This case study examines how AI tools respond to regulatory questions put by Technology & Data teams at Payment Institutions firms operating across international jurisdictions. The testing focused on one regulation — the Guidance on Cyber Resilience for Financial Market Infrastructures published jointly by CPMI and IOSCO in 2016 — and identified three aggregated questions on which AI assistants produced materially incorrect responses.
Across these findings, AI tools consistently overstated the specificity and external alignment of the 2016 guidance, misattributed quotations to the wrong documents, and characterised a relatively high-level framework as containing operational detail that was only provided by a later, separate publication. Technology & Data teams at Payment Institutions firms that rely on AI responses to navigate this guidance face a real risk of building internal policies, vendor assessment criteria, and incident response procedures on inaccurate foundations.
Technology & Data teams at Payment Institutions firms regularly consult AI tools when drafting or reviewing internal security policies, mapping their firm's existing controls against international standards, and preparing materials for regulators, auditors, or board-level governance committees. The CPMI-IOSCO 2016 Cyber Resilience Guidance is a natural reference point for these teams: it is widely cited in supervisory expectations, third-party assurance frameworks, and industry self-assessments across multiple jurisdictions.
Teams will ask AI assistants to explain the guidance's scope, clarify which external frameworks it draws on, locate specific phrases they have encountered in regulatory correspondence, and confirm whether detailed incident response requirements sit within the 2016 document or in a later publication. All three of these question types appeared in this research — and all three produced incorrect AI responses.
The corporate use-cases that sit on top of these queries are substantial. A Technology & Data team might use an AI-assisted summary of the CPMI-IOSCO guidance to populate a regulatory-mapping spreadsheet used by product, compliance, and operations simultaneously. They might use an AI-generated description of the guidance's framework alignment to brief a supplier or correspondent bank. They might rely on an AI answer about the document's incident response content when drafting the firm's cyber incident response plan — a document that regulators in most international jurisdictions will expect to be technically sound and traceable to recognised guidance.
When the AI's answer is wrong, the firm — not the individual employee — absorbs the consequences. A regulatory mapping that incorrectly describes the CPMI-IOSCO guidance's external citations or its operational detail can propagate errors into compliance attestations, internal audit findings, and supervisory submissions. Where a payment institution operates across multiple jurisdictions, regulators that cross-reference each other's expectations may identify inconsistencies that trace back to a single flawed AI response. The corrective cost — remediation of affected work-products, re-engagement with regulators, and potential reputational exposure — falls on the firm's leadership and the Technology & Data function.
The three findings share a common failure mode: AI tools convert structural similarity or contextual proximity into confident factual claims. Where the 2016 CPMI-IOSCO guidance is organised in a way that resembles an external framework, AI assistants assert an explicit citation that the document does not contain. Where a phrase is associated with the general area of BIS cyber work, AI tools attribute it to a specific document without confirming the source.
Where a topic area — incident response — appears in the 2016 guidance at a high level, AI assistants describe the document as containing detailed operational requirements that were only provided by a later publication. In each case the error is not a random invention; it is a plausible-sounding overstatement driven by the AI filling gaps in its knowledge with confident inference.
All three findings cluster on a single regulation — the CPMI-IOSCO 2016 Cyber Resilience Guidance — which is both widely referenced and relatively difficult to verify quickly. The guidance is a joint CPMI-IOSCO publication available from BIS, but its relationship to other documents in the cyber resilience ecosystem (later FSB publications, BIS speeches, CPMI wholesale payments fraud work) requires careful navigation. That complexity is precisely the kind of context in which Technology & Data professionals are most likely to trust an AI shortcut — and most likely to receive a misleading answer.
The systemic risk for a Payment Institutions firm is that these errors are mutually reinforcing. A team that asks three related questions about the same guidance and receives three confidently incorrect answers may never notice that anything is wrong, because each answer is consistent with the others. The firm's regulatory-mapping document, its board cyber report, and its incident response plan may each carry a different variant of the same underlying misunderstanding.
When a regulator or external auditor probes the firm's understanding of the CPMI-IOSCO framework, the firm faces not a single correction but a cross-cutting remediation across multiple work-products — a cost that grows in proportion to how deeply the Technology & Data team embedded AI output into its governance materials.
3 findings in this case study. Click any to see its full evidence card.
The default position for any Technology & Data team at a Payment Institutions firm should be that AI tools are a starting point, not a primary source, when researching regulatory obligations under the CPMI-IOSCO cyber resilience framework or related BIS and FSB guidance. The three failures documented here all occurred on questions that appear straightforward — what frameworks does the document cite, where does a specific phrase come from, how detailed is a given section — and in each case the AI produced a confident, internally coherent answer that was nonetheless wrong.
Confidence and coherence are not proxies for accuracy in this domain.
At the firm level, practical safeguards should include a written policy that identifies AI tools as an unreliable source for questions of regulatory document content, citation, and historical attribution in these rule areas. Any AI output that influences a work-product — a regulatory mapping, an incident response plan, a board paper, a supplier assessment — should be accompanied by an audit trail showing that the underlying regulatory text was consulted directly. Sign-off requirements before AI-drafted or AI-summarised regulatory content enters firm-wide use reduce the risk of a single incorrect answer propagating across multiple documents.
Teams should also distinguish clearly between content where the AI drafted language that the team then verified, and content where the AI's answer was taken as authoritative: those are different risk profiles and should be treated differently in governance records.
There are areas where AI tools can be used more safely within a Technology & Data workflow. Drafting non-regulatory copy — internal communications, training slide text, project documentation — carries lower risk because the output is not being relied on as an accurate description of a legal obligation. Generating a first set of research questions that a team member will then investigate using primary sources is a legitimate use of AI assistance.
Summarising long documents that the team can verify section by section is another reasonable application, provided the summary is treated as a working draft rather than a final record. The discipline required is to hold the boundary between those uses and the higher-risk activity of treating AI responses as accurate descriptions of what regulators require.
RegLeg's published hallucination research gives Technology & Data teams at Payment Institutions firms a free, immediately usable reference before relying on any AI response in the areas covered here. The findings documented across the CPMI-IOSCO cyber resilience framework — and across the broader body of international financial regulation that RegLeg has tested — allow a team to check whether the specific question they are asking AI tools falls into a known high-error zone. That check takes minutes and can prevent a firm from embedding an incorrect regulatory characterisation into documents that will be difficult and costly to correct later.
For firms that want to go further, RegLeg offers bespoke regulator deep-dives that map which AI-supported workflows within a Payment Institutions firm carry the highest hallucination exposure. That work is calibrated to the firm's specific jurisdiction mix, its regulatory perimeter, and the Technology & Data processes that most frequently involve regulatory research — whether that is third-party cyber due diligence, incident reporting obligations, or framework alignment for new product launches. The output is a prioritised picture of where the firm's current AI use creates the greatest risk of relying on inaccurate regulatory information.
RegLeg can also provide a confidential review of a firm's existing AI-use policy against RegLeg's catalogue of documented failure modes, with prioritised remediation recommendations. For firms building or refreshing internal training, RegLeg's research translates into CPD-aligned content that Technology & Data teams can use to build practical awareness of how AI hallucinations arise on regulatory questions — not as abstract warnings, but as documented patterns drawn from real testing on the regulations the team works with.
Teams that understand the specific shape of AI failure in their domain are better placed to use AI tools productively while keeping the firm's governance record sound.