Executive Summary
The Principles for Financial Market Infrastructures (PFMI), published by the Committee on Payments and Market Infrastructures (CPMI) and IOSCO, set the international baseline for the safety and efficiency of systemically important financial market infrastructures — covering payment systems, central counterparties, central securities depositories, and trade repositories. Public auditors operating across international jurisdictions increasingly encounter PFMI compliance questions when assessing whether an FMI meets the required governance, risk management, and financial resilience standards.
Across four aggregated findings in this cell, AI tools were unable to provide verbatim content, exact thresholds, or specific cross-references from key PFMI-related regulatory documents, even when web search was available. Every failure in this cell took the same form: the AI correctly identified a document's existence and general subject matter but could not retrieve, quote, or verify its internal provisions — leaving auditors without the precise textual basis they need for workpapers, opinions, and attestations. The consistent pattern across findings means Public Auditors cannot rely on AI tools to substitute for direct document review of CPMI-IOSCO publications.
How AI gets this regulation wrong
When tested on PFMI-related questions, AI tools consistently encountered the same limitation: they could acknowledge that a document existed and describe its general purpose, but were unable to retrieve or confirm verbatim provisions, specific numerical thresholds, or precise paragraph cross-references from the underlying PDFs. This was true even when the AI had access to live web search, because the relevant documents are binary PDFs whose internal text is not reliably accessible to AI systems at the paragraph level. The table below breaks down where these gaps appeared across the PFMI regulatory corpus.
| AI's Failure Mode | Count | Affected findings |
|---|---|---|
| Blind Spot | 4 | Finding#1 · Finding#2 · Finding#3 · Finding#4 |
What that means for your practice
For Public Auditors, the dominant risk arising from AI's gaps on PFMI is producing a wrong deliverable — an audit workpaper, attestation, or advisory opinion that cites provisions, thresholds, or cross-references the AI was never able to verify against the source text. Because PFMI assessments require auditors to ground findings in the specific language of the principles and their supporting guidance, a gap in verbatim access translates directly into a gap in evidentiary reliability. The table below maps how those risks materialise across different audit and assurance tasks.
| Risk Impact | Count | Affected findings |
|---|---|---|
| Wrong deliverable | 4 | Finding#1 · Finding#2 · Finding#3 · Finding#4 |
When this affects Public Auditors
Public auditors in international jurisdictions encounter PFMI most directly when scoping or executing assessments of systemically important financial market infrastructures — central counterparties, payment systems, central securities depositories, and trade repositories. In those engagements, auditors need to map FMI practices to specific PFMI principles, identify the applicable quantitative thresholds (such as the six-month liquid net assets standard under Principle 15), and cross-reference the core PFMI text against supporting guidance documents such as the CCP resilience and recovery consultative reports.
When an auditor uses an AI tool to draft an issue description, populate a control matrix, or check a threshold figure, an AI gap in verbatim document access means the draft is built on text the AI acknowledged it could not verify.
The risk extends to engagements where the IOSCO co-published versions of PFMI documents are the operative reference — for example, where a securities regulator requires reliance on the IOSCO publication rather than the BIS version. AI tools we tested were unable to provide verbatim content from the IOSCO-branded PFMI publications (IOSCOPD377 and IOSCOPD396), even though those documents carry the same substantive provisions. An auditor who does not personally retrieve and read the source PDF may produce a finding letter or attestation that paraphrases provisions the AI summarised from general knowledge rather than from the text of the operative document.
The problem is compounded for more recent CPMI-IOSCO guidance. The November 2025 Level 3 assessment on general business risks is beyond the training data horizon of current AI tools, meaning AI cannot describe its findings at all from its own knowledge — and even web search produced only summary-level awareness rather than verbatim content. Auditors scoping engagements against the most current regulatory expectations are therefore at risk of relying on AI-generated content that reflects an earlier — and potentially superseded — picture of what CPMI-IOSCO expects.
The findings at a glance
The table below summarises each finding in this cell — the question asked, what the AI was and was not able to provide, and the resulting exposure for Public Auditors working with PFMI.
