Responsible AI in Financial Services: What MAS Expects Your Firm to Evidence

The MAS Instruments Behind Responsible AI in Financial Services

A firm that treats every row as optional misses the Notices in the last row, and a firm that treats every row as law overstates what MAS has actually issued.

Why Adopting FEAT Falls Short of Evidencing It

What MAS found in its 2024 review of banks’ AI model risk management

MAS reviewed banks’ AI and generative AI model risk management in mid-2024 and published the good practices it observed on 5 December 2024. The Information Paper on AI Model Risk Management groups those practices under three headings: governance and oversight, key risk management systems and processes, and the development and deployment of AI. The review looked at banks, but MAS “encourages all financial institutions to reference these good practices”, and MAS sent the paper to insurers the same day under circular ID 18/24. The November 2025 consultation announcement states that the proposed Guidelines “build on MAS’ supervisory thematic review of key banks’ use of AI in 2024”, so the information paper is the plainest preview of what supervisors look for.

Model Risk, Explainability, Oversight and Fairness: What a Firm Evidences

The 14 rows fall into four groups, with Fairness holding four principles, Ethics two, Accountability five and Transparency three. Ownership runs through the evidence column, because every record names a person or a committee, and a record with no owner answers none of the principles.

Model risk: the inventory and the materiality score

Model risk management for AI starts with knowing every AI use in the firm, which is why MAS’s proposal puts identification and an up-to-date inventory ahead of any individual control. The November 2025 announcement asks firms to rate each use for risk materiality across three dimensions: impact, complexity and reliance. For example, a customer-facing credit model that staff rely on without challenge rates high on impact and reliance, while an internal drafting assistant reviewed line by line rates lower, and the depth of testing, monitoring and approval scales with that rating. UK data shows how ratings spread in practice: the Bank of England and FCA survey of 21 November 2024 found firms rate 62% of their AI use cases low materiality and 16% high materiality. In Synthesised.ai’s view, the inventory works best as the same register the model risk function already keeps, which the firm extends to cover generative AI and vendor models, and it complements ISO/IEC 42001, the AI management system standard that some firms adopt voluntarily as a management layer.

Explainability the customer can use

FEAT calls for explainability to the customer on request: principles 13 and 14 ask for clear explanations of the data behind a decision, how that data affected it and what the decision means for the person. FEAT also balances transparency against confusion and the risk of individuals exploiting or manipulating models, with the materiality of the decision relevant to the level of transparency (paragraph 8.1); its fraud-detection illustration (paragraph 8.4) describes a firm deciding not to provide explanations about a fraud model for those reasons. That standard sits closer to a reason-code letter than to a model interpretability chart. A firm evidences it with an explanation template per decision type, a log of requests answered, and a disclosure line (principle 12) telling customers AI is in use. Internal explainability for validators and auditors runs alongside and uses different artefacts, such as model documentation and feature importance reports.

Human oversight and internal approval

Human oversight in FEAT terms starts with a named internal authority approving each AI-driven decision process (principle 7) and ends with the firm telling management and the Board what AI it runs (principle 9). The proposed Guidelines list human oversight as a lifecycle control that firms apply in proportion to materiality, so in an illustrative design a high-materiality use carries a human checkpoint before a decision takes effect and a low-materiality use carries sampling review. The approval record belongs in the inventory entry, next to the materiality rating behind it.

Fairness testing and bias review

Fairness under FEAT means an AI-driven decision puts no group at a systematic disadvantage unless the firm justifies it, and the firm reviews data and models regularly to minimise unintentional bias. The justification test matters: some differences in outcome are legitimate, such as a credit decision that tracks verified income, and FEAT asks the firm to record the justification for a difference rather than to remove every difference. Personal attributes used as inputs need their own recorded reason under principle 2.

How do banks test AI models for fairness?

Banks test AI models for fairness by naming the groups a decision affects, choosing metrics that fit the use case and checking results against thresholds the firm justifies. The Veritas FEAT Fairness Principles Assessment Methodology, from the consortium’s first phase to 6 January 2021, sets out that method with worked credit risk scoring and customer marketing cases, according to the MAS Veritas page. A fairness result carries weight when the firm fixed the metric, the groups and the threshold before the test ran.

How Veritas Turns the FEAT Principles Into an Assessment

The documents are MAS-hosted PDFs on the Veritas page; the step labels and order are Synthesised.ai’s implementation summary, following Document 3’s advice to start with ethics and accountability and to run fairness and transparency together.

Step 1: Scopes the use case and the decision it drives

The assessment starts by fixing the use case, the decision the model drives and the individuals or groups affected, because every later test depends on that boundary. A credit scoring model and a marketing propensity model need different affected-group definitions, and the Veritas case studies treat them as separate use cases.

Step 2: Assesses ethics and accountability

The ethics and accountability step checks that the use matches the firm’s own standards and that a named authority owns it, covering principles 5 to 11. UOB, AXA and Accenture developed this methodology using customer marketing and insurance fraud detection cases.

Step 3: Assesses fairness

The fairness step tests FEAT principles 1 to 4 against metrics chosen for the use case and records the justification for any remaining difference in outcomes. In Phase 2, Swiss Re and Accenture developed the fairness methodology for insurance predictive underwriting.

Step 4: Assesses transparency

The transparency step, which runs alongside the fairness step, designs the disclosure and explanation the firm gives customers and internal reviewers, covering principles 12 to 14. Standard Chartered, HSBC and Truera developed the transparency methodology on credit risk scoring and customer marketing.

Step 5: Integrates the assessment into existing governance

The final step folds the Veritas assessment into the firm’s existing model risk and governance processes so it repeats with each material change. Phase 3’s pilots at seven institutions tested this integration, and MAS reports working with solution providers such as IBM and SAS to integrate the toolkit into their AI products. The same build-stage discipline, building controls in rather than reviewing them at the end, drives responsible AI implementation, where bias controls, oversight and audit trails go into the build.

Red Lines That Stop a Financial AI Deployment

A consequential decision with no approving authority

An AI-driven decision process with no named internal approver falls short of FEAT principle 7 before any model question arises. The fix is administrative: name the authority, record the approval against the inventory entry, and tie it to the materiality rating.

A decision the customer cannot question

A customer-affecting AI decision with no enquiry, appeal or review route falls short of FEAT principle 10, and one with no explanation on request falls short of principle 13 unless the firm has recorded a materiality or anti-manipulation reason for withholding it, as FEAT’s fraud-detection illustration describes. The review route also has to take account of verified new information the customer supplies, under principle 11.

A vendor model nobody inside the firm owns

A vendor model with no internal owner falls short of FEAT principle 8, which holds firms accountable for externally sourced models on the same footing as models built in-house. The proposed Guidelines list third-party risks as a lifecycle control, and the MindForge AI Risk Management Toolkit, which a consortium of 24 banks, insurers, capital market firms and other industry partners developed and MAS announced on 20 March 2026, organises its Operationalisation Handbook into four sections that align with the proposed Guidelines.

Can a bank outsource responsibility for an AI model?

A bank keeps responsibility for an AI model it buys: FEAT principle 8 holds firms accountable for both internally developed and externally sourced models. Where a vendor arrangement counts as an outsourced relevant service, MAS Notice 658, effective 11 December 2024, imposes binding risk-management requirements on banks in Singapore. Whether a given AI vendor arrangement falls within the Notice turns on its definitions, which a bank’s legal and outsourcing functions confirm case by case. The vendor supplies the model; the bank keeps the accountability.

Turn the Principles Into Evidence Before the Guidelines Are Final