Eu AI Act Specialist

EU AI Act Compliance Specialist

Production-ready compliance patterns for Regulation (EU) 2024/1689 -- the EU Artificial Intelligence Act. Covers risk classification, provider/deployer obligations, GPAI model requirements, conformity assessment, and AI governance.

AI System Inventory and Classification Workflow

The agent classifies AI systems under the EU AI Act's risk-based framework and maps applicable obligations.

Workflow: Classify and Map Obligations

1. Inventory all AI systems -- for each system, document: name, provider/developer, description, intended purpose, deployment status, affected persons, geographic scope, data processed, and decision impact level.
2. Apply classification decision tree to each system:
   • Does it meet the Art. 3(1) definition of an AI system? If no, document exclusion.
   • Does it fall under a prohibited practice (Art. 5)? If yes, flag as UNACCEPTABLE RISK -- must be discontinued.
   • Is it a safety component of an Annex I product? If yes, HIGH-RISK (product legislation path).
   • Does it fall under an Annex III category? If yes, apply Art. 6(3) exception analysis. If exception does not apply, HIGH-RISK.
   • Does Art. 50 transparency obligation apply? If yes, LIMITED RISK. Otherwise, MINIMAL RISK.
3. Map obligations based on classification -- assign compliance owners for each obligation.
4. Run gap analysis using scripts/ai_compliance_checker.py to identify compliance gaps.
5. Prioritize remediation by deadline urgency, penalty severity, and number of affected persons.
6. Validation checkpoint: Every AI system classified; prohibited practices flagged for immediate action; high-risk systems have assigned compliance owners and remediation timelines.

Example: AI System Classification Output

{
  "system_name": "Resume Screener v2.1",
  "provider": "Internal ML Team",
  "intended_purpose": "Screen job applications and rank candidates for recruiter review",
  "ai_act_classification": "HIGH-RISK",
  "classification_rationale": "Annex III Category 4 - Employment: AI for recruitment and screening of job applicants",
  "art_6_3_exception": false,
  "exception_rationale": "System directly influences which candidates proceed to interview stage - not a narrow procedural task",
  "applicable_obligations": [
    "Risk management system (Art. 9)",
    "Data governance (Art. 10)",
    "Technical documentation (Art. 11)",
    "Record-keeping / automatic logging (Art. 12)",
    "Transparency and information to deployers (Art. 13)",
    "Human oversight (Art. 14)",
    "Accuracy, robustness, cybersecurity (Art. 15)",
    "Quality management system (Art. 17)",
    "Conformity assessment (Art. 43)",
    "CE marking (Art. 48)",
    "EU database registration (Art. 49)",
    "Post-market monitoring (Art. 72)"
  ],
  "compliance_deadline": "2026-08-02",
  "assigned_owner": "Head of AI Governance"
}

Risk Classification System

The AI Act uses a risk-based approach with four tiers.

Tier 1: Prohibited Practices (Art. 5) -- Banned from 2 February 2025

Tier 2: High-Risk AI Systems (Art. 6, Annex III)

An AI system is high-risk if it falls under Annex III categories OR is a safety component of a product covered by Annex I harmonization legislation.

Annex III Categories:

Tier 3: Limited Risk -- Transparency Obligations (Art. 50)

Tier 4: Minimal Risk

No mandatory requirements. Voluntary codes of conduct encouraged (Art. 95).

Provider Obligations for High-Risk AI

Providers of high-risk AI systems must comply with all of the following:

Deployer Obligations (Art. 26)

General-Purpose AI Models (GPAI)

GPAI Provider Obligations (Art. 53) -- Effective 2 August 2025

Systemic Risk GPAI Models (Art. 51, 55)

Classified as systemic risk if: high impact capabilities, AI Office designation, or trained with >10^25 FLOPs (rebuttable presumption).

Additional obligations: Model evaluation with adversarial testing, red-teaming proportionate to risk, systemic risk assessment and mitigation, incident tracking and reporting, cybersecurity protection, energy consumption reporting.

Conformity Assessment Workflow

The agent guides organizations through conformity assessment for high-risk AI systems.

Workflow: Internal Control (Annex VI)

1. Establish QMS per Art. 17 -- document design, development, testing, data management, and post-market monitoring processes.
2. Compile technical documentation per Art. 11 -- system description, development process, risk management, data governance, performance metrics.
3. Implement all Chapter III Section 2 requirements -- verify each obligation is addressed.
4. Conduct internal assessment:
   • Verify risk management system addresses all identified risks (Art. 9)
   • Verify data governance meets Art. 10 requirements
   • Verify technical documentation is complete and current (Art. 11)
   • Verify logging capability (Art. 12)
   • Verify transparency and instructions for use (Art. 13)
   • Verify human oversight design (Art. 14)
   • Verify accuracy, robustness, cybersecurity (Art. 15)
   • Confirm QMS covers all required elements (Art. 17)
5. Sign EU Declaration of Conformity (Art. 47), affix CE marking (Art. 48), register in EU database (Art. 49).
6. Implement post-market monitoring (Art. 72) and maintain documentation updates.
7. Validation checkpoint: All 12 provider obligations verified; declaration signed; CE marking affixed; EU database registration complete; post-market monitoring operational.

