Engineering and data teams in payment institutions land on d224 and d230 at the API gateway, the ISO 20022 schema and message-validation layer, the address-normalisation pipeline and the data-lineage register. Two AI failures on this regulation hit those surfaces directly. Opus 4.7 returned a reconstructed stakeholder taxonomy against d224's 10 recommendations, and Sonnet 4.6 committed to a November 2026 ISO 20022 structured-address cutover that does not appear in the d230 source text. A backlog or capability map built off either AI output routes tickets to the wrong squad and schedules engineering capacity against an undocumented regulatory deadline.
What the AI got wrong, and why it matters here
Both failures inject content into the two deliverables engineering and data write fast: a capability map and a readiness backlog. The cost of catching either later, after sprints have committed, is materially larger than the cost of a primary-source check up front.
Finding 1: Reconstructed stakeholder taxonomy
Opus 4.7 returned a clean stakeholder taxonomy across d224's 10 recommendations, built from category labels rather than the recommendation text. A backlog scoped off the AI taxonomy routes tickets to the wrong squad and drops recommendations the AI silently elided.
Citation: RLB-H-INT-BIS-CPMI-API-HARMONISATION-CROSS-BORDER-2024-Q008-Opus47.
Finding 2: Fabricated November 2026 ISO 20022 cutover
Sonnet 4.6 committed to a hard November 2026 structured-address cutover for ISO 20022 cross-border payment messages, framed as a d230 commitment. The d230 source describes only standardisation and regulatory developments since 2023 and a separate technical annex. A readiness epic or backlog ticket scoped against the AI line books engineering capacity against a regulatory deadline that does not exist.
Citation: RLB-H-INT-BIS-CPMI-API-HARMONISATION-CROSS-BORDER-2024-Q009-Sonnet46.
When this hits the engineering calendar
Engineering and data pull CPMI material on three artefacts: the d224 capability map that drives backlog scoping, the ISO 20022 readiness backlog and address-format epics, and the data-lineage register for cross-border payment fields.
| Standing item | Where the AI risk surfaces | Failure mode |
|---|---|---|
| d224 capability map | Stakeholder-to-recommendation routing | Finding 1 |
| ISO 20022 readiness backlog | Cutover-date commitments | Finding 2 |
| Data-lineage register | Both | Both |
Aggregate impact on the team
The two failures together corrupt both the capability map (Finding 1) and the readiness ticketing (Finding 2). The downstream cost is wasted engineering capacity and a capability map that breaks in architecture review.
| Risk Impact | Count | Affected findings |
|---|---|---|
| 0 |
What this team should do
Tag the d224 stakeholder taxonomy and the d230 ISO 20022 cutover date as known-failure outputs. Any AI draft that touches either must be returned through a primary-source check before it enters the backlog or the data-lineage register.
Detection patterns to add to AI-review
- Stakeholder-to-recommendation mappings on d224 must be verified against the recommendation text.
- ISO 20022 cutover-date assertions against d230 must be verified against the d230 text and technical annex.
How RLB can help
RLB tracks AI failures on d224 and d230 and refreshes the catalogue against live AI subjects on rotation. PI engineering and data can wire the catalogue into the backlog-grooming review step so these two failure shapes never reach a sprint commitment.