Most banks that deploy AI models have some form of monitoring in place. The real question is whether that monitoring is connected to governance — whether a monitoring alert triggers a defined response with accountable owners, escalation paths, and documentation — or whether it is a technical dashboard that nobody acts on until a complaint or an audit finding makes the degradation visible after the fact. We have seen both versions sitting on the same desk: a beautifully built Grafana dashboard with every threshold line drawn in, and nobody who checks it more than once a quarter.
SR 11-7's model risk management requirements, and the direction RBI's FREE-AI recommendations are pushing supervisory practice, both treat monitoring as a governance function, not a technical one. The model's monitoring plan must be documented before production deployment, not built after deployment reveals a gap. Thresholds must be pre-defined, not set retrospectively after degradation has already occurred. And monitoring must cover fairness and drift, not just accuracy — a model that holds its average accuracy steady while developing disparate impact on protected groups is degrading in a way accuracy monitoring alone will never surface.
Six Production AI Monitoring Dimensions
1. Performance metrics. For classification models (credit scoring, fraud detection, AML): AUROC, Gini coefficient, KS statistic, precision and recall by class. For regression models: RMSE, MAE, and systematic over- or under-prediction bias. Track both overall and by key subpopulations — population-level accuracy masking subgroup degradation is a common and easily missed failure mode. Monitoring frequency: monthly minimum for credit models, daily for real-time fraud detection, with trigger thresholds pre-defined per model in the monitoring plan.
2. Data drift (covariate shift). The statistical distance between current production input distributions and training data distributions. The Population Stability Index (PSI) is the banking standard: under 0.1 is stable, 0.1–0.25 needs review, above 0.25 needs a model review with escalation to model risk. Monitor PSI at the feature level, not just as a summary statistic — a feature-level breakdown shows which features are drifting, which may point to data quality issues, customer behavior shifts, or product mix changes, each with a different governance response.
3. Concept drift (label shift). A change in the relationship between inputs and outcomes, detected by comparing model-predicted outcome rates against actual outcome rates on a rolling basis. If the model predicts a 5% credit default rate but the realized rate has moved to 8%, the model's learned associations no longer match current credit behavior. Concept drift typically requires full or partial retraining with updated outcome data — unlike data drift, which recalibration can sometimes fix.
4. Fairness monitoring. Differential performance across protected groups, tracked at the same frequency as accuracy monitoring. For India banking: differential approval rates, differential false positive rates in fraud detection, and differential pricing across India proxy variable dimensions (geography/pincode cluster, religion-proxy UPI counterparty patterns, language/device proxy). Fairness monitoring needs subgroup sample sizes large enough to detect a meaningful differential — monthly monitoring on small subpopulations may need quarterly aggregate reviews to reach statistical power. Fairness can drift even while overall accuracy holds perfectly steady, which is exactly why relying on the accuracy dashboard alone misses it.
5. Output distribution and calibration. A credit model calibrated to predict realistic default probabilities at validation can drift toward systematically biased probability estimates in production — predicted probabilities clustering at the high or low end in ways that don't reflect realized outcomes. Calibration monitoring tracks Expected Calibration Error and reliability diagrams. This matters most for risk-weighted pricing decisions, where the absolute value of the predicted probability drives the financial calculation, not just the rank ordering.
6. Override and incident tracking. The rate at which human reviewers override AI recommendations, and the frequency of AI-related incidents. A high override rate can indicate model degradation (reviewers no longer trust the model), deployment mismatch (the model is being used outside its intended scope), or a governance gap (the model is used without the human review it should have).
Monitoring Threshold Governance: What Triggers What Action
| Monitoring signal | Level 1 (watch) | Level 2 (review) | Level 3 (escalate/suspend) | Governance action |
|---|---|---|---|---|
| Credit model AUROC | Decline >3% from validation benchmark | Decline >5%, or below absolute pre-set threshold | Decline >10% or underperforms simple baseline | L2: model risk owner investigation + compliance notification. L3: suspension or shadow operation pending revalidation. Board notification if customer-facing. |
| PSI (data drift) | 0.1–0.15 on any top-10 feature | 0.15–0.25 on any significant feature | >0.25 on any significant feature, or >0.1 on multiple correlated features | L2: root cause investigation. L3: model review committee, retraining assessment. Document findings. |
| Fairness: differential approval rate | Adverse impact ratio 0.75–0.80 (emerging gap) | Below 0.75 | Below 0.60, or statistically significant differential | L2: fairness investigation with compliance. L3: suspension or operational safeguards pending remediation; regulatory notification if required; board report. |
| LLM policy violation rate | Rises >20% above baseline week-on-week | Doubles from baseline; specific category rises materially | Exceeds pre-defined absolute threshold; evidence of systematic attack | L2: security review — organic usage or adversarial? L3: suspension or input filtering escalation; incident report. |
| Human override rate | Rises >10% above baseline | Rises >25% above baseline, sustained 30 days | Doubles, or indicates systematic model disagreement | L2: model owner investigation. L3: model risk committee review; assess continued production viability. |
India-Specific Production Monitoring Requirements
PMJDY segment monitoring. AI models deployed across the full retail customer base must produce disaggregated monitoring reports for the Pradhan Mantri Jan Dhan Yojana (PMJDY) financial inclusion segment, which has systematically different transaction patterns, income profiles, and credit behavior from conventional retail customers. A credit or fraud model that looks acceptable overall can be significantly miscalibrated for this segment — and monitoring that only reports aggregate numbers will never show it.
India proxy variable fairness drift monitoring. Fairness monitoring must track differential performance on proxy dimensions that have no equivalent in standard international monitoring frameworks: geographic cluster, UPI counterparty name patterns, language/device type, disability-proxy interaction patterns. Fairness drift on these dimensions can emerge from retraining even when explicit proxy variables are excluded from the feature set — the model can relearn the proxy association from correlated features. Monthly proxy dimension fairness checks should be a standing item for any credit or fraud model trained on data that reflects India's social geography.
Monsoon and crop cycle seasonal drift. Agricultural credit models have structural seasonal patterns — application volumes, income patterns, and default rates cycle with the monsoon and crop seasons. These shifts look statistically like drift but are expected variation, not degradation. Monitoring programs for agricultural credit AI must distinguish the two, or PSI thresholds tuned for non-seasonal products will trigger false alerts at every kharif-to-rabi transition. The fix is season-stratified PSI baselines or documented seasonal exclusion periods, not a blanket threshold that gets silently ignored twice a year.
Model Monitoring Action Checklist
- Confirm every production model's monitoring plan — thresholds, frequency, escalation owner — was documented before deployment, not assembled after the fact.
- Add feature-level PSI breakdowns where only a summary drift statistic currently exists.
- Stand up fairness monitoring on India proxy variable dimensions at the same cadence as accuracy monitoring, not as a separate quarterly exercise.
- Document the actual source of every borrowed numerical threshold (0.75 adverse impact ratio, 0.25 PSI) rather than attributing it to an uncited circular.
- For agricultural credit models, build season-stratified baselines before the next kharif-to-rabi transition generates a wave of false drift alerts.
Related: AI model lifecycle governance
Production AI monitoring connected to governance — thresholds set at validation, actions triggered automatically, board-level reporting built in
Zytra's monitoring module connects production metrics to governance workflows. Thresholds defined at validation persist through deployment. Breach events trigger governance actions — not just log entries — with named owners and response timelines. India proxy variable fairness monitoring and PMJDY segment monitoring included for India BFSI deployments.