AI-Salesman: Towards Reliable Large Language Model Driven Telemarketing
Qingyu Zhang, Chunlei Xin, Xuanang Chen, Yaojie Lu, Hongyu Lin, Xianpei Han, Le Sun, Qing Ye, Qianlong Xie, Xingxing Wang
2025-11-18
Summary
This paper focuses on building an AI that can have persuasive conversations, like a telemarketer, while staying truthful and planning ahead. It tackles the difficulty of creating these kinds of AI systems, which current large language models struggle with.
What's the problem?
Creating an AI that can successfully persuade someone over multiple turns of conversation is really hard. Existing AI models often lack the specific training data needed for tasks like sales, and they can be easily thrown off course or even make up false information. Basically, they aren't very good at strategically planning a conversation and sticking to the facts.
What's the solution?
The researchers created a new dataset called TeleSalesCorpus, which contains real-world sales dialogues. They then built a system called AI-Salesman, which works in two steps. First, it learns effective sales strategies from the dataset using a special learning method that can handle imperfect data. Second, it uses a 'dynamic outline' – a sort of script library – to guide the conversation turn-by-turn, ensuring it stays on track and uses good sales techniques. They also developed a way to evaluate the AI's performance using both automated tests and human feedback.
Why it matters?
This work is important because it moves us closer to building AI assistants that can handle complex, real-world tasks requiring persuasion and factual accuracy. A successful AI-Salesman could have applications beyond just sales, like negotiation, customer service, or even providing helpful advice, and it addresses a key limitation of current AI technology.
Abstract
Goal-driven persuasive dialogue, exemplified by applications like telemarketing, requires sophisticated multi-turn planning and strict factual faithfulness, which remains a significant challenge for even state-of-the-art Large Language Models (LLMs). A lack of task-specific data often limits previous works, and direct LLM application suffers from strategic brittleness and factual hallucination. In this paper, we first construct and release TeleSalesCorpus, the first real-world-grounded dialogue dataset for this domain. We then propose AI-Salesman, a novel framework featuring a dual-stage architecture. For the training stage, we design a Bayesian-supervised reinforcement learning algorithm that learns robust sales strategies from noisy dialogues. For the inference stage, we introduce the Dynamic Outline-Guided Agent (DOGA), which leverages a pre-built script library to provide dynamic, turn-by-turn strategic guidance. Moreover, we design a comprehensive evaluation framework that combines fine-grained metrics for key sales skills with the LLM-as-a-Judge paradigm. Experimental results demonstrate that our proposed AI-Salesman significantly outperforms baseline models in both automatic metrics and comprehensive human evaluations, showcasing its effectiveness in complex persuasive scenarios.