EU AI Act risk tiers: a practitioner primer

How the risk pyramid works in practice

The Act is a risk-based law. Heavier uses carry heavier duties; banned practices have no compliance path. The Commission's overview is the shortest official entry point before you dig into articles (European Commission — AI Act). Keep the Official Journal text open when a vendor slide conflicts with primary law (Regulation (EU) 2024/1689).

Dates move. The Digital Omnibus package shifted several high-risk application dates after the original calendar proved unworkable for standards and guidance. Treat any date table — including ours — as a snapshot and re-check a maintained timeline before you commit budget (EU AI Act implementation timeline). "We planned for August 2026" is not a defense if the binding date is later — and it is also not a reason to freeze work.

Provider vs deployer in one sentence

Providers place the system on the market or put it into service under their name; deployers use it under their authority. Many enterprise agent stacks mix both roles. Classify the system first, then assign who owns which duty. Mixing the two conversations is how teams "finish" a checklist that nobody can sign.

Unacceptable risk: banned outright

Article 5 bans practices such as social scoring by public authorities, certain real-time remote biometric identification in publicly accessible spaces (with narrow exceptions), and emotion recognition in workplaces and schools in the contexts the article defines. These prohibitions applied from 2 February 2025. There is no conformity pathway: if your system falls here, it cannot be placed on the EU market.

Product teams sometimes invent a "limited mode" for a banned practice. That does not create a legal path. If the intended purpose is the banned practice, stop. If a feature could be used that way as a foreseeable misuse, document the misuse and the technical or contractual controls that block it — and get counsel when the boundary is fuzzy.

Checklist — prohibited-use screen

- Intended purpose statement reviewed against Article 5 categories - Foreseeable misuse list written (not just happy-path demos) - Kill-switch or disablement plan if a feature drifts into banned territory - No sales language that implies a "compliant version" of a banned practice

High-risk: the documentation and monitoring load

A system is high-risk through two paths in Article 6: it is a safety component of a product covered by EU harmonization law listed in Annex I, or its intended purpose falls under one of the Annex III domains — including biometrics, critical infrastructure, education, employment, access to essential private and public services (credit and insurance are frequent triggers), law enforcement, migration, and administration of justice. Full high-risk chapter obligations for stand-alone Annex III systems now apply from 2 December 2027 under the Digital Omnibus amendments (the original general date was 2 August 2026). Annex I product-embedded high-risk systems apply from 2 August 2028. Confirm the live timeline before you print a board slide.

When you are high-risk, the load includes a risk-management system (Art. 9), data governance (Art. 10), technical documentation (Art. 11), automatic logging (Art. 12), transparency to deployers (Art. 13), human oversight (Art. 14), and accuracy / cybersecurity expectations (Art. 15). Our sister article walks the documentation gaps teams miss most often.

Article 6 test — how we run it in a workshop

- Write the intended purpose in one plain paragraph. Ban marketing adjectives. - Ask: is this a safety component of an Annex I product? If yes, follow the Annex I path and sectoral product law interactions. - Else walk Annex III domains one by one. Employment screening, worker management, creditworthiness, and education scoring are the enterprise-agent magnets. - Record the decision, the owner, and the date. Revisit when the purpose or the deployer class changes. - If the answer is "maybe," treat it as high-risk for planning until counsel says otherwise. Ambiguity is not a free pass to do nothing.

Tier | Typical examples | Core duty shape | Calendar note (verify live) Unacceptable | Social scoring by public authorities; banned biometric / emotion uses in Art. 5 | Ban — no compliance path | From 2 Feb 2025 High-risk (Annex III) | Hiring agents, credit scoring aids, education assessment tools | Docs, logging, oversight, risk mgmt, conformity path | From 2 Dec 2027 (Omnibus) High-risk (Annex I) | AI as safety component in regulated products | Same high-risk chapter + sectoral product law | From 2 Aug 2028 (Omnibus) Limited-risk | Many chatbots; synthetic media with transparency duties | Disclose AI / label synthetic content (Art. 50 family) | Transparency still on 2026 track — re-check timeline Minimal-risk | Spam filters; many internal drafting aids | No mandatory AI Act duties; voluntary codes only | Still classify — do not assume

Limited risk: transparency obligations

Chatbots, deepfake generators, and emotion recognition outside high-risk contexts often land here. The main duty is transparency: users must know they are interacting with AI, and synthetic content must be labeled where the Act requires it. There is no conformity assessment for ordinary limited-risk chatbots. Penalties for certain information and transparency failures sit at a lower band than prohibited-practice fines (Article 99 uses tiers; see the regulation text for the exact ceilings).

