The decision that gates everything
The EU AI Act is risk-tiered: what you must do depends entirely on which tier your system falls in. A prohibited system cannot ship at all. A high-risk one carries a heavy compliance programme. A limited-risk one needs only disclosure. A minimal-risk one needs nothing. So the first, gating task for any AI system is classification - and it is a decision you can learn to make quickly and defend. This session builds that reflex with a checker you drive.
First match wins 7 min live
Classification runs in order: check prohibited first, then high-risk, then limited, else minimal - and assess GPAI as a parallel layer. The first tier that matches is the answer.
LiveThe order, and why order matters3 min▶
- In scope? (Art 2) - if not (no EU market, no EU output), the Act does not apply. Everything below assumes you are in.
- Prohibited? (Art 5) - the eight banned practices. If it matches, stop: it cannot ship, and the fine tier is the highest (7% / EUR 35M).
- High-risk? (Art 6 + Annex III) - the eight use-case areas (employment, credit, education, biometrics, law enforcement...). This is where most consequential business AI lands.
- Limited-risk? (Art 50) - chatbots, deepfakes, synthetic content. Disclosure and labelling only.
- Minimal - everything else. No mandatory obligations.
- GPAI (parallel): if the artefact is a general-purpose model, it carries its own obligations regardless of the tier of systems built on it.
Live★ Build-along: classify systems4 min▶
Pick systems in the checker and read the tier, the obligations, and the basis. Start with "screen job applicants" (Vantage's tool -> high-risk, Annex III(4)), then a prohibited one (social scoring), a limited one (chatbot), and a GPAI model.
Downgrade, and the profiling trap 8 min live
Two rules decide the close calls in the high-risk tier. One can pull a system down out of high-risk; the other slams it back in. Miss them and you will misclassify.
LiveThe Article 6(3) downgrade3 min▶
Not every system in an Annex III area is automatically high-risk. Article 6(3) lets a system escape the high-risk tier if it does not pose a significant risk because it only:
- performs a narrow procedural task; or
- improves the result of a completed human activity; or
- detects decision patterns/deviations without replacing human review; or
- performs a preparatory task.
Example: a tool that only formats and forwards applications to a recruiter, without scoring or ranking, may qualify for the downgrade. But the moment it ranks, scores, or filters candidates, it is doing the consequential work - and it is high-risk.
LiveThe profiling trap3 min▶
- The override: if the system performs profiling of natural persons, the Article 6(3) downgrade does not apply - it is high-risk, full stop.
- Why it matters: teams reach for the downgrade to avoid the compliance burden, then forget that any real evaluation of a person (scoring, ranking, predicting behaviour) is profiling. The escape hatch closes.
- The honest test: "does this system make or shape a judgement about a specific person?" If yes, assume high-risk and prove otherwise, not the reverse.
Vantage classifies its hiring tool. Vantage hoped its CV screener might qualify for the Article 6(3) downgrade as a "preparatory" tool. But it ranks candidates - that is profiling. No downgrade: it is high-risk under Annex III(4). The classification decided Vantage's entire compliance roadmap for the rest of the track.
Classify your real systems ★ 10 min · the checker above
Run your own AI on the classifier.
Your systems. For each in-scope AI system, pick the closest scenario in the checker and record the tier + basis.
Test a downgrade. Take a borderline Annex III system and ask honestly: is it narrow/preparatory, or does it profile a person? Decide whether the downgrade genuinely applies.
Find a prohibited one. Check whether anything you do touches an Article 5 ban (emotion recognition at work, social scoring). If so, that is a stop-now item.
Reflect. In one line: which of your systems is high-risk and least prepared? That is where sessions b2-b7 pay off first.
Before session b2 ◐ 40 min total
- Classify every in-scope AI system with the checker and record tier + basis for each.
- For any you want to downgrade under Art 6(3), write the honest profiling test result - does it judge a person?
- Flag anything touching an Article 5 prohibition as an immediate stop-and-review.
- Read: EU AI Act Article 6 and Annex III - the high-risk definition and the use-case list.
Official sources covered
Taught from the EU AI Act text. This page covers ~80% of the working content on classification.
Three questions before you go 🎯 ◐ 90 seconds
1 · You classify a system in order. What do you check FIRST (after confirming it is in scope)?
First match wins, and prohibited is checked first: if it matches Article 5 it cannot ship, and no lower tier is relevant. Then high-risk, then limited, else minimal.
2 · An Annex III system might avoid the high-risk tier via the Article 6(3) downgrade - UNLESS it...
The downgrade covers narrow/preparatory systems, but profiling of people overrides it: any system that scores, ranks, or predicts about a person is high-risk, no downgrade.
3 · An AI CV-screening tool that ranks candidates, used on EU applicants, is...
Employment/recruitment is Annex III(4). Ranking candidates is profiling, so the Article 6(3) downgrade does not apply - it is high-risk, with the full obligations.