ML model evaluating consumer creditworthiness for loan origination decisions. Processes 15K applications/month across retail and SME lending.
Yes — risk management framework established Q4 2025, reviewed quarterly by the AI governance committee. Covers age, gender, ethnicity, and disability bias vectors.
EU AI ActFundamental rights impact assessment completed. Key risks: financial exclusion, discriminatory outcomes in loan pricing. Mitigations include fairness constraints and borrower appeals process.
EU AI ActYes — 5 years of lending data documented. Demographic breakdowns provided across age, income, geography, and ethnicity cohorts.
NIST AI RMFYes — monthly disparate impact ratio tests run for all protected classes. Current ratios within 80% four-fifths rule threshold.
Colorado AI ActYes — adverse action notices include top contributing factors via SHAP values, compliant with ECOA requirements.
EU AI ActYes — all automated denials are reviewable by senior loan officers. Override capability with audit trail is built in.
EU AI ActAUC-ROC: 0.87, KS stat: 0.42 — monitored daily. Fairness metrics (equalized odds, calibration) tracked weekly.
EU AI ActYes — all decision logs, model artifacts, and feature inputs retained for 5 years per EU AI Act Art. 12 and ECOA requirements.
EU AI Act3 gaps identified
Bias monitoring for protected characteristics (ethnicity, disability) not yet fully automated in production scoring pipeline.
Recommendation: Deploy statistical parity and equalized odds monitoring with automated alerting across all protected classes.
Annual algorithmic impact assessment not yet scheduled for renewal cycle.
Recommendation: Schedule annual renewal and assign compliance officer as owner.
Intended use documentation does not cover all deployment contexts (SME vs. retail lending).
Recommendation: Update intended use documentation to differentiate retail and SME lending contexts with separate risk profiles.