Efficient Fraud Management with Agentic AI

Situation

A leading global bank was struggling to stay ahead of rising fraud risks. Their existing third-party, machine learning model required heavy manual intervention and couldn’t adapt quickly to new or emerging fraud patterns. The bank needed a smarter, faster, and more adaptive way to detect fraud in real time.

Challenge

The client’s current fraud detection system relied on 31 pre-defined features, many of which were outdated or irrelevant to evolving fraud tactics. Key challenges included:

• High false positive rates, causing customer friction and operational inefficiency.
• Manual feature engineering, which consumed analysts’ time and delayed response to new fraud types.
• Limited adaptability, as the model couldn’t learn from emerging data without human supervision.
• The need to ensure fair, unbiased detection across customer segments while maintaining compliance accuracy.

The goal was to automate feature engineering, reduce false positives, and enhance the overall precision and speed of fraud detection.

Our Approach

Using Mu Sigma’s Decision Science framework and Agentic AI, we developed a new “Challenger Model” to outperform the existing “Champion Model.”

• Automated Feature Engineering: Leveraged Agentic AI through muTalos to derive and select new, high-impact features from raw data, reducing manual effort and bias.
• Enhanced Model Architecture: Combined Random Forest and Isolation Forest algorithms to improve detection of anomalous behavior.
• Continuous Learning: Embedded Continuous Service as a Software, enabling adaptive learning to identify emerging fraud patterns proactively.
• Bias Mitigation: Applied model evaluation to ensure consistent, unbiased fraud detection across all customer segments.

Out of hundreds of candidate features, the top 25 were retained, 17 newly derived, and 4 enhanced from the original model, creating a more refined and intelligent detection system.

Impact

The enhanced “Challenger Model” transformed the bank’s fraud detection capabilities:

• 21% improvement in fraud detection accuracy through smarter feature selection.
• 27-30% increase in fraud recovery rates.
• 20% reduction in financial losses due to faster, earlier fraud identification.
• Significant drop in false positives, improving customer experience.
• Less manual intervention, freeing analysts to focus on complex cases and strategic oversight.