AI risk appetite is the governance mechanism that connects board-level direction to day-to-day AI governance decisions. Without defined appetite thresholds, governance teams make ad hoc materiality judgments — whether a monitoring alert is serious enough to escalate, whether a fairness concern warrants a deployment pause, whether an AI incident requires board notification. Those judgments will be inconsistent, and inconsistency is exactly what an examiner probes for: ask the same escalation question of three different committee members and get three different answers, and the appetite statement is exposed as decorative.
With defined appetite thresholds, governance becomes mechanical in the useful sense: a model with a Population Stability Index above 0.25 is a board notification event because appetite says so, not because someone judged it serious; a credit AI with an adverse impact ratio below 0.75 must not be deployed because appetite says so; an LLM customer service system that produces financial advice within a defined period constitutes an appetite breach requiring a pre-agreed response. The governance team stops exercising ad hoc judgment about materiality and starts applying thresholds the board already signed off on — which is also the only version of this that holds up when an examiner asks "who decided this was serious enough to escalate, and when."
AI Risk Appetite by Domain: Tolerance Levels and Thresholds
| Risk domain | Appetite level | Specific threshold | Governance action at threshold | Board notification? |
|---|---|---|---|---|
| Model accuracy degradation (credit AI) | Moderate | AUROC decline >5% from validation benchmark; absolute AUROC below pre-defined minimum; Gini below policy floor | Model review required within 30 days; deployment restricted pending review outcome | Yes — if decline sustained >60 days without resolution |
| AI fairness — adverse impact | Zero tolerance | Adverse impact ratio below 0.75 on any monitored protected dimension; statistically significant differential detected at 95% confidence | Immediate escalation to AI governance committee; deployment pause or operational safeguard until remediated; regulatory notification assessment | Yes — board notification within 5 business days |
| LLM financial advice boundary violation | Zero tolerance | Any confirmed instance of individualized investment advice, insurance recommendation, or unauthorized lending recommendation reaching customer | Immediate escalation; incident report; system review; SEBI/RBI notification assessment; customer remediation if harm occurred | Yes — immediate board notification for material incidents |
| Data drift (production AI) | Moderate | PSI >0.25 on any significant feature; PSI >0.10 across multiple correlated features simultaneously | Root cause investigation; model review committee notification; retraining assessment within defined SLA | No — unless model review determines material accuracy impact |
| AI security — prompt injection to customer | Zero tolerance | Any confirmed prompt injection attack that reaches the customer with altered bank messaging; any confirmed PII exfiltration event through LLM channel | Immediate incident response; system suspension pending security review; CERT-In notification; regulatory notification | Yes — board notification within 24 hours for security incidents |
| Agentic AI authorization breach | Zero tolerance | Any confirmed instance of agentic AI taking actions outside its declared authorization scope, including unauthorized API calls, unauthorized transaction execution, or unauthorized data modification | Immediate agent suspension; incident investigation; authorization scope review; customer impact assessment | Yes — board notification for any confirmed authorization breach |
| AI operational resilience | Moderate | Unplanned downtime exceeding defined threshold for critical AI systems; recovery time exceeding RTO defined at deployment | Incident response per BCM plan; root cause analysis; review of resilience controls | No — unless AI downtime causes material customer impact or regulatory reporting obligation |
| Third-party AI change without governance approval | Low | Any vendor AI system change deployed to production without bank governance approval; any change notification received less than contractual minimum notice period | Retrospective impact assessment; vendor governance escalation; contract review; potential rollback if impact assessment identifies risk | No — unless change caused or is likely to cause appetite breach in another domain |
India-Specific AI Risk Appetite Requirements
India proxy variable fairness appetite (mandatory for India BFSI). Standard AI fairness risk appetite frameworks monitor protected characteristics defined in applicable anti-discrimination law — in the US, race and gender; in the EU, comparable demographic categories. India's fairness exposure is shaped by its constitutional non-discrimination framework and by the social geography of caste, religion, and language. India banking AI risk appetite should include fairness thresholds for proxy dimensions that have no equivalent in imported frameworks: geographic or pincode cluster (a caste proxy in many regions), UPI counterparty name patterns (a religion proxy), language or device type (a language-group proxy), and disability-proxy interaction patterns. A bank whose risk appetite statement covers standard demographic dimensions but not these has a fairness appetite gap specific to India — and copying a global bank's template will reproduce exactly this gap.
Financial inclusion AI appetite. India's financial inclusion mandate creates a risk appetite dimension largely absent from international frameworks: the bank's tolerance for AI systems that systematically disadvantage Pradhan Mantri Jan Dhan Yojana (PMJDY) account holders or other financial inclusion segments. An AI credit model that is accurate overall but materially worse for PMJDY customers is not just a fairness concern — it is a financial inclusion policy concern that can attract regulatory attention on its own, independent of the Fair Practices Code. India BFSI risk appetite should define a tolerance threshold for AI performance differentials on the PMJDY segment, separate from the general fairness appetite threshold.
Examination evidence for risk appetite. RBI examination practice is moving toward expecting banks to show that their AI risk appetite is operational, not just documented. Evidence worth assembling in advance includes: board minutes showing appetite approval and periodic review; governance committee records showing appetite thresholds actually applied in deployment decisions; monitoring records showing appetite threshold lines overlaid on production monitoring data; and incident records showing governance actions triggered by appetite breaches. A bank with a well-written appetite statement but no evidence it was ever applied has a policy-to-practice gap that is trivially easy for an examiner to find — ask for the last breach, and ask what happened next.
AI Risk Appetite Action Checklist
- Get board sign-off on domain-level thresholds individually, not on a single aggregate "AI risk appetite" statement — aggregate statements are what examiners push back on hardest.
- Document the source and rationale for every borrowed numerical threshold (0.75 adverse impact ratio, 0.25 PSI) as an internal benchmark decision, not a cited regulatory figure.
- Add India proxy variable thresholds (pincode cluster, UPI counterparty pattern, language/device type) to the fairness appetite statement if they are not already there.
- Set a separate, named tolerance threshold for AI performance on the PMJDY and financial inclusion segment.
- Pull the last three appetite breaches from your incident log and confirm each one produced the governance action and board notification the appetite statement promises — if it didn't, fix the process before the next examination finds it.
Related: board-level AI governance
AI risk appetite operationalized in governance workflows — domain thresholds applied in monitoring, deployment gates, and incident management
Zytra operationalizes board-approved AI risk appetite: thresholds defined once, applied consistently in deployment gates, production monitoring alerts, and incident escalation triggers. India proxy variable fairness appetite and PMJDY financial inclusion appetite thresholds are supported natively.