Risk classification: which of your organization's systems is high-risk and proving why
The short answer
Classification is the hinge of the whole Act
The tier a system lands in decides everything after: banned, heavily regulated, transparency-only, or largely free. This topic sorts every system you run into a tier and proves each sort, because every later obligation attaches to the right systems only if the classification is right.
What you will be able to do
- Distinguish the four risk tiers of the EU AI Act (prohibited, high-risk, limited or transparency risk, and minimal risk) and place a given AI system in the correct tier with a reason.
- Analyze an AI system against the two doors into high-risk classification: Article 6(1) (a safety component of a product covered by the Union harmonisation legislation listed in Annex I) and Article 6(2) (a use case listed in Annex III).
- Map a system to the eight Annex III areas, and recognize when a use case such as evaluating eligibility for public benefits (Annex III, point 5(a)) puts a system squarely in high-risk territory.
- Apply the Article 6(3) filter correctly: judge whether an Annex III system escapes high-risk because it performs only a narrow procedural, confirmatory, pattern-detecting, or preparatory task and poses no significant risk, and recognize that any system that profiles natural persons is always high-risk regardless.
- Separate a prohibited system from a high-risk one, using the Denmark case to see why the same welfare model can raise both the Article 5 social-scoring ban and the Annex III high-risk classification.
- Identify who bears the classification duty across the provider, deployer, and developer roles, and when a deployer becomes a provider by modifying a system or putting its own name on it (Article 25).
- Produce a risk classification record for your own AI systems inventory that names each system's tier, cites the Annex III point or Article 5 practice that decides it, and, for anything you call not-high-risk, documents the Article 6(3) reasoning, so the placement is provable and not merely asserted.
The ideas, one by one
Four tiers, tested strictest first
Prohibited (Article 5), high-risk (Article 6), transparency (Article 50), minimal. Test the top rung first and stop at the first that decides, so a banned system never hides as merely high-risk and a high-risk system never falls to minimal.
High-risk has exactly two doors
Article 6(1) (a safety component of an Annex I product needing third-party conformity assessment) and Article 6(2) (a use in one of the eight Annex III areas). Most organizations meet the use-case door, and the everyday high-stakes uses (hiring, credit, benefits, policing) are behind it.
The Article 6(3) exit is narrow, and profiling closes it
An Annex III system escapes high-risk only if it does a narrow procedural, confirmatory, pattern-detecting, or preparatory task and poses no significant risk, and any system that profiles people is always high-risk regardless. The "there is a human in the loop" defense does not lower the tier of a profiling system.
Prohibited and high-risk are categorically different
A prohibited system may not be used at all; no conformity file rescues it. The Denmark welfare models are argued to be prohibited social scoring and are separately high-risk on their benefit-eligibility function. Classify to the strictest tier the use triggers, and never build a compliant record for something that is simply banned.
Classification attaches to a role, and the role can switch
The provider makes the primary classification; deployers must still know the tier to meet their own duties and to catch an under-classifying vendor. Under Article 25, a deployer that rebrands or substantially modifies a system becomes its provider, inheriting the heavy obligations. Record your role, and re-run when you modify a system.
"Proving why" is the second half of the job
Article 6(4) requires a provider that calls an Annex III system not-high-risk to document the assessment and produce it on request. A tier is not a decision until it is a documented decision; never record a tier without its reason, and never claim not-high-risk on an Annex III system without the written 6(3) argument.
Classify to the sub-point, not the area
The precise wording carries the carve-outs: Annex III point 5(b) credit scoring exempts fraud detection, while point 5(a) benefits eligibility does not. Reading the general area misses the exact line the legislator drew.
The high-risk timeline moved, and it is runway, not a reprieve
The Digital Omnibus (Council adoption 29 June 2026) deferred stand-alone Annex III high-risk obligations to 2 December 2027 and embedded Annex I high-risk to 2 August 2028. Classification is due now, the prohibitions and transparency and literacy duties already apply, and you cannot make compliant a system you have not classified.
