AI Guide For Boards in Singapore

AI GUIDE FOR BOARDS IN SINGAPORE

Section III: Governance and Oversight

7.3.2 Establishing key risk indicators

framework, the board may be governing AI indirectly rather than governing it deliberately.

Boards should move beyond qualitative assurances, and require management to establish measurable AI-related key risk indicators (See: Key Risk Indicators: Examples ) linked to dened tolerance thresholds.

e purpose of key risk indicators is not to eliminate all AI risk, but to provide boards with visibility into whether risks remain within acceptable operating boundaries.

If AI risks are not visible within the organisation’s core enterprise risk

Key Risk Indicators: Examples

Tolerance Boundaries (Example)

Category

Key Risk Indicators

Model stability

Escalation and review if > 5% of core models dri

Percentage of production models experiencing signicant AI dri per quarter Human override rates for high-consequence automated outputs Percentage of AI training sets lacking certied provenance or privacy data lineage Number of near-misses or automated exceptions agged by automated guardrails

Human oversight

Flag if override rate drops below 2% (signals rubber- stamping) or exceeds 25% (signals model failure) Hard limit at 0% for Tier-1 customer-facing models

Data integrity

Incident management

Trigger committee escalation if more than 3 Tier-2 incidents in a month

64

Made with FlippingBook Digital Proposal Creator