Dernière mise à jour :
A quantifiable metric (e.g., model accuracy drift, bias remediation time) used to monitor and report on AI governance and compliance objectives.
A specific, measurable value that reflects how effectively an organization or AI project meets governance goals - such as overall model accuracy stability over time, average time to detect and fix bias issues, or percentage of models with current impact assessments. Good KPIs are tied to strategic objectives, have clear thresholds, and are tracked via automated dashboards to prompt timely interventions when metrics stray outside acceptable ranges.
Exemple concret :
A retail bank tracks the KPI “monthly model-drift percentage,” which measures the drop in predictive accuracy of their credit-risk model compared to its last calibration. When drift exceeds 3%, an automated alert triggers a model-retraining pipeline, ensuring credit decisions remain reliable and compliant with internal accuracy standards.
