MAS and AI in financial institutions
AI Model Risk Management §6.4 — Observations from a Thematic Review
Banks use “validation” and “review” interchangeably, and the MAS information paper says so at footnote 18. It records them as two different things: independent validation is an objective assessment by someone with no stake in the outcome, and a peer review is a colleague who did not build the model. The label a firm uses does not settle which one it holds.
§ 1 — Who it binds
Who this reaches, and on whose rating
Financial institution
The information paper binds nobody: it records observed practice among selected banks and extends its relevance to other financial institutions by encouragement at §1.2. Which tier a given system falls into is set by the firm’s own risk materiality rating, so the reach of this expectation is only as sound as the assessment producing it. Nothing in the paper fixes the threshold between independent validation and peer review; that boundary is the firm’s to draw and to record.
§ 2 — In practice
Independence is about stake, not reporting line
§6.4.1 states what independent validation is for — an objective and unbiased assessment of the suitability, performance and limitations of AI, acting as an important challenge to developers and establishing that the relevant standards and processes were adhered to during development. §6.4.4 records the range MAS found. One bank required all AI to be independently validated, with the depth and rigour varying by risk materiality rating. Most others required independent validation only above a materiality threshold, with the remainder subject to peer review — and even below it, the involvement of either a validator or a peer reviewer allowed some degree of challenge. Two things follow. The tier boundary belongs to the firm and has to be written down. And neither tier is nothing: the paper records no population going unchallenged. MAS has since proposed a two-tier arrangement of this kind as an expectation, at paragraphs 4.19 and 4.20 of the draft Guidelines — read there in its agentic setting. That consultation closed on 31 January 2026 and the Guidelines had not been issued as at 30 August 2026.
Footnote 65 anchors independence to a formulation turning on interest rather than structure: validation performed by individuals not responsible for development or use, who do not have a stake in whether a model is determined to be valid. A reporting line into risk does not by itself satisfy that. Someone sitting inside the business unit whose profit and loss the model serves has a stake; so does someone whose own objectives include the delivery date. The boundary case the two-tier split creates is the one nobody staffs — a system just under the threshold receives a peer review from a colleague uninvolved in development, and in a team of four the population of qualified, uninvolved colleagues is frequently empty. Uninvolved has to mean uninvolved in ongoing deployment as well, not merely absent from the build.
§6.4.2 gives validators something firms often do not honour. Where issues are found, the adjustments or compensating controls are typically proposed by developers and agreed to by validators before deploying AI — an agreement condition, not an advisory opinion. Where deployment can proceed over an open finding, the challenge function is a recommendation function whatever the policy calls it. The other half of the problem is evidential: a validation report supports reliance only if somebody else could work from it and arrive at the same conclusion, which is the reperformance bar set out on evidence sufficiency for AI controls. §6.4.3 supplies the competence limb — banks invested so that validation staff held the relevant technical expertise for AI — and that is about method rather than awareness. The question is whether the reviewer could have found the defect, which is what auditor competence turns on too.
§ 3 — What discharges it
What the validation file must show
The artefacts an assessor asks to see, and what makes each one sufficient rather than merely present.
01
The validation standard, with the tier boundary written down
Which tier applies at which materiality, and what each tier requires of the reviewer. An unwritten boundary is settled case by case by whoever has capacity that month, which is the opposite of a standard.
02
One independent validation report with the stake declared
For a high-materiality system, stating the reviewer’s reporting line, their involvement in development, and their involvement in ongoing deployment on the face of the report. Independence asserted in a policy is not independence evidenced in a file.
03
A finding the development team did not simply accept
The disagreement, the remediation proposed, and the record of the reviewer agreeing it before the system was deployed. A history with no contested finding evidences either a very good team or a very quiet one.
04
The exception log for systems deployed without their tier’s review
Which systems went live without what their rating required, who approved that, and what compensating control was applied instead. Exceptions are where a standard is shown to be real or decorative.
§ 4 — Worked example
Worked example — a reviewer who owns two features
A bank’s model risk policy requires independent validation above a materiality threshold and peer review below it. A pricing model rated just below the line is reviewed by a senior data scientist in the same eight-person team, who did not write the code but attends the model’s fortnightly performance meeting and maintains two of its features in the shared repository. Her report identifies a data leakage issue, recommends a change, and notes that the team will address it in the next release. The model went live on schedule.
Which of the two things in the policy has actually happened?
Neither, cleanly. It is not independent validation: the reviewer has a stake in the model being determined valid, because she attends its performance meetings and maintains part of it, so the interest limb of footnote 65’s formulation is not met whatever the reporting line says. It is not quite peer review either, since uninvolved has to reach ongoing deployment and feature ownership is involvement. The second failure is the deployment decision. The remediation was meant to be agreed by the reviewer before the system went live; instead it was deferred to a future release and the model shipped, so the challenge produced a recommendation rather than a condition. The fix is not more process. It is a written tier boundary saying what each materiality buys, a declared statement of stake on the front of every review, and a rule that an open finding blocks a release unless a named person accepts it in writing.
§ 5 — What a weak answer looks like
A challenge function that has never challenged
A validation function that has never stopped anything. The reports are thorough, the templates complete, the coverage against the model inventory total, and every one of them concludes that the model is fit for its intended use subject to recommendations. Asked when a finding last held up a release, the firm cannot name an occasion; asked what would happen if one did, it describes an escalation route nobody has used. Challenge that has never been exercised is indistinguishable, from outside, from challenge that could not be.
§ 6 — Elsewhere
The same independence question elsewhere
Where another instrument addresses the same obligation. These are correspondences, not comparisons — the Council does not rank one framework against another.
IIA Standard 14.1 fixes sufficiency to a reperformance test. It is the bar a validation file has to meet before anyone else — internal audit, a supervisor, a group function — can rely on its conclusion.
That page addresses competence for the person performing the work. §6.4.3 records banks investing so that validation staff held technical expertise specific to AI, which is the same requirement on the validation side.
The draft Guidelines address formal independent validation and the reviewer’s involvement in ongoing deployment. That page carries the draft’s treatment; this one carries the published observations it would consolidate.
A correspondence indicates that two instruments address the same underlying obligation. It is not a mapping endorsed by either body, not a statement that one satisfies the other, and not a judgement about which is more demanding.
§ 7 — When it applies
A publication date, not a deadline
5 December 2024
The publication date of the information paper — the day the observations appeared, not a commencement date. Nothing was required of any institution on it, and an information paper has no legal effect in MAS’ instrument taxonomy.
§ 8 — provenance
The provision itself
This page sets out what the instrument requires and what discharges it. The official text is the authority — these go straight to it.
- Information paper, Artificial Intelligence Model Risk Management — Observations from a Thematic Review, December 2024. Section 6.4 Validation, with footnote 18 on interchangeable usage and footnote 65 on the independence formulation.
- Landing page publication date, 5 December 2024. The paper states no effective date, because it imposes no duty.
- Consultation landing page for P017-2025, published 13 November 2025 and closed 31 January 2026. The two-tier validation proposal sits at paragraphs 4.19 and 4.20 of the draft Guidelines at section 6.
§ 9 — Also read
Also read
Evidence sufficiency for AI controls — IIA Standard 14.1
Certification
Assessed on the same standard of evidence
Every Council credential is examined on applied judgement against a published anchor, set out the way the obligations on this page are. The Certified Applied AI Associate, the Council’s flagship, is open to any adult today, its fee waived by a coupon code from the AI-Readiness Check, and the register lists what is open for enrolment.