Chapter 3 of 8

The Article 6(3) Exemption: When Your Annex III Use Case Is Not High-Risk

15 min read

The Article 6(3) exemption takes Annex III systems out of high-risk status — never profiling. Four conditions, one documentation duty, fines to EUR 15M or 3%.

Article 6(3) of Regulation (EU) 2024/1689 is the only exit from high-risk classification for an AI system that matches a use case listed in Annex III. By way of derogation from Article 6(2), such a system is not considered high-risk where it does not pose a significant risk of harm to the health, safety or fundamental rights of natural persons, including by not materially influencing the outcome of decision making, and where it meets at least one of four narrow conditions. The exemption is never available to a system that performs profiling of natural persons. Claiming it also leaves two duties standing: documentation before market placement under Article 6(4) and registration in the EU database under Article 49(2).

The rest of this chapter sets out the four conditions, the profiling override that removes most HR and credit use cases from contention, the documentation you must build before launch, and what an over-claimed exemption costs once a market surveillance authority reads it differently.

Annex III Is a Presumption, Not a Verdict

How Article 6(2) and Article 6(3) Interact

Chapter 2 of this guide established the default rule: under Article 6(2), an AI system intended for any of the use cases listed in Annex III is classified as high-risk. If you have not yet confirmed whether your system matches a listed use case, start with Annex III explained and the Article 6 classification rules. Article 6(3) only becomes relevant once an Annex III match exists.

Article 6(3) is the derogation from that presumption, and its defining structural feature is who decides. No authority grants or approves the exemption. It is a provider self-assessment: you analyse, you conclude, you document, and you carry the entire burden of being right. That structure shapes everything else in this chapter. The conditions are drafted narrowly, the documentation duty is unconditional, and the enforcement procedure in Article 80 exists because the legislator expected some providers to over-claim.

The Significant-Risk Test in the First Subparagraph

The first subparagraph sets the threshold: an Annex III system is not high-risk where it "does not pose a significant risk of harm to the health, safety or fundamental rights of natural persons, including by not materially influencing the outcome of decision making."

The filter is deliberately narrow. It captures ancillary systems that sit alongside an Annex III decision, not the decision engine itself. If your system produces the score, the ranking, the shortlist or the assessment that the Annex III use case turns on, the derogation was not written for you.

The Four Derogation Conditions, Precisely

Meeting any one of the four conditions suffices; the text applies the derogation "where any of the following conditions is fulfilled". The flexibility ends there. You must name the specific condition you rely on and defend it in writing, because the Article 6(4) file (covered below) has to show your reasoning.

Condition (a): Narrow Procedural Task

The system is intended to perform a narrow procedural task. The draft Commission guidelines give examples such as converting unstructured data into structured data, classifying incoming documents into categories, or detecting duplicates among a large set of applications. The defining feature is that the task involves no evaluation of any person. The system handles the paperwork around the decision, not the person the decision is about.

Condition (b): Improving a Previously Completed Human Activity

The human did the substantive work first, and the system only polishes it. The standard example is a tool that improves the language or tone of an already-drafted document. The qualifier to respect is that the human activity must be genuinely completed, not nominal. A tool that "improves" a half-finished assessment is, in substance, doing the assessment.

Condition (c): Detecting Decision-Making Patterns or Deviations

The system detects decision-making patterns or deviations from prior decision-making patterns, and it is "not meant to replace or influence the previously completed human assessment, without proper human review". This is monitoring and flagging only, and the proper human review qualifier is where claims fail. If the flag effectively dictates the outcome because nobody with real authority reviews it, the condition is not met.

Condition (d): Preparatory Task to an Annex III-Relevant Assessment

The system performs a preparatory task to an assessment relevant for the purposes of the Annex III use cases: indexing case files, preparing dossiers, translating documents, running smart search across submitted material. The work stays clearly upstream of the assessment, and the assessment itself stays human.

Conditions Table: What Qualifies, What Fails

Condition	Statutory wording (condensed)	Qualifying example	Disqualifying example
(a)	Narrow procedural task	Classifying incoming documents; duplicate detection	Any scoring or evaluation of a person
(b)	Improves a previously completed human activity	Polishing the language of a drafted decision	"Improving" an assessment a human never finished
(c)	Detects decision patterns or deviations	Audit flag reviewed by a human who can change the outcome	A flag that de facto dictates the outcome
(d)	Preparatory task to a relevant assessment	Indexing, translation, file preparation	Pre-ranking that shapes the assessment itself

The examples in this table come from the draft Commission guidelines of 19 May 2026 and may change in the final version; their status is covered below.

A common misreading is worth correcting. The four conditions describe the system's intended purpose, not its technical sophistication. A simple model used to score people fails the test, while a complex model doing pure document triage can pass it.

The Profiling Override: No Exemption, Full Stop

What Counts as Profiling — Article 3(52)

The third subparagraph of Article 6(3) is absolute: an Annex III AI system "shall always be considered high-risk" where it performs profiling of natural persons. None of the four conditions can rescue a profiling system. The exemption is structurally unavailable.

