Executive Summary
Finance teams at Payment Institutions firms operating under PFMI oversight — whether directly designated or held to equivalent standards by their home supervisors — rely on Principle 15 to calibrate their liquid net assets funded by equity (LNAFE) buffers, sign off the Basel capital carve-out treatment, and defend those calculations to internal audit and regulators. Across three findings on this regulation, AI assistants we tested produced materially incorrect answers on the specific mechanics of PFMI Principle 15 Key Consideration 3: the minimum LNAFE floor, how it is structured, and the precise condition governing Basel-eligible equity.
In every case the AI answered with apparent confidence before retracting or contradicting itself when pressed, a failure pattern that is particularly dangerous when a Finance professional treats the first response as authoritative. The errors are not fringe misreadings — they go to the quantitative floor your team uses to set the buffer and the capital-eligibility condition your treasury function uses to avoid double-counting Basel CET1.
A firm that acts on any of these AI responses risks submitting a deficient LNAFE calculation to its supervisor or miscalibrating the Basel carve-out in a way that either understates the required buffer or inappropriately offsets capital that should count separately.
How AI gets this regulation wrong
Every failure on this regulation shares the same pattern: AI assistants invented rule text that does not exist, presented it confidently, and then reversed or contradicted themselves when the question was rephrased or the source was cited back to them. The errors concentrate on Principle 15's internal architecture — specifically, the boundary between Key Considerations 2 and 3 — where AI tools consistently merged obligations from separate provisions, invented composite tests, or misattributed which KC contains which quantitative requirement.
| AI's Failure Mode | Count | Affected findings |
|---|---|---|
| Exposed Fabrication | 2 | Finding#1 · Finding#2 |
What that means for your team
All three failures on this regulation carry regulatory enforcement risk — the kind that surfaces when a submitted LNAFE calculation, a capital adequacy attestation, or a supervisory self-assessment contains a materially incorrect interpretation of Principle 15. For Finance teams at Payment Institutions, where the buffer quantum and the Basel eligibility determination are both Finance-owned sign-off items, the risk travels directly from an incorrect AI answer into a document your CFO or Head of Treasury has already signed off on.
| Risk Impact | Count | Affected findings |
|---|---|---|
| Regulatory enforcement | 2 | Finding#1 · Finding#2 |
When this affects your department
Finance teams at Payment Institutions firms consult AI on this regulation most frequently during three moments: (1) drafting or updating the internal LNAFE policy and capital buffer framework — particularly when a new CPMI-IOSCO assessment cycle prompts a policy review; (2) preparing regulatory submissions or self-assessments under Principle 15, where the KC3 quantitative floor and the Basel carve-out treatment must be precisely described; and (3) briefing non-Finance stakeholders — board risk committee, external auditors, or a new Head of Treasury — who need a technically accurate summary of what the firm is required to hold and why.
The specific regulatory arithmetic at stake is not abstract. KC3 sets a hard six-month operating-expense floor for LNAFE, and separately permits equity held under international risk-based capital standards (Basel CET1) to count toward that floor where relevant and appropriate to avoid duplicate capital requirements. If your Finance team asks an AI assistant to explain that carve-out and receives a response that either invents a supplementary KC4 liquidity condition or flatly denies the carve-out exists, the resulting internal guidance memo or policy schedule will misstate the eligibility criteria.
A treasury function calibrating the buffer against incorrect eligibility criteria may exclude Basel CET1 that legitimately qualifies — overstating the required additional liquidity — or, worse, include equity on terms the regulation does not support.
Similarly, if a junior analyst uses AI to draft a technical briefing on KC3 and the AI attributes the six-month floor to KC2 (or invents a "greater of" dual-track requirement combining scenario analysis with the six-month minimum), the error will typically propagate unchallenged. The KC2/KC3 boundary is not something a non-specialist reader will catch, and internal reviewers who have not read the verbatim PFMI text in detail are unlikely to spot the misattribution before the document is circulated.
In a supervisory examination, that kind of structural error in a submitted policy document or self-assessment response is exactly the finding that triggers a follow-up request for a full corrective review.
The findings at a glance
All three findings below concern PFMI Principle 15's liquid net assets funded by equity mechanics — the minimum buffer floor, how it is calculated, and what equity qualifies under the Basel carve-out provision in KC3.
Aggregate impact
The three failures cluster with unusual tightness: all three concern PFMI Principle 15 Key Consideration 3, and all three involve AI assistants mischaracterising the structure of a single, relatively compact provision. That clustering tells you something operationally important. The AI tools tested do not have a general problem with Principle 15 — they have a specific problem with the internal architecture of KC3, and in particular its boundary with KC2.
