EU AI Act Risk Classification: How to Work Out Your Category in 7 Questions

Every AI system sold or used in Europe lands in one of four risk categories — and the category decides your obligations, your deadlines and your fine exposure. These seven questions, asked in order, sort most systems honestly. Where the answer stays unclear, that is exactly what a written classification is for.

The short answer: classification follows what the system does and whom it can affect — never what the vendor calls it. Run the prohibited check first, then the high-risk check, then transparency. Whatever is left is minimal. A general-purpose model underneath your product is a separate, fifth track with its own duties.

Why the order of the questions matters

The EU AI Act is a funnel, not a quiz where points add up. One “yes” near the top outweighs ten “no” answers further down: a system that performs a banned practice is prohibited even if everything else about it is harmless, and a system in a listed high-risk area carries high-risk duties even if its interface is a friendly chatbot. So work top-down, and stop at the first category that fits.

The 7 questions

Question 1

Does it do anything on the prohibited list? Prohibited

Check the banned practices first: social scoring of people, manipulative or deceptive techniques that distort behaviour and cause harm, exploiting vulnerabilities of age or disability, untargeted scraping of faces to build recognition databases, emotion inference in workplaces and schools (with narrow exceptions), and real-time remote biometric identification in public spaces for law enforcement outside the strict exceptions. From 2 December 2026 two more bans join the list: generating non-consensual intimate imagery and generating child sexual abuse material.

If yes: stop. The practice is banned — it applied from February 2025 — and fines reach up to €35 million or 7% of global turnover. No documentation package fixes a prohibited use; the use itself has to go.

Question 2

Is it a safety component of a product already regulated in the EU? High-risk

If your AI is a safety component of a regulated product — medical devices, machinery, vehicles, aviation and similar — or the product itself needs a third-party conformity assessment under EU harmonisation law, the AI system is high-risk by that route alone.

If yes: plan for the full high-risk duty set: risk management, data governance, technical documentation, record-keeping, human oversight and a conformity assessment before the system goes on the market.

Question 3

Does it operate in one of the Annex III areas? High-risk

Annex III lists the uses the Act treats as high-risk: biometrics; critical infrastructure; education and vocational training (admissions, assessment, proctoring); employment and worker management (recruitment, task allocation, evaluation, termination support); access to essential private and public services — including credit scoring and pricing of life and health insurance; law enforcement; migration and border control; and the administration of justice and democratic processes.

If yes: the system is likely high-risk. A narrow exception exists where the system performs a narrow procedural or preparatory task and does not materially influence decisions — but claiming that exception requires documented reasoning, not a feeling. Fines for most violations reach up to €15 million or 3% of global turnover.

Question 4

Does it act on its own — sending, booking, approving, paying? Raises the answer

Autonomy does not create a separate category, but it pushes a system toward the serious end of whichever category it sits in. An agent that only drafts text is one thing; an agent connected to hiring, credit, eligibility, pricing or safety decisions can pull the whole setup into high-risk territory, because its decisions have consequences for people. Autonomy also makes logging non-negotiable: if a system acts on your behalf, you need records of what it did and why.

Remember: using a vendor’s agent does not hand your obligations to the vendor. The business that deploys the system keeps the deployer’s duties.

Question 5

Does it interact with people, or produce content people will see? Limited risk

If the system talks to people (chatbots, voice agents), generates or manipulates content (images, audio, video, text presented as authentic), performs emotion recognition, or creates deepfakes, specific transparency duties apply: people must be told they are interacting with AI, and AI-generated or manipulated content must be disclosed as such.

If yes — and questions 1–3 were no: you are most likely in the limited-risk band. The duties are real but narrow, and they are mostly about honest disclosure, designed so a person is never misled about dealing with a machine.

Question 6

Is there a general-purpose AI model underneath? Fifth track

Foundation and other general-purpose AI (GPAI) models carry their own obligations — technical documentation, training-data summaries and, for the most capable models, systemic-risk evaluation — regardless of your product’s category. If you build on someone else’s model, your duties as a deployer remain yours; if you substantially modify or rebrand a model as your own, some provider duties can move to you. GPAI duties applied from August 2025.

