Regulation (EU) 2024/1689 · Artificial Intelligence Act ● Annex III: — days

The EU AI Act for Banks: Credit Scoring, Fraud Detection and Annex III Point 5

A single paragraph of Annex III decides whether a lender runs a high-risk AI system. What counts as evaluating creditworthiness, where the fraud carve-out ends, and what banks owe as deployers, providers or both.

Renomira Team · 11 min read
Classical bank buildings and modern towers in a financial district
Photo: Elisa.rolle / CC BY-SA 3.0 via Wikimedia Commons

A credit score decides a person’s economic future, and the EU AI Act treats it as one. Everything a bank, lender, broker or insurer needs sits in Annex III point 5 — essential private and public services: a creditworthiness limb, a life-and-health insurance limb, and a fraud-detection exception narrower than most compliance teams assume.

Why point 5 is the financial sector’s paragraph

Point 5 covers AI systems intended for evaluating the creditworthiness of natural persons or establishing their credit score — excepting systems used for detecting financial fraud — plus risk assessment and pricing in life and health insurance. The first two are yours:

Use case, as it actually appears Risk tier Key articles
Consumer, mortgage or personal-guarantee scoring High risk — Annex III 5(b) Art. 6(2), 26, 27, 86
Bank-built engine used on its own book High risk — provider and deployer Arts. 8–17, 43, 47, 48, 72, 73 + 26, 27
Life or health insurance risk assessment and pricing High risk — Annex III 5(c) Arts. 26, 27, 86
Fraud scoring (carve-out); AML, sanctions, PEP (not listed) Not high risk under Annex III GDPR, DORA, CRR, AMLD
Motor, property or travel insurance pricing Outside point 5 (life and health only) GDPR, IDD

What “evaluating creditworthiness” actually covers

In scope: any AI estimating whether a natural person can repay — a regional bank’s mortgage engine, a fintech lender’s buy-now-pay-later underwriting, an affordability or income-verification model, payroll- or open-banking-linked lending, a pre-approval engine deciding whether an application proceeds. No word “score” required: if it evaluates a person’s economic situation and changes what happens to them, it counts.

Natural persons only. Corporate lending sits outside point 5; personal guarantees, directors’ undertakings and sole traders pull it back to individuals. Purpose decides: a “revenue quality model” used to refuse a loan is a creditworthiness system; sized to a marketing budget, it is not. Write the intended purpose down; it is what your classifier reads first.

The fraud exception — and where AML sits

The carve-out is real and exact. Point 5(b) excludes systems “used for the purpose of detecting financial fraud”: application-fraud models, transaction monitoring and first-party fraud scoring stay out — classified by intended purpose, not by the model card’s label.

It collapses when fraud signals drive a credit decision. A merged risk score deciding whether someone receives credit is still evaluating creditworthiness. Keep the two models, purposes and records separate; that separation makes the exception defensible.

AML/CTF, sanctions and PEP screening are not listed in Annex III at all. No classification to argue — but no free pass: GDPR, AML law, DORA and supervisory expectations still apply.

Insurance: point 5(c) covers life and health only

The trigger is narrow. Point 5(c) covers risk assessment and pricing for life and health cover on natural persons. Motor, property, travel and liability pricing sit outside it — the broker quoting car insurance has no classification there, while a life quotation driven by a health-informed model is high-risk.

Brokers and mutuals are deployers, rarely providers. Configuring a vendor’s engine and applying its output to a quotation makes you the deployer under Article 26; you become a provider only where you develop a system, or substantially modify one, under your own name.

Health data is a second problem. GDPR Article 9 and a DPIA apply regardless of classification — and Article 27(4) lets the fundamental-rights assessment cross-refer to those DPIA sections rather than duplicate them.

Deployer, provider or both: the in-house problem

Bought and configured → deployer. Your Article 26 list: use it as instructed; assign named human oversight with authority to intervene; check input-data relevance; keep automatically generated logs (Article 26(6), six-month floor); co-operate with the authority; honour Article 86.

Built in-house, or substantially modified under your own name → also provider. On top come Articles 8–15 (risk management, data governance, logging, transparency), Annex IV technical documentation, Article 43 conformity assessment without a notified body, CE marking, Article 49 registration, Article 72 monitoring and Article 73 incident reporting.

Both at once is normal for a bank with a modelling team — it doubles the evidence, not the logic. A model-risk register only looks like Annex IV until you line the fields up.

Why the Article 6(3) derogation rarely rescues a scoring engine

A chapeau and four conditions. The system must not pose a significant risk of harm or materially influence the outcome of decision-making, and must do only a narrow procedural task, improve a previous human result, detect decision patterns without replacing human assessment, or act as a preparatory step.

Then the profiling wall closes it. Article 6(3) makes an Annex III system that profiles natural persons always high-risk — and credit scoring is textbook profiling under GDPR Article 4(4): automated evaluation of a person’s economic situation to predict repayment. The derogation reaches document-anomaly detection feeding a human underwriter; it does not reach a score that decides.

