Who can halt an AI action, and how does the system respond?

Designing human intervention

An oversight process needs defined decision rights and a way to act on them. We identify the AI use cases, consequential actions and risk thresholds that require intervention. The design states who receives an escalation and what they are authorised to decide.

A Suspended Handoff State holds an action at a defined threshold pending human ratification. Its design includes the context delivered to the reviewer, time limits and an agreed fail-safe outcome. The mechanism is selected according to the use case and its risk materiality.

The authority loop must reach an outcome when a reviewer is unavailable or an approval is invalid. Scenario testing exercises those paths and records where the process or configured controls need further work.

The engagement connects intervention authority to the governance structure responsible for operating it.

What this engagement produces

The deliverables cover authority design, operating procedures and the records needed for oversight.

Override authority matrix. A mapping of intervention authority for each use case and its systems. It identifies decision limits, trigger conditions and the escalation path when the designated owner is unavailable.

Suspended Handoff State definition. The conditions that hold execution, the evidence a ratifier must see and the outcome when approval is rejected, delayed or absent.

Governance oversight charter. A written mandate covering scope, decision rights and relationships with senior management and the board. Existing oversight bodies can perform these roles where appropriate.

Role definitions and selection criteria. The knowledge, authority and independence needed for each role, with responsibilities for reviewing decisions and responding to exceptions.

Meeting cadence and trigger-based review protocols. Scheduled reviews and conditions for an earlier review, including risk breaches, significant changes and incidents. Each agenda identifies the decisions and records required.

Board reporting format. A template covering the strategic direction for AI use, risk appetite, material findings, interventions and unresolved exceptions. The records show how oversight decisions were reached and followed through.

Onboarding and inaugural review. A briefing on the mandate, authority matrix and use cases in scope. Facilitation of the first review establishes the operating process and its decision records.

How it works

The eight-week engagement begins with stakeholder consultation and authority mapping. Design sessions define the handoff state, roles and reporting. Scenario testing then checks intervention paths before onboarding and the inaugural review.

The scope sets out the use cases, system access and participants needed for the design and testing sessions.

Oversight evidence for Singapore operations

The final AIRG sets supervisory expectations for financial institutions. FEAT principles and the 2024 information paper remain separate references. The engagement documents who approves AI use, how risks reach the board and how decisions are implemented.

A group framework must meet the AIRG locally, and local senior management remains accountable to MAS. Aivance checks local responsibilities, visibility into group decisions and escalation paths. The MAS AIRG guide provides the detailed section mapping and phased dates.