Profiling takes its definition from Article 3(52) of the AI Act, which points to Article 4(4) GDPR: any form of automated processing of personal data to evaluate personal aspects of a natural person — work performance, economic situation, health, personal preferences, reliability, behaviour, location or movements.

Why the Override Removes HR and Credit Use Cases

Most systems under Annex III point 4 (employment) and point 5(b) (creditworthiness) exist to evaluate personal aspects of people. A CV scorer evaluates work performance and reliability; a credit model evaluates economic situation. For those use cases the exemption is rarely available, so treat any vendor or internal claim to the contrary with suspicion.

One edge case closes a likely gap: a system that profiles for a "narrow procedural" sub-step is still profiling. The text contains no de minimis carve-out. The override turns on whether profiling happens, not on how much of the pipeline it occupies.

If you now suspect your system profiles, check if your AI is high-risk before investing any more effort in the exemption argument.

Two Duties Survive the Exemption

Document the Assessment Before Market Placement — Article 6(4)

Claiming the exemption is not a silent opt-out. Article 6(4) requires a provider who considers an Annex III system not to be high-risk to document its assessment before that system is placed on the market or put into service, and to provide that documentation to national competent authorities upon request.

Sequence matters. The obligation is pre-market by its terms. A file written after launch does not cure the breach; it records it.

The statute does not enumerate the contents, but a defensible assessment covers the Annex III point engaged, the single condition (a)-(d) relied on and the reasoning behind it, the profiling analysis, evidence of the system's intended purpose, and a re-assessment trigger for substantial modifications or changes of intended purpose.

Register in the EU Database — Article 49(2)

Under Article 49(2), the provider, or its authorised representative, must also register itself and the system in the EU database referred to in Article 71 before placing it on the market or putting it into service, supplying the information set out in Annex VIII. Your exemption claim becomes publicly visible the moment you register. The mechanics are covered in Article 49 EU database registration.

The trade is straightforward: the Article 6(3) exemption swaps the full high-risk obligation stack for a permanent, inspectable record. It is lighter, not weightless.

What the Draft Commission Guidelines Say

Status: Draft, Consultation Open Until 23 June 2026

On 19 May 2026 the Commission published draft guidelines on high-risk classification under Article 6(5). A targeted consultation is open until 23 June 2026, still open as of this writing, and the final version is expected later in 2026.

The guidelines are not legally binding. Authoritative interpretation of Article 6(3) rests with the Court of Justice of the EU. Every reading below is attributed to the draft and may change in the final text.

The draft is structured in three parts: (i) general classification principles, (ii) Article 6(1) and Annex I product-embedded systems, and (iii) Article 6(2) and Annex III, with practical in-scope and out-of-scope examples. Part (iii) is the one that matters for the exemption.

How the Draft Reads Conditions (a) Through (d)

Four readings from the draft are worth carrying into your assessment:

The conditions are interpreted narrowly, against the backdrop of the Annex III presumption, so doubt resolves toward high-risk.
Narrow procedural task excludes anything that touches the substance of the assessment, however small the touch.
Condition (c) requires human review that is real and capable of changing the outcome. Review theatre does not count.
Condition (d) preparatory tasks must remain clearly upstream of the assessment, with no feedback into it.

Revisit this page once the final guidelines land later in 2026. The conditions-table examples above come from the draft and will be updated if the final text moves.

Decision Tree: Can You Claim Article 6(3)?

Run the seven steps in order. Any STOP means the exemption is off the table.

Annex III match? If the system does not fall under an Annex III use case, Article 6(3) is irrelevant, so run the standard analysis instead.
Profiling of natural persons? If yes, STOP. The system is always high-risk and the exemption is unavailable.
Significant risk or material influence? If the system poses a significant risk of harm to health, safety or fundamental rights, or materially influences the outcome of decision making, STOP. No condition can help.
Name your condition. Identify the single condition (a)-(d) the system satisfies and write down why. If none fits cleanly, do not claim.
Document under Article 6(4) before placing the system on the market or putting it into service.
Register under Article 49(2) in the EU database.
Set a re-assessment trigger. Any substantial modification or change of intended purpose reopens the analysis.

A useful test: if you cannot finish step 4 in one defensible sentence, you are high-risk.

Not sure where your system lands? Run the 90-second high-risk check.

The Cost of Over-Claiming

Authorities Can Reopen Your Classification — Article 80

Article 80 creates a dedicated procedure for AI systems classified by their provider as not high-risk in application of Article 6(3). Market surveillance authorities may evaluate the system, and where they find it misclassified, the provider must reclassify it and take corrective action. Your self-assessment is provisional in a way an authority-granted approval is not: it stands only until someone with enforcement powers reads your file. The exposure is not only yours. Deployers who built their own compliance posture on your classification inherit the fallout of a forced reclassification.

€15 Million or 3% — Article 99(4)

Misclassification exposes the provider to administrative fines under Article 99(4): up to €15 million or 3% of total worldwide annual turnover, whichever is higher. The full structure is in the Article 99 penalty tiers.

