Short answer. A bank classifies an AI system as high-risk by matching its intended purpose against the 8 areas listed in Annex III of the EU AI Act. Credit scoring and creditworthiness assessment fall under Annex III point 5(b) and are high-risk by default, with a narrow carve-out for fraud detection. If a system fits an Annex III use case it is high-risk under Article 6(2), unless the bank documents before deployment that it meets one of the four Article 6(3) exceptions. Full obligations for these systems apply from 2 August 2026, with penalties up to 35 million EUR or 7% of global turnover.
Classification is a decision you make one system at a time, based on what the system is used for, not on how it was built. A gradient-boosted model and a large language model land in the same high-risk category when both decide who gets a loan. The intended purpose sets the class, so the same model reused for a different task can carry a different classification.
The practical work is triage. A mid-market bank in Central Europe typically runs dozens of models and automations across credit, collections, fraud, marketing, and internal operations. Each one needs a recorded answer to a single question: does its intended purpose match an Annex III use case? Most banks find that a small share of their systems are high-risk and the majority are not, but they cannot prove that split without a documented review.
For banks and insurers the classification decision is the gate to every other EU AI Act obligation. A system classified high-risk must carry a risk management system, data governance controls, technical documentation, automated logging under Article 12, human oversight, and a conformity assessment before it goes live. Get the classification wrong in the permissive direction and the bank ships a high-risk system without those controls, exposed to fines up to 35 million EUR or 7% of worldwide annual turnover. Get it wrong in the cautious direction and the bank spends conformity-assessment effort on systems that never needed it.
The timing is fixed. Obligations for Annex III high-risk systems apply from 2 August 2026, and the European Commission has confirmed the rules reach systems already in production, not only new builds. In parallel, DORA already requires financial entities to inventory and govern the ICT and third-party services behind these models. A bank that has built a clean AI system inventory for DORA has most of the raw material for EU AI Act classification, which is why the two programmes belong in one operating model rather than two.
Classify each AI system by its intended purpose against the 8 Annex III areas, not by the technology it uses.
A system is high-risk when its intended purpose matches one of the use cases in Annex III, or when it is a safety component of a product covered by separate EU product law. For financial services the relevant Annex III entries sit under point 5, access to essential private services. Point 5(b) covers AI used to evaluate the creditworthiness of natural persons or to establish their credit score, and point 5(c) covers risk assessment and pricing for life and health insurance. A single explicit exemption applies: systems used to detect financial fraud are not high-risk on the point 5(b) trigger.
Two points trip banks up. First, business-to-business lending decisions about companies are not automatically caught, because Annex III point 5(b) is written around natural persons. Second, a system that profiles individuals, meaning it automatically evaluates aspects of a person’s economic situation, behaviour, or reliability, stays high-risk even where a bank argues its output is only advisory. Profiling removes access to the Article 6(3) exceptions described below.
The classification of the system does not change, but your obligations do. A provider builds or substantially modifies the system and owns the heavy duties: Annex IV technical documentation, the conformity assessment under Article 43, and registration in the EU database. A deployer uses a third-party system without material change and owns human oversight, staff AI literacy under Article 4, and a fundamental rights impact assessment. Most banks are deployers for foundation models and providers for the scoring logic they train on their own portfolios.
The line moves. Under Article 25, a deployer that fine-tunes a licensed model on its own data, or puts its own name on the system, can become a provider and inherit the full provider obligations. A bank that fine-tunes a general-purpose model on its loan book to score creditworthiness has almost certainly crossed that line. Recording the provider-or-deployer answer for each high-risk system is therefore part of classification, not a later step.
Article 6(3) lets a provider treat an Annex III system as not high-risk when it does not pose a significant risk to health, safety, or fundamental rights. It applies in four situations: the system performs a narrow procedural task, improves the result of a completed human activity, detects decision patterns without replacing human judgement, or performs preparatory work for an assessment. A model that automates a self-contained data-cleaning step inside a larger manual credit review can qualify.
The exception carries an evidence burden. The provider must document the assessment before the system is placed on the market or put into service, keep it available for supervisory authorities, and register the system in the EU database anyway. Any system that profiles individuals is barred from the exception. Treat Article 6(3) as a documented derogation with proof attached, not as a reason to skip classification.
Most of a bank’s AI estate sits outside Annex III. A collections-prioritisation tool, a document-processing model for onboarding, and an internal knowledge assistant are usually minimal-risk and carry no high-risk obligations, though good practice still applies. Two categories need a second look. Systems used in recruitment or to make decisions about staff fall under Annex III point 4 and are high-risk. Any system that interacts with people or generates content carries the Article 50 transparency duties regardless of risk class.
General-purpose AI models add a separate track. When a bank builds on a general-purpose model, that model carries its own provider obligations upstream, and the bank inherits deployer duties for the high-risk system it builds on top. Classification therefore records two things for each system: its own risk class, and whether it depends on a general-purpose model whose documentation the bank needs to obtain.
Start with an AI system inventory. List every model and automation in use or in build, its intended purpose, its owner, the data it uses, and whether it depends on a third-party or general-purpose model. Against each row, record the Annex III match, the resulting risk class, the provider-or-deployer status, and any Article 6(3) assessment. This register is the single artefact a supervisor, a DORA auditor, and the board all ask to see, and it is the fastest way to separate the few high-risk systems from the many that are not.
Classification is where EU AI Act programmes either start clean or start confused. Ableneo has shipped 34 production AI projects, roughly 4 of 5 reaching production, for financial and insurance clients including banks in the same regulatory world as ČSOB, Erste, and UNIQA, so the classification register we build is grounded in systems that actually run under supervision, not a paper exercise. We tie the AI system inventory to the ICT inventory a bank already maintains for DORA, so classification, governance, and audit evidence live in one operating model. See our AI transformation FAQ for the connected questions on governance, DORA sign-off, and conformity assessment.
Key takeaways
Planning AI in a regulated business? Ableneo takes systems from classification to governed production.