Tier 3: Safety and resilience controls

Human oversight and human-in-the-loop controls

Design and operational controls that ensure humans can understand, intervene, override, and stop AI decisions when risk thresholds or safety conditions require it.

Board-defensible evidence

• Oversight design documentation defining where humans review outputs, what they are expected to check, and what conditions require mandatory override or escalation.
• Role-based training and competency records showing that humans assigned to oversight have the authority, knowledge, and procedures to intervene effectively.
• Override and intervention logs showing when humans intervened, who intervened, what decision was changed, why it was changed, and what follow-up actions were taken.
• UI and workflow evidence showing that humans receive meaningful context (confidence, rationale, limitations) and that "rubber-stamping" risk is addressed through controls.
• Periodic effectiveness reviews showing oversight performance metrics (override rates, error catch rates, escalation timelines) and documented improvements when oversight fails.

Why this matters

When an AI-driven harm occurs, oversight is judged by whether humans could realistically prevent it, not whether a policy said they could.