The risk assessment assigns every AI system to a risk class. It is rule-based and uses no AI: the same answers always produce the same result, and the reasoning names the provisions it rests on. Since rule set V3.0, "high risk" only arises on the two routes Art. 6 AI Act provides: via Annex III or via Annex I.

At a glance

Included from: Starter
Where in the app: Classify & assess → Risk assessment
Output: Risk class with reasoning, applicable articles and version history
Legal basis: Art. 5, 6, 50, 53, 55 AI Act
API: /api/v1/risk-assessments

Result: four risk classes

The check runs in a fixed order. The first level that applies decides.

Class	When
PROHIBITED	At least one prohibited practice under Art. 5 applies. Overrides every other answer.
HIGH	An Annex III area is selected and the Art. 6(3) exception does not apply (Art. 6(2)), or the system is a safety component or product under Annex I requiring third-party conformity assessment (Art. 6(1)).
LIMITED	Not high risk, but a transparency obligation under Art. 50 applies. For high-risk systems the transparency obligations apply in addition (Art. 50(6)).
MINIMAL	None of the levels above applies.

The questionnaire

Yes/no questions can also be answered "unsure". "Unsure" counts as "yes". When in doubt the class is never set too low. The only exception is a8: the Art. 6(3) exception only applies with an explicit "yes".

Section	Questions	Effect
0: Prohibited practices (Art. 5)	p1–p8: manipulative techniques, exploiting vulnerabilities, social scoring, biometric categorisation by sensitive traits, emotion recognition at work and in education, untargeted scraping of facial images, crime prediction based solely on profiling, real-time remote biometric identification for law enforcement. p9: intimate deepfakes without consent. p10: child sexual abuse material.	A "yes" results in PROHIBITED (for p9/p10 see below).
A: High risk (Art. 6)	a5: Annex III area (biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, justice and democracy). a6/a7: Annex I product and third-party conformity assessment. a8: reliance on Art. 6(3). a9: profiling.	The only route to HIGH.
B: Transparency (Art. 50)	b1: interaction with people, b2: content generation (both para. 1/2, providers), b3: risk of confusion (documented only), b4: emotion recognition or biometric categorisation (para. 3, deployers), b5: deepfakes or texts on matters of public interest (para. 4, deployers).	LIMITED unless HIGH.
C: Decisions about people (GDPR)	a1: decisions about natural persons, a2: automated or materially prepared, a3: legal or similarly significant effect, a4: special categories under Art. 9 GDPR.	Does not change the risk class. Produces a GDPR hint (Art. 22, 35 GDPR) and a check hint for Annex III points 4/5.
D: Context & use	c1: supporting only, c2: internal only, c3: sensitive business area.	Documentation only, no classification.
E: GPAI models	d1: you provide a general-purpose AI model yourself, d2: systemic risk (more than 10^25 FLOPs or designated by the Commission).	Adds Art. 53 or Art. 55.

p9 and p10: mandatory from 2 December 2026

The prohibitions under Art. 5(1)(ba) and (bb) were inserted by Regulation (EU) 2026/1744 and apply from 2 December 2026. Until then p9 and p10 are optional and a "yes" appears as a hint. From 2 December 2026 both questions are mandatory, and a "yes" results in PROHIBITED. Ideally, answer them now.

The Art. 6(3) exception

An Annex III system is not high risk if it poses no significant risk to health, safety or fundamental rights. The app maps this as follows:

Only on an explicit "yes" (a8)

The exception only applies if you rely on it explicitly. "Unsure" is not enough. The reasoning then cites Art. 6(4) (documenting the assessment) and Art. 49(2) (registration).

Never with profiling (a9)

If the system profiles natural persons, it is always high risk under Art. 6(3) subpara. 3, even if a8 is answered "yes". "Unsure" on a9 counts as profiling.

Annex III only

There is no such exception for Annex I systems (Art. 6(1)).

Check hint for a supporting role

If an Annex III system is used in a supporting role only (c1) and the exception was not claimed, the app suggests checking it; with profiling it states instead that the exception is excluded.

Role: provider or deployer

The risk class alone does not say which obligations apply to you. That depends on your role. The role is derived by the legal logic (rule set V1.1): whoever develops a system or materially modifies a third-party system (for example by fine-tuning it or integrating it into their own system) and makes it available to others or operates it under their own authority is a provider (Art. 3(3) and (11) AI Act). This also applies if the system is only used in-house.

Versioning and older assessments

Every assessment stores the rule set it was created with (currently EU_AI_ACT_V3.0), along with answers, reasoning and decision logic.
Assessments made under older rule sets keep their risk class and reasoning. They are not silently recalculated but carry the notice "Rule set updated, please reassess".
When you reassess, the previous answers are pre-filled. Renamed questions from V2.3 (b4, b5, b6) are carried over to the current questions.
When use, data or system boundaries change, create a new assessment. The history stays auditable.

Effect on the compliance score

A system classified as PROHIBITED scores 0 in the compliance score, and the organisation’s overall score is capped at 49 while such a system is in use. Details on the dashboard page.

Information

The classification is rule-based support and does not replace legal advice. It is only as good as the answers. Describe purpose and context of use precisely.