This case study examines how AI tools perform when Legal teams at Corporate Banking firms in Singapore query them about regulatory capital adequacy requirements administered by the Monetary Authority of Singapore (MAS). One regulation was tested — MAS Notice 637 on risk-based capital adequacy for banks — and one aggregated question was identified where AI assistants produced an incorrect or misleading response. The error documented here involves the regulatory framework applicable to financial holding companies incorporated in Singapore, a structurally important distinction for Corporate Banking groups operating through holding-company structures.
Although a single finding may appear narrow in scope, it reflects a pattern of AI tools constructing plausible-sounding but unverifiable regulatory labels that can mislead Legal teams into treating fabricated designations as authoritative. For Singapore-based corporate banking groups, where capital adequacy compliance and group-level regulatory mapping are core Legal responsibilities, this type of error carries material institutional risk.
Legal teams at Corporate Banking firms in Singapore routinely consult AI tools when scoping regulatory requirements for new product launches, drafting internal compliance policies, preparing regulatory mapping documents for group entities, and responding to questions from business lines about which MAS rules apply to a proposed structure or transaction. Capital adequacy is a topic that surfaces frequently in these workflows — particularly when the firm is restructuring its corporate group, onboarding a new regulated entity, or advising on the regulatory perimeter of a financial holding company.
An AI tool that gives a confident but incorrect answer about which MAS notice governs a financial holding company can anchor the Legal team's research in the wrong framework from the outset.
The corporate use-cases sitting on top of this topic are significant. Group-level capital adequacy mapping feeds directly into regulatory reporting obligations, board-level risk disclosures, product structuring advice given to corporate clients, and internal approvals for acquisitions or restructuring. Legal teams are often the first point of contact when a business line or finance function asks whether a proposed holding-company arrangement meets MAS requirements. If the AI tool consulted at that stage supplies a fabricated regulatory reference, the error will propagate into briefing notes, legal opinions, board papers, and potentially external submissions to MAS.
The firm — not the individual lawyer who ran the query — absorbs the consequences when an AI-generated error enters firm work-products. Those consequences include regulatory action by MAS for mischaracterising applicable rules, the cost of remediation if internal policies or structures were built on an incorrect regulatory foundation, reputational harm if the error surfaces in a regulatory review or client dispute, and the operational disruption of unwinding decisions that relied on flawed AI output.
In Singapore's corporate banking sector, where MAS supervisory expectations are exacting and group structures are closely scrutinised, the threshold for regulatory error is low and the cost of correction is high.
Across the question tested against MAS Notice 637, the error produced by AI assistants follows a specific and instructive pattern: the AI correctly identified that a separate regulatory instrument applies to financial holding companies, then fabricated the specific notice designation. This is not a case of the AI being entirely wrong — the structural observation (that a distinct FHC-specific notice exists) is directionally accurate. The danger lies precisely in this partial correctness. An AI answer that is partly right is far more likely to be trusted and acted upon than one that is obviously wrong.
The fabricated notice label is plausible enough to survive a superficial review, and it will not be caught unless the reader independently verifies the designation against MAS's published materials.
The finding clusters on a single regulator (MAS) and a single regulation (MAS Notice 637), but the underlying failure mode — constructing a regulatory label by analogy from a known reference — is not unique to this instrument. Legal teams should treat this finding as illustrative of a broader vulnerability: AI tools tested on Singapore capital adequacy rules show a tendency to generate specific-sounding regulatory identifiers that do not correspond to any published instrument. This is particularly risky for Legal work because regulatory labels carry legal weight.
A notice number cited in an internal policy, board paper, or client advice has an implied claim to existence and authority that a fabricated label cannot satisfy.
The systemic risk to a Corporate Banking firm compounds quickly. If a single AI-assisted research note containing a fabricated notice designation is used as the basis for an internal compliance policy, that policy may cite a non-existent instrument. Downstream work-products — regulatory mapping documents, training materials, product-approval memos, and external submissions — may all inherit the same error. Correcting this after the fact requires identifying every work-product that relied on the original note, which is operationally intensive and may require disclosures to MAS if the error influenced a regulatory filing or correspondence.
1 finding in this case study. Click any to see its full evidence card.
The default position for Legal teams at Corporate Banking firms should be that AI tools are a starting point for regulatory research, not a primary or authoritative source. This is especially true for questions involving specific regulatory instrument names, notice numbers, or the precise scope of MAS rules. When an AI tool supplies a specific notice designation or regulatory label — particularly one that the team cannot immediately locate on the MAS website — that designation must be independently verified against MAS's published materials before it appears in any internal or external work-product.
A plausible-sounding label is not the same as a confirmed one.
At the firm level, practical safeguards should include a regulatory-verification policy that explicitly identifies AI output as an unreliable source for specific rule references in the Singapore capital adequacy space; a requirement to maintain audit trails when AI output influences a research note, legal opinion, or policy document; and a sign-off requirement before any AI-sourced regulatory reference enters firm-wide use or is shared with the business.
Where AI output is used to draft regulatory-facing material, the team should clearly distinguish between content that has been AI-drafted and then verified, and content that has only been AI-summarised and remains subject to confirmation. In both cases, the responsible lawyer or compliance officer must be able to point to a primary source for every regulatory reference.
AI tools are genuinely useful in Legal workflows for tasks that do not require the AI to supply specific regulatory identifiers as verified facts: drafting non-regulatory sections of policies and agreements, summarising lengthy documents that the team can then check against source materials, generating lists of questions for further primary-source research, and producing first drafts of internal training content. These uses leverage the AI's language capabilities without exposing the firm to the specific failure mode documented here, which arises when the AI is asked to supply regulatory authority rather than assist with drafting or summarisation.
RegLeg publishes hallucination research covering the specific regulations and regulatory questions most likely to surface in Legal workflows at Corporate Banking firms in Singapore. This research is available as a free reference check that Legal teams can consult before relying on an AI answer in a regulatory context. Rather than asking a team to build its own testing regime from scratch, RegLeg's published findings give practitioners a ready reference for the rule areas where AI tools are known to produce incorrect or unverifiable output — including the MAS capital adequacy framework covered in this case study.
For firms that want to go further, RegLeg offers bespoke regulatory deep-dives that map which AI-supported workflows in a Corporate Banking firm carry the highest hallucination exposure. These engagements are structured around the firm's actual Legal workflow — the points at which AI tools are currently consulted, the work-products those queries feed into, and the downstream consequences if the AI answer is wrong. The output is a prioritised view of where the firm's current AI use creates regulatory risk, and where it can be continued with confidence subject to appropriate verification steps.
RegLeg also provides confidential review of a firm's existing AI-use policy against its failure-mode catalogue, with prioritised remediation recommendations tailored to the Legal function in a Corporate Banking context. For teams looking to build internal capability, RegLeg can supply training material and CPD-aligned content that Legal teams can use to develop their own practitioners' judgment about when to trust, verify, or discard AI-generated regulatory output. These resources are designed to complement, not replace, the firm's own compliance and risk frameworks.