AI on Implementation Monitoring of the PFMI: Level 3 Assessment on General Business Risks for Accountants (CA/PA) in international jurisdictions

Executive Summary

The November 2025 CPMI-IOSCO Level 3 assessment on general business risks subjects FMI compliance with Principle 15 — particularly the liquid net assets funded by equity (LNAFE) standard — to international scrutiny that directly affects how accountants advise CCPs, CSDs, and payment systems on capital adequacy and going-concern buffers. When Accountants (CA/PA) in international jurisdictions use AI tools to interpret Principle 15's Key Consideration 3 requirements, those tools consistently produced wrong answers across three tested questions — all categorised as hallucinations where the AI initially stated incorrect rules with confidence, then retracted or contradicted itself when pressed.

The failures concentrated on the precise quantitative structure and qualifying conditions of the KC3 LNAFE floor: AI tools variously misattributed the six-month operating expense minimum to the wrong Key Consideration, invented a "greater of" dual-track test that does not exist in the rule text, and fabricated an overlay KC4 liquidity condition onto a Basel/CRD carve-out that has its own standalone qualifier in KC3.

For practitioners signing off on LNAFE calculations, drafting regulatory opinions for FMI clients, or reviewing the compliance implications of the 2025 assessment findings, these are not academic misreadings — they produce materially incorrect advice at the level of the numbers and qualifications that regulators will test.

How AI gets this regulation wrong

Every failure on this regulation followed the same pattern: AI tools gave confident, detailed answers that were structurally wrong, and retracted or contradicted themselves when challenged on the precise rule text. The errors were not random — they clustered on the internal architecture of Principle 15 KC3, with AI tools routinely collapsing distinct Key Considerations into synthetic composites, reassigning provisions to the wrong KC, or inventing dual-track tests from elements that live in separate parts of the standard.

What that means for your practice

All three failures in this cell carry liability and professional indemnity exposure — the risk category that lands squarely on the accountant, not just the FMI client. Where the AI's error flows into a written opinion, a sign-off on an LNAFE calculation, or a briefing to a CCP risk committee, the practitioner has adopted a misstatement of the rule as a professional work product.

When this affects Accountants (CA/PA)

The most direct exposure arises when an accountant is retained by an FMI — or by its auditor — to verify or opine on the sufficiency of LNAFE under Principle 15 KC3. The 2025 Level 3 assessment has given regulators an explicit benchmark to test FMI compliance against: firms that cannot demonstrate a correctly-structured six-month operating expense floor, held in qualifying assets with a compliant Basel carve-out where applicable, will face supervisory escalation.

An accountant who has relied on an AI tool to characterise the KC3 standard — and who then frames their opinion around an invented "greater of" test, or around a KC4 liquidity overlay that has no basis in KC3 — is handing the regulator a discrepancy between the sign-off and the actual rule. That is a professional liability event, not just a drafting issue.

The same risk materialises in scoping new engagements. When onboarding a CCP client in the wake of the 2025 assessment, an accountant may use AI to sketch the Principle 15 compliance landscape before the first substantive meeting. If the AI places the six-month LNAFE floor in KC2 rather than KC3, or misframes the Basel equity carve-out as a liquidity condition, the scope memo and the initial risk register inherit those errors. The client's own risk team — who have read the actual CPMI text — will notice. At a minimum, it damages credibility in the opening phase of an engagement.

At worst, it locks incorrect assumptions into the engagement scope before anyone has looked at the original document.

Accountants who are training or supervising junior team members face a compounding risk. A detailed, confident AI answer that internally contradicts the rule text is exactly the kind of output a junior will not challenge — the structure sounds authoritative, the terminology is correct, and the error is in the precise framing of a condition, not in the vocabulary. The result is that incorrect interpretations of KC3 propagate through team working papers and client deliverables until someone senior enough to spot the discrepancy — or a regulator review — forces a correction.

The findings at a glance

The three findings below cover the specific Principle 15 KC3 questions where AI tools produced hallucinations — each representing a distinct structural misreading of the LNAFE standard that a Accountants (CA/PA) in international jurisdictions would encounter in practice.

Aggregate impact

The three failures here are not isolated retrieval errors — they reflect a systematic misrepresentation of how Principle 15's Key Considerations are structured. AI tools appear to treat the KC framework as a loose grouping rather than a set of distinct, separately-numbered obligations, and they resolve ambiguity by merging adjacent KCs into synthetic rules.

