How to classify AI systems by risk, including EU AI Act categories

You tier AI systems, not models: the same model can power a minimal-risk coding assistant and a high-risk hiring tool.

12 min read
In short: Risk tiering sorts AI systems by the harm they could cause so review and controls scale with it. The EU AI Act is usually read as four levels: prohibited practices, high-risk systems under Annex I and Annex III, transparency duties, and minimal risk. Classification is a legal judgment; record the tier and role on each AI system and revisit them when its use changes.

What is AI risk tiering?

Risk tiering sorts AI systems by the harm they could cause, so review depth and controls scale with it. The EU AI Act is usually read as four levels: prohibited practices, high-risk systems (Annex III and regulated products), transparency duties, and minimal risk. Record the tier on each AI system, and let it decide its review cadence and controls.

Classification under the EU AI Act is a legal judgment. Whether a system is high-risk depends on its intended purpose, the exceptions in Article 6(3) and how you use it, and the consequences of getting it wrong are legal ones. Use this guide to prepare the question and the evidence for your counsel, not to answer it alone.

An AI system is an application or agent together with the model providers and tools it uses. You tier systems, not models: the same model can sit in a minimal-risk coding assistant and in a high-risk hiring tool. For the wider program that tiers feed, see what AI governance is. For the obligations that follow from a high-risk tier, see the EU AI Act for deployers.

What are the EU AI Act risk categories?

The EU AI Act sorts AI uses into prohibited practices, high-risk systems, systems with transparency duties, and everything else. The Act doesn't use the phrase "four levels" itself. It sets out each category in its own articles, and a system can fall into more than one, such as a high-risk system that also talks to people.

CategoryWhere in the ActExamplesWhat follows
Prohibited practicesArticle 5

Social scoring, emotion recognition at work or school outside medical and safety uses, untargeted scraping of facial images, manipulative techniques that cause significant harm

Banned outright
High-risk systemsArticle 6, Annex I and Annex III

CV screening, credit scoring, exam proctoring, safety components of regulated products

Provider duties in Articles 8 to 22 and deployer duties in Article 26

Transparency dutiesArticle 50

Chat agents that talk to people, generators of synthetic content, deep fakes, emotion recognition

Disclosure and machine-readable marking
Minimal riskNo specific article

Coding agents, spam filters, internal search, most productivity tools

No AI Act duties beyond AI literacy (Article 4); voluntary codes of conduct

General-purpose AI models sit on a separate track in Chapter V, with extra duties for models with systemic risk. Those duties fall on the companies that train and release the models. A company that builds on a model through an API isn't a general-purpose AI model provider for that reason alone.

The Act's dates differ by category. Prohibitions have applied since February 2, 2025, with two added prohibitions from December 2, 2026. Transparency duties apply from August 2, 2026. After the July 2026 amendment by Regulation (EU) 2026/1744, high-risk duties apply from December 2, 2027 for Annex III systems and from August 2, 2028 for Annex I systems. These dates are from the amended Article 113 of Regulation (EU) 2024/1689, read on EUR-Lex on November 16, 2026. They have changed once, so confirm the consolidated text before you rely on one.

Which AI systems are high-risk under Annex III?

Annex III lists eight areas where an AI system is high-risk because of what it's used for. A system is also high-risk under Article 6(1) if it's a safety component of a product covered by the EU product laws in Annex I, such as medical devices or toys, and that product needs a third-party conformity assessment.

