How does Maditon classify an AI system under the EU AI Act?

The flow is: register the system, run the interview, receive a draft, accept or override it.

The interview asks about purpose, users, the data involved, the sector, and the effect on people. It is deliberately free of regulatory vocabulary — you describe your product, not your legal position.

The classification checks the answers against the whole risk structure of the regulation, not just the tier it expects: the Article 5 prohibitions, the Article 6 and Annex III high-risk routes, the Article 50 transparency duties, and the residual minimal-risk case. It returns a tier, the reasoning, and citations that quote the passages the tier rests on.

The verification pass then re-checks each conclusion against the retrieved legal text and reports any claim it could not support, with a confidence score. Low-confidence results are raised for expert review instead of being presented as settled.

Acceptance is a separate, human step. Until someone takes it, the result is a draft.

Classification runs asynchronously: start it, then poll the job. A free Solo account runs three of these a month.