Over-classification has a real cost
Calling a limited-risk chatbot high-risk steals the scarce compliance effort the genuinely high-risk systems need and trains the organization to treat the label as noise. Correct classification cuts both ways, up for the welfare model, down for the chatbot.
The record is a living artifact that travels
Your conformity file is built on it (see Topic 5.6), the deployer impact assessment attaches to it (see Topic 10.4), the dossier inherits it (see Topic 13.1), and a hostile review will attack it (see Topic 11.1). Build it from the real use of each system and a written reason, and it holds; label by gut and it fails at the worst moment.
Classify by use, not by system, and split what does more than one thing
A single system can carry several uses in several tiers, and a single use can answer to several Annex III points at once. Split a system by its distinct uses, classify each, record every applicable hook, and govern the whole to the strictest tier any use triggers. The model layer (a general-purpose model's own obligations, owned by (see Topic 5.5)) and the use layer (the Article 6 tier) are separate questions; the use sets the tier, not the model.
Courts already test the "we only assist" defense, and it fails when the human leans on the output
In the SCHUFA credit-scoring ruling the Court of Justice of the EU held that an agency generating an automated repayment score is itself making an automated decision when lenders draw strongly on it, rejecting the claim that it only performed a preparatory act (Case C-634/21, 7 December 2023). It is a data-protection case, not an AI Act one, but it lands where the profiling override lands: you cannot score a person and disown the decision. "A human decides" is a factual claim about reliance, not a label that lowers the tier.
Listen to it
The same lesson, talked through by two hosts as episode 36 of the podcast.
Real cases
These examples show risk classification done well and badly, with the legal reasoning made explicit. The deep anchor is Denmark's welfare AI; the others sharpen one point each and are owned in depth by their own topics.
Example 1 (the anchor): Denmark's welfare fraud models and the tier that was skipped. Between the pension administrator ATP and the welfare authority UDK, Denmark ran as many as sixty fraud-detection models over the merged personal data of millions of residents, including "Really Single" (flagging "unusual" living arrangements in pension and childcare schemes) and "Model Abroad" (flagging beneficiaries with strong ties to non-EEA countries), and Amnesty International's November 2024 report "Coded Injustice" documented that these systems risk discriminating against people with disabilities, low-income people, migrants, and marginalized racial groups (Amnesty International, "Coded Injustice: Surveillance and Discrimination in Denmark's Automated Welfare State," November 2024). Read against this topic's four questions, the classification these systems demanded is stark. Question 1: Amnesty argues the apparatus functions as social scoring, which Article 5 prohibits; if that argument holds for a given model, the model is banned and the analysis ends. Question 2: even setting the prohibition aside, a system a public authority uses to evaluate benefit eligibility and to reduce, revoke, or reclaim benefits is high-risk under Annex III point 5(a) by name, and because these models profile people from residency, citizenship, and family data, the Article 6(3) profiling override keeps them high-risk regardless of the "we only support caseworkers" defense. The lasting lesson: the harm did not begin with a bad model, it began with a system operated as if it were a minimal-risk internal analytics tool when its use put it on the top two rungs of the law. Classification was the control that would have forced the safeguards, and it was the control that was skipped.
Example 2 (the same area, the credit sub-point): consumer credit scoring under Annex III point 5(b). Annex III point 5(b) makes AI used to evaluate the creditworthiness of natural persons or set their credit score high-risk, with a narrow carve-out for AI used to detect financial fraud. A bank running an AI model to approve or price consumer loans is squarely high-risk on this hook, and the classification is usually uncontested. The instructive subtlety is the fraud carve-out: an AI used purely to detect financial fraud in transactions is excepted from the credit-scoring category, which is a deliberate line the legislator drew, and it is a reminder that classification is exact wording, not theme. Note the contrast with the welfare case: welfare-benefit fraud detection is not credit scoring, so the 5(b) fraud carve-out does not rescue it; it lives under 5(a), which has no such carve-out. Reading the precise sub-point, not the general area, is the skill. The deep treatment of specific enforcement in these areas belongs to other Module 5 topics; here the point is that neighboring sub-points carry different carve-outs and you must classify to the sub-point.