Annex III areaExamples of high-risk uses (paraphrased)
1. Biometrics

Remote biometric identification, categorisation by sensitive attributes, emotion recognition

2. Critical infrastructure

Safety components in critical digital infrastructure, road traffic, or water, gas, heating and electricity supply

3. Education and vocational training

Admissions, grading learning outcomes, deciding the level of education a person gets, proctoring tests

4. Employment and workers' management

Recruiting and filtering applications, decisions on promotion and termination, allocating tasks by behavior, monitoring and evaluating performance

5. Essential private and public services

Eligibility for public benefits, creditworthiness, life and health insurance pricing, triage of emergency calls

6. Law enforcement

Victim risk assessment, polygraph-like tools, evaluating evidence, profiling in investigations

7. Migration, asylum and border controlRisk assessments of travelers, examining applications
8. Justice and democratic processes

Helping courts research and apply the law, influencing elections or voting behavior

Article 6(3) lets an Annex III system fall outside high-risk when it doesn't pose a significant risk of harm. That covers four cases: a narrow procedural task, improving the result of work a person already finished, detecting patterns without replacing human review, and preparatory work for an assessment. An Annex III system that profiles people is always high-risk. A provider that relies on the exception must document its reasoning before release (Article 6(4)).

For software companies, area 4 is the one to watch. A coding agent is minimal risk while engineers use it to write code. Using the same agent's activity to rank engineers, allocate their work by behavior, or inform promotion decisions puts that use in Annex III.

How do you classify an AI system by risk?

Classify each AI system by its intended purpose and actual use, record the result with its reasoning, and revisit it when either changes. The following steps prepare the decision for your legal team and keep it attached to the system it describes.

  1. Describe the system. Write its purpose, the people it affects, the data it handles and the actions it can take. A purpose that says "summarizes support tickets for agents" classifies more easily than "AI assistant."
  2. Check the prohibitions. Confirm that no use falls under Article 5. If one does, stop using it.
  3. Check Annex I. Ask whether the system is a safety component of a regulated product, or is the product.
  4. Check Annex III. Compare the purpose and the actual use with the eight areas. Include uses that people have adopted without a decision.
  5. Check the exceptions. If an Annex III area applies, ask whether an Article 6(3) case covers it, and whether the system profiles people, which overrides the exceptions.
  6. Check transparency. Ask whether the system talks to people, generates content or recognizes emotions (Article 50).
  7. Set your role. Decide whether you're its provider, its deployer or both, and whether Article 25 makes you a provider.
  8. Record the tier, the role and the reasoning. Put them on the system's inventory record, with the date and the person who decided.
  9. Revisit on change. Re-classify when the purpose changes, when the system reaches a new provider or tool, or when its observed behavior drifts from its purpose.

Step 4 is where most mistakes happen. A survey asks what a system is for. The answer that matters is how people use it, which you see only by observing it. For that gap, see AI purpose drift.

How should teams outside the EU tier AI systems?

Use an internal tier based on impact and autonomy, and map it to the EU AI Act categories only if you have EU exposure. NIST AI RMF expects your risk tolerance to set how much review each system gets (GOVERN 1.3) and each system to be categorized (MAP 2), but it doesn't prescribe tiers.

The following table shows one workable scheme. Adjust the criteria to your own risk tolerance.

Internal tierTypical criteriaExamples
High

Makes or shapes decisions about people, acts on production systems or money, or handles regulated data such as PHI

Agent that issues refunds, hiring screen, agent with production deploy rights

Medium

Acts with write access inside your environment, or handles customer data, with a person reviewing the result

Coding agent with shell and repository access, support reply drafter

Low

Reads or drafts only, uses no customer data, and every output is reviewed by a person

Internal search, meeting summaries, writing assistant

Autonomy moves systems up. A chat assistant that only answers questions is low. The same assistant given tools that write to a database or call an external API is medium or high, because its actions take effect before anyone reviews them.

What does a risk tier change in practice?

A tier sets how much review, oversight and evidence a system needs. If the tier doesn't change anything, it's a label with no reader. The following table shows the controls that commonly scale with it.