Workflow: Third-Party Assessment (Annex VII)

Required for biometric identification systems (Annex III point 1) and cases where harmonized standards are insufficient.

1. Complete all internal control steps above.
2. Select and engage notified body with relevant AI system expertise.
3. QMS assessment -- notified body reviews and assesses QMS; issues certificate or requires corrective action; annual surveillance.
4. Technical documentation assessment -- notified body reviews documentation, tests system, issues type-examination certificate.
5. Sign EU Declaration of Conformity with notified body identification number on CE marking.
6. Maintain ongoing compliance -- notified body surveillance, notify of significant changes, maintain all documentation.
7. Validation checkpoint: Notified body certificates issued; CE marking with NB number affixed; ongoing surveillance scheduled.

Bias Detection and Fairness Testing

The agent performs bias detection per Art. 10 data governance requirements.

Workflow: Bias Testing

1. Define protected attributes -- age, gender, ethnicity, disability, religion, and other relevant characteristics for the system's context.
2. Analyze data distribution -- check representation ratios (target: 0.8-1.25 vs. population), class imbalance ratios (>0.5), and coverage of all known groups.
3. Evaluate outcome fairness using these metrics:
   • Demographic parity: P(positive outcome) equal across groups (within 80% / four-fifths rule)
   • Equalized odds: TPR and FPR equal across groups (within 80%)
   • Predictive parity: PPV equal across groups (within 80%)
   • Calibration: Predicted probabilities accurate for all groups (within 5pp)
4. Identify proxy variables -- check for features correlated with protected attributes.
5. Implement mitigation -- data augmentation, re-sampling, re-weighting, adversarial debiasing, threshold adjustment, or reject option classification as appropriate.
6. Validate -- re-run analysis to confirm improvement.
7. Document -- record all findings, measures taken, and residual bias levels in technical documentation.
8. Validation checkpoint: All protected attributes tested; fairness metrics within thresholds or residual bias documented with justification; mitigation measures recorded.

Example: Bias Detection Command

# Analyze dataset statistics for bias indicators
python scripts/ai_bias_detector.py --input dataset_stats.json \
  --protected-attributes gender,age_group,ethnicity

# Output as JSON for integration with compliance documentation
python scripts/ai_bias_detector.py --input dataset_stats.json --json

Implementation Timeline

Penalties (Art. 99)

SMEs and startups receive proportionate treatment (lower of absolute or percentage).

AI Model Documentation Templates

Template: AI System Description

AI SYSTEM DESCRIPTION
=====================
System Name:
Version:
Provider:
Date:

1. GENERAL INFORMATION
   - Intended purpose:
   - Target users (deployers):
   - Affected persons:
   - Geographic scope:
   - AI Act classification:
   - Annex III category (if applicable):

2. TECHNICAL ARCHITECTURE
   - Model type:
   - Input data modalities:
   - Output description:
   - Key design choices and rationale:

3. TRAINING AND DATA
   - Training data sources:
   - Data volume and characteristics:
   - Data preparation methods:
   - Bias examination results:

4. PERFORMANCE
   - Accuracy metrics:
   - Robustness testing results:
   - Known limitations:
   - Performance across demographic groups:

5. HUMAN OVERSIGHT
   - Oversight level: [human-in-the-loop / on-the-loop / in-command]
   - Override mechanism:
   - Automation bias safeguards:

Template: Risk Management Documentation

RISK MANAGEMENT SYSTEM -- AI SYSTEM
====================================
System Name:
Version:
Risk Management Lead:
Date:

1. RISK IDENTIFICATION
   | Risk ID | Description | Likelihood | Severity | Risk Level |
   |---------|-------------|------------|----------|------------|
   | R-001   |             |            |          |            |

2. RISK CONTROL MEASURES
   | Risk ID | Measure | Type | Verification | Status |
   |---------|---------|------|-------------|--------|
   | R-001   |         |      |             |        |

