EU AI Act Risk Classification Checklist

Why Classification Matters

Your obligations under the EU AI Act are entirely tier-dependent. A minimal-risk spam filter has no specific obligations. A high-risk credit-scoring model carries Articles 8 through 17, Annex IV technical documentation, CE marking, conformity assessment, and post-market monitoring. A prohibited system cannot be deployed at all and exposes your organisation to fines of up to €35 million or 7% of global annual turnover — whichever is higher.

The two failure modes are equally costly. Under-classify and you ship a high-risk system without the conformity assessment, the Article 12 logging, or the Article 14 human oversight that the Regulation requires — and inherit the liability. Over-classify and you burn six to twelve months of engineering and compliance budget on obligations that do not apply to you, while your competitors reach the market first. The tier decision is therefore not a legal afterthought; it is an engineering decision with six- and seven-figure consequences.

This checklist is a first-pass filter. It is structured around the Regulation's own tier logic — Article 5 (prohibited), Annex III (high-risk), Article 50 (limited-risk transparency), and Articles 51 through 55 (GPAI). Annex III of Regulation (EU) 2024/1689 is the authoritative list. This tool helps you form a working classification. Get written sign-off from qualified counsel before you rely on it for a deployment decision.

Section 1 — Prohibited Practices Check

Article 5 lists categorically banned AI practices. If your system falls into any of the eight categories below, you cannot place it on the Union market, put it into service, or use it — regardless of technical safeguards, consent, or business justification. Work through each item honestly. Edge cases belong to your counsel, not to this checklist.

Section 2 — High-Risk Check

Annex III lists the eight category groups that constitute high-risk AI systems when the system is used in the listed context. Work through each group. Any single yes in any group triggers Articles 8 through 17 and the full conformity-assessment regime.

Pay attention to the framing: what matters is the use, not the technology. A generic language model deployed to score job applicants is high-risk employment AI. The same model summarising meeting notes is not. Classification follows deployment context.

Section 3 — Limited-Risk Transparency Check

Even if your system is not high-risk, Article 50 may still impose transparency obligations. The purpose is to ensure users know when they are interacting with AI or consuming AI-generated content. These obligations apply in addition to any other classification — a high-risk system may also carry Article 50 duties.

Section 4 — GPAI Obligations Check

General-Purpose AI (GPAI) obligations run on a parallel track to the tiered classification above. A GPAI provider may simultaneously place a high-risk system on the market and carry GPAI obligations at the model layer. These questions determine whether you are a GPAI provider, and whether your model is considered GPAI with systemic risk.

Result Interpretation Table

Summary of what each classification actually means in operational terms:

Red Flags Most Companies Miss

Most mis-classifications we see in advisory work share a common pattern: the company reads the Annex III list, decides "that's not us," and moves on. The following traps catch organisations that thought they were safe.

• Deploying a third-party LLM in a high-risk use case makes you the deployer with Article 26 obligations. If you wrap OpenAI, Anthropic, or Google's API to make hiring decisions, you are the high-risk deployer. The model provider's compliance does not substitute for yours.
• Credit scoring is ALWAYS high-risk. The Annex III(5)(b) entry does not have an exception for "also does other things." If any part of your system evaluates creditworthiness of natural persons or establishes a credit score, the system is high-risk. Fraud-detection carve-out is narrow and must be the sole purpose.
• Fine-tuning a GPAI for a specific high-risk purpose makes you a high-risk provider, not just a deployer. When you substantially modify a general-purpose model for a high-risk intended purpose, you assume provider obligations under Article 25(1)(b). The upstream GPAI provider does not shield you.
• "EU market placement" includes offering via API to EU customers. Physical presence in the Union is not required. If EU-based organisations can pay for and use your AI system, you are placing it on the Union market. The Act has explicit extraterritorial reach (Article 2(1)(c)).
• Prohibited practices apply regardless of tier on the rest of your stack. A minimal-risk product can still contain a prohibited component. Article 5 is not a tier — it is a categorical ban on eight specific practices that cuts across the rest of the Regulation.
• "High-risk" is determined by intended purpose, not by what the model "could" do. You document your intended purpose; regulators will hold you to it. Selling a generic model while turning a blind eye to customers using it in high-risk contexts will not keep you out of the high-risk regime.

What to Do Next

If Section 1 returned yes, stop. Engage counsel today. If Section 2 returned yes, your critical path runs to 2 August 2027 and conformity assessment alone typically consumes six to twelve months. If Section 3 returned yes, your deadline is 2 August 2026 and transparency controls are largely engineering work. If Section 4 returned yes on item #1, your deadline is also 2 August 2026 and you should already be drafting your Annex XI technical documentation.