Example 3 (the employment door, referenced): AI hiring screens under Annex III point 4. AI used to recruit, screen, or select candidates, or to make promotion and termination decisions, is high-risk under Annex III point 4. The 2025 litigation over Workday's hiring-screening AI, where a federal court allowed an age-discrimination collective action to proceed and treated the vendor's tool as an agent of the employers, shows why the tier fits: a screening system materially influences who gets a job, a textbook fundamental-rights effect (deep treatment and the case citation are owned by Topic 5.4 (see Topic 5.4)). For classification, the lesson is two-fold: the employment area is one of the most common ways an ordinary company's AI becomes high-risk, and the role question bites hard here, because an employer that fine-tunes or configures a vendor screen to its own criteria may cross Article 25 into being the provider of a high-risk system, not merely its deployer.
Example 4 (the filter used honestly): a genuinely procedural Annex III system. Not every system that touches an Annex III area is high-risk, and Article 6(3) exists for real cases. Consider an AI that does nothing but convert scanned benefit application forms into structured text for a human assessor who then independently decides eligibility, reading the original documents. If the extraction does not evaluate the person, does not profile them, and does not materially influence the outcome (the assessor decides from the source documents, not the AI's summary), it can qualify for the Article 6(3) narrow-procedural or preparatory-task exit and be classified not-high-risk. The discipline is that this classification is a claim you document under Article 6(4), and it holds only as long as the facts hold: the moment the extraction starts scoring applicants or the assessors start deferring to it, the profiling and significant-risk tests reengage and the system climbs back to high-risk. The honest use of the filter is narrow and evidenced; the dishonest use is calling a scoring system "just support."
Example 5 (transparency risk, not high-risk): a customer-service chatbot. A retail company runs a generative chatbot that answers product questions. It does not decide anything in an Annex III area, so it is not high-risk. But it interacts with people who might think they are talking to a human, so Article 50 requires it to disclose that it is an AI system. Classify it as transparency-risk, record the Article 50 hook, and the obligation is a clear notice, not a conformity file. This example matters because it is the opposite error from Denmark: over-classifying a genuinely limited-risk system as high-risk wastes the scarce compliance effort you need for the systems that are actually high-risk. Correct classification cuts both ways, up for the welfare model and down for the chatbot.
Example 6 (minimal risk, and why "minimal" is still a decision): a demand-forecasting model. A logistics firm runs an AI model that forecasts warehouse demand to optimize stock. It touches no Annex III area, interacts with no consumer, and generates no content, so it is minimal risk with no mandatory obligations under the Act. The reason to still record it, with the reason "no Annex III use, no Article 50 trigger," is that classification is a living decision: if the firm later repurposes the same model to, say, score gig workers for shift allocation, it has entered Annex III point 4 and must be reclassified. A minimal-risk record with its reason is what makes that later change visible; an unclassified system is one whose creep into a higher tier no one notices until an incident. Minimal is a tier you assign on purpose, not a synonym for "we did not look."
Example 7 (a court closes the "we only prepare, the human decides" defense): the SCHUFA credit-scoring ruling. The Court of Justice of the European Union ruled in December 2023, in the SCHUFA case, that when a German credit-reference agency generates an automated probability score of whether a person will repay, and lenders "draw strongly" on that score to grant or refuse credit, the agency is itself engaged in automated individual decision-making under Article 22 of the General Data Protection Regulation (Court of Justice of the EU, Case C-634/21, Judgment of 7 December 2023, SCHUFA Holding). SCHUFA had argued that it merely produced a score and that the bank made the decision, so the rules on automated decisions fell on the bank, not on it. The Court rejected that: where the human decision-maker draws strongly on the score, the score is the decision in substance, and the scorer cannot disclaim it as a mere preparatory act. This is a data-protection case, not an EU AI Act ruling, but read next to the Article 6(3) filter the lesson is exact. The "we only support a human who really decides" defense is not a magic phrase; it is a factual claim that a court will test against how much the human actually relies on the output. For classification it reinforces two things at once: creditworthiness scoring of natural persons is the textbook example of a consequential, profiling decision (the Annex III point 5(b) hook from Example 2), and the "the AI only assists, the human decides" argument fails exactly when the human leans on the AI, which is the very test that closes the Article 6(3) exit under the AI Act. Two different laws, data protection and the AI Act, arrive at the same place the profiling override reaches: you cannot score a person and then disown the decision.
