The AI Act's risk tiers, explained without the legal jargon

Most explanations of the EU AI Act open with a list of obligations: documentation, human oversight, conformity assessments. That list is real, but it is the second question. The first question is simpler and gets skipped constantly: which of the four risk tiers does your AI system actually sit in? Get that wrong and you either panic over rules that do not apply to you, or miss rules that do.

The AI Act classifies systems, not companies

This is the part that trips up small businesses most: the AI Act does not ask how big you are before deciding what applies. It asks what the system does. A five-person recruitment agency using a biased CV-screening tool faces the same obligations as a 5,000-person one using the identical tool. Size only enters the picture later, in how the fine is calculated (Article 99(6)) and in the lighter documentation route for small and medium enterprises (Article 11(1)). The classification itself is risk-based, full stop.

That means the useful question is never "am I big enough to worry about this," it is "what does this specific tool do." A decision tree answers that faster than a legal summary does.

The beslisboom: four questions in order

A workflow diagram on paper. The classification is a decision tree, not a judgment call.

Question 1: Does the system do one of a short list of forbidden things?

Article 5 bans a specific, narrow set of practices outright, for everyone, in every sector. Not "regulated heavily," banned. The list includes:

• Subliminal or manipulative techniques that distort someone's behaviour in a way that causes harm
• Exploiting vulnerabilities tied to age, disability, or a specific social or economic situation
• Social scoring: evaluating or ranking people based on social behaviour, with unrelated consequences attached
• Inferring emotions in workplaces or schools (with narrow medical or safety exceptions)
• Untargeted scraping of facial images from the internet or CCTV to build a recognition database
• Real-time remote biometric identification in public spaces, for law enforcement, outside a short list of narrow exceptions

If your system does one of these, the answer is not "add documentation." The answer is "do not build this," and no company size changes that. In practice, almost nothing a recruitment or logistics business builds lands here. This tier exists mostly to rule itself out quickly, which is exactly what it should do for you.

Question 2: Is it on the high-risk list?

If a system is not outright forbidden, the next question is whether it appears on one of two lists.

Annex III lists eight categories of standalone high-risk use, including biometrics, critical infrastructure, education, and, relevant to Radical's own sector, employment and worker management: CV-screening tools, tools that decide who gets shortlisted, tools that monitor or evaluate staff performance. A recruitment business building or deploying any of these needs to take Annex III seriously, not as a hypothetical.

Annex I covers a different situation: AI embedded as a safety component inside a product that already needs third-party certification, think machinery or medical devices. This is a smaller, more specialised group, and it is the one with the longest runway: 2 August 2028.

Annex III's original deadline of 24 months after entry into force has been pushed back by the Digital Omnibus Regulation (EU) 2026/1744, in force since 27 July 2026, to 2 December 2027. That is a real extension, not a rumour, and it is worth knowing so you are not rushing to meet a date that no longer applies.

High-risk systems carry the heaviest obligations: a risk management system, data governance, technical documentation, human oversight design, and a conformity assessment before the system goes into use.

Question 3: Does it just need to be honest about being AI?

Most systems that are not high-risk still fall under Article 50 if people interact with them directly. Chatbots need to disclose they are AI unless it is obvious from context. AI-generated or manipulated audio, image, or video content needs a machine-readable marker. This tier does not require a conformity assessment or a risk management file. It requires telling people what they are talking to. The compliance date here is 2 August 2026, already in force.

Question 4: Everything else is minimal risk

Spam filters, recommendation engines with no employment or safety angle, internal drafting tools: the large majority of what a small or mid-sized business actually uses in daily operations sits here, with no specific AI Act obligation attached beyond ordinary good practice.

A danger sign. A small number of AI practices are simply forbidden, regardless of company size or sector.

The four tiers side by side

This table is the destination, not the starting point. Reading it top to bottom without walking the decision tree first is how companies end up either over-engineering a spam filter or under-documenting a CV-screening tool. The table tells you what applies once you know the tier. The tree is what tells you the tier.

