This case study examines how AI tools perform when Compliance teams at Investment Banking firms in Singapore consult them on capital adequacy requirements under MAS Notice 637. Across three aggregated questions drawn from that single regulation, AI tools returned incorrect or unverifiable answers in every instance tested. The errors span regulatory instrument labelling, document-formatting conventions with operative legal significance, and the structural content of specific notice divisions. For Singapore-licensed investment banks, where capital adequacy compliance sits at the centre of MAS supervisory scrutiny, each of these failure points carries direct regulatory and operational exposure.
A Compliance team at an investment bank in Singapore will typically turn to AI tools when a question about capital adequacy rules arises quickly — during a product approval process, while updating internal capital policy documentation, or when a business line asks whether a proposed transaction structure changes the firm's regulatory capital treatment. MAS Notice 637 is a technically dense instrument, and Compliance officers frequently need to identify which version of the notice is in force, which provisions apply to their entity type, and what specific sections require action.
AI tools appear well-suited to answering these threshold questions rapidly, and in practice many teams now use them as a first-pass source before engaging in more formal regulatory research.
The corporate use-cases that sit on top of these topics include drafting capital adequacy policy statements for board approval, preparing responses to MAS thematic reviews, training materials for front-office staff on capital implications of new product types, and regulatory mapping exercises conducted during mergers or subsidiary restructurings. In each of these contexts, a Compliance team member may cite or paraphrase an AI-generated answer — perhaps a notice reference, a document annotation, or a description of a notice division — and that answer flows into a firm work-product without independent verification.
If the AI's answer is wrong, the firm bears the consequences. An incorrect notice designation submitted to MAS in a regulatory query or capital return creates a compliance record that may need correction under MAS scrutiny. Policies or training materials built on misidentified provisions may expose the firm to supervisory criticism if those materials influence internal conduct. Where a misstated capital rule affects how a product is structured or how a position is reported, the downstream risk is material misreporting — an outcome that can trigger MAS enforcement action, including mandatory remediation, undertakings, or public censure.
The individual staff member who used the AI tool is rarely personally liable for the error; the firm's Compliance function and its leadership absorb that cost.
All three findings in this case study concern MAS Notice 637, Singapore's primary risk-based capital adequacy framework for banks. The errors are not random — they cluster around the edges of the notice where precision matters most: the scope of application across entity types, document-level formatting conventions that carry operative legal significance, and the structural content of specific notice divisions. In plain terms, AI tools tested on this regulation tended to fill gaps in their knowledge with plausible-sounding inferences rather than acknowledging uncertainty.
They constructed regulatory instrument names by analogy, offered generic editorial explanations for document markings that have a specific technical meaning, and described a notice division's coverage in terms that could not be verified from authoritative sources.
The pattern is one of confident imprecision. Two of the three findings involved more than one AI tool producing an incorrect or unverifiable response to the same question, which means the problem is not isolated to a single tool or query formulation. For a Compliance team at an investment bank, the implication is that even cross-checking an AI answer against a second AI tool — a common informal verification step — provides no reliable protection when both tools share the same underlying knowledge gaps about a regulation.
The systemic risk to the firm compounds because these are not obscure edge questions. They are the kind of threshold and structural questions a Compliance officer would reasonably ask early in any capital adequacy review — which notice applies to our entity, what do the amendment markings mean, what does a particular division require. If AI outputs on these questions feed into a capital policy document, a regulatory submission, or a board briefing, the error propagates across every downstream work-product that relied on it.
Correcting a foundational misstatement after it has been embedded in multiple firm documents is significantly more costly — in time, legal fees, and regulatory goodwill — than preventing it at the point of first reliance.
3 findings in this case study. Click any to see its full evidence card.
The default position for Compliance teams at investment banks should be that AI tools are a starting point for research on regulatory topics, not a primary or verifiable source. This case study shows that AI tools can produce incorrect notice designations, misstate document-level formatting conventions with legal significance, and describe regulatory provisions based on inference rather than retrieved text — and that they may do so while sounding authoritative.
For capital adequacy questions under MAS Notice 637, any AI output that names a specific regulatory instrument, characterises a document marking, or describes the content of a specific provision should be verified against the MAS portal before it enters any firm work-product.
At the firm level, practical safeguards should include a regulatory-verification policy that explicitly identifies AI tools as unreliable sources for instrument-level and provision-level questions in areas like capital adequacy, and that requires independent verification before AI output is used in regulatory submissions, board papers, or internal policy documents. Compliance teams should maintain audit trails showing where an AI output was used and what verification step was taken. Where AI-drafted content reaches firm-wide use — in policy templates, training decks, or regulatory mapping documents — a sign-off requirement from a qualified Compliance officer should be standard.
Teams should also distinguish between AI-drafted content (where the AI has generated text that may carry embedded errors) and AI-summarised content (where the AI has condensed a source document the team can itself review), and apply proportionate scrutiny to each.
AI tools remain useful in Compliance workflows for tasks that do not require precision on specific regulatory text: drafting non-regulatory internal communications, generating first-draft interview or audit questions for later human review, summarising lengthy consultation papers where the team will read the underlying document, and formatting or structuring documents that Compliance officers will populate with verified content. The discipline is using AI to accelerate research and drafting while keeping a qualified human between any AI output and any regulatory-facing conclusion.
RegLeg's published hallucination research gives Compliance teams at investment banks a practical, no-cost first check before relying on any AI answer in Singapore capital adequacy and related MAS rule areas. The findings are organised by regulation, question type, and failure pattern, so a Compliance officer can look up whether the specific question they asked an AI tool sits in a documented high-risk zone before using the response in a work-product. This is not a replacement for reading the MAS Notice — it is a filter that helps teams identify where to invest their verification effort.
For investment banking Compliance functions that want a more structured picture of their exposure, RegLeg offers bespoke regulator deep-dives that map which AI-supported workflows in the firm carry the highest hallucination risk. Capital adequacy, prudential reporting, and regulatory scope questions are among the areas where AI tools show consistent difficulty — and a targeted mapping exercise can give the Compliance team and its leadership a clear view of where existing AI-use practices need stronger controls.
RegLeg can also conduct a confidential review of the firm's current AI-use policy against its failure-mode catalogue, and identify the highest-priority gaps for remediation without requiring a full policy overhaul.
RegLeg also develops training material and CPD-aligned content that Compliance teams can use internally to build staff awareness of AI limitations in regulatory research. These materials are designed for practical use — covering how to recognise AI output that carries hallucination risk, how to verify regulatory references efficiently, and how to document AI use in a way that satisfies audit and supervisory expectations. If your team is developing or refreshing its AI governance framework, RegLeg welcomes the conversation.