Question 7

None of the above — is it truly none of the above? Minimal risk

Spam filters, inventory forecasts, game AI, basic internal tools: most everyday AI is minimal-risk and carries no extra mandatory duties under the Act. Two cautions before you settle there. First, “internal” tools stop being low-stakes the moment they screen CVs, score customers or set prices — re-run questions 3 and 4. Second, minimal is a conclusion you should be able to show, not just feel: partners, insurers and enterprise customers increasingly ask for the reasoning in writing.

Three worked examples

Illustrative examples, written for this guide — not client cases.

Example A — Support chatbot

A help-desk chatbot answers order-status questions and drafts replies a human sends. Q1–Q3: no. Q5: yes — people must know it is AI. Likely limited risk. If it later gains the power to approve refunds on its own, re-run Q4 and re-check.

Example B — CV screening tool

A tool ranks job applicants for employers. Q3: yes — employment and recruitment is an Annex III area. Likely high-risk, with documentation, oversight and conformity duties, whatever the vendor’s brochure calls it.

Example C — Inventory forecaster

A retailer forecasts stock levels from past sales. Q1–Q6: no — it decides nothing about people and produces no public content. Likely minimal risk. Keep the reasoning on file; that note is what an insurer or partner will ask for.

5 mistakes that misclassify a system

  1. Trusting the vendor’s label. “Our tool is compliant / minimal risk” is marketing until the use case is checked — classification follows your deployment, not their slide deck.
  2. Classifying the model instead of the system. The Act classifies what the whole system does in context. The same model is minimal in a game and high-risk in hiring.
  3. Forgetting the agent upgrade. A chatbot that gains tools — payments, bookings, approvals — changes category in practice while nobody re-runs the questions.
  4. Assuming US companies are out of scope. If people in the EU use it, or you sell into the EU, the Act applies regardless of where the company is registered.
  5. Treating “minimal” as “nothing to write down.” An undocumented conclusion cannot be shown to a partner, an insurer or a regulator. The written reasoning is the product.

Summary checklist

Want the answer in writing, for your system?

The EU AI Act Classification Report maps your specific system against the Act: risk category with reasoning, the obligations that follow, your deadlines and a step-by-step plan in plain English. $49 one-time, delivered by email. Optional Act Watch is $19/month and re-checks your classification as the rules move. Not legal advice.

Get My Report — $49 Answer the free questionnaire first

Frequently asked questions

How do I know if my AI system is high-risk under the EU AI Act?

Work through the seven questions in order. Annex III use — biometrics, education, employment, essential services such as credit or insurance pricing, law enforcement, migration, justice — or being a safety component of a regulated product makes a system likely high-risk. A written classification settles borderline cases.

Does the EU AI Act apply to a US company?

Yes, if your AI system is used by people in the EU or you sell into the EU. The Act follows the market and the use, not the company’s postcode.

Is a customer support chatbot high-risk?

Usually not on its own — it normally sits in the limited-risk band, where the main duty is telling people they are talking to AI. The answer changes once it starts approving, refunding or deciding on its own; then it needs a proper classification.

When do the high-risk obligations apply?

Banned practices applied from February 2025 and GPAI duties from August 2025. The Article 50 transparency duties have applied since 2 August 2026. The Digital Omnibus — Regulation (EU) 2026/1744, in force since 27 July 2026 — moved the high-risk duties: Annex III high-risk obligations apply from 2 December 2027 and Annex I high-risk obligations from 2 August 2028. On 2 December 2026 the grace period for marking AI-generated content from generative systems already on the market ends, and two new Article 5 bans take effect — on generating non-consensual intimate imagery and child sexual abuse material (information as of November 2026).

Is this guide legal advice?

No. This guide and the classification report are educational information based on the public text of the EU AI Act. They do not replace a lawyer, and a formal conformity assessment is a separate, regulated process.

SkillQuesters is independent and is not affiliated with the European Commission. Primary sources: Regulation (EU) 2024/1689 and the European Commission’s AI Act pages. This page contains no affiliate links for the services above — the Classification Report and Act Watch are SkillQuesters’ own services.