Skip to main content

EU AI Act Conformity Assessment: A Step-by-Step Procedure

Operational guide to conformity assessment for high-risk AI systems. Internal evaluation, Notified Body involvement, EU Declaration of Conformity…

What do you think about this article?

EU AI Act Conformity Assessment: A Step-by-Step Procedure
EU AI ActPhotos:RDNE Stock projectPexels

At a Glance

Operational guide to conformity assessment for high-risk AI systems. Internal evaluation, Notified Body involvement, EU Declaration of Conformity…

Conformity assessment under the EU AI Act is the mandatory evaluation process for high-risk AI systems before market release in Europe. The procedure involves evaluating the AI system's technical documentation internally and optionally through an independent Notified Body. The result is an EU Declaration of Conformity and CE marking. This article walks through the concrete operational steps, the roles of stakeholders (provider, deployer, Notified Body), registration in the NEBA database, post-market monitoring, and serious incident management. Target audience: compliance officers, technical leads, legal advisors implementing conformity assessment in an organization.

What is Conformity Assessment in the EU AI Act?

Conformity assessment is the systematic evaluation process that demonstrates compliance with the requirements of Regulation (EU) 2024/1689 (EU AI Act). It is not an absolute safety certification, but an attestation of meeting mandatory requirements: transparency, documentation, risk management, human oversight, and compliance with harmonized standards where available.

Unlike traditional CE marking (machinery, medical devices), AI CE marking has a unique feature: the provider self-declares conformity without mandatory external involvement in most high-risk AI cases — but for specific high-risk systems (biometrics, critical infrastructure), a Notified Body may be engaged on a voluntary basis or required by national authorities.

The procedure varies by system type: Limited-risk AI (chatbots, image generators) follow reduced transparency requirements; high-risk AI (credit scoring, biometric surveillance) require complete internal evaluation with optional Notified Body involvement.

Which Organization Must Perform Conformity Assessment?

According to the EU AI Act, the party responsible for conformity assessment is the provider — the organization that develops, trains, fine-tunes, and distributes the AI system in Europe. If the system was already on the market before the law applied (transition period until 2025), the provider has 36 months to comply.

In practice:

  • Provider: responsible for conformity assessment, documentation, EU declaration, CE marking
  • Deployer/Importer: verifies CE marking, applies additional controls if configuring the system
  • Notified Body: independent assessment of conformity (optional for some systems, required for specific high-risk categories)

Step 1: Classify the AI System and Identify Requirements

The first step is to classify the system according to the EU AI Act. The law defines 4 risk categories:

  • Prohibited risk: systems with certain negative safety effects (manipulative biometrics, racial classification, unvalidated predictive policing). Banned from February 13, 2025.
  • High-Risk: systems impacting fundamental rights, critical safety, or resource allocation. Examples: credit scoring, biometric surveillance, critical infrastructure access control. Require complete conformity assessment.
  • Limited-Risk: chatbots, image generators, text-to-speech systems. Require transparency only (disclosure).
  • Minimal risk: spam filters, content moderators. No specific constraints.

To identify the category, use the NIST AI Risk Management Framework 1.0 or the risk assessment grid from EU AI Act Annex III (high-risk) and Annex IV (limited-risk).

Step 1 Checklist:

  • ☐ Document the intended use of the system
  • ☐ Apply the EU AI Act risk grid Annex III/IV
  • ☐ Classify into one of 4 categories
  • ☐ List all applicable mandatory requirements (transparency, documentation, monitoring)
  • ☐ Assign owner responsible for compliance

Step 2: Build the Risk Management System

The risk management system is the heart of conformity assessment. It is a cyclical process (plan, do, check, act) that identifies and mitigates system risks throughout the lifecycle (development, training, deployment, monitoring).

The reference frameworks are ISO 31000:2018 (Risk Management) and ISO/IEC 42001:2023 (AI Management System). In practice, the risk management system must:

  • Identify hazards (e.g., training data bias)
  • Assess severity + probability for each risk
  • Define mitigations (e.g., data cleaning, human review, monitoring alerts)
  • Implement and test mitigations
  • Monitor continuously during post-market phase

Concrete example: Credit scoring AI system (High-Risk)


Risk Management System - Credit Scoring AI

Identified Hazards:
  - H1: Training data distorted (missing demographics)
  - H2: Model predicts differently for minorities
  - H3: Lack of human review before loan denial
  - H4: Data leakage to third parties

Risk Assessment (Severity × Probability):
  - H1: Critical (privacy impact), High probability
  - H2: Critical (discrimination), High probability
  - H3: Critical (lack of human oversight), Medium probability
  - H4: Critical (data breach), Medium probability

Mitigation Strategy:
  - H1: Augment training data, stratified sampling, bias audit pre-deployment
  - H2: Fairness testing (demographic parity), disparate impact testing, regular audits
  - H3: Implement human-in-the-loop: loan officers review all denials >85% confidence
  - H4: Data encryption, anonymization, access logs, audit trail
  - Overall: AI 6-month monitoring, quarterly fairness reports
  

Step 3: Prepare Required Technical Documentation

Conformity assessment is based on technical documentation required by EU AI Act Article 11 and Annex IV. This documentation demonstrates that the system was developed per the requirements.

