Classify a system: the guided flow, screen by screen
Classification decides your risk tier and role, which decide what you must do. One question at a time, with follow-ups only where an answer calls for them — Art. 5, GPAI, Annex I, Annex III, Art. 6(3), your role, Art. 25, Art. 50 and GDPR — then a review.
Classification is the hinge of the whole EU AI Act. It decides your risk tier and your role, which together decide how much you have to do. Registration and classification are one guided flow — a short description of the system, then one question at a time, then a review — and nothing is created until you confirm on the last screen. Aria can draft the reasoning at the end; the answers, and the classification, are yours, and the human decision is what your audit trail records.
Veritome never hard-codes obligation lists. Every duty is derived by the obligation engine from the system's role, risk tier, Annex III area, its behavioural flags and its GDPR answers. Change any of them by re-classifying and the obligations follow.
The four risk tiers
| Tier | What it means | Roughly what is required |
|---|---|---|
| Prohibited | Banned under Art. 5 (social scoring, manipulative techniques, untargeted facial-image scraping, emotion recognition in the workplace or education, and the rest of the list). | It cannot be placed on the market or used. |
| High-risk | A safety component of an Annex I product (Art. 6(1)) or a use listed in Annex III (Art. 6(2)). | The full set — risk management, data governance, documentation, human oversight, conformity assessment, registration. |
| Limited (transparency) | Interacts with people, generates content, or recognises emotions. | Art. 50 transparency: tell people they are dealing with AI; label synthetic content. |
| Minimal | Everything else. | No mandatory obligations beyond Art. 4 literacy; voluntary codes encouraged. |
First — the system
Name the system and say what it does, who uses it, and what it decides or supports. This is the text the engine and Aria read back on the review screen, so make it descriptive. Register a system even if you are not sure it is in scope: running it through classification is the cleanest way to find out.
Then — the questions
The questions are the same ones the public check, Aria and a system's facts use, asked one at a time with the provision each rests on. A progress list beside them shows where you are; a follow-up appears only when an earlier answer makes it matter, so most systems see fewer than the full set.
- Prohibited practices (Art. 5) — does the system do anything on the prohibited list? The list is shown in full, sourced from the in-app regulatory registry, including the two practices added by Regulation (EU) 2026/1744, which apply from 2 December 2026. A yes ends the questions: the system is classified Prohibited and recorded as decommissioned, because it cannot lawfully be placed on the market or used. Not sure is treated as not prohibited and flagged for review. A no is written onto the Art. 5 obligation, so the dossier shows the screening was done.
- General-purpose model — is a GPAI model involved? Building on someone else's model (Mistral, OpenAI, Anthropic) creates no Chapter V duty; that belongs to the model provider. Providing the model yourself brings Art. 53; an open-source release with public weights is exempt from Art. 53(1)(a) and (b); a model trained with more than 10²⁵ floating-point operations is presumed to carry systemic risk and brings Art. 55.
- Product law (Art. 6(1), Annex I) — is it a product, or a safety component of a product, under the Annex I laws (medical devices, machinery, toys, vehicles and the rest)? It is high-risk on this ground only where that law requires third-party conformity assessment; a follow-up then asks which Annex I section applies.
- Annex III area — the single area the system works in, or none (biometrics; critical infrastructure; education; employment; access to essential services including creditworthiness; law enforcement; migration; justice and democratic processes). For biometrics and essential services a follow-up asks which use is closest — credit scoring and life or health insurance pricing, for example, carry duties of their own.
- Art. 6(3) exception — asked only if you named an area: does the system only do one of the narrow things Art. 6(3) carves out? A yes takes it out of the high-risk tier — unless it profiles natural persons, which the final subparagraph of Art. 6(3) says is always high-risk. The flow applies that override for you.
- Your role — per system, not per company: Deployer (you use it as supplied), Provider (you built it or put your name on it, including your own feature with someone else's model behind it), Importer, Distributor or Authorised representative. Each role has its own obligation set.
- Art. 25 changes — for a deployer, distributor or importer: did you put your own name or trade mark on it, substantially modify it, or change its intended purpose so that it is now high-risk (Art. 25(1)(a)–(c))? Any one makes you the provider for the rest of the flow — you inherit the provider obligation set.
- Follow-ups for your role — a provider (or a model provider) is asked where the organisation is established, since one outside the EU appoints an authorised representative (Art. 22, Art. 54). A deployer of a high-risk system is asked whether it is a public authority or a private organisation providing a public service, which decides the fundamental-rights impact assessment (Art. 27) and registration (Art. 26(8)). Where the product already knows the answer from your organisation profile, it is filled in for you to confirm.
- Personal data — does personal data flow through the running system, and are you controller, processor or joint controller for it?
- What it does — tick everything that applies: people talk to it (Art. 50(1)); it generates text, images, audio or video (Art. 50(2)); it produces realistic depictions of real people, places or events, or you publish its text to inform the public (Art. 50(4)); it recognises emotions or sorts people by biometric data (Art. 50(3)); it is used on workers (Art. 26(7)). Where personal data flows, the same list asks about solely automated decisions (GDPR Art. 22), special categories (GDPR Art. 9) and transfers outside the EU/EEA (GDPR Art. 44–49). The Art. 50 ticks make an otherwise-minimal system Limited; the GDPR answers decide the GDPR rows the engine adds — see GDPR for AI systems.
- Market status — for a high-risk system, or a provider's content generator: is it planned, in real-world testing, or already on the market? A system already live is asked the Art. 111 transition questions that decide when its duties start.
Review
The last screen shows the engine's reading in three tiles — Role (and whether Art. 25 upgraded it), GPAI, and Risk tier (with the Annex III area or the Art. 6(3) claim) — then every answer with a Change link, and the obligations it will generate. Aria drafts a short reading of your answers (Why this tier), which is kept with the classification as drafted by Aria and confirmed by you. Tick the declaration (the classification is produced by a deterministic rules engine from your answers, is not legal advice, and the answers describe the system accurately) and press Create system. The system lands on the register, its obligations are generated, its classification dossier opens with a sealed first entry, and a classification evidence record is minted.
The formal Art. 6(3) self-assessment
The Art. 6(3) answer is what the engine reads. The documented assessment that Art. 6(4) expects — the criteria, the reasoning, who assessed it — is produced by the dedicated wizard at the system's Art. 6 obligation (open it from the Obligations tab; the record links to the wizard). It is listed on Assessments as the Art. 6(3) self-assessment. See The Article 6(3) exception for the test itself.
Reclassification
Open System actions → Review classification in the system header. The same flow opens with the current answers filled in; change what is different today and press Save reclassification. Until you do, the current classification stays in force. On confirmation the engine replaces the obligation set, the dossier records the change (for example Reclassified High-risk → Limited-risk), and the classification evidence record is refreshed. The permission to reclassify a system is separate from the permission to edit one.