarios.

Time is precious in sales, and salespeople often face the dilemma of "wasting time on low-interest customers by continuing to follow up, or missing opportunities by giving up too early". The optimal stopping problem is a classic problem in operations research; however, traditional mathematical models are difficult to handle sales scenarios due to their dynamic dialogue and high-dimensional text information.

The stopping agent is a decision-making agent designed for large language models, which learns directly from dialogue text without manual feature/rule definition. Technically, it is based on an imitation learning framework: the model is trained by inferring the optimal stopping strategy after the fact, predicting the profit difference between continuing and exiting based on dialogue history, and suggesting exit when the difference is negative, thus avoiding the exploration-exploitation dilemma of reinforcement learning.

Applied to outbound call data from a large European telecommunications company, the results show: failed call time reduced by 54%, sales revenue almost fully retained, and overall sales efficiency increased by 37%. It was also found that human salespeople have cognitive biases, overvaluing obvious rejection signals while ignoring subtle interest indicators.

The stopping agent can data-drivingly correct human cognitive biases and provide more accurate real-time decision-making suggestions. In addition to sales, it is also applicable to scenarios such as customer service (escalation to human agents), online education (student intervention), medical consultation (urgency assessment), and recruitment interviews (candidate matching judgment).

The stopping agent, which combines classic theory with large language models, is a technological innovation that reveals the potential of AI to enhance human decision-making (not replace, but assist in correcting biases). In the future, more specialized agents are expected to emerge, forming an efficient AI-assisted decision-making ecosystem.