AI Governance for Banks: Credit Scoring Is High-Risk, Fraud Detection Is Not

A bank's AI estate is broad — credit decisioning, AML transaction monitoring, customer chatbots, branch-network forecasting, document processing — and the EU AI Act does not treat any two of those the same. The mistake most banks make in their first governance pass is to sweep everything into one "AI policy" and assume the strictest rules apply to all of it. They do not.

"AI governance" itself is a management practice, not a clause in Regulation (EU) 2024/1689. The Act imposes duties — screen for prohibited use, classify by use, manage risk on high-risk systems, disclose where Article 50 bites — and leaves the organisation to you. The job of a bank-grade governance model is to route each of those duties to a named owner, and to do it without bolting a parallel structure onto the model-risk-management machinery a regulated bank already runs. This guide maps the AI Act onto that existing machinery.

---

What is high-risk in a bank, and what is not

The single classification that matters most for a bank sits in Annex III, point 5(b): AI used to evaluate the creditworthiness of natural persons or establish their credit score. Consumer lending decisions, credit-card approvals, retail mortgage scoring, overdraft and limit decisions — if an AI system materially shapes whether a person gets credit and on what terms, it is high-risk. It does not need to make the decision alone; a score a credit officer relies on is still in scope.

Two carve-outs change the picture sharply:

• Fraud detection is explicitly excluded from Annex III point 5(b). A transaction-monitoring, payment-anomaly or account-takeover model is not high-risk on this basis, however sophisticated. The legislature drew the line deliberately: fraud models protect customers and the system; they do not gate access to credit.
• AML and sanctions screening likewise are not high-risk under Annex III. They sit under the EU AML framework and, where they make solely-automated decisions about a person, under GDPR Article 22 — but not the AI Act high-risk stack.

Life and health insurance risk-and-pricing models are high-risk under Annex III point 5(c) — relevant to bancassurance arms, less so to a pure deposit-and-lending bank.

Getting this map right is the governance model's first deliverable. Mis-tagging a fraud model as high-risk is not harmless caution — it tells an examiner your classification method is unreliable and wastes the budget the credit model actually needs.

---

The customer chatbot: Article 50, not Annex III

Retail banks run chatbots and voice assistants across servicing and onboarding. These are limited-risk: Article 50 requires that a person be told, clearly, that they are interacting with an AI system unless it is obvious. That is a disclosure obligation — a banner, an opening line, a voice prompt — not a conformity assessment. Article 50 has applied since 2 August 2026 — it was not delayed by the Digital Omnibus. A breach is the €15M/3% penalty tier (Article 99(4)), the same tier as most high-risk breaches — so disclosure is cheap to do and expensive to miss.

---

Provider, deployer, or both

Most banks are deployers under Article 26 of credit models they license from a vendor: they embed the score in origination and use the output. A bank that builds its own credit model in-house is also a provider under Article 16 and carries the full provider stack. And Article 25 means the line can move: a bank that puts its own brand on a vendor model, substantially modifies it, or retrains it on its proprietary book has likely become a provider for that system.

The practical governance rule: the use-case register records, for each system, both its classification and its role — because the role decides the obligation list.

---

Mapping AI Act duties onto model risk management

A regulated bank already runs model risk management — the discipline the ECB's TRIM exercise, the ECB Guide to internal models, and the EBA's guidelines on internal governance and on loan origination (EBA/GL/2020/06) have shaped for years. The AI Act does not ask a bank to build a second, parallel apparatus. It asks the bank to extend the one it has to cover the AI Act's specific requirements, and to prove it.

Mapping the high-risk provider/deployer duties onto existing functions:

The gaps that are genuinely new — and so need a named owner from day one — are the Article 43 conformity assessment, the Article 49 EU-database registration, the Article 27 FRIA for credit-scoring deployers, and Article 73 reporting to the market-surveillance authority. Everything else is an extension of an existing control, not a greenfield build. ECB/EBA model-risk evidence is reusable; it is not, on its own, AI Act conformity.

---

The DORA interplay

DORA (the Digital Operational Resilience Act, applying since 17 January 2025) governs ICT risk and third-party oversight for financial entities. A vendor credit model is simultaneously a DORA ICT third party (subject to the register of information and the contractual requirements of DORA Articles 28–30) and an AI Act high-risk system the bank deploys. The documentation is distinct, but the vendor due-diligence process can run once and feed both: a single intake that captures the DORA third-party register fields and the AI Act provider evidence (the Article 13 instructions, the Article 47 declaration of conformity) is the efficient design. Governance models that keep DORA and AI Act vendor onboarding in separate silos duplicate work and miss the systems that trigger both.