If your provider claims it, ask for the paperwork. Article 6(4) requires that assessment to be documented before market placement and produced on request — a claim without the document is a procurement gap.

What supervisors already expect: CRR, EBA and ECB model risk

The artefacts already sit in your model-governance file. CRR’s internal-ratings rules demand documented methodologies, data-quality controls and validation; EBA loan origination guidelines require documented creditworthiness assessment; ECB model-risk expectations require inventory and clear ownership.

The AI Act asks a different question. Validation asks whether the model measures risk accurately; the Act asks whether the use respects fundamental rights, whether a named human can intervene and whether Article 27 was completed first. A passed validation does not answer Article 27.

Use the supervisory seams. Article 74(6) makes your national financial supervisor the market surveillance authority for AI used under Union financial services law — the inspector you already report to. Articles 19 and 26(6) let logs sit inside existing sector record-keeping, and Article 72(4) folds monitoring into current plans.

The overlap stack: GDPR, ECOA and the Article 86 explanation

GDPR Article 22 already governs the decision. Refusing credit solely by automated processing with legal or similarly significant effects requires human intervention on request, the chance to express a point of view and the right to contest — plus logic transparency (Articles 13–15) and a DPIA.

Article 86 adds an AI-specific explanation. Anyone subject to a decision taken on the output of an Annex III high-risk system, which they consider adverse to their rights, may demand clear and meaningful explanations of the system’s role and the main elements of the decision. Design intake, identity verification, a drafted response and a named owner.

US-facing lenders carry ECOA too. Adverse-action notices must state specific principal reasons under Regulation B. One reason-code pipeline — each code traceable to the features that drove the outcome — serves Article 86, GDPR and Regulation B at once.

The obligations, in the order they bite

Live now: AI literacy (Article 4, February 2025), the prohibited-practice screen (Article 5), transparency when customers interact with an AI system (Article 50, since August 2026), and penalties — up to €15M or 3% of worldwide turnover for most breaches, €35M or 7% for prohibitions.

2 December 2027: the Annex III provider and deployer duties, moved from August 2027 by the Digital Omnibus — Article 26 oversight and logging, the Article 27 assessment, Article 49 registration for providers, and Article 86 operational rather than theoretical.

The FRIA trigger is uniquely yours. Deployers of Annex III 5(b) and 5(c) systems are the only private-sector deployers forced to assess fundamental rights before deployment — processes, affected groups, risks, oversight, remediation — then notify the results to the authority. Your credit model is in that set; your CV-screening tool is not.

How to check your own position

Six questions per system, in writing: (1) does it evaluate a person’s creditworthiness or price life or health insurance? (2) is fraud detection genuinely its purpose in live operation? (3) who develops it — a vendor, or your team under your own name? (4) does it profile individuals, closing the Article 6(3) derogation? (5) does the Article 27 FRIA trigger fire before deployment? (6) could you answer an Article 86 request today? Two signed, dated pages per system is a defensible file.

Frequently asked questions

Is credit scoring AI high risk under the EU AI Act?

Yes. Evaluating the creditworthiness of natural persons or establishing a credit score is Annex III point 5(b), making the system high-risk under Article 6(2). The Article 6(3) derogation never applies to systems that profile natural persons, and credit scoring is profiling by definition — so the provider or deployer regime applies in full, including the Article 27 fundamental-rights impact assessment.

Does the fraud-detection exception cover our AML and transaction monitoring?

Fraud detection is expressly excluded from point 5(b), so application-fraud and transaction-monitoring scoring are not high-risk on that basis. AML/CTF, sanctions and PEP screening are not listed in Annex III at all. The exception is purpose-bound: if fraud signals feed a model deciding whether someone receives credit, the carve-out does not follow it.

Does a bank that builds its own scoring model become a provider?

If you develop it, or substantially modify a vendor’s model and put it into service under your own name, you are both provider and deployer. Provider duties add Annex IV documentation, Article 43 conformity assessment, declaration of conformity, CE marking, Article 49 registration, Article 72 monitoring and Article 73 incident reporting — alongside the Article 26 duties you already owe.

Does the AI Act cover motor and property insurance pricing?

Not under Annex III point 5, limited to risk assessment and pricing for life and health cover on natural persons. Motor, property, travel and liability pricing have no high-risk classification on that basis, though GDPR, the Insurance Distribution Directive and conduct rules still apply — classify per product line, not per company.

Next step

Put this guide to work on your own estate.

Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.

  • Free forever plan
  • No credit card
  • EU-hosted · GDPR-native
Run a free scan Sign up free

Build a live AI register in minutes — upgrade only when you need the audit pack.

ai act for bankscredit scoring ai act high riskannex iii point 5 financial servicesai act for insurance companiesai in creditworthiness assessment ai actai act financial sector compliancefraud detection ai act risk classification

Keep reading