This case study examines how AI tools respond to regulatory questions relevant to Finance teams at Retail Banking firms operating in Singapore. The testing covered MAS Notice 637 on Capital Adequacy for Banks, one of the foundational prudential frameworks that Finance functions must navigate accurately in their day-to-day capital reporting and compliance work. Across the aggregated questions tested, AI assistants produced at least one materially unreliable answer on this regulation — an answer that a Finance professional consulting an AI tool could plausibly act on without realising it was wrong.
The findings below detail where those failures occurred, how the AI behaved, and what the practical consequences are for a firm whose Finance team relies on AI-assisted research in this regulatory space.
Finance teams at Singapore retail banks encounter MAS Notice 637 routinely — in capital adequacy calculations, ICAAP documentation, regulatory reporting to MAS, and the internal policy work that underpins those processes. When a new product is being assessed, when reporting templates are updated, or when a business line asks Finance to clarify what capital treatment applies to a particular instrument or exposure, it is entirely natural for a team member to turn to an AI tool for a quick structural overview of the relevant Notice.
The same happens during onboarding of new Finance staff, when preparing training materials on capital rules, or when a senior manager needs a briefing on how a particular Part or Division of the Notice applies to a proposed transaction.
AI tools are also consulted when Finance teams are building regulatory mappings — cross-referencing Notice requirements against internal policies, or scoping what a regulatory change means for existing procedures. In these contexts, the AI's answer is not merely informational: it seeds a document that may be reviewed by Compliance, signed off by a CFO, or submitted as part of MAS supervisory engagement. The further downstream that AI-originated content travels, the more embedded the error becomes.
If the AI's answer is wrong — for instance, mischaracterising what a specific Division of Part VI covers — the firm risks building capital-reporting procedures on a false understanding of the regulatory framework. MAS has broad supervisory powers under the Banking Act and Notice 637 itself, including the ability to require remediation, impose additional capital buffers, or take formal enforcement action where reporting failures are identified. Beyond regulatory consequences, errors in capital adequacy processes carry direct financial risk: miscalculated capital ratios, misstated regulatory returns, and the remediation costs that follow.
The Finance team does not bear personal liability for an AI tool's error, but the firm, its leadership, and its capital position absorb the consequences in full.
The finding in this case study illustrates a failure mode that is particularly difficult for Finance teams to detect: the AI tool produced an answer that was plausible in structure and confident in tone, but its factual basis was uncertain. Rather than declining to answer or clearly flagging that the specific divisional content of Part VI could not be confirmed, the AI offered a characterisation — that Division 4 covers capital instrument submission requirements — while burying a qualification so hedged as to be easy to overlook.
This pattern, where an AI tool fills a gap in its knowledge with a reasonable-sounding inference rather than an honest admission of uncertainty, is especially hazardous in technical regulatory content where the structural detail of a Notice matters as much as its general subject matter.
The errors here cluster specifically around the internal architecture of MAS Notice 637 — which Division covers what, and what obligations sit where within the Notice's Parts. This is precisely the level of granularity that Finance professionals need when they are navigating the Notice in practice: not just knowing that Part VI relates to capital definitions, but knowing which Division governs which aspect of that framework. An AI tool that conflates structural inference with retrieved fact gives the team a false sense of navigational certainty.
The systemic risk to the firm compounds quickly. A single AI-generated mischaracterisation of Notice structure, if unchecked, can propagate into internal guidance notes, training materials, and regulatory mapping documents. Finance staff who consult those materials in turn have no reason to re-verify the original AI claim. By the time an error surfaces — in a supervisory review, an internal audit, or a reporting discrepancy — it may have influenced multiple downstream work-products, each of which then requires separate remediation. The cost is not just the correction itself; it is the credibility cost of having produced materially inaccurate regulatory documentation.
1 finding in this case study. Click any to see its full evidence card.
The default position for Finance teams at Singapore retail banks should be that AI tools are a starting point for orientation, not a primary or authoritative source for any regulatory question that touches the specific text, structure, or requirements of MAS notices. This applies with particular force to structural questions about Notice architecture — which Part covers what, which Division governs which obligation — because these are exactly the questions where AI tools are most likely to produce plausible but unverified inferences.
Any AI-generated summary of a regulatory notice should be treated as a first-draft prompt for further research, not a reliable answer in itself.
At the firm level, Finance leadership should consider establishing a short-form regulatory-verification policy that explicitly names AI tools as an unreliable primary source for MAS notice content, and that sets out the verification steps required before AI output influences any firm work-product. Practically, this means maintaining an audit trail for AI-assisted regulatory research, requiring a sign-off step before AI-drafted or AI-summarised content enters internal guidance, regulatory mappings, or materials that will be reviewed by Compliance or senior management.
Where AI output is used, the work-product should clearly distinguish between content the team has independently verified against the published notice and content that originated from an AI summary.
There are areas where AI tools do add genuine value in the Finance workflow without the same verification burden: drafting non-regulatory correspondence, generating a structured list of questions for a Compliance or legal review, or producing a first-draft summary of a long internal document that the team will then check directly. The key distinction is whether the AI's answer needs to be factually correct about specific regulatory text. Where it does, verify against the MAS-published source before relying on it.
RegLeg's published hallucination research gives Finance teams at Singapore retail banks a free, practical reference before relying on any AI answer in MAS-regulated areas. The research maps specific regulatory questions — including those drawn from MAS Notice 637 — against the actual responses AI tools produce, so your team can see in advance where the failure points are, what a wrong answer looks like, and how to spot the hedging patterns that distinguish an AI inference from a retrieved fact.
Using this research as a pre-check when onboarding AI tools into Finance workflows costs nothing and can prevent the kind of downstream propagation described in this case study.
For firms that want a more structured view of their exposure, RegLeg offers bespoke regulator deep-dives tailored to the workflows specific to retail banking Finance functions in Singapore. These engagements map which AI-supported processes in your team — capital reporting, regulatory mapping, policy drafting, training material production — carry the highest hallucination risk given the regulations you work with most, and produce a prioritised view of where human verification is non-negotiable.
RegLeg can also conduct a confidential review of your firm's existing AI-use policy against our failure-mode catalogue, identifying gaps and providing prioritised remediation recommendations that are proportionate to your actual workflow risk.
Finance teams that want to build internal capability can draw on RegLeg's training material and CPD-aligned content, designed to help practitioners understand how AI tools fail on regulatory content, what the warning signs look like in practice, and how to build verification habits into team workflows without adding disproportionate overhead. The goal is a Finance function that uses AI tools confidently and accurately — knowing where to trust them, and where to check.