The MAS Guidelines call for risk-based controls across the entire AI lifecycle. Our own governance journey shows why effective control requires more than a checklist: ownership, evidence, testing, monitoring and the ability to intervene must work together.
From policy statements to operational discipline
AI risk changes as data, models, configurations, users and operating environments change. Controls must therefore begin at use-case inception and continue through design, deployment, monitoring, change and retirement. We apply a common internal governance foundation across the firm and use it to support our MAS-regulated service line’s alignment with relevant lifecycle expectations. Based on our experience, the following ten control outcomes should be prioritised.
1. Maintain meaningful human oversight
Human oversight must be designed into the use case, not added as a final sign-off. People assigned to review AI outputs need adequate competence, authority to challenge and intervene, clear escalation routes and awareness of automation bias.
Our implementation example: We require professional review, verification and validation of AI-generated output before use in client deliverables or significant internal decisions. For critical agentic-AI actions, we require explicit human approval.
MAS cross-reference: MAS 4.10
2. Govern data before it reaches AI
Data used by AI should be fit for purpose, representative, accurate, complete, current, appropriately classified and protected. Privacy, intellectual-property, retention and lineage requirements should be addressed across the lifecycle.
Our implementation example: We prohibit client or sensitive organisational data from being entered into public or unapproved AI services. Approved processing must occur in a secure environment or validated private-cloud arrangement, with anonymisation or pseudonymisation where possible.
MAS cross-reference: MAS 4.5
3. Make AI use transparent and outcomes explainable
The level of transparency and explainability should reflect the AI’s risk, autonomy and effect on customers or risk-management outcomes. Higher-impact use cases require clearer rationale, disclosures and channels for redress.
Our implementation example: Our policy requires material AI use in a client engagement to be disclosed where appropriate and favours systems that support explanation of outputs and decisions. AI-derived content must be reviewed and identifiable to reviewers.
MAS cross-reference: MAS 4.6-4.7
4. Define fairness and test for harmful bias
Organisations should define what fair outcomes mean for the use case, identify relevant attributes and test whether AI creates systematic disadvantage. Results and mitigation actions should be documented.
Our implementation example: Our core principles include fairness and non-discrimination. Our sandbox validation requires outputs to be tested for unacceptable bias before production deployment.
MAS cross-reference: MAS 4.8-4.9
5. Assign lifecycle ownership and approval authority
Each AI use case needs an accountable business owner and clear responsibilities across business, technology, risk, compliance, data and cybersecurity. Approval, escalation, intervention and suspension authority should be explicit.
Our implementation example: Our process identifies a sponsor, Head of MIS review, DPO or GRC involvement and AI Governance Committee decisions. The committee may approve, approve with conditions, reject or request more information.
MAS cross-reference: MAS 4.1-4.2; 2.4-2.6
6. Apply due diligence to third-party AI
Using a vendor does not transfer accountability. Due diligence should address transparency, fairness, data handling, security, supply-chain dependencies, concentration, change notification, contingency arrangements and contractual rights.
Our implementation example: Our AI Governance Committee oversees AI vendor due diligence. Approved platforms processing client data must operate with contractual data-protection safeguards, and third-party services are subject to review before use.
MAS cross-reference: MAS 4.11
7. Test before deployment and independently challenge high-risk AI
Testing should use measurable thresholds and realistic conditions, including edge cases and adversarial scenarios. It should address performance, hallucination, harmful content, bias, data leakage and unauthorised actions. High-risk AI should undergo formal independent validation.
Our implementation example: We require secure sandbox testing of functionality, security, data-protection controls, output accuracy and bias. Agentic-AI logging and kill-switch mechanisms must also be validated before final sign-off.
MAS cross-reference: MAS 4.14-4.15; 4.18-4.22
8. Secure the AI environment before go-live
AI should be deployed in secured, well-governed environments with hardened configuration, segregation, encryption, input validation, API security, least privilege, multi-factor authentication, logging and separation of duties.
Our implementation example: We require security by design, strict scoping of agent permissions, comprehensive logging and approved secure environments for sensitive data.
MAS cross-reference: MAS 4.16; 4.22
9. Monitor performance, incidents and changes continuously
Monitoring should cover performance, reliability, data quality, fairness, drift, unusual behaviour and scope adherence. Thresholds should trigger escalation, remediation, override or deactivation. Material changes should be assessed, tested and reapproved.
Our implementation example: We assign ongoing monitoring responsibility to business and MIS functions, require incident escalation, periodic reassessment and change-management controls, and provide for kill switches or overrides for agentic and higher-risk use.
MAS cross-reference: MAS 4.23-4.25
10. Preserve auditability and retire AI safely
An independent reviewer should be able to understand how the AI was selected, developed, tested, approved, monitored and changed. Retirement should address dependencies, data retention, secure removal and stakeholder communication.
Our implementation example: We require retention of relevant prompts and outputs for client work, formal risk and validation records, immutable logging for agentic AI and periodic review of continued suitability. Our current framework provides a strong basis, with secure decommissioning controls representing an area for continued maturation.
MAS cross-reference: MAS 4.17; 4.26
Three management priorities
- Start with higher-risk AI: Focus first on use cases with significant customer, financial, regulatory, operational or data consequences.
- Establish control gates: Require documented risk assessment, testing, security review, independent challenge where appropriate and accountable approval before deployment.
- Demand ongoing assurance: Give management visibility into AI performance, incidents, changes, exceptions and residual risks against risk appetite.
How we can help
We can help financial institutions translate MAS expectations into proportionate, auditable controls through lifecycle framework design, control-gap assessments, use-case risk reviews, third-party AI assessments, pre-deployment assurance, monitoring design, training and internal audit support. Our advice is informed by our implementation experience within the firm and by applying a common governance foundation that supports our service line regulated by MAS.
The central lesson from our journey is that alignment is sustained through an ongoing organisational capability. Policies matter, but trust is created by controls that operate, evidence that can be reviewed, accountable people who know when to intervene, and governance that remains responsive to regulatory and technological change.
-cropped.svg)