ControlLowMediumHigh
Named owner and purposeRequiredRequiredRequired
ReviewEvery 180 daysEvery 180 days, and on driftEvery 180 days or sooner by policy, on drift and on any change
Runtime policy on consequential actionsOptionalRequired, tested in dry run firstRequired, with a second approver to publish changes
Human oversightOutput reviewDeny rules on risky toolsHold on consequential actions, routed to the owner
Sensitive-information detectionRecommendedRequiredRequired
Decision log retentionPlan defaultPlan default or longer

At least six months; longer where other law requires (see log retention)

An EU AI Act high-risk classification adds its own duties on top of an internal "High": instructions for use, notices to workers and affected people, and, for some deployers, a fundamental rights impact assessment. The EU AI Act checklist lists them.

What Arcjet does not cover

Arcjet records a tier. It doesn't classify a system, and a recorded tier isn't a finding that you meet the EU AI Act. The following stays with you:

  • The classification itself. Deciding a system's AI Act category and role is a legal judgment.
  • Impact assessments. Fundamental rights impact assessments (Article 27), data protection impact assessments, and any impact assessment your management system calls for.
  • Article 6(4) documentation. A provider's written reasoning for treating an Annex III system as not high-risk.
  • Systems Arcjet can't see. The inventory covers enrolled macOS and Windows devices, agents and applications that report to Arcjet, and coding agents with hooks installed. A system outside those sources has no record to carry a tier.
  • Uses that leave no trace in traffic. Arcjet sees the providers, tools and data a system touches. It doesn't see a person using a system's output to make a hiring decision.

Where Arcjet fits

Arcjet records the risk tier and EU AI Act role on each AI system in the AI inventory, next to the owner, the purpose and the review date. Team administrators set them; other team members can read them. The tier sits on an inventory built from observation: Discovery finds the AI providers people use on managed macOS and Windows devices, and the SDK and coding-agent hooks report the agents you build and run.

The controls in the preceding table are ones Arcjet runs. Reviews fall due every 180 days, and the weekly owner digest lists overdue reviews and violations. Drift sends the owner its own notice. Arcjet uses one 180-day interval for every system, so if your policy reviews high-tier systems more often, schedule those reviews yourself and record each one with Mark reviewed. Guard puts a policy on tool calls, runs it in dry run, requires a passing test before a rule goes live, and can require a second approver to publish. Hold decisions route to the system's owner. Purpose drift compares declared data classes and tools with what Arcjet observes, which is the signal that tells you to revisit a classification.

The evidence page maps the tier and role, with the rest of each system's record, to NIST AI RMF, ISO/IEC 42001 and the EU AI Act, listing each requirement as supported or not covered. To see how the frameworks line up, read NIST AI RMF vs ISO 42001 vs EU AI Act.

Frequently asked questions

Is a coding agent a high-risk AI system under the EU AI Act?

Not while engineers use it to write code, because software development isn't an Annex III area. It becomes high-risk if you use it, or its activity data, to evaluate engineers' performance, allocate their work by behavior or inform promotion and termination, which falls under Annex III point 4.

Who decides the risk classification of an AI system?

Your organization does, usually legal or compliance with the system's owner. Classification under the EU AI Act is a legal judgment based on intended purpose and actual use. A provider that treats an Annex III system as not high-risk must document its reasoning under Article 6(4).

Can an Annex III AI system be classed as not high-risk?

Yes, under Article 6(3), if it doesn't pose a significant risk of harm, for example because it performs a narrow procedural task or prepares an assessment that a person makes. An Annex III system that profiles people is always high-risk, whatever the exception.

How often should a high-risk AI system be reviewed?

Neither the EU AI Act nor NIST AI RMF sets a fixed interval. Arcjet uses 180 days for every system. Many teams also review high-tier systems on any change of purpose, provider or tool, and whenever observed behavior drifts from the documented purpose.

Do I need internal risk tiers if I have no EU exposure?

It helps. NIST AI RMF expects your risk tolerance to set how much review each system gets (GOVERN 1.3), and an internal tier based on impact and autonomy is a common way to do that, with or without the EU AI Act categories.

AI runtime security in your code

Protect your AI agent workflows with Arcjet

A tier means something only when it changes the controls. Remote policies let you tighten the rules on a high-tier system centrally, without an application deploy.