Dieser Artikel ist derzeit auf Englisch verfügbar.
The CRO's Guide to AI Risk: Building a Framework That Satisfies Regulators and the Board
Chief Risk Officers in financial services face AI risk from three directions simultaneously: model risk, conduct risk, and operational risk. Traditional risk frameworks were not built for this. Here is how to adapt them.
Key Takeaways
AI risk does not fit cleanly into traditional risk taxonomy — it creates simultaneous model risk, conduct risk, operational risk, and reputational risk that interact in non-linear ways.
Regulators (APRA, FCA, MAS, ACPR) are not asking for AI-specific risk frameworks — they are asking how AI risk is integrated into existing enterprise risk management. The answer must be specific and evidenced.
Model risk management for ML/AI requires updating the SR 11-7 framework — the core concepts (validation, monitoring, documentation) still apply but the implementation must account for model complexity, data drift, and explainability.
The three AI risks that are most likely to materialise in 2026-2027: model drift in credit and fraud detection, discriminatory outcomes in automated decisions, and third-party AI supplier failures.
A practical AI risk register template and the six metrics every CRO should be monitoring on AI systems in production.
"Nur zu Informationszwecken. Dieser Artikel stellt keine rechtliche, regulatorische, finanzielle oder professionelle Beratung dar. Konsultieren Sie einen qualifizierten Spezialisten für spezifische Beratung."
Why traditional risk frameworks struggle with AI
Enterprise risk frameworks were built around human decision-making with technology as a support tool. The fundamental assumption is that a human is ultimately responsible for every significant decision — the human may use data, models, or systems to inform that decision, but accountability is clear. AI changes this assumption in ways that create genuine framework gaps.
When an AI system makes thousands of credit decisions per hour, or flags hundreds of suspicious transactions per day, the human oversight model becomes nominal rather than real. No team can meaningfully review every AI decision. The risk framework needs to shift from transaction-level oversight to system-level oversight — monitoring the AI's performance as a whole rather than individual outputs. Most traditional risk frameworks have not made this shift.
The three AI risk types and how they interact
Model risk — the risk that a model produces incorrect or biased outputs — is familiar territory for financial services CROs. The SR 11-7 framework provides the conceptual architecture. What has changed is the complexity of the models and the difficulty of validation. A logistic regression credit model can be fully understood and validated. A gradient boosted ensemble or a neural network used for the same purpose cannot be validated in the same way. The validation methodology must change, and most model risk functions have not updated their approaches.
Conduct risk from AI is less familiar but increasingly material. When an AI system systematically charges higher prices to one demographic group, or denies services at higher rates to another, this is a conduct failure regardless of whether it was intentional. The AI does not know it is discriminating — it is optimising for an objective function that produces discriminatory outputs. Regulators treat this as conduct risk, not technology risk. The CRO needs to own it as conduct risk.
Operational risk from AI includes the failure modes traditional operational risk frameworks recognise — system outage, data quality failure, cyberattack — plus new categories: model drift (performance degradation over time as the world changes), adversarial inputs (deliberate attempts to manipulate AI outputs), and third-party AI failures (when your supplier's AI fails in ways that affect your operations).
What regulators are actually asking for
APRA, the FCA, MAS, and ACPR have all issued guidance or conducted thematic reviews on AI risk in financial services. Their asks are more consistent than the volume of guidance suggests. They want to see: an inventory of AI systems used in regulated activities; a governance structure with named accountability; a validation and monitoring regime appropriate to the complexity of each system; evidence that the board and senior management understand and oversee AI risk; and a plan for AI system failures. They are not asking for bespoke AI risk frameworks — they are asking how AI risk is integrated into existing governance.