Umair Zaffar, Gründer und CEO der sifamo GmbH
Umair Zaffar October 28, 2026 0 Comments

EU AI Act Article 6: High-Risk Classification, Implemented in Practice

EU AI Act Article 6 decides which AI systems count as high-risk, and at SIFAMO we implement that classification as a decision record per use case in the registry, with three signatures, a written reasoning and triggers for re-classification.

Article 6 is the article everything else hangs on. Whether a system needs risk management, logging, human oversight and a conformity assessment is decided here. It addresses the provider, who classifies the system and documents that assessment under Art. 6(4). But it affects every deployer, because deployer obligations under Article 26 apply only to high-risk systems, and a deployer becomes a provider under Article 25 faster than most expect. I am an engineer, not a lawyer. My concern is how the classification becomes a workflow rather than a gut feeling. Deadlines and roles are explained in EU AI Act in practice.

For context: the EU AI Act has been in force since 1 August 2024. Prohibitions and AI literacy have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. The obligations for high-risk systems under Annex III apply, after the postponement by amending Regulation (EU) 2026/1744 (Digital Omnibus), from 2 December 2027, and for Annex I products from 2 August 2028. Article 6 applies from the same dates. The Omnibus moved deadlines; it removed no obligation.

What Article 6 requires

Article 6 opens Chapter III. It does not regulate what a high-risk system may do, but which systems fall into that class. Two routes lead in, one leads out:

  • Art. 6(1) (Annex I): A system is high-risk if it is a safety component of a product covered by the harmonisation legislation in Annex I, or is itself such a product, and that product additionally requires a third-party conformity assessment. The consolidated text clarifies in paragraphs 1a to 1c what counts as a safety component.
  • Art. 6(2) (Annex III): In addition, the systems listed in Annex III are considered high-risk. For banks and insurers, two of the eight areas matter. Point 4 covers recruitment and selection, including targeted job advertisements, filtering applications and evaluating candidates, as well as promotion, termination, task allocation and performance monitoring. Point 5(b) covers evaluating the creditworthiness of natural persons or establishing their credit score, excluding the detection of financial fraud. Point 5(c) covers risk assessment and pricing in life and health insurance, not other lines of insurance.
  • Art. 6(3) (the exception): An Annex III system is not high-risk if it does not pose a significant risk of harm to health, safety or fundamental rights, including by not materially influencing the outcome of decision-making, and meets one of four conditions: a narrow procedural task (a), improving the result of a completed human activity (b), detecting decision-making patterns or deviations without replacing the human assessment (c), a preparatory task for an assessment under Annex III (d). Hard limit: a system that profiles natural persons “shall always be considered to be high-risk”.
  • Art. 6(4) (documentation duty): A provider that uses the exception documents the assessment before placing the system on the market or putting it into service, registers itself and the system in the EU database under Art. 49(2), and provides the documentation to national authorities on request.
  • Art. 6(5) to (8): The Commission had to provide guidelines with examples by 2 February 2026 and may change the conditions in paragraph 3 by delegated act, without lowering the level of protection.
  • Art. 80 (the consequence): If a market surveillance authority considers a non-high-risk classification to be wrong, it orders corrective action within a set period, for all similar systems of the provider. A provider that fails to comply, or that misclassified deliberately to circumvent Chapter III Section 2, faces fines under Article 99.

The technical example

A bank has six AI use cases in its portfolio and classifies them before the first production call. One record per use case, not per model: the same language model can carry two risk classes in two use cases.

  • Credit pre-check. The model reads the pseudonymised application and proposes a route: fast lane, standard review, specialist review. Annex III point 5(b) is touched. The exception in Art. 6(3)(d) is tempting but does not hold: the route shapes the outcome, and the model evaluates economic attributes of the applicant, which is profiling. Classification: high-risk.
  • Fraud detection. The model flags transactions and triggers queries or blocks. Point 5(b) expressly excludes financial fraud. Classification: not an Annex III system, but internal tier 2, because a blocked card is a decision about a person. If the score feeds into creditworthiness rules, the exclusion is gone.
  • Document classification. Documents are assigned a type and routed to the right file. The system works inside the credit process but assesses nothing about the applicant. Recital 53 names exactly this case as a narrow procedural task. Classification: exception under Art. 6(3)(a), documented under paragraph 4, registered under Art. 49(2).
  • Customer-service RAG. The assistant answers questions about products and forms with sources. No Annex III area; Article 50 applies. The boundary is the scope: if the assistant tells a customer whether they will get a loan, it is in point 5(b). The gateway blocks that statement, not just the prompt.
  • HR screening. The system filters applications and produces a ranking. Point 4(a), and a ranking of people is profiling, so no exception. We rejected running the CV extraction separately under paragraph 3(a), because it feeds straight into the ranking. Classification: high-risk as one system.
  • Marketing segmentation. Customers are grouped for campaigns. No Annex III area, but profiling in the GDPR sense. Two triggers in the record: if segments determine who receives a credit offer on which terms, point 5(b) is touched; targeted job advertisements would be point 4(a).

Result: two high-risk systems, one documented exception, three systems outside Annex III with other obligations. The data flow is always the same: the business unit fills in the intake, legal and compliance assess, the risk function assigns the internal tier, the record is signed and written to the registry. Only then is there a use-case ID, and only with the ID does the gateway release a route. The obligation bites at the first call.

