Can you translate NIST AI RMF and SR 11-7 into operational risk management? They want evidence of framework implementation, not just theoretical knowledge.

1. Risk appetite definition — work with CRO to define AI risk tolerance aligned with enterprise risk appetite. 2. AI system inventory — catalogue all AI/ML models with risk tiers based on EU AI Act classification. 3. Framework alignment — map controls to NIST AI RMF functions (GOVERN, MAP, MEASURE, MANAGE) and SR 11-7 requirements. 4. Validation program — establish independent model validation with credible challenge. 5. KRI monitoring — define risk indicators for model drift, bias, accuracy, and operational performance. 6. Reporting structure — executive dashboards and regulatory-ready documentation.

This is the defining concept for banking AI risk roles. Do you understand second-line-of-defense independence, or will you rubber-stamp first-line model development?

Credible challenge means providing independent, rigorous, evidence-based pushback to first-line model developers. In the three-lines-of-defense model: first line (business units) develops AI models, second line (risk management, your role) provides oversight and challenge, third line (internal audit) provides independent assurance. Challenge must be substantive: reviewing model assumptions, testing for bias, validating performance metrics, questioning data quality, and assessing deployment conditions. The output is a formal validation report with risk ratings and required remediations.

GenAI has unique risk characteristics that don’t fit traditional model risk frameworks. Can you extend your risk assessment to cover hallucination, prompt injection, and data provenance?

Extend traditional model risk assessment with GenAI-specific risk categories: hallucination risk (factual accuracy in outputs), prompt injection (security boundary violations), data exfiltration (IP leakage through prompts), training data provenance (copyright and bias in training data), content provenance (attribution and watermarking), and agentic risk (autonomous action boundaries). Use NIST AI 600-1 GenAI Risk Profile as the assessment framework. Each risk gets likelihood, impact, and mitigation controls.

Can you quantify AI risk in terms that CROs and regulators can act on? Generic answers like ‘accuracy’ won’t cut it.

Build a multi-dimensional KRI framework: Performance KRIs — model accuracy, precision, recall against baseline thresholds; response latency against SLAs. Drift KRIs — prediction drift (PSI/CSI), feature importance changes, data distribution shifts. Fairness KRIs — demographic parity ratio, equalized odds ratio, disparate impact metrics across protected groups. Operational KRIs — model exception rates, override frequencies, escalation counts. Regulatory KRIs — validation finding closure rates, audit issue resolution timelines, compliance gap counts by framework.

Risk communication is the most critical interpersonal skill. Can you translate model risk into business impact language?

Frame AI risk in three business-impact dimensions: 1. Financial exposure — potential loss from model failure (credit risk understatement, fraud miss rate, fair lending violations) quantified in dollar terms. 2. Regulatory exposure — non-compliance consequences mapped to specific regulations (SR 11-7 findings, EU AI Act penalties, consent decree risk). 3. Reputational exposure — bias incidents, customer harm, and public trust erosion quantified through comparable incident analysis. Use risk heatmaps and trend dashboards rather than technical metrics.