This case study examines how AI tools perform when Legal teams at Retail Banking firms in Singapore consult them on capital adequacy regulation under the Monetary Authority of Singapore's supervisory framework. Testing covered MAS Notice 637 (Risk Based Capital Adequacy Requirements for Banks), a core prudential instrument that Legal teams in this sector regularly work with. Across two aggregated question areas, AI assistants produced incorrect or misleading responses — in both cases the errors concerned specific regulatory designations or document conventions that carry direct compliance significance.
The findings below document the nature of those failures and their practical consequences for a Legal function operating in Singapore's retail banking environment.
Legal teams at Retail Banking firms in Singapore consult AI tools on capital adequacy topics in a range of routine workflows: scoping the regulatory perimeter when a firm reviews its group structure or considers establishing a financial holding company, drafting internal compliance policies that reference the applicable MAS notices, and preparing regulatory mapping documents for new product launches or business line expansions.
AI tools are also frequently used to generate first-draft training materials for non-Legal staff, to brief senior management or board committees on the regulatory framework, and to support responses to business lines that come to Legal with threshold questions about which rules apply to which entities.
The two finding areas in this study sit squarely in that territory. Questions about which notice applies to a financial holding company arise whenever a firm's group structure is under review, when a new holding entity is being incorporated, or when Legal is advising on the perimeter of MAS supervisory obligations. Questions about the meaning of highlighting conventions in an amendment notice arise whenever Legal is working from the amendment document directly — for example, when preparing a gap analysis against existing internal policies, or when advising on effective dates for newly introduced provisions.
If the AI's answer is wrong in either of these contexts, the consequences fall on the firm, not the individual. Acting on a fabricated regulatory reference — such as a notice designation that does not exist — can result in a firm misjudging the applicable capital adequacy framework for an entity in its group, which carries the risk of regulatory breach and MAS supervisory action. Misreading the significance of highlighted provisions in an amendment document can cause a firm to mis-sequence implementation, applying provisions from the wrong operative date or missing that certain requirements take effect on a different timetable.
Both pathways carry the risk of MAS enforcement, remediation costs, and reputational consequences in a jurisdiction where regulatory standing is a core business asset.
Both findings in this study concern MAS Notice 637, and both involve AI tools substituting plausible-sounding but incorrect information for precise regulatory content. In the first finding, an AI tool constructed a notice designation that does not exist — reasoning by analogy from the bank notice number rather than from verified source material. In the second, multiple AI tools offered generic or inferred explanations for a specific document convention, replacing a technically precise drafting signal with vague descriptions that would mislead any reader relying on the AI to understand the amendment document.
The pattern across both findings is the same: AI tools fill gaps in their knowledge with confident, structurally coherent answers that look authoritative but are wrong on the specific detail that matters.
Both errors cluster on the same regulation and the same regulator, which is significant. It suggests that AI tools' coverage of MAS Notice 637 — including its amendment instruments and the group-level entities it sits alongside — is unreliable in ways that are not visible from the surface quality of the AI's response. A Legal team that has no prior reason to doubt the AI's answer on one of these questions is unlikely to spot the error without independent verification against the MAS source materials.
The systemic risk to a Retail Banking firm is amplified by how these errors propagate. A wrong notice designation inserted into a regulatory mapping document, a policy, or a board briefing does not stay in one place — it gets cited, incorporated into downstream work-products, and relied upon by other teams. A misunderstood highlighting convention in an amendment document can corrupt an entire gap-analysis exercise, causing the firm to sequence implementation incorrectly across multiple provisions. When a single AI response is the shared input to several parallel work-streams, the cost of the error multiplies with each downstream use.
2 findings in this case study. Click any to see its full evidence card.
The default position for any Legal team in Singapore's retail banking sector should be that AI tools are a starting point, not a primary source, for regulatory questions touching MAS notices, their designated scope, or the technical conventions of amendment instruments. The findings in this study illustrate that AI tools can produce responses that are structurally sound and confidently expressed while being wrong on the specific detail — a notice designation, an effective-date convention — that determines whether the firm's analysis is correct.
Where the error is not visible from the surface of the response, the risk of acting on it without verification is material.
At the firm level, practical safeguards should include a documented regulatory-verification policy that explicitly identifies AI as an unreliable source for questions about MAS notice designations, scope of application, and document-drafting conventions, and that requires independent verification against published MAS materials before any AI output on these topics enters a work-product. Any AI output that influences a regulatory mapping document, internal policy, board briefing, or gap analysis should be logged in an audit trail, with a record of the verification step taken.
Work-products that are AI-assisted should be clearly distinguished from those that are AI-drafted, and sign-off requirements should apply before AI-assisted content is relied upon in regulatory-facing or firm-wide material.
AI tools remain useful in the Legal workflow for tasks where the risk of undetectable error is lower: drafting non-regulatory communications, generating first-draft questions for further research, summarising long documents where the team will verify the substance independently, or producing initial checklists that experienced Legal staff will review and correct. The key discipline is matching the level of verification to the regulatory consequence of getting the answer wrong.
RegLeg's published hallucination research is available as a free reference check for Legal teams before they rely on any AI-generated answer in the MAS Notice 637 area or the broader Singapore capital adequacy framework. The research documents the specific question areas where AI tools have produced incorrect responses, the nature of the errors, and their regulatory consequence — giving a Legal team a concrete basis for deciding which AI outputs warrant closer scrutiny before use.
For firms that want a more structured picture of their exposure, RegLeg offers bespoke regulator deep-dives that map which AI-supported workflows in a Retail Banking Legal function carry the highest hallucination risk across Singapore's regulatory landscape. These are practical working sessions, not general briefings — they are scoped to the firm's actual regulatory perimeter, its existing use of AI tools, and the specific MAS frameworks its Legal team works with most frequently. The output is a prioritised picture of where verification discipline matters most and where AI use can proceed with lighter oversight.
RegLeg can also conduct a confidential review of a firm's existing AI-use policy against our failure-mode catalogue, identifying gaps in how the policy treats AI output on regulatory questions and providing prioritised remediation recommendations. For firms building out internal capability, we produce CPD-aligned training content that Legal teams can use directly — covering AI limitation patterns, verification workflows, and practical guidance for staff who use AI tools in regulatory research and drafting. All engagement is conducted on a confidential basis and is designed to complement, not replace, the firm's existing compliance and legal risk framework.