How Bias Enters
Training data reflects historical patterns. If historical conversions favored certain demographics or regions, the model learns that preference. Scoring new leads perpetuates the pattern.
Detection
Disparate impact analysis across protected attributes. Statistical tests for proportional outcomes. Monitoring scores across segments continuously. Early detection beats post-deployment discovery.
Mitigation
Remove protected attributes from features. Audit proxy variables (zip code, first name). Retrain with balanced data. Fairness-aware ML techniques. Human review of high-stakes decisions.
EU AI Act
Lead scoring affecting access to essential services is Annex III. Bias audit required before deployment. Documentation of mitigation. Enforcement starts August 2, 2026. Assess your scoring models now.