A worked example: the CV-screening tool a recruitment agency actually uses

Take a concrete case, close to Radical's own sector. A recruitment agency uses a tool that ranks incoming CVs against a job description and produces a shortlist score. Walk it through the four questions in order.

Question 1. Does it manipulate behaviour, exploit a vulnerability, assign a social score, infer emotion, scrape faces, or identify people in real time in public? No. It ranks documents against a job description. Cleared.

Question 2. Is it on the high-risk list? Yes, immediately: Annex III names employment decision-support tools by name, and a shortlist score is exactly that, a tool that materially affects whether someone gets an interview. This is not a borderline case and it is not a size question. A five-person agency using this tool carries the same Annex III obligations as a national staffing firm.

That means: a risk management system covering the ranking logic, data governance over the CV data feeding it, technical documentation describing how the score is produced, a human oversight step before a rejection becomes final, and a conformity assessment before the tool is used for real candidates. The deadline for having this in place is 2 December 2027, not next week, but not indefinite either.

Question 3 and 4 do not apply here; the system already landed in Question 2.

Now compare that to a second tool at the same agency: an internal assistant that drafts outreach emails to candidates, with a human reviewing and sending every message. Question 1: no. Question 2: no, it does not decide who gets shortlisted, it drafts a message a person still approves. Question 3: if candidates could mistake it for a human writing to them without disclosure, Article 50 applies, so the agency adds a line noting AI assistance where relevant. Otherwise it lands in Question 4, minimal risk. Two tools, same company, same week, two completely different obligation sets, because the classification tracks what the tool decides, not who built it or how big the company is.

Why this order matters more than the list itself

A legal summary presents these four tiers as parallel categories. In practice they are sequential filters, and each one only matters if the system survived the previous question. Walking a real tool through the tree, in order, takes minutes and ends with a specific answer instead of a vague sense of dread. Walking a spreadsheet of forty obligations without first asking which tier applies takes hours and usually ends with the same vague sense of dread, just later.

The practical version we use with clients: name the tool, name exactly what decision or output it produces, and check it against Question 1 before touching anything else. Nine times out of ten for a business outside biometrics, critical infrastructure, or employment decisions, the tool clears the forbidden-practices question and the high-risk list in the same short conversation, and lands in Article 50 or minimal risk. The tenth time is worth catching early, because Annex III's obligations are not something you retrofit in a week.

Someone working at a laptop. Most AI systems a small business actually uses land in the two lowest tiers.

Where the general-purpose AI rules fit in

One source of confusion worth clearing up separately: the rules for general-purpose AI models, the large models underlying tools like ChatGPT or Claude, are not a fifth tier sitting alongside the four above. They are a parallel track that applies to the model provider, not to the tier of the application built on top of it. A GPAI provider has its own obligations around technical documentation, a training-data summary, and, for models above a systemic-risk compute threshold, additional safety testing. Those obligations became enforceable on 2 August 2026, the same date as the Article 50 transparency rules. For a small business building a CV-screening tool on top of a general-purpose model, both tracks can apply at once: the application is classified through the four-question tree above, and separately, whichever GPAI model it is built on carries its own provider-side obligations that the business does not have to duplicate, only be aware of when choosing a vendor.

What this does not replace

This is a classification exercise, not legal advice, and it does not replace AI-beleid as the document that governs how your organisation actually uses AI day to day, or AI-geletterdheid as the baseline understanding your team needs regardless of which tier a given tool falls into. Article 4's AI literacy requirement applies across all tiers, not just the high-risk ones, so classification and literacy are two separate jobs, both needed. A team that can correctly answer question 1 through 4 for its own tools has, in effect, already met a meaningful part of what Article 4 asks for: a working understanding of how the AI it uses actually gets treated under the law.

About this page

The four-tier structure, the Article 5 practice list, and the 2 December 2027 and 2 August 2028 dates come directly from the AI Act text and the Digital Omnibus Regulation (EU) 2026/1744, cross-checked against the Article 99(6) SME fine rule and the Article 11(1) simplified documentation route mentioned above. Written by Radical's own team; no client data was used.