Short answer. A high-risk AI system is one the EU AI Act places under its strictest controls because it can affect a person’s safety, access to services, or fundamental rights. Article 6 and Annex III define the categories, and for banks and insurers they include AI that scores creditworthiness and AI that prices life or health insurance. Any system in scope must meet six obligations, from risk management to human oversight, before it runs in production, and breaches can draw fines up to 15 million EUR or 3% of worldwide turnover.
The Act sorts AI into four tiers: unacceptable, high, limited, and minimal risk. High risk is the tier that carries real compliance weight for financial services, and Annex III names the use cases directly. Two sit at the core of banking and insurance operations:
Annex III is not an automatic verdict. Article 6 adds a filter: a listed system can fall out of the high-risk tier when it performs a narrow procedural task and does not materially shape the outcome of a decision. One exception closes that escape route. Any system that profiles a person, meaning automated processing to evaluate their economic situation, behaviour, or reliability, stays high-risk. Credit scoring is profiling by definition, so for most lending models the high-risk label holds.
Credit scoring and life and health insurance pricing are high-risk under Annex III.
When a system falls in scope, Articles 9 to 15 require six controls: a documented risk management system that runs across the full lifecycle, data governance that keeps training data relevant and representative, technical documentation, automatic logging for traceability, human oversight exercised by named staff, and safeguards for accuracy, robustness, and cybersecurity. Before the system goes live, the provider also runs a conformity assessment, draws up an EU declaration of conformity, affixes the CE marking, and registers the system in the EU database.
One distinction decides who carries the work. The provider builds the system and proves conformity. The deployer, the bank or insurer that puts the system to use, still owns how it is used in practice, runs the human oversight, monitors the system in operation, and reports serious incidents to the national authority. A bank that buys a model from a vendor cannot outsource its compliance to that vendor.
The high-risk obligations for Annex III systems carry an application date of 2 August 2026. A late-2025 proposal, the Digital Omnibus, would defer that date to December 2027 for some systems, but it still needs endorsement by the Council and the Parliament, so August 2026 stays the planning anchor until the change is adopted. One related duty already binds: the AI literacy requirement under Article 4 has applied since 2 February 2025, so staff who operate AI systems must already hold a working level of AI literacy.
For CEE banks and insurers, the AI Act does not stand alone. It layers on top of DORA’s ICT risk rules, GDPR’s limits on automated decisions, and the model-risk governance these institutions already run. The European Banking Authority found in its November 2025 analysis that most high-risk controls overlap with frameworks supervised institutions operate today, so the task is mapping systems and closing gaps rather than starting from zero. Enforcement sits with national market surveillance authorities, not only the central AI Office, which means a Slovak, Czech, or Austrian supervisor can apply the rules differently from its neighbours.
The practical path is the same for a bank in Bratislava or an insurer in Vienna. Build an inventory of every AI system in use, classify each one against Annex III and the Article 6 filter, then run a gap analysis against the six requirements. Systems that score credit or price life and health insurance move to the top of the list. Documentation is the part teams underestimate, since a conformity assessment asks for evidence that the controls work, not a statement that they exist. Starting the evidence trail early matters, because logging and oversight records cannot be reconstructed after the fact, and a system without a history is hard to certify under deadline.
The cost is concrete. Breaches of the high-risk requirements can draw fines up to 15 million EUR or 3% of total worldwide annual turnover, whichever is higher. The quieter risk is misclassification in the other direction: a system treated as low-risk that a supervisor later judges high-risk arrives at the deadline with none of the six controls in place, no conformity assessment, and no logging history to reconstruct. Because enforcement sits with national market surveillance authorities, a Slovak, Czech, or Austrian supervisor can reach that judgement independently. The safe path is to classify conservatively, document the reasoning behind each call, and keep the evidence, so a borderline decision can be defended rather than rebuilt under deadline pressure.
Classification is the first decision, and it sets every step that follows. Ableneo shipped 34 production AI projects in 2025, and 94% of them used large language models, so most of our recent work sits inside the scope the AI Act now governs. That production record is why we read Annex III as an operating spec rather than a legal abstraction: each requirement maps to a control you can build, test, and observe. Our work with banks and insurers in Slovakia, the Czech Republic, and Austria starts with an AI inventory and a risk classification, then moves to the governance, logging, and documentation a conformity assessment will ask for. See our AI transformation services for how we take a system from classification to governed production.
Key takeaways
Planning AI in a regulated business? Ableneo takes systems from classification to governed production.