Planned actions (e.g., bias remediation, retraining, feature re-engineering) to address identified AI risks and compliance gaps.
A catalogue of interventions mapped to risk types: data-drift triggers → retraining pipelines; fairness violations → bias-mitigation algorithms; security threats → input-sanitization layers; performance degradation → architecture tuning. Governance entails selecting appropriate strategies, documenting implementation, and verifying their effectiveness through follow-up assessments and tests.
After a bias audit finds that loan approvals favor one demographic, the team applies re-weighting of training samples and adds fairness constraints. They then conduct post-mitigation testing to confirm that approval rates now meet the bank’s equity thresholds.
