Glass Box — How we classify AI tools
Every classification in our index is deterministic. No LLM in the classification path. This page shows the full decision tree, live test cases from the index, and the classifier version so you can audit or reproduce any result.
100
tools classified
16
high risk
34
limited risk
50
minimal risk
Classifier decision tree
Applied in order, top to bottom. The first matching condition determines the risk tier. Based on EU AI Act (2024/1689) as amended by the Digital Omnibus Regulation (EU) 2026/1744 (in force 27 July 2026).
Does the system match any Article 5 prohibited practice?
Article 5
Does the use case match any Annex III category (as amended by Digital Omnibus)?
Article 6 + Annex III
Does the system generate synthetic content, run as a chatbot, or perform emotion recognition?
Article 50
No mandatory EU AI Act obligations apply for the typical deployment pattern.
Recital 47
Score modifiers within each tier
| Factor | Weight | What it measures |
|---|---|---|
| Decision impact on individuals | +0 to +2 | How consequential the AI output is for real people. Autonomous employment or credit decisions score +2. |
| Provider EU compliance posture | +0 to +1 | Whether the vendor has published EU AI Act documentation. Absence of provider docs adds 1 point within the tier. |
| Deployment autonomy | +0 to +1 | Fully autonomous outputs (no human review required) score +1 vs decision-support tools. |
Final score = base tier floor + sum of modifiers, capped at the tier ceiling. Full factor definitions: /methodology.
Live test cases from the index
These are real entries from the 100-tool index, pulled live. They serve as regression anchors — any future classifier update must produce the same output for all of these cases.
verified
2026-06-12
Generative AI Assistant
Article 50 — chatbot transparency (if customer-facing)
verified
2026-06-12
verified
2026-06-14
Financial AI
Annex III, Category 5 — AI for credit scoring and creditworthiness assessment
verified
2026-06-14
Synthetic Media AI
Article 50(4) — synthetic audio-visual content depicting persons (deepfake labeling)
verified
2026-06-14
verified
2026-06-14
All 100 entries at /vendors. Machine-readable via POST /api/classify.
Download the full methodology pack
The full pack includes: complete Article-by-Article obligation matrix for deployers, all 100 test cases with score derivations, Annex III category mapping table, and Digital Omnibus deferral analysis. PDF, CC BY licensed.
No spam. One email with the pack. Unsubscribe any time.