---

Worked example: a regional bank's governance model

Donaubank AG — ~750 employees, ~€210M turnover, retail and SME lending in Germany and Austria — runs four material AI systems: an in-house consumer credit-scoring model, a licensed fraud-detection engine, a customer-servicing chatbot, and an internal document-classification tool.

Its governance model, hub-and-spoke:

• Accountable executive: the CRO, who already owns model risk, takes AI risk onto the risk map and approves high-risk go/no-go calls.
• AI governance lead: a second-line role that owns the use-case register and the Article 5/6 classification method, sitting beside model validation.
• Classification result: the credit model is high-risk (Annex III 5(b)); as the bank both built and uses it, Donaubank is provider and deployer. The fraud engine and document tool are minimal; the chatbot is Article 50 limited.
• The high-risk work, by 2 December 2027: extend the existing model documentation to the Article 11 / Annex IV pack, formalise the Article 9 RMS against the model-risk framework, complete the Article 43 internal-control conformity assessment, sign the Article 47 declaration and register under Article 49, run the Article 27 FRIA (mandatory for creditworthiness deployers), and confirm Article 14 credit-officer override is documented.
• The duties live now: Article 4 AI literacy (measures to support staff AI literacy, in force since 2 February 2025) and Article 5 prohibited-use screening at intake — both owned today, not in 2027. The chatbot's Article 50 disclosure has been required since 2 August 2026.

Donaubank does not start from zero. It maps the Act onto the model-risk machinery it already runs and fills four genuinely new gaps — conformity assessment, registration, FRIA, incident reporting.

---

The deadline after the Digital Omnibus

For stand-alone high-risk systems such as a bank's credit model, the original application date was 2 August 2026. Under the Digital Omnibus — a Commission proposal of November 2025, now Regulation (EU) 2026/1744 — that moved to 2 December 2027 (and 2 August 2028 for high-risk AI embedded in Annex I products).

The point worth stating plainly: the deferral is law. Regulation (EU) 2026/1744 was published in the Official Journal on 24 July 2026 and has been in force since 27 July 2026, so the amended AI Act itself now reads 2 December 2027 — there is one legal date, not a statutory date and a proposed one. A prudent bank plans against 2 December 2027. Assembling the Annex IV pack and generating an Article 9 evidence trail across a model lifecycle takes 12–18 months even from a mature model-risk base, so the extra runway is for hardening the file, not for waiting.

---

Penalties

For a bank, the relevant tiers under Article 99:

• Article 5 prohibited practices — €35,000,000 or 7% of worldwide annual turnover (Art 99(3)).
• High-risk obligations (Articles 9, 10, 11, 13, 14, 15, 43, 49) and Article 50 disclosure — €15,000,000 or 3% (Art 99(4)).
• Supplying incorrect or misleading information to an authority or notified body — €7,500,000 or 1% (Art 99(5)).

The third tier is 1%, not 1.5%. For groups that meet the SME/start-up definition, Article 99(6) caps the fine at the lower of the percentage or the fixed amount — relevant to smaller institutions, less so to a large banking group whose turnover dwarfs the fixed sums. For small mid-caps (SMCs), Article 99(6a) applies the same lower-of cap to the Article 99(4) and (5) tiers, though not to Article 5 fines.

---

How Confir helps

Confir steps a bank's intake through the Annex III points in plain-English checklists, confirms the credit model as Annex III 5(b) while correctly leaving the fraud and AML engines out of high-risk, derives the Article 16/26 role, and routes the obligation list to an owner. The classification is deterministic and rule-based — the same intake always produces the same finding, and the rule that fired is recorded — which is what a model-risk function and an examiner both want. It generates the Article 11 / Annex IV pack, the Article 47 declaration, and the Article 27 FRIA as structured outputs.

Related guides

• AI governance operating model
• EU AI Act for fintech
• Article 9 risk-management system
• Annex III credit-scoring AI
• Article 27 FRIA and the GDPR intersection
• Vendor management for AI
• EU AI Act deadlines and timeline