How to Scope and Deliver an AI Governance Assessment in Singapore: Method, Evidence and Deliverables

Direct answer

An AI governance assessment is not a one-off audit against a single framework. It is a structured exercise that scopes every AI system and AI agent in use, gathers evidence of controls across the AI lifecycle, and produces deliverables a board or risk committee can act on — a system inventory, a risk-tiered gap analysis, and a remediation roadmap. Done properly, it is assessed against IMDA's Model AI Governance Framework, CSA's Guidelines on Securing AI Systems and, where financial services are involved, MAS's AI risk guidance — not treated as a tick-box exercise against any single document.

What Singapore's regulators and frameworks expect

Singapore does not have a single binding AI law. Instead, organisations deploying AI work against a layered set of guidance from IMDA, CSA, MAS and PDPC, each covering a different part of the governance picture.

IMDA's Model AI Governance Framework

On 23 January 2019, IMDA/PDPC released its first edition of the Model AI Governance Framework for Traditional AI for broader consultation and adoption. IMDA launched the Model AI Governance Framework for Generative AI on 30 May 2024, and most recently extended its coverage to autonomous systems. On 22 January 2026, Minister for Digital Development and Information Josephine Teo announced the launch of the Model AI Governance Framework for Agentic AI at the World Economic Forum, building on the governance foundations of the 2020 framework. This is the world's first governance framework specifically designed for AI agents capable of autonomous planning, reasoning and action. It was subsequently updated: on 20 May 2026, IMDA updated the Model AI Governance Framework for Agentic AI, adding safety and reliability components — access controls, guardrails, human approvals, logging and monitoring — as core components of an AI agent.

CSA's Guidelines on Securing AI Systems

CSA's Guidelines on Securing AI Systems, published in October 2024, support system owners adopting or considering AI by identifying security risks and setting out mitigation guidance across the AI lifecycle, grounded in the principle that AI should be secure by design and secure by default, covering five stages: Planning and Design, Development, Deployment, Operations and Maintenance, and End of Life. The guidelines are non-mandatory but strongly encouraged, and CSA references established international frameworks — including OWASP, MITRE ATLAS and NIST — as complementary resources. As agent-based deployments have grown, CSA has extended this coverage: a companion addendum on securing agentic AI was issued in draft for public consultation on 22 October 2025 and finalised on 17 June 2026, extending the guidelines to identifying and assessing risk according to agent capabilities, mapping workflows to surface exploitable weak points, and setting out controls across the development lifecycle for different levels of autonomy.

MAS AI risk guidance

For financial institutions, MAS guidance sits on top of IMDA and CSA expectations. MAS's FEAT principles — Fairness, Ethics, Accountability and Transparency — were published in 2018 and remain the foundation. More recently, on 7 October 2026 MAS issued Guidelines on Artificial Intelligence Risk Management, setting out what it expects of every financial institution using AI: board oversight, an AI inventory, controls across the AI lifecycle, and skilled staff. These Guidelines take effect on 7 October 2027, with Sections 5 and 6 following by 7 October 2028. At board level, the board or a delegated committee is expected to approve the governance approach, cover AI within the risk appetite, and understand AI sufficiently to provide oversight and challenge. MAS also expects named accountability at system level: institutions should assign clear roles and responsibility for ongoing monitoring and managing of incidents of deployed AI, including the designation of an appropriate accountable owner.

PDPC's advisory guidelines

On 1 March 2024, PDPC published Advisory Guidelines on the use of personal data for the design and deployment of systems that embed machine learning models. These Advisory Guidelines are not legally binding, but in enforcing the Personal Data Protection Act, PDPC is likely to take positions consistent with them, so any AI governance assessment touching personal data should evidence compliance against them alongside the PDPA itself.

← Swipe to compare →

How to scope and deliver the assessment: method, evidence, deliverables

The frameworks above tell you what good looks like. They do not tell you how to run an assessment. A defined method has three stages: scope, evidence, deliverables.