Aggregate impact
All four findings in this cell share a single structural pattern: AI tools correctly declined to fabricate verbatim content from CPMI-IOSCO regulatory publications but could not supply the verbatim content either. The honest refusals are a better outcome than invented text — but they are not a safe outcome for audit work, because the auditor still leaves the interaction without the provision, threshold, or cross-reference they needed.
The gap is systemic across the PFMI corpus: it affects the foundational 2012 PFMI text (in both the BIS and IOSCO co-published editions), the CCP resilience and recovery consultative guidance from 2016, and the most recent Level 3 supervisory assessment from 2025.
The findings cluster on a specific audit need: verbatim textual accuracy. PFMI assessments are not conceptual exercises — they require auditors to cite specific principles by number, quote the operative language, and cross-reference supporting guidance with precision. AI tools can discuss PFMI in general terms and correctly situate documents within the regulatory framework, but they cannot reproduce the internal text of PDF publications at the level of exactness that workpapers and attestations require. That gap is consistent regardless of whether web search is enabled, because search returns summaries and abstracts rather than the full binary text of a regulatory PDF.
For Public Auditors in international jurisdictions, the aggregate implication is that any AI-assisted PFMI workproduct carries a structural verification gap unless the auditor independently retrieves and reads each source document. The risk is heightened where multiple documents interact — for instance, where an auditor needs to triangulate a CCP's liquid net assets position against both Principle 15 of the core PFMI and the specific quantitative standards elaborated in the 2025 Level 3 assessment. AI tools can map the relationship between those documents in general terms, but cannot supply the exact figures or language that an audit conclusion must rest on.
What your team should do
The default position for PFMI audit work should be that AI tools do not substitute for direct document review of any CPMI-IOSCO publication. Every workpaper that cites a PFMI principle, a specific threshold, or a paragraph from a supporting guidance document must be traced to the source PDF by the auditor, not to an AI-generated summary. This is not a precautionary overcorrection — it reflects a limitation the AI tools themselves acknowledge: they do not have reliable verbatim access to these PDFs and, to their credit, say so rather than guess.
The auditor's job is to ensure that acknowledgment does not become a gap in the audit file.
Practical safeguards for PFMI engagements include: maintaining a standing library of the operative CPMI-IOSCO PDFs (retrieving the current version from bis.org and iosco.org at the start of each engagement rather than relying on cached copies); building a principle-by-principle reference matrix from the source text before using AI to help draft issue descriptions or remediation recommendations; and treating any AI-generated threshold figure — such as a liquid net assets percentage or a recovery timeline — as unverified until checked against the PDF.
For assessments that involve IOSCO co-published documents rather than the BIS-only versions, retrieve both editions and confirm they match before relying on either.
AI tools are reliably useful for PFMI work at the framing and orientation stage: mapping which principles apply to a given FMI type, identifying the broad structure of the resilience and recovery framework, or drafting an engagement scope memo that will be reviewed against the source documents. They are also useful for explaining PFMI concepts to non-specialist stakeholders or for comparing PFMI requirements to domestic regulatory frameworks at a high level.
The line to hold is between conceptual guidance — where AI performs well — and textual precision, where it cannot be trusted to produce or verify the exact language that PFMI audit conclusions require.
How RLB Can Help
RegLeg's published Hallucination Research gives public auditors a practical pre-flight check before placing weight on AI-assisted analysis of regulatory questions. The research catalogues the specific failure modes — misquoted thresholds, conflated jurisdictional requirements, fabricated citation trails — that AI tools produce most often in public-sector and cross-border audit contexts. Auditors can use these findings to calibrate their review steps before any AI output enters a working paper, providing a documented basis for the professional scepticism their standards already require.
Where an audit team or firm has multiple practitioners working across the same regulatory portfolio, RLB can deliver bespoke deep-dives on individual regulations. These sessions go beyond the published research to map failure modes specific to the instruments, guidance notes, and enforcement expectations most relevant to the team's current engagements. The output is practical rather than theoretical: teams leave with concrete review checkpoints aligned to the regulations they are actually auditing against.
RLB also develops training material and CPD-aligned content built around the failure-mode catalogue, so that auditors at all experience levels understand what to look for and why. For firms that have already deployed AI tools and drafted internal use policies, RLB offers confidential reviews of those policies against the same catalogue — identifying gaps between what the policy assumes AI tools will do reliably and what the research shows they frequently get wrong.