This case study examines how AI assistants perform when queried on regulatory obligations relevant to Treasury functions at Retail Banking firms operating in Singapore. The analysis draws on findings from one regulation — MAS Notice 637 on Capital Adequacy for Banks — covering one aggregated question where AI tools produced inaccurate or unverifiable responses. While the volume of findings is modest, the subject matter sits at the core of Singapore's prudential framework for banks, making accuracy in this area operationally and legally material.
Treasury teams that rely on AI tools for regulatory research in this domain face a specific risk: AI assistants frequently infer structural details of complex prudential notices rather than retrieving them, and hedge their answers in ways that may not be noticed under normal workflow conditions.
Treasury teams at Retail Banking firms in Singapore regularly engage with MAS Notice 637 as part of their day-to-day capital management responsibilities. Common touchpoints include drafting internal capital adequacy policies, producing regulatory capital computations, supporting business line heads who need to understand how new products or exposures will be treated under the capital framework, and preparing training materials for credit and risk staff.
As AI tools become embedded in knowledge-management and research workflows, it is increasingly likely that a Treasury analyst or manager will put a structural question about MAS Notice 637 to an AI assistant — either directly, or via an internal tool that uses AI to surface regulatory guidance.
The corporate use-cases that sit on top of accurate knowledge of MAS Notice 637's structure are substantial. Internal capital adequacy reporting, ICAAP documentation, capital instrument eligibility assessments, and regulatory submissions to MAS all depend on the team correctly understanding which part of the notice governs which obligation. If AI tools are used to identify the applicable divisional requirements — for example, to locate the rules governing particular capital instruments or to confirm reporting thresholds — an incorrect AI response can propagate silently into internal work-products before anyone verifies the source.
If the firm acts on an incorrect AI answer in this space, the consequences fall on the institution, not the individual employee. MAS holds the bank responsible for the accuracy of its regulatory submissions and the robustness of its capital framework. Errors in capital classification or reporting can attract supervisory scrutiny, mandatory remediation, or formal regulatory action. Reputational consequences within the Singapore banking sector — a tightly interconnected community with active regulatory dialogue — compound the direct financial and operational exposure.
The finding in this case study reflects a pattern that is common across dense, technically structured prudential notices: AI assistants fill structural gaps by inference rather than retrieval. When asked about the specific content of a numbered division within a complex regulatory notice, AI tools tend to reason from the surrounding subject matter — extrapolating what a division is likely to cover based on the regulation's general topic — rather than returning a verified answer. The result is a response that sounds authoritative but is not grounded in the actual text.
What makes this failure mode particularly hard to catch is that some AI tools partially acknowledge their uncertainty in hedging language buried within the response, which teams under time pressure may overlook.
All findings in this case study relate to MAS Notice 637 on Capital Adequacy for Banks, which means the error exposure is concentrated in one of the most consequential parts of a retail bank's regulatory framework. Capital adequacy obligations in Singapore carry direct prudential and supervisory weight, and the detailed divisional structure of MAS Notice 637 matters: different divisions impose different requirements on different instrument types and reporting obligations. An AI tool that misidentifies what a specific division covers can send a Treasury team to the wrong part of the notice, causing them to miss or misapply obligations.
The systemic risk to the firm scales with how embedded AI-assisted research has become in the Treasury function. A single incorrect AI answer that is not verified before it enters a capital policy document, an ICAAP narrative, or a regulatory submission can generate a chain of downstream errors across multiple work-products. The cost of correcting those errors — in staff time, regulatory correspondence, and potential re-submission — typically far exceeds the time that would have been saved by using the AI tool in the first place.
1 finding in this case study. Click any to see its full evidence card.
The default position for Treasury teams at Retail Banking firms in Singapore should be that AI tools are a starting point for regulatory research, not a primary source. This is especially true for MAS Notice 637, which is a technically detailed and periodically revised notice where structural accuracy — knowing which division covers which obligation — is as important as understanding the substantive rules. Any AI response that identifies a specific division, article, or schedule of MAS Notice 637 should be verified against the current published text before it is used in any internal work-product, regulatory submission, or training material.
At the firm level, Treasury leadership should establish a clear regulatory-verification policy that names AI-assisted research as requiring human sign-off before the output influences a work-product of record. This policy should require an audit trail — a note or record showing that an AI-sourced regulatory reference was checked against the regulator's published text before use. Firms should also distinguish between "AI-drafted" content (where the AI has generated substantive regulatory characterisations) and "AI-summarised" content (where the AI has distilled material the team can independently verify).
Regulatory-facing documents — ICAAP submissions, capital policy frameworks, MAS correspondence — should require explicit sign-off confirming that any AI-sourced regulatory content has been verified.
There are areas of the Treasury workflow where AI tools are genuinely useful. Drafting non-regulatory narrative sections of internal documents, generating first-draft questions for further human research, and summarising long background documents that the team will then verify are all appropriate uses. The risk arises when AI tools are used to resolve factual questions about what a specific rule or division says — particularly in a complex prudential framework where being one division off can mean applying the wrong obligation entirely.
RegLeg publishes hallucination research as a free resource that Treasury teams can use before relying on any AI-generated answer in regulated areas. For MAS Notice 637 and other Singapore prudential regulations, the research flags the specific question types and structural queries where AI tools have been observed to infer rather than retrieve — giving your team a practical reference for where to apply extra scrutiny. This is not a subscription service: the published findings are available to any Treasury professional who wants to cross-check an AI response before it enters a firm work-product.
For firms that want to go further, RegLeg offers bespoke regulator deep-dives that map which AI-supported workflows in a Retail Banking Treasury function carry the highest hallucination exposure. These engagements are tailored to the firm's actual working practices — the specific points in the capital management, ICAAP, or regulatory reporting cycle where AI tools are in use — and produce a prioritised risk register your team can act on.
RegLeg can also provide a confidential review of your firm's existing AI-use policy, comparing it against our failure-mode catalogue and identifying gaps where the policy does not adequately address the categories of error most relevant to Treasury and prudential regulation in Singapore.
Treasury teams with CPD obligations or internal training requirements can draw on RegLeg's training materials, which are aligned to professional development standards and cover both the failure modes themselves and the verification disciplines that protect against them. Whether the need is a one-off briefing for the Treasury team or structured content that can be embedded in an ongoing training programme, RegLeg can work with the team to develop material that fits your context and your audience.