MAS and AI in financial institutions
AI Model Risk Management §5.3 — Observations from a Thematic Review
MAS published this as an information paper rather than a Guideline: it records what a thematic review of selected banks found, and obliges nobody. What §5.3 records is a three-dimension materiality assessment — impact, complexity and reliance — which MAS later proposed as an expectation in its draft Guidelines. Reliance is the dimension firms omit.
§ 1 — Who it binds
What the paper reaches, and what it obliges
Financial institution
The paper binds nobody. It is scoped to AI model risk management and draws its own boundary at §3.5: data governance, technology and cyber risk, and third-party risk are not covered here, and the existing notices, guidelines and information papers on those subjects apply instead. Within that scope it reaches AI as the firm itself defines AI, across whatever population the firm’s identification process catches — which makes this assessment the gateway to every control downstream of it.
§ 2 — In practice
Three dimensions, and the one that changes answers
Status first, because the paper is routinely described as guidance. An information paper does not appear in MAS’ instrument taxonomy at all; it is a publication, and this one records observations from a thematic review of selected banks. Its own framing is that MAS encourages all financial institutions to reference these good practices when developing and deploying AI, and at §1.2 that the practices should generally apply beyond the institutions reviewed. Nothing in it is enforceable and non-observance carries no consequence of its own. It does carry weight of another kind: this is what MAS’ supervisors saw across a population of banks and elected to publish, which is a reasonable proxy for what a supervisor expects to be shown when the subject comes up.
§5.3.1 states the purpose plainly — the assessments are what let a firm calibrate its approach across the diverse areas AI is used in, mapping risk materiality to the depth and scope of validation and monitoring required. Most banks used quantitative and qualitative dimensions in three categories: impact on the bank, its customers or other stakeholders, including financial, operational, regulatory and reputational impact; complexity arising from the nature of the model or system, or the novelty of the area or use case; and reliance on the AI, taking account of the autonomy granted to it, or the involvement of humans in the loop as risk mitigants. The same three appear in the Guidelines MAS put to consultation, at paragraph 3.10 of the draft — so a firm with a working methodology today is nearer the proposed expectation than one waiting for issuance, and rather few know it.
Impact and complexity are familiar from conventional model risk management and travel across without difficulty. Reliance does not, and methodologies inherited from a model risk framework frequently arrive without it. It is the dimension separating a model that recommends from one that decides, and its absence is testable arithmetically: take a single use case, hold impact and complexity fixed, remove the human reviewer, and see whether the rating moves. Where it does not, the methodology has three column headings and two dimensions. This is the same axis the built page on MAS supervisory expectations for agentic AI turns on, where raising an agent’s autonomy is what changes where it sits.
§5.3.2 records two review activities where firms usually run one. Banks established processes to confirm that the materialities assigned remain appropriate over time — the annual re-rating everybody does — and also reviewed the measures and methods themselves, the example given being financial-impact measures updated when the business changed. A methodology written in 2023 against a book that has since doubled will keep producing ratings and will have quietly stopped discriminating between them. The distribution is the diagnostic: where nearly every system lands in one band, the assessment has become a formality with a scoring rubric attached, and the shape of the population shows it faster than any review of the rubric would.
Two further confusions are worth naming. Materiality is not risk appetite. Materiality decides how much control a system attracts; appetite decides whether what remains is acceptable, and a firm running one scale for both cannot tell an assessor which question it answered. And the rating has to do something: §5.3.1’s own formulation maps materiality onto validation depth and monitoring scope, so a band changing neither the validation tier, the monitoring frequency, the documentation standard nor the approval level is a label rather than a calibration. It is also worth being clear that this is a methodology and not a legal test — a mapping onto the listed categories in Article 6 answers a classification question somebody else set, and a firm that has done one has not thereby done the other.
§ 3 — When it applies
Two publication dates, and no duty
5 December 2024
Publication date of the information paper. Not a commencement date: an information paper has no legal effect, nothing was required of any institution on this date, and the observations are published as good practice MAS encourages rather than as a duty.
