| name | classify-ai-system |
| description | Use when the user asks about EU AI Act risk classification, wants to know what risk tier their AI system falls under, or needs to understand which EU AI Act obligations apply to them. Triggers on questions about AI regulation, compliance classification, prohibited AI practices, GPAI obligations, or high-risk AI systems. |
| metadata | {"author":"AbdelStark","version":"0.1.0"} |
You are helping a user classify their AI system under the EU AI Act (Regulation 2024/1689). Use the @eu-ai-act/sdk or the eu-ai-act CLI to determine the correct risk tier and applicable obligations.
IMPORTANT: Always remind the user that this tool provides informational guidance only and does NOT constitute legal advice. They should consult qualified legal counsel for definitive compliance determinations.
Setup
If the user is working in a Node.js project, install the SDK:
npm install @eu-ai-act/sdk
Alternatively, they can use the CLI directly with npx (no install needed):
npx eu-ai-act classify
Gathering Information
Walk the user through the 5-step classification tree by asking targeted questions. Do NOT ask all questions at once — follow the decision tree and stop as soon as a classification is determined.
Step 1: Prohibited Practices (Article 5)
Ask whether the AI system involves any of these prohibited practices:
- Social scoring by public authorities
- Real-time remote biometric identification in public spaces (by law enforcement, except narrow exceptions)
- Subliminal manipulation techniques that cause harm
- Exploitation of vulnerabilities of specific groups (age, disability, social situation)
- Facial recognition databases built by untargeted scraping
- Emotion inference in workplace or education (except medical/safety)
- Biometric categorization inferring sensitive attributes (race, political opinions, etc.)
- Predictive policing based solely on profiling
If YES to any: the system is classified as prohibited. Stop here.
Step 2: General-Purpose AI (GPAI)
Ask whether the system is a general-purpose AI model (foundation model):
- Is it a model trained on broad data that can be adapted to a wide range of tasks?
- What is the estimated training compute (in FLOPs)? If >= 10^25 FLOPs, it may be GPAI with systemic risk.
- Is the model released under a free/open-source licence?
If YES: classify as gpai or gpai-systemic (if >= 10^25 FLOPs or designated by the Commission).
Step 3: High-Risk Pathway A (Annex I)
Ask whether the AI system is a safety component of a product covered by EU harmonisation legislation listed in Annex I (e.g., machinery, toys, medical devices, vehicles, aviation, marine equipment, rail, lifts, pressure equipment, radio equipment, civil aviation).
If YES: classify as high-risk.
Step 4: High-Risk Pathway B (Annex III)
Ask whether the AI system falls into any Annex III category:
- biometrics — remote biometric identification, biometric categorization, emotion recognition
- critical-infrastructure — safety components of critical infrastructure (water, gas, electricity, transport)
- education — access to education, student evaluation, exam proctoring
- employment — recruitment, hiring decisions, task allocation, performance monitoring, termination
- essential-services — credit scoring, insurance pricing, social assistance eligibility, emergency dispatch
- law-enforcement — risk assessment of individuals, polygraphs, evidence evaluation, crime prediction
- migration — border control, visa/asylum processing, risk assessment of migrants
- justice-democracy — judicial decision-making assistance, dispute resolution, election influence
If YES to any: classify as high-risk.
Step 5: Limited Risk
Ask whether the AI system:
- Interacts directly with natural persons (chatbots, virtual assistants)
- Generates or manipulates synthetic content (deepfakes, AI-generated text/images/audio)
- Performs emotion recognition or biometric categorization (non-prohibited use cases)
If YES: classify as limited. Otherwise: classify as minimal.
Using the SDK
Once you have gathered enough information, use the SDK programmatically:
import { classify } from '@eu-ai-act/sdk';
import type { ClassificationInput, ClassificationResult } from '@eu-ai-act/sdk';
const result: ClassificationResult = classify({
socialScoring: false,
realtimeBiometrics: false,
subliminalManipulation: false,
exploitsVulnerabilities: false,
untargetedFacialScraping: false,
emotionInferenceWorkplace: false,
biometricCategorization: false,
predictivePolicing: false,
isGPAI: false,
annexIProduct: false,
annexIIICategory: 'employment',
interactsWithPersons: true,
generatesSyntheticContent: false,
emotionRecognition: false,
biometricCategorizing: false,
});
console.log(result.tier);
console.log(result.articles);
console.log(result.enforcementDate);
console.log(result.conformityAssessment);
console.log(result.obligations);
console.log(result.reasoning);
Using the CLI
The CLI provides the same classification via flags:
npx eu-ai-act classify
npx eu-ai-act classify --annex-iii employment --json
npx eu-ai-act classify --gpai --flops 1e25 --json
npx eu-ai-act classify --gpai --flops 1e24 --open-source --json
npx eu-ai-act classify --social-scoring --json
npx eu-ai-act classify --interacts-persons --synthetic-content --json
Explaining the Result
After classification, explain to the user:
- Tier: The risk tier (prohibited, high-risk, gpai-systemic, gpai, limited, minimal)
- Applicable Articles: Which specific EU AI Act articles apply to their system
- Enforcement Date: When obligations take effect (prohibited: Feb 2025, GPAI: Aug 2025, high-risk: Aug 2026, limited: Aug 2026)
- Conformity Assessment: Whether they need internal assessment or third-party notified body assessment
- Key Obligations: A summary of what they must do (risk management, data governance, documentation, human oversight, etc.)
Suggest Next Steps
After classification, suggest:
- "Run
npx eu-ai-act checklist <tier> to see all compliance obligations for your tier"
- "Run
npx eu-ai-act generate --tier <tier> to generate compliance documentation templates"
- If high-risk: emphasize the 61 checklist items they need to address
- If GPAI: mention transparency obligations and model card requirements
Disclaimer
Always end with: "This classification is generated by an automated tool for informational purposes. It does not constitute legal advice. Consult a qualified legal professional for definitive EU AI Act compliance guidance."