Telecom billing systems process vast volumes of financial transactions daily, making them highly vulnerable to anomalies that can cause revenue loss and compliance risks. These systems are susceptible to a range of anomalies such as overcharges, duplicated entries, missed charges, and unauthorized usage, which can result in substantial revenue loss and erode consumer trust. Traditional supervised learning methods require extensive labeled datasets, which are often unavailable or expensive to produce in the telecom domain due to the rarity and class imbalance of real anomalies. In this paper, we propose a novel anomaly detection framework based on Self-Supervised Learning (SSL), which eliminates the need for labeled anomalies. Our approach combines contrastive learning for latent representation and autoencoder-based reconstruction error to detect outliers. We apply our model (code and data not publicly released at this time) to both synthetic and real-world telecom billing datasets, achieving superior performance compared to baseline models. The real-world dataset spans 18 months of anonymized telecom billing records from over 150,000 users, enabling robust validation of the proposed framework. Furthermore, we integrate SHAP-based explanations to ensure interpretability, which is crucial for operational deployment in billing systems. This method reduces false positives by 28% and demonstrates strong generalizability and operational readiness, offering a practical solution to anomaly detection in large-scale billing systems