This case study examines how AI tools perform when Finance teams at Investment Banking firms in Singapore query them on local regulatory requirements. Testing covered one regulation — MAS Notice 637 on Capital Adequacy for Banks — and identified one aggregated question where AI assistants produced responses that could not be verified against the source text. The errors found are not trivial edge cases: they concern foundational prudential rules that Finance teams routinely reference when preparing capital calculations, reviewing internal valuation frameworks, and supporting business lines on compliance matters.
Firms that rely on unverified AI output in these areas risk embedding incorrect regulatory interpretations into internal processes, management information, and client-facing work-products.
Finance teams at Investment Banking firms in Singapore regularly consult AI tools to accelerate regulatory research. Typical triggers include drafting or refreshing internal capital adequacy policies, preparing training materials ahead of regulatory updates, mapping rule changes to existing valuation frameworks, and supporting front-office or risk colleagues who need a rapid read on a regulatory question. When MAS issues or consolidates a Notice such as Notice 637, Finance staff often turn to AI tools to get an initial overview of the document structure — including which annexes cover which topics — before committing analyst time to a full review.
It is precisely at this early orientation stage that AI-generated errors are most dangerous, because they shape the frame through which the rest of the regulatory document is then read.
The corporate use-cases built on top of this kind of AI output are significant. Capital adequacy reporting, fair-value measurement policies, additional valuation adjustment (AVA) calculations, and Pillar 2 submissions all depend on Finance teams correctly understanding which regulatory provisions apply and where they sit in the regulatory text. If an AI tool misidentifies the scope or content of a key annex, downstream work-products — internal sign-off memos, management committee papers, and regulatory submissions — may rest on a factually incorrect foundation without anyone in the review chain realising the error originated from an AI response.
If the AI's answer is wrong and the firm acts on it, the consequences land squarely on the firm. MAS has broad supervisory powers over capital adequacy compliance under the Banking Act and Notice 637, including the ability to require immediate remediation, impose additional capital buffers, and refer persistent failures for formal enforcement action. Beyond direct regulatory sanction, a firm that discovers an error after it has been embedded in regulatory reporting faces remediation costs, potential restatement, and reputational damage with its primary regulator — costs that outweigh any efficiency gain from using an AI tool in the first place.
The finding in this case study points to a specific and recurring pattern: AI tools reason from general international frameworks — in this instance, the broad Basel III structure — rather than retrieving and citing the actual text of the Singapore-specific regulation. When asked about the content of a named annex within MAS Notice 637, the AI assistant produced an answer that had internal plausibility (prudent valuation and additional valuation adjustments are indeed Basel III concepts) but could not be anchored to the MAS document itself.
The AI tool appears to have inferred what an annex should contain based on its training-data understanding of Basel-aligned capital rules, rather than confirming what MAS Notice 637 actually says. This is a low-visibility failure: the response sounds authoritative, uses correct technical vocabulary, and aligns with the broader regulatory framework — which makes it harder for a reader to flag as potentially wrong.
All of the errors identified in testing cluster on a single regulator and regulation — MAS Notice 637 — which is the primary prudential capital adequacy instrument for banks in Singapore. For Finance teams at Investment Banking firms, this is a high-traffic document: it underpins capital reporting, internal model governance, and valuation controls across the business. The concentration of error on a document that Finance staff consult frequently means the exposure is not theoretical; it is embedded in the daily and monthly rhythms of the department.
The systemic risk compounds quickly. If an AI-generated characterisation of a regulatory annex is accepted without verification and used to frame an internal policy or a training module, every subsequent piece of work that draws on that policy or module inherits the error. A single incorrect AI response about which provisions govern prudent valuation can propagate through capital adequacy procedures, AVA methodology notes, and external audit support packs before anyone traces the error to its source.
For a Finance team operating under MAS supervision, the remediation burden — re-checking, correcting, and re-issuing affected documents — can be substantially larger than the original time saving from using the AI tool.
1 finding in this case study. Click any to see its full evidence card.
The default position for Finance teams at Investment Banking firms in Singapore should be that AI tools are a starting point for regulatory orientation, not a primary source. For MAS Notice 637 and other capital adequacy instruments, any AI-generated statement about a specific provision, annex, or rule should be treated as a hypothesis to be verified against the published MAS text — not a finding that can be passed on to colleagues or incorporated into a work-product without a manual check.
This is especially true for questions about document structure (which annex covers what) because these are exactly the questions where AI tools tend to reason by inference rather than retrieval.
Practical firm-level safeguards should include a written policy that identifies AI as an unreliable primary source for regulatory rule content, with capital adequacy rules explicitly named. Work-products that reference MAS Notice 637 provisions should carry an audit trail showing that the cited provision was verified against the MAS source text, not merely against an AI response. Any AI output that influences a firm-wide document — a policy, a procedure note, a regulatory submission, or a training module — should require explicit sign-off from a qualified person before it enters circulation.
Where possible, firms should distinguish between "AI-drafted" content (where AI generated the text) and "AI-summarised" content (where AI compressed a document the team has independently read), because the hallucination risk profile is different and reviewers need to know which they are looking at.
There are areas of the Finance workflow where AI tools add genuine value with lower risk. Drafting non-regulatory copy such as internal communications or meeting agendas, generating a list of questions to guide further research, or producing a first-pass summary of a long document that the team then reads in full — these are uses where an AI error is caught quickly and the cost of a mistake is low. The dividing line is whether the AI's output will be acted on directly or whether it will be verified before use.
For MAS capital adequacy rules, the answer must always be: verify before use.
RegLeg publishes its hallucination research findings openly, and Finance teams at Investment Banking firms can use this material as a free first check before relying on any AI-generated answer in the capital adequacy space. The research maps specific regulatory documents — including MAS Notice 637 — against the questions where AI tools have demonstrably failed, so teams can quickly identify whether the AI answer they are looking at falls into a known high-risk area.
This is not a replacement for legal or compliance review, but it provides an evidence-based prompt to pause and verify rather than accepting an AI response at face value.
For firms that want to go further, RegLeg offers bespoke regulator deep-dives tailored to Investment Banking firms operating in Singapore. These map which AI-supported workflows in the Finance function carry the highest hallucination exposure — covering capital adequacy reporting, valuation framework documentation, internal model governance, and regulatory submissions — and produce a prioritised view of where the firm's current AI use practices carry the most risk. This work is grounded in the same testing methodology behind the published research, applied to the specific regulatory perimeter that matters to the firm.
RegLeg also offers a confidential review of a firm's existing AI-use policy against its failure-mode catalogue, with prioritised remediation recommendations. For Finance teams that have already embedded AI tools into workflows touching MAS Notice 637 or other prudential rules, this review provides an independent assessment of where gaps exist and what controls would close them. Training materials and CPD-aligned content are available for teams that want to build internal capability — helping Finance staff recognise the characteristics of an AI response that should trigger a verification step, and understand the regulatory consequences of acting on an unverified answer.