The KC2/KC3 conflation (attributing the six-month LNAFE floor to KC2 rather than KC3) and the KC3/KC4 conflation (applying a KC4 liquidity test as a condition on the KC3 Basel carve-out) are both examples of this: the AI has reconstructed a plausible-sounding Principle 15 from the right building blocks, but assembled them in the wrong order. For a regulator testing compliance against the actual KC3 text, the distinction is not academic — the six-month floor, the qualifying conditions for Basel equity, and the scenario-analysis sizing obligation each sit in a specific KC and are assessed separately.

The professional indemnity concentration is notable. All three failures land in the same risk category — liability exposure for the practitioner — because all three errors concern the precise quantitative and qualifying conditions that an accountant would be expected to get right when opining on LNAFE sufficiency. An error in the framing of a Basel carve-out condition, or in the location of the six-month floor, is not a general interpretive disagreement; it is a misstatement of the rule that a peer reviewer or a regulator can identify against the source text in seconds.

The AI's pattern of initial confidence followed by retraction under challenge is also significant: it means the error is not a stable misreading that a practitioner might spot by asking a follow-up question — the AI will contradict itself, leaving the user with no reliable answer and a risk that the first, wrong response has already been relied upon.

For Accountants (CA/PA) advising CCPs and other FMIs across multiple jurisdictions, the 2025 Level 3 assessment has raised the stakes for KC3 compliance opinions specifically. Assessment findings that flag non-compliance with the LNAFE standard will be cross-referenced against practitioners' own work product in regulatory correspondence and remediation plans. An accountant whose prior opinion rested on an AI-generated characterisation of KC3 that turns out to be wrong faces both a reputational and a PI exposure — and the fact that the error originated with an AI tool is not a defence.

What your team should do

The default position for any KC3 LNAFE work on this regulation should be source-first, AI-never-primary. The failures here are not edge cases or unusual interpretations — they concern the core quantitative standard in Principle 15 KC3, which is a short, unambiguous paragraph in the CPMI-IOSCO text. If a team member cannot locate and read the KC3 text directly before drafting any opinion or calculation note, the engagement has a process problem that no AI safeguard will fix.

For Principle 15 KC3 specifically, the original CPMI-IOSCO PFMI document and the November 2025 assessment report are the only authoritative inputs for advice work.

Where AI tools remain useful for this regulation is at the contextual and structural level — mapping the overall Principle 15 framework before diving into the source text, summarising publicly available commentary on the 2025 assessment findings, or drafting non-technical client summaries once the practitioner has verified the underlying rule characterisation independently. AI tools should not be used to characterise the precise conditions of any KC3 qualifier, to establish the minimum LNAFE quantum, or to frame the Basel/CRD equity carve-out.

For briefing work that covers multiple PFMI principles in combination (as a CCP risk committee briefing typically would), the KC-level structure of each principle should be verified against source before the AI's framing is adopted.

For supervision, the practical safeguard is requiring team members to cite the KC and page reference for any quantitative or qualifying condition they state in working papers on Principle 15. An AI-generated answer that cannot be traced to a specific KC location in the CPMI text should not survive peer review. This is not an onerous standard — the KC3 LNAFE provisions are a paragraph — but it surfaces exactly the kind of error documented here: a confident AI answer that attributes the six-month floor to KC2, or invents a KC4 liquidity overlay, will fail the citation test immediately.

How RLB Can Help

RegLeg's published Hallucination Research is available as open reference — use it as a pre-flight check before relying on AI output on regulatory questions that matter to your sign-off. The findings are organised by regulation and failure mode, so if you are working across IFRS application guidance, PCAOB standards, or cross-border group reporting obligations, you can pull the relevant regulation page and see, specifically, where AI tools have fabricated citations, misstated effective dates, or collapsed jurisdiction-specific carve-outs into a single incorrect answer. That is faster and more defensible than discovering the error after the advice has gone out.

For firms running multiple Accountants on the same regulatory portfolio — group reporting, audit quality frameworks, independence requirements across jurisdictions — RegLeg offers bespoke deep-dives. We work through the specific regulations in scope, map the failure modes that surface most consistently in that regulatory space, and produce a structured briefing your team can use as a standing reference. This is not a one-size engagement: the output is scoped to the regulations you are actually using AI tools against, and framed around the workflow decisions those findings affect — materiality judgements, disclosure drafting, cross-border reconciliation.

We also produce training and CPD-aligned material built around the failure modes your team should be stress-testing in their own AI use. Not generic AI literacy content — specific failure patterns documented against the regulations accountants in international practice touch most, presented in a format that maps to the professional judgement calls your team makes daily. If your firm has an existing AI-use policy, we can review it confidentially against RegLeg's failure-mode catalogue and flag where the policy's assumptions about AI reliability are not supported by what the research actually shows.