Agente inteligente para qualificação de leads no setor automotivo utilizando machine learning

Abstract

This Final Year Project details the development, implementation, and comparison of Machine Learning models for intelligent lead- potential customers who have demonstrated initial interest through digital channels- qualification at a luxury car dealership, facilitated by an Artificial Intelligence agent named “Betina”. Motivated by the need to optimize customer service amidst growing digital demand and the consequent overload of the sales team, the solution evolved from a Minimum Viable Product (MVP) on a no-code platform (Zaia) to a robust architecture integrating Make.com, ChatGPT-4o, and a custom ML API. The core of this study lies in the comparison of two predictive approaches trained on a dataset of 72 historical conversations: (1) a Random Forest classifier utilizing TF IDF and behavioral features, interpreted via a surrogate model; and (2) the fine-tuning of Transformer models (BERT and mBERT). The methodology encompassed advanced preprocessing, including speaker separation and data anonymization. While the fine-tuning of BERT demonstrated potential for high accuracy, computational constraints limited its viability in this context. Conversely, the Random Forest model proved to be a robust and interpretable solution, yielding actionable insights. The project validates the system’s effectiveness through empirical case analysis and the generation of strategic insights for the sales team regarding qualified leads, demonstrating the practical application and challenges of deploying AI and ML techniques to optimize business processes.