AI GUIDE FOR BOARDS IN SINGAPORE
Section V: Organisational Resilience and Capability
10.3 Operational Resilience
• Recovery arrangements: Rollback procedures, model version controls and fallback processes where workows cannot tolerate extended downtime.
Operational resilience for AI depends on early detection, clear escalation pathways, and the ability to recover quickly when systems fail, behave unexpectedly, or produce harmful outputs. Boards should expect AI-enabled systems supporting important workows to be managed as live operational capabilities rather than static technology deployments. AI failures may arise from model dri , infrastructure outages, data pipeline failures, adversarial attacks, or integration breakdowns. Organisations should establish: • Early detection mechanisms: Monitoring for degraded model performance, anomalous outputs, application programming interface failures, or organisational irregularities. • Incident response protocols: Clear escalation procedures, de ned responsibilities and response playbooks for common AI failure scenarios.
10.3.1 Incident response
Incident response arrangements should also extend beyond technical failures. Organisations may require protocols for handling harmful outputs, misuse of connected systems, governance failures, or inappropriate AI behaviour in sensitive operational contexts. Resilience requires continuous learning. Management should use incidents, near misses and operational feedback to strengthen safeguards, improve monitoring and re ne deployment practices over time.
Boards should ask whether the organisation can rapidly identify, escalate, and respond when AI-enabled workows stop performing as intended.
93
Made with FlippingBook Digital Proposal Creator