How to Classify Your AI System Under Article 6: A Step-by-Step Self-Assessment
Classification is the load-bearing decision of the entire Act — every obligation, deadline and document follows from it. The five-step Article 6 procedure, the derogation test, and the two-page record that proves you did it.
Every obligation in the AI Act hangs off one decision: what tier is this system? Get it wrong and everything downstream is wrong — the documentation you skipped, the oversight you never assigned, the registration you thought didn’t apply. Classification is also the first artefact every enterprise buyer and every market-surveillance authority asks to see.
The good news: the procedure is a flowchart, not a consultation. Article 6 gives you the sequence, and a disciplined self-assessment takes under an hour per system.
The five-step procedure
Step 0 — The prohibition screen (Article 5)
Run first, always. Eight banned practices, no compliance route. Workplace emotion inference, social scoring, scraped face databases: if any fires, the flow ends — the system leaves the estate. One dated line per system records the screen: no Article 5 practice identified.
Step 1 — Is it a safety component under Article 6(1)?
The system is an AI component of a product covered by Annex I — machinery, medical devices, lifts, radio equipment, toys, and the rest of the product-safety acquis — and that product requires third-party conformity assessment under the relevant Annex I legislation. If yes: high-risk, full stop, no Annex III analysis, no derogation. Obligations bite from 2 August 2028.
Most SaaS answers here is no. Move on.
Step 2 — Map against Annex III
Walk the eight areas against what the system actually does in your context — not its marketing category. The same model classifies differently per deployment: an LLM summarising support tickets is minimal; an LLM ranking CVs is Annex III 4; an LLM scoring callers’ emotion from voice is Article 5 prohibited.
For each area, cite the specific sub-point you considered. “Assessed against Annex III” is not a record; “Annex III 4(b) — screening job applications — triggered, therefore high-risk” is.
Step 3 — The Article 6(3) derogation test
If Step 2 matched but the tier matters, test the four prongs. The derogation holds where the system does only:
- (a) a narrow procedural task;
- (b) an improvement to a previous human activity’s result;
- (c) detection of decision patterns or deviations without replacing or influencing human assessment;
- (d) a preparatory task — or pattern recognition on input data.
Two walls override any passing prong: profiling persons with NLP to infer traits kills the derogation outright; and member states may withhold it where the system influences outcomes in education, employment, essential services or their appeal procedures. If either wall stands, the answer at Step 2 holds.
Step 4 — Record the decision
Two pages, five fields, a signature and a date:
| Field | Content |
|---|---|
| System + intended purpose | What it does, for whom, in what context |
| Areas considered | Every Annex III area reviewed with sub-point citations |
| Derogation reasoning | Which prongs argued, walls checked, outcome |
| Tier + obligations triggered | Prohibited / high-risk / limited / minimal, with the article list |
| Review trigger | Re-classify when purpose, data, or model version changes |
That document is what “we assessed it” means in an audit. Keep it beside the system, not beside the strategy deck.
Limited risk is a tier, not a shrug
The fourth outcome — limited — carries the Article 50 transparency package: disclose AI interaction, mark synthetic content, retain provenance evidence where required. Live since 2 August 2026. Two postures fail reviews constantly: treating “limited” as “basically compliant” (it is a duty list), and leaving the tier blank because nobody ran the flow. A blank tier fails a questionnaire as hard as a wrong one.
Who runs this, and how often
Trigger events, not calendars: new system onboarded; model version changes behaviour; intended purpose widens; a vendor changes classification on you. Each trigger reopens Step 2 with the original record as the baseline — deltas take ten minutes.
Ownership: the person who understands the system’s behaviour signs; legal or compliance reviews. Empty a sign-off field to a name, because “the committee” reads as nobody.
Vendor systems: you cannot classify what you cannot see. Procurement captures the provider’s classification claim, their Annex III reasoning, and a change-notification commitment — three fields that keep their answer auditable as theirs.
How to check your own position in four minutes
The free exposure scan asks twelve questions about your AI usage and governance and returns an exposure score, your ranked exposures with the exact article that fired, and your first three moves — mapped to the dates that apply. Confirm your email and the full report unlocks.
Frequently asked questions
What is the AI Act classification process under Article 6?
Five moves: screen against the Article 5 prohibitions; check whether the system is a safety component of an Annex I product (Article 6(1) — if yes, high-risk); map what it actually does against the Annex III areas; test any high-risk match against the Article 6(3) derogation prongs; then record the decision with the areas considered, the reasoning, the tier, the obligations triggered and a review trigger. The record is as important as the outcome.
What is the Article 6(3) derogation in plain English?
If a system matches an Annex III area but only performs a narrow procedural task, improves a prior human result, detects patterns without influencing a human assessment, or acts as a preparatory step — it is not high-risk. Two walls block it: profiling people to infer traits, and any use influencing decisions in education, employment, essential services or their appeals.
Can we classify our own AI, or do we need an auditor?
Self-classification is the default for most systems — Annex III uses follow self-assessment with documentation, while certain product-integrated systems (Article 6(1)) need third-party conformity assessment through a notified body. What you never self-serve is the record: the dated reasoning must exist regardless of who concluded it.
How often do we need to reclassify an AI system?
On trigger, not on schedule: purpose change, model or version change, new data sources, vendor classification change, or a rule-pack update from the Commission. Keep the baseline document and record deltas — a reclassification triggered by a model upgrade should take minutes, not a project.
Where does AI Act classification appear in enterprise questionnaires?
As three fields: the tier, the Annex III areas considered, and the date/reviewer of the classification record. Vendors who answer “minimal — no Annex III applies” with the sub-points they checked pass; vendors who answer with a marketing sentence fail and trigger a follow-up.
Put this guide to work on your own estate.
Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.
- Free forever plan
- No credit card
- EU-hosted · GDPR-native
Build a live AI register in minutes — upgrade only when you need the audit pack.
Keep reading
AI in healthcare under the EU AI Act: clinical decision support and high-risk duties
11 min
Risk & ClassificationThe Annex III High-Risk AI Systems List: All Eight Use Cases Explained with Examples
11 min
Risk & ClassificationThe Eight Prohibited AI Practices Under Article 5: A Compliance Checklist
10 min