Do not treat "limited-risk" as "ignore the calendar." Several transparency obligations stayed on the 2026 track even when high-risk dates moved. Watermarking and related technical marking rules for systems already on the market have their own sub-dates on the public timelines. Read the implementation timeline page before you tell a customer "everything slipped to 2027."

Checklist — limited-risk honesty

- User-facing disclosure that the interlocutor is AI (where required) - Synthetic media labeling plan for generated audio/video/images - No claim of "high-risk certified" for a limited-risk chatbot - Separate note if a chatbot feature later feeds an Annex III decision

Minimal risk: essentially unregulated under the Act

Spam filters, many recommendation engines, and most internal drafting assistants sit here. The Act imposes no mandatory obligations, though voluntary codes of conduct exist. Most enterprise AI agents that make or influence consequential decisions do not land in this tier — that is the gap teams underestimate when they point at a writing assistant and conclude the whole company is "out of scope."

If your agent only drafts and a human always retypes into another system with no AI-influenced decision, you may be closer to minimal risk. If the agent's score, rank, or recommendation is the de facto decision, revisit Annex III. Purpose and use beat architecture diagrams.

Penalties in plain numbers (still verify in Art. 99)

Article 99 sets ceilings by severity: prohibited practices up to EUR 35 million or 7% of global annual turnover (whichever is higher); high-risk non-compliance up to EUR 15 million or 3%; certain incorrect information / transparency failures up to EUR 7.5 million or 1.5%. These are maxima, not automatic invoices. Still, they are why classification is a board-level topic, not a Friday afternoon chore.

For process vocabulary outside the EU statute, teams sometimes borrow the NIST AI Risk Management Framework. That is fine for internal governance language (NIST AI RMF). It does not change your tier under Regulation (EU) 2024/1689. The EU's broader digital-policy context also sits around the Act; the Commission's "Europe fit for the digital age" pages help place AI rules next to neighboring digital files (Europe fit for the digital age), but tiering still starts from Articles 5–6 and the annexes.

Common classification mistakes

- Classifying the model instead of the use. The same foundation model can power a spam filter and a hiring ranker. The use decides the tier. - Assuming "internal only" means minimal risk. Internal employment tools are a classic Annex III trigger. - Freezing the calendar at 2 August 2026 for all high-risk duties. Omnibus moved Annex III / Annex I high-risk application dates; transparency duties did not all move with them. - Letting sales choose the tier. Sales wants "low risk." Engineering and legal own the map. - Skipping a written decision. If you cannot show why you are not high-risk, a reviewer will assume you never looked.

Decision criteria: what to do this quarter

If you are building or buying EU-facing AI in 2026:

- Inventory systems that influence employment, credit, education, or essential services - Run the Article 6 test and store the written result - Keep prohibited-use and transparency work on the near calendar - Use the 2027/2028 high-risk runway to build documentation, logging, and oversight — do not wait for the week before the date - Point every external claim at EUR-Lex or the Commission page, not at a podcast summary

What this is NOT

This primer is not legal advice, not a conformity assessment, and not a promise that VendorVet (or any checklist tool) guarantees correct classification or "never misses" a high-risk use. We do not invent customer metrics or claim a fine you will "definitely avoid." Dates in this article reflect the Digital Omnibus shift for high-risk application as commonly reported after mid-2026 adoption; always verify against the Official Journal and a maintained implementation timeline before you bind a compliance program.

Closing

Map the tier first. Then open the right checklist. Banned practices stop; high-risk systems collect evidence; limited-risk systems tell the truth to users; minimal-risk systems still need an honest inventory. When the documentation load starts, read the technical documentation checklist next — that is where most teams discover the gaps that classification alone cannot fix.

References

- Regulation (EU) 2024/1689 (EUR-Lex) - European Commission — AI Act - EU AI Act — Implementation Timeline - NIST AI Risk Management Framework - European Commission — Europe fit for the digital age

Related articles

- Technical documentation under the AI Act: what teams usually miss - Explore the full AI micro-SaaS topic matrix →