Is my AI system high-risk under the EU AI Act?

Everything expensive in the EU AI Act — Annex IV documentation, Article 9 risk management, conformity assessment, registration — hangs on one question: is your system high-risk? This walkthrough is the engineering team's version of that analysis. It is not legal advice; it is the homework that makes the conversation with counsel ten times cheaper.

The two routes into high-risk

Route 1 — Annex I products (Art. 6(1)). Your AI is a safety component of a product already covered by EU product legislation (machinery, medical devices, vehicles, toys…). If that is you, you likely have a notified body relationship already; this guide is secondary.

Route 2 — Annex III use cases (Art. 6(2)). The list that catches software companies. Eight categories, condensed:

Categories 4 and 5 are where most SaaS teams discover themselves. An HR tech product that ranks applicants is squarely in category 4. A fintech that scores creditworthiness is category 5.

The Article 6(3) escape hatch — use with care

A system in an Annex III category is *not* high-risk if it does not pose a significant risk of harm — the Act gives four qualifying conditions, such as performing a narrow procedural task or merely preparing an input for a human assessment. Two warnings:

• Profiling kills the exemption. If the system profiles natural persons, it is high-risk regardless of the four conditions.
• You must document the exemption analysis and register the system as exempt. Claiming the escape hatch silently is itself a compliance gap.

Honest test: if your sales deck says "AI-powered decisioning" and your exemption memo says "merely preparatory", an authority will notice the tension.

Common edge cases

• "A human reviews every output." Human review does not remove the classification — human oversight is an *obligation* of high-risk systems (Art. 14), not an exemption from being one.
• "We only provide the model; our customer deploys it." Providers and deployers both have obligations. If you place the system on the EU market, the Annex IV documentation duty is yours.
• "It's just an LLM wrapper." The classification follows the *use case*, not the architecture. A GPT-4 wrapper that screens résumés is a category-4 high-risk system.
• "We're not in the EU." The Act applies if the system's output is used in the EU. Selling to EU customers puts you in scope.

What to do with the answer

If you are high-risk: the obligation stack is Annex IV technical documentation (before market placement), Article 9 continuous risk management, data governance, logging, human oversight, accuracy/robustness measures, registration in the EU database, and conformity assessment. Start with the documentation — it is the artefact every other obligation references, and its mechanical half (component inventory, risk register, infrastructure description) can be generated from your repository on every commit.

If you are not high-risk: document *why* (the analysis has to exist somewhere when a customer or authority asks), watch the transparency obligations in Article 50 (chatbots must disclose they are AI; synthetic media must be labelled), and expect enterprise procurement to ask for the same evidence anyway.

Either way: the classification is per-system and per-use-case, and it drifts. A feature launch can move you into Annex III without anyone noticing — which is an argument for making the AI inventory continuous rather than a one-off audit. That inventory is exactly what an AI-BOM scanner produces from your CI pipeline, free, on every commit.