Two distinct failure modes are visible: one AI tool invented a condition that does not appear in KC3 (a KC4 liquidity test for the Basel carve-out), while another flatly denied the carve-out exists at all. A third tool placed the quantitative six-month floor in KC2 rather than KC3. In each case the AI corrected itself, partially, when challenged — but the corrections were themselves inconsistent, suggesting the model has no stable internal representation of which obligation sits in which Key Consideration.
For Finance teams at Payment Institutions, the systemic risk is that these are not exotic edge-case questions. They are the exact questions a junior finance analyst or a newly-joined treasury specialist would ask when getting up to speed on the firm's LNAFE framework. The KC3 floor calculation and the Basel carve-out eligibility determination are routine Finance sign-off items — they appear in capital adequacy frameworks, in treasury policy documents, in regulatory reporting footnotes, and in responses to supervisory information requests.
A firm that has incorporated an AI-derived misstatement of KC3 into its standing policy documentation faces not just a one-off error but a systemic misalignment between its stated framework and the regulator's text, the kind of gap that emerges prominently in a Principle 15 supervisory deep-dive.
The regulatory enforcement risk is direct and traceable. CPMI-IOSCO's Level 3 assessments are specifically designed to test whether FMIs' LNAFE frameworks comply with KC3 in practice — not just on paper. A Payment Institution whose Finance function has documented the wrong floor structure, or whose Basel carve-out eligibility criteria do not match KC3's verbatim condition, faces a deficiency finding that requires a formal remediation plan, remediation evidence, and typically a follow-up supervisory review.
The cost is not just the remediation itself but the supervisory relationship damage and the internal credibility loss when Finance has to explain why a core capital policy was built on a misreading of a CPMI document.
What your team should do
The default position for Finance teams using AI on Principle 15 KC3 questions should be: read the verbatim PFMI text before accepting any AI characterisation of the floor structure or capital eligibility conditions. The failures we found are not subtle — an AI telling you that KC3 contains no Basel carve-out, or that the minimum is a "greater of" dual-track figure, is contradicting text that is publicly available on the BIS website. The check takes less than five minutes and eliminates the entire category of risk represented by these three findings.
If your firm's LNAFE framework documentation was drafted with AI assistance and has not been cross-checked line-by-line against the PFMI source text, treat that as an open action.
AI tools are safer on this regulation when used for orientation rather than authoritative interpretation. Asking an AI to summarise the general scope of Principle 15 general business risk, or to outline the categories of FMI operational risk, is lower-stakes: the cost of a soft characterisation being slightly off is manageable. The risk concentrates sharply when the question is quantitative or structural — the exact floor, the exact KC, the exact eligibility condition.
Those are the questions where AI consistently reached for plausible-sounding but fabricated specificity, and where Finance teams are most likely to rely on a precise answer without independently verifying it.
For regulatory submissions, supervisory self-assessments, and any internal document that will be cited in audit or regulatory defence, require the relevant PFMI KC text to be quoted directly alongside the firm's interpretation — not paraphrased from an AI summary. This is both a control against AI error and a demonstrably good practice in any CPMI-IOSCO assessment context, where supervisors are looking precisely at whether firms can articulate the regulatory basis for their calibration choices.
If a team member used AI to draft the KC3 section of your LNAFE policy, the remediation is a direct comparison of the policy language against the verbatim KC3 text — a targeted 30-minute exercise that should be standard procedure before any Principle 15 document leaves Finance for sign-off.
How RLB Can Help
RegLeg's published Hallucination Research gives your team a concrete pre-flight check before you rely on AI-assisted output on regulatory questions — capital adequacy calculations under PSD2/EMD2, safeguarding thresholds, DORA operational resilience obligations, cross-border settlement exposure under SWIFT or ISO 20022 migration rules. The research documents, by regulation and failure mode, where AI tools have demonstrably got it wrong: wrong thresholds, inverted obligations, fabricated carve-outs. Finance teams at payment institutions are running lean; the research is free, specific, and faster to apply than a compliance sign-off cycle you haven't built yet.
Where the published findings don't cover your specific regulatory footprint, we run bespoke regulator deep-dives scoped to the workflows your Finance function actually owns — treasury reporting under local settlement finality rules, IFRS 9 provisioning against regulatory capital floors, FX exposure aggregation for e-money passporting, PSR and national competent authority fee and levy calculations. The output is a prioritised map of which AI-supported steps in those workflows carry material hallucination exposure, expressed in terms your team can act on without translating through a methodology deck.
If your firm already has an AI-use policy — or is drafting one under pressure from a group risk or audit function — we can run a confidential review against our failure-mode catalogue and return a prioritised remediation list: which policy clauses are underspecified for Finance's actual AI touchpoints, which failure categories the policy doesn't address, and where the gap between what your policy says and what AI tools actually do is large enough to constitute a control weakness.
We also produce CPD-aligned training content your team can use internally — structured around the failure modes that are specific to payment institution Finance rather than generic AI literacy material your people will sit through once and ignore.