Lawyers advising on the CPMI Harmonised ISO 20022 Data Requirements (Updated Report) are increasingly using AI to draft client memos on Fedwire postal address compliance, validate threshold language in correspondent banking opinions, and prepare partner-level briefings on the harmonised ISO 20022 governance lineage. The same tools support first-pass advice on cross-border payment obligations under the updated CPMI data model.
Two frontier AI models tested by the RLB Specialist Panel on the workflows lawyers use to support advice on the CPMI Harmonised ISO 20022 Data Requirements (Updated Report) produced three discrete hallucinations bound to regulator-issued source text. The Panel records two distinct failure classes, Numeric Drift and Schema Over-Specification across the set. Questions were prepared by the Specialist Panel based on real practical AI usage in the workflows lawyers use AI for, and each finding is bound to verbatim regulator-issued source text held as primary substrate.
For Lawyers, each hallucination has a direct operational consequence in the regulatory opinion, partner-level memo, or correspondent banking advice. The Panel's testing surfaces ISO 20022 adoption rate conflation (RTGS vs faster payments), Fedwire hybrid postal address schema over-specification, and ISO 20022 adoption rate conflation (RTGS vs faster payments). Where these errors flow into a deliverable, the exposure is PI exposure, client correction, and discoverable error in opinion drafts that propagate to multiple counterparties.
The pattern is uniform across the set: the AI returns a confident, sourced-looking answer that conflicts in a load-bearing specific with the regulator's verbatim text, and the error survives a first-pass review precisely because the surface form is plausible. The Panel records each hallucination with the regulator's primary substrate held as the anchor, so the corrective text is available alongside the failure.
The Specialist Panel records the citation IDs as follows: RLB-H-INT-BIS-CPMI-ISO-20022-HARMONISATION-UPDATED-2026-Q006-Opus47 (Claude Opus 4.7 (web search on), Numeric Drift); RLB-H-INT-BIS-CPMI-ISO-20022-HARMONISATION-UPDATED-2026-Q010-Opus47 (Claude Opus 4.7 (web search on), Schema Over-Specification); RLB-H-INT-BIS-CPMI-ISO-20022-HARMONISATION-UPDATED-2026-Q006-Sonnet46 (Claude Sonnet 4.6 (web search on), Numeric Drift). Each citation links to the verbatim regulator-issued source text, the tested AI question, and the recorded AI response, so the Panel's assessment is traceable end to end. For lawyers, the citation IDs operate as a reference index: when an AI answer in the working draft matches a known Panel finding, the cited regulator text is already available as the corrective anchor.
The full per-finding analysis cards, including the audience-specific impact statement, sit on the cell's detail surface for sign-off use.
This is the consolidated view of findings. Click the Citation IDs or 'see details →' on any item for the full details for each finding.
Advising a client on the current state of ISO 20022 adoption using a single 79% figure, when the authoritative source reports materially lower RTGS adoption than FPS adoption, misrepresents the market in a way that matters for timing advice, benchmarking arguments, and regulatory comparisons across system types. When the AI was challenged on this figure, it acknowledged it was reconstructed and likely conflated across years, meaning there is no underlying source the lawyer could trace the number back to. Any regulatory submission or opinion that includes adoption statistics derived from this response would be built on an invented figure.
The Fedwire hybrid postal address format specifies optional free-format lines of 70 characters each for the unstructured component, not optional structured fields such as Street Name, Building Number, and Post Code. AI tools we tested inverted this, substituting structured optional elements drawn from general CBPR+ address schema knowledge. That substitution would be invisible in a first-read review of an advice memo or compliance annex.
A lawyer who includes the AI's description in implementation guidance delivered to a Fedwire participant, or in a sign-off on a compliance specification, has delivered a technically wrong answer about a named system's format requirement, with PI exposure that is difficult to disclaim once the description appears in a client deliverable.
Claude Sonnet 4.6 with web search exhibited the same blended 79% adoption figure as the Opus 4.7 test on the matching question, collapsing the regulator's separately-stated faster-payment and RTGS rates into a single composite. The structural exposure for this audience is identical to the Opus variant: advising a client on the current state of ISO 20022 adoption using a single 79% figure, when the authoritative source reports materially lower RTGS adoption than FPS adoption, misrepresents the market in a way that matters for timing advice, benchmarking arguments, and regulatory comparisons across system types.
When the AI was challenged on this figure, it acknowledged it was reconstructed and likely conflated across years, meaning there is no underlying source the lawyer could trace the number back to. Any regulatory submission or opinion that includes adoption statistics derived from this response would be built on an invented figure.
Every finding on this page compares an AI subject's account of the rule against the regulator's verbatim text from the regulator's own portal. Both are linked. Each delta, its root causes, and impact analysis are documented and published with immutable Citation IDs.