AI Guide For Boards in Singapore

AI GUIDE FOR BOARDS IN SINGAPORE

Section V: Organisational Resilience and Capability

10.4 Model Risk Management

• Fallback mechanisms where workows cannot tolerate AI failure.

Managing model risk is not a one-time approval exercise. As business conditions, user behaviour and operating environments evolve, organisations need processes to determine whether AI systems remain suciently reliable and eective. A practical approach is to dene upfront what “good enough” performance means for a given use case. is may include: • Dening the intended business purpose of the AI system. • Establishing measurable performance outcomes. • Identifying unacceptable errors or behaviours. • Maintaining benchmark examples or gold standards for ongoing evaluation. Boards should expect management to monitor AI systems continuously for changes in performance, reliability, accuracy and suitability for their intended business purpose.

Oversight should focus on whether model performance, reliability and fairness are monitored over time. Appropriate human and operational safeguards include dening a clear authority for restricting, retraining or retiring models.

Boards should require clear thresholds dening when AI systems are no longer considered t for purpose.

10.4.2 Enterprise risk management

As AI becomes increasingly embedded in business operations, boards should ensure that AI oversight is integrated into existing frameworks for digital resilience, cyber security and enterprise risk governance. is includes understanding how infrastructure will support AI at scale, how AI-related cyber threats are managed, how operational failures and harmful outputs are addressed, and how model performance is monitored over time.

10.4.1 Continuous monitoring

AI systems are not static soware assets. Risks may emerge over time through: • Concept dri: Changes in relationships between inputs and outcomes. • Data dri: Changes in the characteristics or distribution of input data. • Model failure: Infrastructure, integration, or pipeline failures that cause inaccurate or unreliable outputs. Key governance considerations include: • Continuous monitoring of model accuracy and reliability. • Regular retraining using updated data. • Human oversight for high-stakes decisions.

Model risk management should be treated as part of broader operational resilience and enterprise risk governance.

94

Made with FlippingBook Digital Proposal Creator