Learning When to Quit in Sales Conversations

Abstract: Salespeople frequently face the dynamic screening decision of whether to persist in a conversation or abandon it to pursue the next lead. Yet, little is known about how these decisions are made, whether they are efficient, or how to improve them. We study these decisions in the context of high-volume outbound sales where leads are ample, but time is scarce and failure is common. We formalize the dynamic screening decision as an optimal stopping problem and develop a generative LLM-based sequential decision agent - a stopping agent - that learns whether and when to quit conversations by imitating a retrospectively-inferred optimal stopping policy. Our approach handles high-dimensional textual states, scales to LLMs, and works with both open-source and proprietary LLMs. When applied to calls from a large European telecommunications firm, our stopping agent reduces the time spent on failed calls by 54% while preserving nearly all sales; reallocating the time saved increases expected sales by up to 37%. Upon examining the linguistic cues that drive salespeople's quitting decisions, we find that they tend to overweight a few salient expressions of consumer disinterest and mispredict call failure risk, suggesting cognitive bounds on their ability to make real-time conversational decisions. Our findings highlight the potential of artificial intelligence algorithms to correct cognitively-bounded human decisions and improve salesforce efficiency.

• The paper models quitting sales calls as an optimal stopping problem using generative language models and imitation learning.
• It reports a 54% reduction in failed call time and a 37% increase in expected sales by reallocating efforts effectively.
• The methodology outperforms traditional reinforcement learning baselines and remains robust across different campaign data.

Learning When to Quit in Sales Conversations

The paper "Learning When to Quit in Sales Conversations" (2511.01181) explores the efficiency of salesforce management by modeling the decision to quit a call as an optimal stopping problem. By leveraging generative LLMs and imitation learning, the authors introduce a "stopping agent" that enables objective decision-making to enhance productivity in telemarketing scenarios, specifically by determining optimal points to discontinue a sales call if its success probability is low.

Optimal Stopping Framework

The core contribution of the paper is the formalization of the dynamic qualification decision. This decision is framed as an optimal stopping problem where the main challenge is to decide whether to continue or abort a conversation based on its probability of success. The decision framework employs a generative LLM to interpret high-dimensional textual states from call transcripts.

The agent integrates imitation learning, allowing it to mimic inferred optimal policies from historical salesperson actions. This approach addresses the inefficiencies associated with reinforcement learning methods in high-dimensional state spaces, specifically unstable optimization and hyperparameter sensitivity. Instead, by reformulating the problem as supervised learning, the stopping agent reaps the advantages of pre-trained LLMs.

Empirical Evaluation

The stopping agent was applied to a real-world dataset comprising 11,627 outbound calls from a European telecommunications firm. Notable findings include:

• A 54% reduction in time spent on failed calls.
• Up to 37% increase in expected sales through the reallocation of saved time.

Figure 1: Total time spent (left) and total number of sales made (right) by each salesperson, both with (y-axis) and without (x-axis) the stopping agent with T=2 decision opportunities.

A significant challenge salespeople face is cognitive bias, where certain phrases such as "no me interesa" ("I'm not interested") overly influence their decision to quit, leading to time inefficiencies.

Linguistic Analysis

The study utilized extensive linguistic feature analysis to compare the reaction of salespeople to conversation cues against the AI-driven stopping agent. While salespeople often followed a narrow set of linguistic triggers, the stopping agent demonstrated a more nuanced and evolving approach to decision-making, adapting its criteria as the conversation progressed (Figure 2).

Figure 2: Top unigrams and bigrams (by logit coefficients) by decision-maker and decision opportunity.

Benchmarking and Robustness

The stopping agent's effectiveness was evaluated against baseline models, including simple decision rules and reinforcement learning approaches. Notably, the agent outperformed the baselines that ignored call content and time costs (Figure 3).

Figure 3: Expected number of sales by the stopping agent vs. several content and cost-agnostic baselines.

Furthermore, testing the agent without retraining on a subsequent campaign demonstrated its robustness and adaptability to distribution shifts, achieving a 16% increase in expected sales.

Implications and Future Work

The findings have significant implications for salesforce management, demonstrating how AI can augment human decision-making by complementing human cognitive limitations. The proposed framework can be adapted to various domains beyond sales, such as customer support and consultation services.

Future research directions include extending the action space of stopping agents to incorporate proactive engagement strategies during calls and exploring the integration of multimodal signals such as voice tone and sentiment analysis to further refine decision-making processes.

Conclusion

This study illustrates the potential of leveraging AI-driven strategies to optimize conversation management in sales, highlighting how imitation learning and LLMs can transform human-AI collaboration in decision-making processes. The application of stopping agents reveals the promise of AI in enhancing operational efficiency, reducing time wastage, and increasing productivity in high-volume sales environments.