How do you know if your fintech AI model is high-risk under the EU AI Act
You check whether your model's intended purpose matches one of the specific use cases listed in Annex III of Regulation (EU) 2024/1689, because that list, not how sophisticated or risky the model feels to you, is what decides the legal classification. For most fintechs the honest answer comes down to one question: does this system evaluate a natural person's creditworthiness, or price their life or health insurance? If yes, you are almost certainly high-risk; if no, you probably sit in a lighter tier, but you still need to check.
This article is the map. If you want the deep dive on the credit-scoring deadline specifically, read our piece on the credit-scoring deadline. If you want the mechanics of human oversight once a system is confirmed high-risk, read our Article 14 breakdown. Here, we zoom out.
The four tiers, in plain terms
Regulation (EU) 2024/1689 sorts every AI system into one of four tiers, and the tier decides the paperwork:
- Prohibited (Article 5). A short list of practices banned outright: manipulative or deceptive techniques causing significant harm, exploiting vulnerabilities, social scoring, certain biometric categorisation and untargeted facial scraping, among others. These apply since 2 February 2025.
- High-risk (Article 6, read with Annex I and Annex III). Two routes in. Annex I covers AI embedded as a safety component in products already regulated under EU harmonisation law (machinery, medical devices, and similar). Annex III lists eight specific use-case areas, independent of any product. For fintech, Annex III is the door that matters.
- Limited risk / transparency (Article 50). No risk tier label changes, but systems that interact with people, generate synthetic content, or produce deepfakes must disclose that a person is dealing with AI, or that content is AI-generated.
- Minimal risk. Everything else. No specific obligation beyond the general AI literacy duty in Article 4.
Article 6(2) is what actually pulls a system into the high-risk tier if it falls under an Annex III heading. Article 6(3) then provides a narrow filter: even inside an Annex III area, a system is not treated as high-risk if it performs a narrow procedural task, improves a previously completed human activity, detects patterns without replacing human judgment, or does preparatory work only, provided it does not profile natural persons. Profiling of individuals closes that filter automatically. Document the Article 6(3) assessment if you rely on it; do not just assume it.
Why credit scoring is high-risk by default
Annex III, point 5(b) names it directly: AI systems intended to evaluate the creditworthiness of natural persons, or to establish their credit score, are high-risk. That covers consumer lending, buy-now-pay-later, retail margin lending, overdraft decisioning, anything where the output decides a person's access to credit. It is natural persons only; SME and corporate credit scoring sit outside this specific entry.
The fraud-detection carve-out, stated precisely
Point 5(b) contains its own exception, written into the Annex itself: the high-risk designation does not apply to "AI systems used for the purpose of detecting financial fraud." Recital 42 backs this up, distinguishing risk analytics that assess the likelihood of fraud on suspicious transactions from the kind of individual profiling the Act is worried about.
The qualifier matters. The exemption protects the fraud-detection function, not every system a lender happens to run. If the same model's output also feeds a creditworthiness decision, identifies a person biometrically, or is used by a public authority to decide access to benefits, that other use is judged on its own terms, and it can be high-risk even while the fraud-detection part of the same stack is not. Keep the two functions documented separately: the question "is this transaction genuine" and the question "is this person creditworthy" get different answers under the Act, even when the same model touches both.
Other Annex III entries fintechs run into
- Life and health insurance pricing (Annex III, point 5(c)). AI used for risk assessment and pricing of natural persons in life or health insurance is high-risk. Motor, property and commercial lines are not named here, so they fall outside this specific entry, though GDPR and sectoral insurance law still apply.
- HR and workforce tools (Annex III, point 4). CV screening, candidate ranking and some employee-monitoring systems are high-risk, which catches fintechs in their capacity as employers, not as financial firms.
- Biometric identification (Annex III, point 1). One-to-one identity verification at onboarding (confirming someone is who they say they are) is generally excluded; one-to-many remote identification is a different story.
What is notably absent: algorithmic trading, robo-advice, AML and sanctions screening, and KYC document matching do not appear in Annex III at all. They carry real obligations under MiFID II, AML law and GDPR, just not the AI Act's high-risk regime.
What this means for your roadmap
Classify system by system, not product line by product line. A single platform can hold a minimal-risk fraud engine, a high-risk credit-scoring model and a transparency-only chatbot, each with a different compliance timeline and a different set of Article 8 to 17 obligations once Annex III status is confirmed. Getting the classification wrong in either direction costs you: overclassifying burns budget on a conformity programme you do not owe, underclassifying leaves you exposed when a regulator reads the Annex the same way we just did.
If you want help mapping your actual model inventory against Annex III and building the governance trail to back it up, that is the kind of groundwork we do at Novafin's AI automation practice.
Questions about where your models land, reach us at team@novafin.co.