Where people go wrong
- "There is a human in the loop, so it is not high-risk." This is the Denmark defense, and it fails twice. Article 6(3)'s pattern-detection exit requires that the AI not replace or improperly influence the human assessment, and a caseworker who defers to the flag is being influenced; and the profiling override makes any system that profiles people high-risk regardless of human review. A nominal human does not lower the tier of a profiling system.
- "We decided it is low risk." A tier is not a decision until it is a documented decision. Article 6(4) specifically requires a provider that calls an Annex III system not-high-risk to document the assessment and produce it to authorities on request. An undocumented "low risk" is not a classification; it is the exposure the law was written to catch.
- "It is just internal analytics, so the Act does not care." The Act grades use and effect, not where the system sits on your org chart. Denmark's models were internal fraud analytics; their use, deciding who keeps a benefit, is what put them on the top two rungs. An internal tool that evaluates people for a consequential decision is classified by that use, not by the label "internal."
- "The model is simple, so it cannot be high-risk." Classification grades consequence, not sophistication. A plain statistical model that decides benefit eligibility is high-risk; a frontier model that suggests recipes is minimal. Do not let "it is only a logistic regression" talk you down a tier the use case forces you up to.
- "We bought it, so classification is the vendor's job." The provider makes the primary classification, but a deployer must know the tier to meet its own Article 26 duties and to catch an under-classifying vendor, and under Article 25 a deployer that modifies a system or puts its name on it can become the provider. Buying does not outsource the classification decision; it just changes whose primary duty it is, and modification can hand it to you.
- "Annex III means automatically high-risk." Almost, but not quite: Article 6(3) genuinely exits a narrow-procedural, confirmatory, pattern-detecting, or preparatory system that poses no significant risk and does not profile. The mistake in both directions is real: treating every Annex III touch as automatically high-risk over-classifies harmless systems, while treating the 6(3) exit as a general hatch under-classifies dangerous ones. The exit is narrow and evidenced, and profiling closes it.
- "High-risk does not apply until December 2027, so we can wait." The deferral (Annex III high-risk to 2 December 2027, Annex I high-risk to 2 August 2028) is runway to get compliant, not permission to ignore. You cannot prepare a system you have not classified, and the prohibitions and transparency and literacy duties already apply. Classification is due now; the obligations bite later.
- "Prohibited and high-risk are basically the same, both mean lots of rules." They are categorically different. A prohibited system may not be used at all; no conformity file, no oversight, nothing rescues it. A high-risk system is permitted with obligations. Classifying a prohibited system as merely high-risk and building it a conformity file is building a compliant record for something that is simply banned.
- "We classified everything last year, so we are done." Classification is a living record. A modification can change the tier (a minimal model wired into a hiring decision becomes high-risk) and the role (Article 25 can turn a deployer into a provider), and the law itself moves (the Omnibus changed the timeline). Re-run classification when a system changes purpose, when you modify it, and when the instrument changes.
- "Reading the general Annex III area is enough." The precise sub-point carries the carve-outs. Point 5(b) credit scoring exempts fraud detection; point 5(a) benefits eligibility does not. Classifying to the area rather than the sub-point misses exactly the line the legislator drew, and a fraud model can be excepted under one sub-point and squarely high-risk under its neighbor.
