Journal of African Development

Public procurement digitalization multiplies the number of transactions that can be seen, but at the same time makes the risks of fraud even bigger, faster, and more complex for control bodies to assess. This study proposes an AI-based digital procurement fraud detection system for public sector organisations, and demonstrates analytical logic using a reproducible set of benchmark data, not by 'claimed' field data. The architecture incorporates the procurement red flags, supervised classification, anomaly detection, explainability, human audit triage and governance controls. A data set comprising 20,000 procurement transactions was created based on risk patterns identified in the literature and subsequently split into training, validation and held out test partitions. The following six models were tested: logistic regression, random forest, XGBoost, isolation forest, and a weighted ensemble were compared based on precision, recall, F1, ROC-AUC, PR-AUC, calibration, and interpretability diagnostics. The performance of the supervised models was significantly superior to that of the unsupervised models for anomaly detection, and logistic regression was the best discriminator in the benchmark. The following factors were identified as risk indicators that are significant in relation to the restrictions in the procedure, single bidding, concentration of suppliers, shortened advertisement periods and changes to the contract. The results validate a risk ranking architecture, where AI can guide review without leaving the possibility for determining guilt, maintain explainability, due process, auditability, and human accountability in operational procurement oversight and responsible public administration....