Learning When to Quit in Sales Conversations

TL;DR

This paper reframes the dynamic qualification decision in high-volume outbound sales as an optimal stopping problem, introducing a stopping agent that uses evolving transcripts to decide when to quit. It solves the high-dimensional state challenge via imitation learning: inferring an expert-like quitting policy from historical conversations and fine-tuning a large language model to generate actions conditioned on transcript state, with a backward-induction threshold mechanism to convert probabilities into deterministic decisions. Empirically, the approach yields substantial efficiency gains in real campaigns, reducing time spent on failed calls and increasing expected sales (e.g., up to 37% improvement with a 60/90s horizon, and durable gains in out-of-sample campaigns), while outperforming simple heuristics and RL baselines. The work also uncovers cognitive constraints in human quitting decisions, such as over-reliance on the salient phrase "no me interesa" and under-weighting early failure indicators, illustrating how AI decision-support can boost frontline efficiency and inform managerial practice.

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 language model-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 large language models, and works with both open-source and proprietary language models. 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.

Figures (15)

• Figure 1: We transform conversations $\mathcal{C}$ into "expert" demonstrations $\mathcal{D}$ (Section \ref{['sec:expert_demonstrations']}), and train an imitation learning policy by fine-tuning an LLM $\pi_\theta$ to generate the "expert's" action $a_t$ for each state $s_t$ wrapped in a prompt (Section \ref{['sec:finetuning']}). We threshold $\pi_\theta(a|s)$ to obtain deterministic actions (Section \ref{['sec:threshold_tuning']}).
• Figure 2: Empirically assessing (a) how well the predicted failure risk correlates with the actual failure rate, and (b) whether and how salespeople react to early indicators of eventual call failure.
• Figure 3: Total time spent (left) and total number of sales made (right) by each salesperson, both with ($y$-axis) and without ($x$-axis) our stopping agent with $T=2$ decision opportunities, at $t=60$ and at $t=90$.
• Figure 4: Expected number of sales by our stopping agent with $T=2$ decision opportunities at $t\in\{60,90\}$.
• Figure 5: Expected number of sales by our stopping agent with $T=2$ and $T=3$ decision opportunities.