Anti-money laundering (AML) efforts are critical not just for financial stability but also for global security, as money laundering supports various criminal activities like terrorism, human trafficking, and drug trade. Fraud detection, as part of AML compliance processes, is essential for companies to protect their customers’ transactions and accounts from fraudulent activity. Fraud and Money Laundering intersect in two different areas: laundering of the proceeds of the fraud, or using fraudulent accounts for money laundering. Machine learning has shown promising results in multiple domains. However, in an AML setting, particularly in supervised learning, it faces important challenges associated with limited available labels, difficulty in detecting emerging criminal tactics, and scarce representation of suspicious cases. Despite the high volume of data available, the instances are typically unlabeled and unbalanced; the real suspicious cases (true positives) are rare, and the labeling process is costly, labor-intensive, and requires domain expertise. To address the above issues, this paper introduces ITERADE, a novel iterative anomaly detection ensemble that integrates dynamic clustering to isolate anomalous and normal cases effectively while allowing the end-user to set the desired budget of cases to be returned. In a completely unsupervised manner, ITERADE provides a strategic starting point for labeling and instances’ inspection by domain experts. This budget-agnostic solution outperforms existing alternatives while offering flexibility in adjusting labeling budgets according to end-user preferences. ITERADE can select a subset of the entire dataset within a flexible budget, achieving a higher percentage of positive cases by a factor of 3 to 15. This capability is particularly beneficial for highly imbalanced datasets, especially when working with a smaller labeling budget.