13 November 2025
Publication of Consultation Paper P017-2025, which proposed the same three dimensions as an expectation at paragraph 3.10 of the draft Guidelines. The consultation closed on 31 January 2026 and the Guidelines had not been issued as at 30 August 2026. A proposal, not a duty.
§ 4 — What discharges it
What the assessment must produce
The artefacts an assessor asks to see, and what makes each one sufficient rather than merely present.
01
The methodology with all three dimensions as scored inputs
Written scoring rules rather than the standing judgement of the model risk team, with each of the three producing a score traceable into the rating. A dimension appearing in the policy and not in the calculation has not been implemented.
02
Two assessments differing only in autonomy
Comparable use cases with the first two dimensions held constant and the human reviewer present in one and absent in the other, showing the rating move. It is the shortest available demonstration that the third dimension does work.
03
The mapping from band to control intensity
What each band actually buys: validation tier, monitoring frequency, documentation standard, approval level. Without it the rating is a label, and the assessment cannot be said to calibrate anything at all.
04
The review of the measures themselves, not the ratings
What changed in the scoring measures, when, and why — kept distinct from the annual re-rating of systems. Most firms can produce the second immediately and have never performed the first.
05
The distribution of ratings across the AI population
A count by band across the whole estate. Where nearly everything sits in one band the method has not discriminated, and the shape of the distribution says so before anyone reads the rubric.
06
A re-rating triggered by a change in the system
One case where a rating moved because the system changed — autonomy raised, population widened, a human step removed — rather than because an annual cycle came round. It is the difference between a live control and a calendar entry.
§ 5 — Worked example
Worked example — two models, one rating
A bank runs two credit-decisioning models on one portfolio. Model A produces a score that a credit officer reviews before every decision. Model B, introduced last year, approves applications below a threshold outright and refers the rest. Both use the same algorithm on the same data with the same expected financial impact. The bank’s methodology scores impact and complexity, both taken from the model documentation, and rates both models medium.
What has the methodology failed to measure?
Reliance — and it is the only respect in which the two differ. Model B has been granted autonomy over a class of decisions Model A refers to a person, so the human mitigant present in one has been removed from part of the other’s population, which is exactly what the third dimension exists to capture. Holding the first two constant and varying only the autonomy ought to move the rating; here it does not, which shows the dimension is missing rather than that the two models happen to be equivalent. The consequence is not the rating but everything downstream of it: if medium buys the same validation depth, the same monitoring frequency and the same approval level for both, the model that decides alone is being controlled as though somebody were still checking it. Producing two assessments that differ only in autonomy is the quickest demonstration either way, and it is what an assessor will ask for.
§ 6 — What a weak answer looks like
A rating that buys nothing
A rating produced once, at the point of approval, and never asked to earn its place. The method is documented, the scores are complete, and every system in the estate was assessed on the day it was proposed. Nothing has been re-scored since, no measure has been revisited, and the bands map to nothing an assessor can inspect: the same validation and the same monitoring apply whichever band a system falls in. It reads as a control and behaves as an intake form, and the give-away is that no rating has ever changed anything about how a system is run.
§ 7 — Elsewhere
Classification questions elsewhere
Where another instrument addresses the same obligation. These are correspondences, not comparisons — the Council does not rank one framework against another.
That page addresses the draft Guidelines’ agentic reading, including how the autonomy granted to an agent bears on where it sits. This page addresses the published assessment method the draft would consolidate.
Article 6 classifies a system by listed use case as a matter of law. The assessment recorded here is a methodology a firm sets for itself, and the two answer different questions about the same estate.
Article 9 requires a continuous risk management process across a high-risk system’s lifecycle. Materiality is what decides how much of such a process any individual system receives.
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.
§ 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 5.3, risk materiality assessment, including the three dimensions and the review of measures at §5.3.2.
- Landing page publication date, 5 December 2024. The paper carries no effective date because it imposes nothing.
- Consultation landing page for P017-2025: published 13 November 2025, closing 31 January 2026, status closed. The draft Guidelines sit at section 6 of the paper with their own numbering.
§ 9 — Also read
Also read
SAFR — runtime governance for agentic finance
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.