In the digital banking era, effective marketing lead generation depends on leveraging hetero-geneous, multimodal customer data. Traditional predictive models primarily rely on tabular attributes, overlooking the relational and contextual information inherent in customer networks. This paper proposes a Graph Neural Network (GNN)-based framework that integrates multi-modal data—including structured CRM attributes, transactional records, and unstructured call transcript text—to predict customer lead conversion in banking. The proposed system models customers as nodes in a heterogeneous graph with relationships based on transactional similarity and communication patterns. Using a multimodal embedding strategy, the model learns customer representations via Graph Convolutional and Attention layers. Empirical results on the UCI Bank Marketing dataset demonstrate an ROC-AUC of 0.87 and accuracy of 0.886, with significant improvements over logistic regression and XGBoost baselines. Extended experiments using a heterogeneous multi-source graph (MovieLens, Last.FM, Amazon co-purchase, OGB-MAG) further confirm the framework’s superiority: accuracy 0.893 and F1-score 0.596 versus a logistic regression baseline that degenerates to F1= 0.000, AUC= 0.500. The paper details system design, dataset structure, implementation, graph construction methodology, and performance evaluation.