3. RESIDUAL RISK ASSESSMENT
   - Acceptability determination:
   - Overall risk-benefit analysis:

4. POST-MARKET DATA INTEGRATION
   - Review frequency:
   - Trigger conditions for update:

Tools

AI Risk Classifier

# Classify AI system from JSON description
python scripts/ai_risk_classifier.py --input system_description.json

# Classify from inline JSON
python scripts/ai_risk_classifier.py --inline '{
  "name": "Resume Screener",
  "description": "AI system that screens job applications and ranks candidates",
  "domain": "employment",
  "uses_biometrics": false,
  "decision_type": "automated_with_review",
  "affected_persons": "job applicants",
  "eu_deployment": true
}'

# JSON output for programmatic use
python scripts/ai_risk_classifier.py --input system.json --json

AI Compliance Checker

# Full compliance check with gap analysis
python scripts/ai_compliance_checker.py --input compliance_status.json

# Check deployer obligations only
python scripts/ai_compliance_checker.py --input compliance_status.json --role deployer

# JSON output with remediation steps
python scripts/ai_compliance_checker.py --input compliance_status.json --json

AI Bias Detector

# Analyze dataset for bias indicators mapped to Art. 10
python scripts/ai_bias_detector.py --input dataset_stats.json

# Specify protected attributes explicitly
python scripts/ai_bias_detector.py --input dataset_stats.json \
  --protected-attributes gender,age_group,ethnicity --json

Reference Documentation

Troubleshooting

Success Criteria

• All AI systems inventoried and classified -- every system assessed against the risk-based framework with documented classification rationale, including Art. 6(3) exception analysis where applicable
• Prohibited practices identified and discontinued -- all Art. 5 prohibited practices flagged by February 2, 2025, with documented evidence of discontinuation or lawful exception application
• High-risk systems fully compliant by August 2, 2026 -- all 12 provider obligations verified, EU Declaration of Conformity signed, CE marking affixed, and EU database registration complete
• Bias testing completed for all high-risk systems -- demographic parity, equalized odds, and predictive parity metrics within four-fifths threshold for all protected attributes, or residual bias documented with justification
• GPAI model obligations met by August 2, 2025 -- technical documentation maintained, downstream provider information supplied, copyright compliance verified, and training data summary published
• Conformity assessment completed per correct route -- internal control (Annex VI) or third-party assessment (Annex VII) selected based on system classification, with all certificates issued and filed

Scope & Limitations

In Scope:

• AI system risk classification across all four tiers (Prohibited, High-Risk, Limited Risk, Minimal Risk)
• Annex III category analysis and Art. 6(3) exception evaluation
• Provider and deployer obligation mapping with compliance gap analysis
• GPAI model classification including systemic risk determination (10^25 FLOPs threshold)
• Bias detection and fairness testing mapped to Art. 10 data governance requirements
• Conformity assessment workflow guidance (Annex VI internal control and Annex VII third-party)
• Implementation timeline tracking with penalty exposure assessment

Out of Scope:

• Actual ML model training, retraining, or adversarial testing -- this skill provides compliance frameworks, not ML engineering tools
• Notified body selection, engagement, or audit execution
• National regulatory sandbox applications or experimental AI system exemptions
• Detailed GPAI Code of Practice implementation beyond obligation mapping
• CE marking physical affixation or EU database registration system interaction
• Legal advice on liability, insurance, or contractual allocation of AI Act obligations

Important Notes:

• The EU AI Act compliance deadline of August 2, 2026 for high-risk systems is firm -- organizations should begin classification and gap analysis immediately
• Penalties are severe: up to EUR 35 million or 7% of global turnover for prohibited practices, EUR 15 million or 3% for high-risk non-compliance
• SMEs and startups receive proportionate penalty treatment (lower of absolute or percentage)

Integration Points

Tool Reference

ai_risk_classifier.py

Classifies AI systems into EU AI Act risk categories based on a JSON system description.

Input Fields: name, description, domain, sub_domain, uses_biometrics, biometric_type, biometric_context, interacts_with_persons, generates_content, content_type, decision_type, affected_persons, is_safety_component, product_legislation, eu_deployment, social_scoring, manipulates_behavior, targets_vulnerable_groups, predictive_policing_individual, untargeted_scraping, is_gpai, training_compute_flops, critical_infrastructure, infrastructure_type.

ai_compliance_checker.py

Validates AI system compliance against all provider and deployer obligations with gap analysis.

Output: Overall compliance score (0-100), per-obligation status, gap analysis with Art. references, and prioritized remediation recommendations.

ai_bias_detector.py

Analyzes dataset statistics for bias indicators mapped to Art. 10 data governance requirements.

Thresholds: Representation ratio 0.8-1.25 (within 20% of population), class imbalance >0.5, four-fifths rule (0.8) for disparate impact, proxy correlation >0.5 for proxy variable detection.

Regulation Reference: Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 Last Updated: March 2026 Version: 1.0.0