This case study examines how AI tools perform when Operations teams at international investment banking firms rely on them for regulatory guidance under the CPMI-IOSCO framework on cyber resilience. Across one aggregated finding drawn from testing of the 2016 CPMI-IOSCO Guidance on Cyber Resilience for Financial Market Infrastructures, AI assistants produced responses that overstated the operational detail contained in the source document.
The single question where AI tools went wrong concerns the boundary between the 2016 CPMI-IOSCO guidance and a later FSB publication that substantially extends it — a distinction that carries direct operational consequences for firms building or auditing their cyber incident response frameworks. Investment banking operations functions that use AI tools to interpret this regulatory landscape face a specific risk: acting on incomplete or overstated guidance as though it were authoritative, without verifying whether a subsequent document has superseded or supplemented the original rule text.
Operations teams at investment banking firms routinely engage with cyber resilience frameworks when reviewing or drafting internal incident response plans, mapping regulatory obligations for new trading platforms or clearing arrangements, and responding to queries from risk, legal, and compliance colleagues who need to understand the operational requirements underpinning FMI-connected activity. These tasks intensify during periods of infrastructure change — when a firm joins or deepens its participation in a payment or settlement system, when regulators issue new guidance, or when internal audit programmes flag gaps in the firm's existing cyber resilience posture.
AI tools are an increasingly common first port of call for orientation questions about what a specific piece of CPMI-IOSCO guidance actually requires and how it interacts with related international frameworks.
The corporate use-cases sitting on top of these topics include drafting or reviewing Business Continuity Plans and cyber incident response runbooks, scoping due-diligence questionnaires for technology vendors and critical third parties, preparing regulatory submissions or engagement letters, and developing training content for operations staff. When AI tools feed answers into these work-products — particularly where a junior analyst or project manager takes an AI-generated summary at face value — the error propagates into firm-wide artefacts before anyone has checked it against the source text.
The problem is compounded in a multi-document regulatory family like CPMI-IOSCO and FSB cyber guidance, where the relationship between an earlier framework and a later, more detailed publication is precisely the kind of nuance that is easy to miss and costly to get wrong.
What is at stake for the firm is not merely an embarrassing correction. Regulatory bodies overseeing FMI participants expect robust cyber resilience governance aligned with the prevailing international standards landscape in its entirety, including supplementary guidance issued after the primary framework. If an operations function builds its incident response framework on a misreading of where the detailed operational requirements actually sit — believing the 2016 CPMI-IOSCO guidance is self-contained when the authoritative operational detail appeared in a later FSB publication — the firm may find its response and recovery procedures fall short during a supervisory review or actual incident.
The consequences range from regulatory findings and required remediation at cost to the firm, through reputational damage if the gap surfaces publicly, to potential operational failures during a real cyber event that the incomplete framework failed to anticipate.
Across the single finding aggregated here, the pattern of error is consistent and specific: AI tools overstate the operational completeness of the 2016 CPMI-IOSCO Cyber Resilience Guidance, presenting it as a self-sufficient source of detailed incident response and recovery requirements when the authoritative operational detail for the Response and Recovery phase was not published until the FSB's 2020 guidance.
The AI response does not simply omit the later document — it actively characterises the 2016 text as providing detailed expectations, listing specific elements such as incident response plans, recovery time objectives, secondary site requirements, and communication protocols in a way that implies they are fully addressed by the older guidance alone. This is a version-boundary error: the AI conflates two distinct documents across a four-year gap, attributing the fuller framework to the earlier text.
The error clusters within the CPMI-IOSCO and BIS regulatory family and within the cyber resilience topic area. For Operations teams at investment banking firms — who are more likely to encounter this framework when reviewing FMI-connected infrastructure than colleagues in other departments — this is a high-exposure zone precisely because it sits at the edge of teams' day-to-day regulatory familiarity.
The technical language of FMI cyber resilience guidance can make AI-generated summaries appear authoritative, reducing the likelihood that an analyst will instinctively sense the need to cross-check the answer against primary sources or investigate whether later guidance supplements the document being described.
The systemic risk to the firm is compounded by how these errors travel. An AI-produced summary characterising the 2016 guidance as comprehensive may inform an internal policy document, a vendor questionnaire, a training slide deck, and a senior management briefing — all from a single query. Each of those downstream artefacts then reflects the same misreading, and correcting them once the error is identified requires coordinated remediation across multiple functions.
In an area where regulators expect firms to demonstrate clear-eyed awareness of the international standards landscape as a whole, that kind of systemic gap can be particularly costly to explain and to remedy.
1 finding in this case study. Click any to see its full evidence card.
Operations teams should treat AI tools as a research starting point rather than an authoritative source when working with international cyber resilience frameworks. For regulatory topics like the CPMI-IOSCO guidance and related FSB publications, the version landscape evolves across multi-year periods and the interactions between documents are precisely the kind of nuance that AI assistants routinely flatten or misrepresent. Any AI-generated summary of what a specific regulatory document requires should be verified against the source text before it is used in a firm work-product, and that verification step should be recorded rather than left to individual judgement.
At the firm level, practical safeguards include establishing a clear policy that AI output on regulatory matters is classified as a first draft requiring sign-off from a qualified person before it enters any document that will be reviewed by risk, compliance, audit, or a regulator. Teams should maintain audit trails distinguishing AI-drafted content from verified content in regulatory-facing material.
Where AI tools feed answers into recurring processes — vendor questionnaire templates, BCP review checklists, regulatory mapping exercises — those templates should carry a standing note that the AI source has not been independently verified against the primary and any supplementary regulatory texts. Senior leadership should be aware that the absence of such controls can result in systemic errors propagating across multiple work-products simultaneously, creating a remediation burden disproportionate to the original time saving.
There are areas where AI tools add genuine value for operations teams without creating material regulatory risk. Using AI to generate a first-draft list of questions for a deeper research session, to reformat or summarise internal documents that the team can itself verify, or to draft non-regulatory operational copy are all lower-risk applications. The appropriate boundary is straightforward: AI output that enters a regulatory or compliance context requires human verification against the primary source — and against any later guidance that supplements it — before use.
That rule is simple enough to apply consistently and protects the firm from the most common and costly category of AI error in this domain.
RegLeg's published hallucination research provides Operations teams at investment banking firms with a ready-made reference before they rely on any AI answer in areas covered by international cyber resilience frameworks. For the CPMI-IOSCO and FSB guidance family specifically, our research identifies the precise question areas where AI tools consistently misrepresent scope, version boundaries, and the interaction between documents — giving your team a concrete basis for deciding when an AI-generated answer warrants immediate verification and when it is more likely to be reliable.
This material is freely available and designed to function as a practical pre-check rather than an academic resource.
For firms that want a more tailored assessment, RegLeg offers bespoke regulatory deep-dives that map which AI-supported workflows within an investment banking operations function carry the highest hallucination exposure. This work covers not just the CPMI-IOSCO framework but the full range of international regulatory instruments that operations teams encounter, with findings structured by workflow type — incident response planning, vendor due-diligence, regulatory mapping for new products and infrastructure — so that your team can prioritise controls where the exposure is greatest rather than applying uniform caution across all AI use.
We also offer confidential review of a firm's existing AI-use policy against RegLeg's failure-mode catalogue, with prioritised recommendations identifying where current controls are sufficient and where gaps exist. For teams building internal capability, RegLeg provides training material and CPD-aligned content that operations staff can use to develop calibrated judgement about AI outputs in a regulatory context — not blanket scepticism, but the specific literacy needed to identify the errors that matter most to your firm and your regulators.