Model Risk Manager AI

Invoked when: a model version changes, a material finding is made in monitoring, or the quarterly model card review cycle is due

• –Reads the current model card alongside the new performance data, any validation findings since the last card update, and any changes to the model's inputs, training data, or deployment scope — and produces an updated model card that reflects the model's current state accurately.
• –Tracks model card completeness against the model risk policy standard — required fields include: model purpose and scope, development methodology summary, training data description, input feature list with data sources, known limitations and failure modes, bias testing results, performance metrics with last-updated date, approval status and approver, and scheduled review date.
• –Flags model cards that are outdated relative to the model's current production performance — a card that shows Gini 72% when the model is currently running at Gini 65% is misleading to any stakeholder who reads it and constitutes a model risk governance gap.

Invoked when: quarterly bias testing cycle runs, or a Fair Lending AI finding triggers an ad-hoc bias review for a specific model

• –Runs the bias testing suite across all production models using the configured bias dataset — approval rate disparity by geography (state/district), income band, and any proxy variables identified by the Fair Lending AI. Computes adverse impact ratios and disparate impact metrics for each protected-characteristic proxy.
• –Distinguishes between bias in the model (a feature that proxies for a protected characteristic and drives differential outcomes) and bias in the data (historical outcome data that reflects prior discriminatory practices). The former requires model remediation; the latter requires careful feature engineering decisions. The bias report labels which type is detected.
• –For models where bias is detected above the policy threshold: produces a bias finding report with the specific feature, the proxy relationship, the magnitude of disparate impact, and the options for remediation — feature removal, fairness constraint, re-weighting, or model replacement. Does not remediate the model; documents the finding for the data science team to address.

Invoked when: monthly performance data triggers a drift alert, or an intra-month approval rate or NPA rate anomaly warrants a mid-cycle check

• –Reads all active drift alerts from the monitoring pipeline — PSI, CSI, Gini degradation, NPA rate divergence — and triages them by severity, urgency, and likely cause. A PSI of 0.12 that has been stable for 3 months is a "monitor" triage; a PSI of 0.12 that jumped from 0.04 in a single month is an "investigate immediately" triage.
• –Aggregates drift signals across models to identify systemic patterns: if three models are simultaneously showing population drift in the same direction, the common cause is likely a portfolio composition change or a macro shift, not individual model failure — and the response is different from addressing three separate model-specific issues.
• –Updates the model risk register with each triage outcome, the evidence, and the recommended next action — so the governance committee always has a current picture of the model risk landscape, not a point-in-time snapshot from the last quarterly review.