What financial firms get wrong.
The routes the Act grants, with their conditions.
Three financial deployments.
A lender running a bought credit-scoring model.
- Annex III 5(b): high-risk by listing, with Art. 26 deployer duties.
- Art. 27 rights assessment applies, feeding the DPIA view.
- Art. 26(5) monitoring folded in, declared rather than assumed.
12 obligations apply to a deployer of a high-risk system.
An insurer pricing life cover with a model it built.
- Annex III 5(c) and the provider set, Art. 9 to Art. 49.
- Art. 17(4) folds most Art. 17(1) aspects into internal governance.
- The reliance recorded, naming the instrument it rests on.
A bank running fraud detection and scoring on related models.
- The fraud model sits inside the Annex III 5(b) carve-out.
- The scoring model does not, even on shared architecture.
- Classification asks about purpose, so the two separate cleanly.
What sits beside the sector duties.
Questions on the EU AI Act for financial services.
Which financial use cases are high-risk?
Annex III point 5(b): creditworthiness and credit scoring, except to detect financial fraud. Point 5(c): risk assessment and pricing in life and health insurance. Both are high-risk by listing; what matters is how you evidence the duties.
We already run internal governance under CRD or Solvency II. Does that count?
Yes, in two places. Article 17(4) lets a provider that is a financial institution fold the Article 17(1) aspects, except (g), (h) and (i), into its internal governance under Union financial-services law. Article 26(5) does the same for deployer monitoring. Neither is a general exemption.
Does a fundamental rights impact assessment apply to us?
Article 27 binds deployers of the creditworthiness and insurance-pricing systems, alongside public bodies. It can build on your DPIA (Art. 27(4)); Veritome lays its answers under the DPIA's headings so the two stay consistent.
Is fraud detection caught?
Annex III point 5(b) carves out fraud detection, and the carve-out attaches to purpose. A model used for fraud detection and separately for scoring is not covered in the second use.
How does this sit with DORA and existing model risk management?
Alongside them. Much of their evidence is reusable as input; the EU AI Act adds a classification, the obligations derived from it and the article behind each.