Effective contract management is critical to organizational resilience, yet overlooked or ambiguous clauses often expose firms to significant legal and financial risks. Traditional manual reviews struggle with the scale and complexity of modern agreements, being slow, costly, and error prone. Recent advances in transformer-based natural language processing (NLP) offer powerful opportunities for automating clause-level risk assessment. This study evaluates BERT and RoBERTa for multi-class classification of contractual risk severity using the Contract Understanding Atticus Dataset (CUAD). Clauses were mapped to Low, Medium, and High-risk levels, preprocessed, and stratified into training and validation sets. Both models achieved strong results, with BERT reaching 82.3\% accuracy and a macro-F1 of $0.795$, while RoBERTa slightly outperformed it with 82.8\% accuracy and a macro-F1 of 0.805. RoBERTa demonstrated greater robustness in distinguishing minority classes, improving classification of Low- and High-risk clauses. Misclassifications primarily occurred between Medium and High-risk levels, reflecting overlapping contractual language. To illustrate practical application, a prototype tool was developed to process text or PDF contracts and generate clause-level risk scores. Although ERP integration was beyond the current scope, this research demonstrates the feasibility of embedding real-time clause risk alerts into enterprise workflows. By validating transformer-based models for contract risk classification, the study contributes a replicable framework and highlights directions for future work, including domain-specific pretraining, multilingual datasets, and seamless organisational integration.