1. Scope by system, not by policy

Start with an inventory of every AI system and AI agent in use — procured, embedded in third-party software, or built in-house — not a review of your AI policy document in isolation. For each system, record its purpose, data inputs, decision authority (recommendation versus autonomous action), and whether it is agentic. This inventory is the backbone referenced in both CSA's lifecycle approach and MAS's expectation that firms maintain AI identification, inventories and risk materiality assessment using impact, complexity and reliance as the basis for proportionate controls.

Each system should then be assigned a risk tier. Low-risk tiers (internal productivity tools with no personal data or customer-facing output) warrant a lighter review; high-risk and agentic systems warrant deeper testing, including the kind of override and human-approval controls IMDA's updated Agentic AI framework now specifies.

2. Gather evidence against the lifecycle, not the policy binder

Evidence should be collected against each stage of the AI lifecycle that CSA sets out — planning and design, development, deployment, operations and end of life — rather than against a static document review. In practice this means:

• Conducting a risk assessment focused on security risks related to each AI system, documented per system rather than at organisation level.
• Assessing and monitoring the security of the AI supply chain, including training data, models, APIs and software libraries.
• Tracking, authenticating, version-controlling and securing AI assets, including models, data, prompts and logs.
• Continuously monitoring and logging inputs and outputs to detect anomalies and potential attacks.
• Developing incident response plans tailored to AI systems, covering a range of potential incidents from minor malfunctions to critical disruptions.

For regulated entities, evidence should also show that AI in the portfolio is re-validated or reviewed where necessary, to ensure it continues to be used appropriately and that associated risks remain appropriately managed. For high-materiality systems, MAS guidance points to further controls: consider implementing kill switches or override mechanisms to rapidly deactivate the AI system if it exceeds risk tolerances.

← Swipe to compare →

3. Produce board-ready deliverables, not just a report

A properly run assessment ends with deliverables a board or risk committee can use without translation: a system-level inventory with risk tiers, a gap analysis mapped against the frameworks relevant to the sector (IMDA, CSA, MAS, PDPC), a prioritised remediation roadmap with owners and dates, and a short board briefing that states residual risk in plain terms. This mirrors what MAS expects of boards directly — that they understand AI enough to provide oversight and challenge rather than receive a technical document they cannot interrogate.

Practical steps for a security lead to prepare

1. Build or refresh the AI system and AI agent inventory before commissioning an external assessment — this shortens the engagement and improves accuracy.
2. Assign a named owner for each AI system, in line with the accountability expectations running through IMDA, CSA and MAS guidance.
3. Classify each system by risk tier using impact, complexity and reliance, not by how the system was procured.
4. Check whether personal data flows through any AI system and map that against PDPC's advisory guidelines separately from the security assessment.
5. If agentic AI is in scope, specifically evidence guardrails, human approval checkpoints and override mechanisms — this is the newest and fastest-moving area of expectation.
6. Decide where ongoing governance sits — a standing committee, a named individual, or a fractional role — before the assessment report lands, so findings have somewhere to go.

How Infracom helps

Infracom runs AI governance assessments for Singapore organisations against IMDA's Model AI Governance Framework, CSA's Guidelines on Securing AI Systems and, where relevant, MAS's AI risk guidance and PDPC's advisory guidelines — scoping the system inventory, gathering lifecycle evidence, and producing board-ready deliverables rather than a generic framework checklist. This work is delivered through our fractional Chief AI Officer service (/grc/vcaio/), with findings often feeding into a broader GRC programme alongside Cyber Essentials or Cyber Trust Mark certification. Infracom governs, assesses and secures AI systems — it does not build or train them. Assessments are delivered by an all-Singaporean SQEP team at SME-accessible rates, with consultants holding the AAISM credential in AI security management.

The AI Governance Pack holds the AI inventory, the risk and impact assessment, the ISO/IEC 27001 to 42001 gap map, the policy set and the board paper as a workbook and document set.