This case study documents findings from testing AI tools against regulatory questions relevant to Accountants (CA/PA) practising in Singapore. Across one regulation — MAS Notice 637 on Risk-Based Capital Adequacy Requirements for Banks — AI assistants produced incorrect, incomplete, or fabricated answers on five aggregated questions. The errors ranged from invented regulatory instrument names to inaccurate descriptions of specific annex content and document-formatting conventions. Accountants (CA/PA) who rely on AI tools to support capital adequacy advisory work, client compliance reviews, or regulatory correspondence face a material risk of acting on answers that cannot be verified against the regulator's published text.
Accountants (CA/PA) in Singapore routinely engage with MAS capital adequacy requirements when advising bank clients, supporting group-level consolidation work, preparing or reviewing regulatory capital disclosures, or assisting with due diligence on financial institutions. AI tools are increasingly used at the research and drafting stage — to answer quick-reference questions about regulatory scope, to check what a specific annex or division of a notice covers, or to understand how a recent amendment interacts with the existing consolidated notice.
Each of the five questions documented here represents exactly the kind of task a practitioner would reasonably delegate to an AI assistant to save time.
The use-cases that sit directly on top of these topics include: scoping whether a holding-company structure is captured by MAS Notice 637 or a separate instrument, briefing a client finance team on what an amendment notice changes, preparing internal training materials on the capital adequacy framework, and reviewing a bank's disclosures against the correct regulatory text. These are tasks where accuracy is not optional — a wrong answer about which notice applies, or what a specific annex covers, can flow silently into client advice, compliance memos, or board-level reporting.
If an AI-generated answer is taken at face value and turns out to be incorrect, the consequences for the practitioner can include professional liability exposure and disciplinary risk with ISCA, the Singapore Accountancy Commission, or MAS itself if the error influences a regulated deliverable. For the practitioner's clients — particularly banks and financial holding companies — a compliance gap traced back to incorrect regulatory guidance can trigger supervisory scrutiny, capital adequacy breaches, or enforcement consequences under the Banking Act.
Restoring client trust after an error of this kind is difficult, and the reputational cost to the advisory firm can be disproportionate to the original research shortcut.
Across the five findings, a consistent pattern emerges: AI tools filled gaps in their knowledge of MAS Notice 637 not with an acknowledgement of uncertainty, but with plausible-sounding inferences constructed from adjacent knowledge. The most striking example is a fabricated regulatory instrument name — an AI tool invented a specific notice designation for financial holding companies by analogy with the bank notice number, producing a reference that does not exist in the MAS register.
In other findings, AI tools described the purpose of document-formatting conventions (yellow highlighting in an amendment PDF) and the content of specific annexes and divisions using general Basel III or regulatory drafting logic rather than anything derived from the actual MAS text. Across multiple findings, the AI's own caveats acknowledged uncertainty — but those caveats appeared after a confident primary answer that many practitioners would record and act on.
All five findings cluster on a single regulation — MAS Notice 637 on Risk-Based Capital Adequacy Requirements for Banks — and on questions about its internal structure: which entities it covers, what its annexes contain, how its amendment conventions work. This is precisely the kind of granular, document-specific knowledge that AI tools handle poorly, because it requires the model to have retrieved and retained the exact text of a specific MAS instrument rather than reasoning from general banking regulation principles.
The Basel III global framework that underpins Notice 637 is well-represented in AI training data; the Singapore-specific implementation details are not, and AI tools systematically conflated the two.
The systemic risk for an Accountants (CA/PA) practice is that these errors are not obviously wrong. An AI answer that names a plausible-sounding MAS notice, or attributes plausible-sounding content to a capital adequacy annex, will not trigger the same suspicion as a response that is clearly absurd. Practitioners who are not already expert in the granular structure of Notice 637 have no immediate basis to challenge the AI's output — which means the error is most likely to be caught only if the practitioner independently reads the regulator's text, the very step that AI use was intended to shortcut.
Without that verification step, incorrect AI answers on these topics would routinely survive into client deliverables.
5 findings in this case study. Click any to see its full evidence card.
The default position for Accountants (CA/PA) working on MAS capital adequacy matters should be that an AI tool's answer is a starting point for research, not a primary source. The findings documented here show that AI tools can produce confident, detailed, and structurally plausible answers about MAS Notice 637 that are nevertheless wrong in ways that matter — fabricated instrument names, mischaracterised annex content, and invented descriptions of document-formatting conventions.
None of these errors is immediately obvious to a practitioner who is not already expert in the granular structure of the notice, which means verification against the regulator's published text is not an optional step but a professional one.
Practical safeguards for any engagement where AI tools contribute to a regulatory deliverable include: always tracing AI-generated regulatory references back to the MAS portal (mas.gov.sg) and confirming the instrument, annex, or division exists and says what the AI claimed; keeping an audit record of any AI use that feeds into a client advice note, compliance memo, or regulatory submission; and never including an AI-generated regulatory citation in a document without independently checking the cited source.
Where an AI tool's own response includes caveats or hedges — as several findings here showed — treat the primary answer as unverified regardless of how confidently it was framed.
AI tools remain genuinely useful in an Accountants (CA/PA) context for tasks that do not require regulatory precision: drafting non-regulatory sections of reports, generating a first-draft list of questions to take into a client meeting, summarising background reading that you will verify against the source, or preparing plain-language explanations of concepts for internal training purposes. The risk is concentrated at the point where AI output touches specific regulatory text — instrument names, annex designations, divisional structure, operative dates, and amendment conventions. Keep those steps human-verified.
RegLeg's published Hallucination Research is available as a free reference for Accountants (CA/PA) who want to check whether a specific regulation or topic area has already been tested before acting on an AI answer. The research covers MAS Notice 637 and other key Singapore financial services instruments, and is updated as AI tools evolve and as regulations change. Before relying on AI output in a capital adequacy context, practitioners can use the published findings to identify known failure points — reducing the risk that a known AI error pattern survives into a client deliverable.
For firms that employ multiple Accountants (CA/PA) working across the same regulatory portfolio — for example, teams advising bank clients on MAS capital adequacy compliance — RegLeg offers bespoke regulation deep-dives. These are structured reference documents that map the specific AI failure modes observed on a given instrument, giving the team a shared, up-to-date picture of where AI output on that regulation can and cannot be trusted.
RegLeg also produces CPD-aligned training content and internal briefing materials that show practitioners what the failure modes look like in practice, how to spot them in AI output, and what verification steps are proportionate to the risk involved.
Firms that have already developed an internal AI-use policy are welcome to bring it to RegLeg for a confidential review against our failure-mode catalogue. The review identifies gaps — topics or use-cases where the policy does not reflect known AI limitations in the Singapore regulatory context — and produces a short-form report that can be used to update the policy or brief senior leadership. The goal is not to discourage AI use but to help Accountants (CA/PA) practices deploy it in a way that is proportionate to their professional obligations and their clients' exposure.