- "Over-classifying is the safe error, so when in doubt call it high-risk." Over-classification has a real cost: it drowns the scarce compliance effort you need for the genuinely high-risk systems in paperwork for harmless ones, and it trains the organization to treat the high-risk label as noise. The customer-service chatbot classified as high-risk is effort stolen from the eligibility model. Correct classification, up and down, is the goal, not maximal caution. (The one exception is a genuine borderline near the high-risk line, where defaulting up while you resolve the doubt is right, as long as you record the open question rather than freeze a guess.)
- "One system gets one tier." A single deployed system can do several jobs, and different jobs fall in different tiers. A platform that both schedules shifts (minimal) and scores workers for promotion (Annex III point 4, high-risk) is not one tier; split it by use, classify each use, record every applicable hook, and govern the whole to the strictest. Treating a multi-use system as one tier hides its highest-risk use, which is usually the one that carries the obligations.
- "The foundation model sets the tier." The general-purpose model layer and the use layer are separate questions. A GPAI (general-purpose AI) model carries its own obligations (owned by (see Topic 5.5)); the Article 6 tier attaches to the system and its use, not to the underlying model. Wiring a general-purpose model into a benefits decision makes that system high-risk regardless of the model's own status, and using the same model to draft marketing copy does not. Do not let "it is just the vendor's model" or "it is a powerful model" set the tier; the use sets the tier.
- "An AI that only produces a score cannot be the decision." The SCHUFA ruling shows a court rejecting exactly this: when the human decision-maker draws strongly on the score, the score is the decision in substance (Court of Justice of the EU, Case C-634/21, 7 December 2023). A scoring system is not saved from high-risk by the presence of a human who defers to its score; that is the same fact the Article 6(3) profiling and influence tests turn on.
- "A pilot or an internal proof of concept is not classified yet." The tier follows the use and its effect on real people, not the project's formal status. A pilot that scores real benefit recipients or screens real candidates is doing the high-risk thing during the pilot. Classify it by what it does to people now, not by "it is not on the market yet," so the safeguards are in place before the harm rather than after the launch.
Questions people ask
- What is risk classification?
- The analysis this topic produces for every AI system you run: sorting each into one of the EU AI Act's four risk tiers, naming the exact legal hook that decides it, recording your role, and documenting the reason, so the placement is provable and not merely asserted.
- What is the four risk tiers?
- The EU AI Act's risk pyramid: unacceptable risk (prohibited, Article 5), high risk (Article 6), limited or transparency risk (Article 50), and minimal risk (everything else). Obligations follow the tier, and you classify to the strictest tier a system's use triggers.
- What is prohibited (unacceptable risk)?
- A practice banned outright under Article 5 (for example social scoring of people, manipulative or exploitative techniques causing significant harm, certain biometric uses). For most of the list, above all social scoring, a prohibited system may not be placed on the market or used at all and no conformity file rescues it; real-time remote biometric identification for law enforcement carries its own narrow, conditioned exceptions, so read the exact sub-clause rather than assume every prohibition is unconditional. In force since 2 February 2025.
- What is high-risk AI system?
- A permitted system carrying the Act's full requirements because it can seriously affect health, safety, or fundamental rights. Defined by Article 6 through two doors: Article 6(1) (safety component of an Annex I product needing third-party conformity assessment) and Article 6(2) (a use in one of the eight Annex III areas). More on High-risk AI system
- What is Annex III?
- The list of eight areas whose use cases make a system high-risk through Door 2: biometrics; critical infrastructure; education; employment and worker management; access to essential private and public services and benefits; law enforcement; migration, asylum, and border control; and administration of justice and democratic processes. More on Annex III
Keep going
This lesson builds EU AI Act obligations and timelines, and that page shows the roles that hire for it. Every Certified AI Governance Professional (CAIGP) lesson.
Watch the lesson and prove it
Proving it happens inside the program: watch, explain it in your own words, and it goes on a record employers can check.
Watch the lesson and prove it →