<p>To mitigate interpretability challenges in business decision-making due to the black-box nature of generative Artificial Intelligence(AI), and to address high information processing costs and inconsistent feature collection standards, a novel marketing lead evaluation framework integrating large language models (LLMs) and classical machine learning algorithms was developed. The framework encompasses three modules: (1) a multi-agent question-answering system leveraging Retrieval-Augmented Generation(RAG) and LLMs; (2) a feature extraction and memory module for precise natural language and public data processing; and (3) a logistic regression (LR) model, trained on 540,000 automotive lead records, with associated calculation logic for decision support. Results indicate that the multi-agent system accurately identifies intentions and routes modules, the feature extraction module reduces manual follow-up costs, and the LR-guided LLM output enhances interpretability. These findings highlight the framework’s potential for auditing abnormal events and advancing marketing intelligence and business systematization.</p>