How we implement it

At SIFAMO we do not treat the classification as an opinion attached to the project charter, but as a field in the registry record that the governance layer reads at runtime. “No registration, no route” gets a preliminary stage: no classification, no registration. The record for the document classifier, simplified:

{
  "use_case_id": "UC-2026-031",
  "name": "Document classification, credit intake",
  "purpose": "Assigns documents to a type and routes them to the file.",
  "annex_iii_area": "point 5(b) (credit process touched)",
  "profiling": "no",
  "exception": {"letter": "Art. 6(3)(a), narrow procedural task",
    "reasoning": "Only the document type is determined. Errors lead to a query to the customer, not to a different credit decision."},
  "decision": "not high-risk, Art. 6(4) documented, registration Art. 49(2)",
  "signed": [
    {"role": "Business owner, retail credit", "date": "2026-09-08"},
    {"role": "Legal and compliance", "date": "2026-09-10"},
    {"role": "Risk function", "date": "2026-09-10"}],
  "scope_enforced": {"data_zones": ["gray"], "actions": ["classify", "route_to_file"]},
  "reclassify_on": ["content extraction", "output feeds creditworthiness rules", "expiry 2026-12-31"]
}

One record per use case

Article 6 turns on the intended purpose, so the record answers four questions per use case in a fixed order: Annex I? Which Annex III area? Profiling? Which condition in paragraph 3, with what reasoning? Whoever uses an exception must be able to say in two sentences why the system does not materially influence the outcome. If that fails, it is high-risk.

Three signatures, business owner first

The text does not say who decides internally. We settled it: the business owner signs the purpose and the actual use, because only the business unit knows what it really does with the output. Legal and compliance sign the legal reading: Annex III area, profiling, exception letter. The risk function signs the internal tier. A vendor statement that the system is “just an assistant” replaces none of these signatures.

Re-classification when the scope changes

The most common source of error is the correct initial classification that nobody touches while the system grows. So every record carries fixed triggers: new data categories, a new affected group, output used for a different decision, a new tool for an agent, a model change, an amendment of Annex III under Article 7, new guidelines, the expiry date. Two of these the gateway detects on its own: it knows the approved scope from the registry and logs every attempt to use data outside the zone or an action that was not approved. When such attempts pile up, a ticket goes to the business owner. Scope drift is a measurement, not a feeling.

What the auditor sees

Under Article 80 three things count: that the record existed before the system was put into service, with timestamp and signatures; that the registration under Art. 49(2) was done; and that actual use matched the declared purpose. The third comes from the evidence log, because every decision is stored with use-case ID and policy version.

Typical mistakes

  • Classification by vendor label. The vendor calls it a chatbot or a copilot, so it runs as minimal risk. Article 6 asks about purpose and influence, not about the product name.
  • “It only recommends.” A recommendation that is followed nine times out of ten materially influences the outcome.
  • Exception without documentation. The team agrees that paragraph 3 applies, writes it down nowhere and registers nothing. Exactly what paragraph 4 requires and Article 80 asks for is missing.
  • Classified once, never again. After the third release the document classifier also extracts income and flags implausibilities. The narrow procedural task has become a preparatory assessment, and the record still says tier 1.

EU AI Act Article 6 checklist

  • Every production AI use case has a decision record before the first call.
  • The record checks in this order: Annex I, Annex III, profiling, condition under Art. 6(3).
  • Every exception used has a reasoning under Art. 6(4) and a registration under Art. 49(2).
  • Business owner, legal and compliance, and the risk function have signed with name and date.
  • The risk class is in the registry, and the gateway reads it together with the approved scope.
  • Re-classification triggers are named, including the expiry date, and scope violations reach the business owner.
  • Systems outside Annex III are still recorded, with an internal tier.

Open questions

  • What does “materially influence” mean? The line between a preparatory task and an assessment is not sharp in paragraph 3. How far the Commission guidelines carry for routing models is open. We choose the stricter reading.
  • Profiling as a stop sign. The AI Act takes over the GDPR notion of profiling. Almost every model that processes attributes of a person evaluates personal aspects. Whether that empties the exception for customer-facing systems is not settled.
  • One model, many purposes. A general-purpose model behind six use cases carries six classifications. How model providers and system providers separate their documentation is open.

Frequently asked questions

Does this apply to us as a deployer too?

The documentation duty in Art. 6(4) falls on the provider. As a deployer you still have to know how a purchased system is classified, because your obligations under Article 26 depend on it, and you have to use it within the declared purpose. If you materially change the purpose or the system, you become a provider yourself under Article 25.

Is the vendor’s classification enough?

As a starting point yes, as evidence no. The vendor classifies its intended purpose; you use the system in your process. We file its documentation in the record and check whether actual use matches it. Where it deviates, we assess ourselves.

What does a wrong classification cost?

Article 80 sets out the path: the authority evaluates, orders corrective action within a period, and fines under Article 99 follow for non-compliance or deliberate circumvention. With a dated record and a registration, a misclassification is a correctable error; without both, it looks like circumvention.

This post is part of the series on the EU AI Act article by article, all articles at a glance. The overall picture of registry, gateway and evidence log is described under AI governance.

Related: EU AI Act · All articles · AI Governance · AI Consultant

Leave Comment