Model Governability Assessment

Can you show how each AI use case is governed?

The Model Governability Assessment examines each AI use case and the systems and models beneath it. It records inherent and residual risk, tests the controls in scope and produces a governability finding for each use case. The evidence supports review by your board, risk function or internal audit.

What "governability" means

The assessment starts with the business use case: its purpose, owners, decisions, data and dependencies. It covers the systems and models that support that use case, including third-party AI. Materiality records address impact, complexity and reliance, with inherent and residual risk documented separately. The governability finding explains which controls hold, which need work and what evidence supports that conclusion.

The assessment is built around five dimensions.

Framework reconciliation. Map the use case to the supervisory expectations, laws and standards that apply. Record overlaps, gaps and conflicts. IMDA’s Model AI Governance Framework for Agentic AI, version 1.5, is voluntary reference material, distinct from the general 2020 framework. This edition was published on 20 May 2026 and updated on 5 June 2026. MAS points financial institutions to the agentic framework as a reference in AIRG footnote 11. The output distinguishes mandatory obligations, supervisory expectations and voluntary references.

Enforcement reliability. Test the configured controls against defined scenarios for the use case. Record what the system allows, denies and escalates, including behaviour at the enforcement threshold. Document human dependencies and any conditions the controls fail to enforce.

Decision lineage. Check whether a consequential decision can be reconstructed through authorisation, execution and constraint. Trace the inputs, model and configuration versions, delegated authority, control decision and outcome. Record evidence gaps that prevent a reviewer from establishing how the decision was made.

Authority loop closure. Test the authority loop: a defined trigger halts the action, a named human receives the context, and execution waits for ratification or an agreed fail-safe outcome. Review the Suspended Handoff State, Human Ratification Gate and Agentic Risk Boundary mechanisms relevant to the use case.

Evidence readiness. Check that the inventory links, assessment rationale, test results and decision records are available for review. Identify missing records, ownership gaps and retention limits. Organise the evidence so a second line, internal audit or supervisor can follow the assessment.

Framework reconciliation Which expectations and obligations apply to this use case?
Enforcement reliability Do the configured controls hold in the defined tests?
Decision lineage Could you reconstruct what governed a specific decision, on demand?
Authority loop closure Does the authority loop reach a defined ratification or fail-safe outcome?
Evidence readiness Are the assessment evidence and decision records ready for review?
The assessment produces a governability finding per use case across all five dimensions.

MAS AI guidelines: readiness for 2027 and 2028

The Monetary Authority of Singapore (MAS) issued the final Guidelines on Artificial Intelligence Risk Management (AIRG) on 7 October 2026, following consultation paper P017-2025 of 13 November 2025. They set supervisory expectations for financial institutions, applied proportionately to their size, AI use and risk profile.

  • From 7 October 2027: Section 3 covers AI oversight. Section 4 covers identification, inventory and risk materiality. The assessment’s use-case materiality record supports Section 4.
  • By 7 October 2028: Section 5 covers AI life cycle controls, including human oversight, third-party AI and monitoring. Section 6 covers AI capability and capacity. My enforcement reliability, decision lineage and authority loop closure findings provide an early view of relevant Section 5 gaps.

Building and testing life cycle controls takes time. Starting now gives owners time to address gaps, collect operating evidence and prepare for review ahead of the 2028 phase.

AI agents and multi-agent systems

The AIRG explicitly includes AI agents and multi-agent systems. For agents in scope, I check inventory coverage of agent-specific identifiers, accessible tools and systems, components and guardrails, as illustrated in AIRG footnote 25. I examine whether monitoring covers relevant reasoning processes, actions taken and tools used under paragraph 5.23(a). These checks connect my Agentic Risk Boundary and Human Ratification Gate work to the inventory, human oversight and monitoring expectations.

MAS intends to consult the financial sector in 2027 on what additional agentic AI guidance would be useful.

Evidence for independent validation

AIRG paragraphs 5.18 to 5.20 expect pre-deployment review by parties outside development and formal independent validation for high-materiality use cases. My evidence pack gives reviewers the control tests, decision records, limitations and remediation findings to inform that work. The assessment does not replace the FI’s independent validation or provide a formal assurance opinion.

Why this runs annually

Use cases change as systems, models, providers and operating conditions change. The annual reassessment reviews those changes and retests the affected controls across the same five dimensions. A material change can trigger an earlier review. Each assessment records a governability finding per use case, with the evidence and unresolved limitations needed to compare it with the baseline.

What you receive

  • A governability finding per use case, with evidence across all five dimensions
  • A framework reconciliation map showing applicability, control coverage and conflicts
  • A decision lineage specification for the consequential decisions in scope
  • An authority loop closure assessment covering every override mechanism in scope
  • An evidence pack organised for regulator, auditor, or board review
  • A prioritised remediation roadmap for any dimension that does not yet hold

How it works

The baseline runs over six weeks: discovery to link each use case to its systems and models, a five-dimension assessment, a reconciliation workshop and a board presentation. I agree the materiality rationale with the relevant owners and document the control function’s review. The three-week annual reassessment focuses on changes and the controls that need retesting.

The baseline fee is agreed before work begins. The scope identifies the use cases, systems, models, testing access and evidence required. The final findings record limitations and remediation priorities alongside the results.

A use-case finding records the controls tested, the evidence observed and the work still required. Repeat assessments show how that finding changes as the use case changes.

Find out how governable your AI use cases are.

I review one AI use case against the MAS AI Risk Management Guidelines in 10 working days, for S$4,800 fixed. The 20-minute call is to choose the use case and confirm it fits.

Book a 20-minute call