Risk teams in retail banking calibrate corridor-concentration, FMI-counterparty exposure and operational-resilience metrics for remittance and consumer cross-border products against a tight set of CPMI numbers: global FPS count, cross-border-enabled subset, planning-pipeline, and central-bank-versus-private operator mix. Two AI failures on this regulation hit that set from opposite directions. Opus 4.7 compresses the FPS universe to 57 and drops the operator-mix breakdown; Sonnet 4.6 holds the 70-plus headline correctly but denies the operator-mix percentages exist. sp231115 supplies the full set. A risk-appetite paper built on either AI answer enters risk committee with an inflated concentration ratio and no operator-mix differentiation.
What the AI got wrong, and why it matters here
Both failures land where retail-bank risk depends on tight denominators and operator-type signal: a fabricated low count, and a denied operator-mix line. Neither has a downstream check before the paper enters committee.
Finding 1: FPS denominator compressed
Opus 4.7 cited the 2025 monitoring survey at 57 (56 in one graph) operational FPS with no operator-type breakdown. sp231115 gives 70-plus operational, 14 cross-border-enabled, 24 in the planning pipeline, 40% central-bank and 35% private. A retail-bank concentration ratio built on the AI denominator inflates the exposure share and strips the operator-mix differentiation.
Citation: RLB-H-INT-BIS-CPMI-API-HARMONISATION-CROSS-BORDER-2024-Q010-Opus47.
Finding 2: Operator-mix denied
Sonnet 4.6 cited the 70-plus FPS headline correctly and denied that a precise central-bank-versus-private operator percentage exists in the Brief 10 summary. sp231115 names 40% central-bank and 35% private. Removing the operator-mix line collapses the central-bank-versus-private differentiation the risk paper depends on.
Citation: RLB-H-INT-BIS-CPMI-API-HARMONISATION-CROSS-BORDER-2024-Q010-Sonnet46.
When this hits the risk calendar
Retail-bank risk pulls CPMI material on three artefacts: the corridor-concentration paper, the FMI-counterparty exposure note, and the annual risk-appetite calibration for cross-border consumer products.
| Standing item | Where the AI risk surfaces | Failure mode |
|---|---|---|
| Corridor-concentration paper | Denominator and operator-mix | Findings 1 and 2 |
| FMI-counterparty exposure note | Operator-mix differentiation | Findings 1 and 2 |
| Annual risk-appetite calibration | Pipeline forward signal and operator-mix | Findings 1 and 2 |
Aggregate impact on the team
The same two failures collapse the operator-mix differentiation and the planning-pipeline forward signal, removing two of the three inputs risk-appetite calibration relies on.
| Risk Impact | Count | Affected findings |
|---|---|---|
| 0 |
What this team should do
Tag the FPS count and the operator mix as known-failure outputs. Any AI draft naming those numbers must be sent through a primary-source check against sp231115 before it lands in a risk paper or a risk-appetite calibration.
Detection patterns to add to AI-review
- FPS counts must trace to sp231115 or to a numbered CPMI brief.
- Operator-mix denials must be cross-checked against sp231115 directly.
How RLB can help
RLB tracks AI failures on the FPS-landscape numerical anchors and refreshes the catalogue against live AI subjects on rotation. Retail-bank risk can wire the catalogue into the risk-paper review step so these two failure shapes are caught before they reach risk committee.