Misleading the Authority — Article 99(5)

A thin or after-the-fact Article 6(4) file compounds the exposure. Supplying incorrect, incomplete or misleading information to notified bodies or national competent authorities falls under Article 99(5): up to €7.5 million or 1% of total worldwide annual turnover, whichever is higher.

For SMEs and start-ups, Article 99(6) caps each fine at the lower of the percentage or the fixed amount. Proportionality exists, but it does not make over-claiming cheap.

The asymmetry argues for conservatism. A wrongly claimed exemption costs more than the compliance work it avoided, and the Article 49(2) database entry means authorities do not have to hunt for your claim. It sits in a register they administer.

Timing, the Digital Omnibus, and the Takeaway

When the High-Risk Regime Bites

As enacted, the application schedule in Article 113 puts the high-risk obligations for Annex III systems into effect from 2 August 2026. That is what the statute still says today.

The Digital Omnibus caveat, stated precisely: a provisional political agreement of 6-7 May 2026, with the COREPER-confirmed text of around 13 May 2026, agreed to defer stand-alone Annex III high-risk obligations (Article 6(2)) to 2 December 2027, and Annex I product-embedded systems (Article 6(1)) to 2 August 2028. As of June 2026 this is agreed but not yet law. It still needs the European Parliament plenary vote, formal Council adoption and publication in the Official Journal. The new dates are fixed calendar dates, and the standards-contingent "stop the clock" proposal was rejected, so do not plan around harmonised-standards availability.

Two things did not move. The Article 5 prohibitions (in force since 2 February 2025) and the GPAI rules (in force since 2 August 2025) are unaffected by anything in this chapter.

The Takeaway: Document It or Don't Claim It

Claim Article 6(3) only with the Article 6(4) file written before launch and the Article 49(2) registration done, and never for a system that profiles natural persons. For the wider decision path, the risk classification hub covers every tier, and Chapter 4 of this guide works through the draft Commission guidelines in detail.

The decision you need to make: for every system that matches an Annex III use case, decide explicitly whether you claim the exemption or accept high-risk status, and document the decision either way. An undocumented "we assumed it was fine" is the one position the Act gives you no way to defend.

How Confir helps

Confir's risk classification module runs the Article 6(3) analysis as a structured workflow: it maps your system against the Annex III use cases, applies the profiling override first, then walks the four derogation conditions one by one. The engine is deterministic and rule-based, using the same logic every time, with no model inference and no hallucination, so two assessors running the same facts reach the same classification.

Where the exemption holds, Confir generates the Article 6(4) assessment record (the Annex III point engaged, the condition relied on, the profiling analysis, the intended-purpose evidence) plus the Article 49(2) registration checklist, both produced before market placement. Substantial modifications reopen the assessment automatically, so the re-assessment trigger in step 7 is built in rather than remembered.

Frequently Asked Questions

What is the Article 6(3) exemption in the EU AI Act? Article 6(3) of Regulation (EU) 2024/1689 lets a provider treat a system listed in Annex III as not high-risk if it does not pose a significant risk of harm to health, safety or fundamental rights and does not materially influence decision outcomes. The provider must show the system meets at least one of four conditions and must document and register that assessment.

When is an Annex III AI system not considered high-risk? An Annex III system is not high-risk when it only performs a narrow procedural task, improves the result of a previously completed human activity, detects decision-making patterns or deviations without replacing human assessment absent proper human review, or performs a preparatory task to a relevant assessment. Profiling of natural persons cancels the exemption in every case.

Can profiling AI systems use the Article 6(3) exemption? No. The third subparagraph of Article 6(3) states that an Annex III system is always considered high-risk where it performs profiling of natural persons. None of the four derogation conditions can override this rule, so any system evaluating personal aspects of individuals must follow the full high-risk regime regardless of how narrow its task is.

Do I still have to register an exempt AI system in the EU database? Yes. Article 49(2) of the EU AI Act requires providers who conclude under Article 6(3) that their Annex III system is not high-risk to register themselves and the system in the EU database before placing it on the market or putting it into service. Skipping registration exposes the provider to penalties under Article 99(4).

What documentation is required for the Article 6(3) exemption? Article 6(4) requires the provider to document its exemption assessment before the system is placed on the market or put into service. The documentation must show which of the four conditions applies and why, and must be provided to national competent authorities upon request. Without this record, the exemption claim has no legal footing.

What happens if a provider wrongly classifies an AI system as not high-risk? Under Article 80, market surveillance authorities can evaluate systems a provider classified as not high-risk and order reclassification. Misclassification exposes the provider to fines under Article 99(4) of up to EUR 15 million or 3 percent of worldwide annual turnover, and supplying misleading information to authorities can add Article 99(5) fines of EUR 7.5 million or 1 percent.

Are the Commission guidelines on high-risk classification legally binding? No. The draft Commission guidelines on high-risk classification, published 19 May 2026 with a consultation open until 23 June 2026, support interpretation but are not legally binding. Authoritative interpretation of Article 6(3) rests with the Court of Justice of the EU, and the final guidelines expected later in 2026 may change the draft reading.