This thesis examines the integration of Support Vector Machine (SVM) into Accounting Information Systems (AIS) for enhanced fraud detection and predictive risk assessment. Fraud remains a persistent organizational threat because fraudulent transactions are often concealed, rare, and difficult to distinguish from legitimate transactions using traditional rule based controls. The study adopts a quantitative experimental design using the IEEE CIS Fraud Detection Dataset, a publicly available dataset containing transaction and identity files joined by TransactionID and a binary fraud target variable, isFraud. The proposed methodology includes data preprocessing, feature engineering, feature selection, class imbalance treatment, SVM model development, and comparative evaluation against Logistic Regression and Random Forest. The system design proposes an AIS architecture with a fraud detection module, risk assessment module, SVM integration layer, dashboard interface, audit review function, and deployment architecture. Because the raw dataset files were not provided in the working environment, the compiled results section presents a reproducible evaluation framework, model comparison table templates, and academic interpretation guidance rather than fabricated numerical results. The study concludes that machine learning can strengthen AIS based fraud monitoring by supporting transaction level classification, fraud alert generation, and risk based audit prioritization. The thesis recommends empirical implementation using the approved dataset, validation with fraud sensitive metrics, and deployment only with model governance, audit trails, and periodic retraining.