Mandatory documents for high-risk systems:

  • System Description Document: intended use, technical architecture, AI and non-AI components, operational workflow
  • Risk Assessment Report: result of risk management system with mitigation plan
  • Data Documentation (Data Sheet): training data, validation data, test data, their sources, quality, bias assessment
  • Model Card: performance metrics (accuracy, precision, recall, AUC), edge case behavior, known limitations
  • Risk Management System Document: cyclical process, implemented mitigations, responsible parties
  • Human Oversight Plan: when and how human operators intervene, required training
  • Audit Trail & Logging: traceability of all critical AI decisions, input timestamp, model version, output, human action
  • Quality Assurance & Testing Report: test coverage, edge case test cases, adversarial testing
  • Post-Market Monitoring Plan: how the system will be monitored in production, triggers for risk assessment review

Minimal System Description template:


# System Description - [System Name]

## 1. Executive Summary
- System name and version
- Intended use (domain, user, geography)
- Risk classification (high-risk category from Annex III)

## 2. Technical Architecture
- Data input sources
- ML model type (regression, classification, NLP, vision, etc.)
- Model training methodology
- Model inference pipeline
- Output specification

## 3. Compliance with Essential Requirements
- [Req 1] Transparency
- [Req 2] Data Quality & Governance
- [Req 3] Record-keeping & Audit Trails
- [Req 4] Human Oversight Mechanisms
- [Req 5] Risk Management System

## 4. Known Limitations & Edge Cases
- Model performance on minority populations
- Adversarial robustness
- Out-of-distribution inputs
- Geographic applicability limits

## 5. Monitoring & Incident Protocol
- Performance metrics tracked
- Alert thresholds
- Incident escalation process
  

Step 4: Execute Internal Audit Assessment

Internal audit is a critical review of documentation and the AI system by a qualified resource outside the development team. The goal is to verify that the documentation is complete, accurate, and the system meets EU AI Act requirements.

Internal assessment criteria:

  • Documentation completeness (all 9 documents present)
  • Technical accuracy (architecture, performance metrics, test results)
  • Consistency (documentation claims match actual implementation)
  • Legal compliance (transparency, audit trail, human oversight, risk mitigation)
  • Risk assessment: have all critical scenarios been considered?

Internal assessment checklist:


Internal Assessment Checklist

Documentation Completeness:
  ☐ System Description Document present
  ☐ Risk Assessment Report present with mitigations
  ☐ Data Sheet with bias analysis
  ☐ Model Card with performance metrics
  ☐ Risk Management System Document
  ☐ Human Oversight Plan
  ☐ Audit Trail & Logging specification
  ☐ Quality Assurance report
  ☐ Post-Market Monitoring Plan

Technical Accuracy:
  ☐ Model performance claims verified with independent test data
  ☐ Data sourcing documented
  ☐ Bias metrics (demographic parity, disparate impact) calculated
  ☐ Edge cases documented and tested
  ☐ Audit trail implementation demo-ed

Legal Compliance:
  ☐ Transparency: output explainability in scope?
  ☐ Human oversight: who decides, when, and escalation mechanism?
  ☐ Data governance: consent, retention, deletion policy?
  ☐ Post-market monitoring: metrics, frequency, incident reporting?
  ☐ Incident management: documented process for serious incidents?

Risk Assessment Validation:
  ☐ All high-risk scenarios from Annex III covered
  ☐ Mitigation strategies realistic and implementable
  ☐ Monitoring plan sufficient to detect risk realization
  ☐ Human oversight sufficient to prevent harm

Sign-Off:
  ☐ Assessor name, date, independence verified
  ☐ Findings documented (pass/conditional pass/fail)
  ☐ Remediation plan for any failures
  

Step 5: Engage Notified Body (Optional but Recommended)

For specific high-risk AI systems (biometrics, critical infrastructure), the provider may engage an independent Notified Body to verify conformity. This is optional in most cases, but recommended to mitigate legal risk and increase credibility of the declaration.

Role of Notified Body:

  • Independently verify technical documentation
  • Examine risk management system and mitigations
  • Test AI system in realistic scenarios (if relevant)
  • Issue an assessment report
  • Provide a conformity certificate (EU Type-Examination Certificate) — optional

How to find a Notified Body:

The NANDO (New Approach Notified and Designated Authorities) database by the European Commission maintains the official list of Notified Bodies. As of 2026, the number of Notified Bodies for AI is still limited (most assess safety, digital rights, cybersecurity). Verify the Notified Body is notified specifically for "AI systems" under module C or D of the EU AI Act.


# Steps to engage a Notified Body

1. Find Notified Body:
   - Visit https://ec.europa.eu/growth/tools-databases/nando/
   - Filter by "AI systems" and EU AI Act
   - Select body for your country (e.g., TÜV, SGS, DEKRA)

2. Send inquiry to Notified Body:
   - Submit System Description + Risk Assessment
   - Request quotation and timeline (typically 4-12 weeks)
   - Sign NDA if required

3. Collaborate during assessment:
   - Notified Body reviews documentation
   - Requests clarifications and additional testing
   - Conducts on-site visit (testing, team interviews)

4. Receive results:
   - Assessment Report (pass/fail/conditional)
   - EU Type-Examination Certificate (if applicable)
   - List of remediation items (if fail)

5. Post-assessment:
   - Remediate any findings
   - Update risk assessment with Notified Body feedback
   - Prepare EU Declaration of Conformity signed
  

Step 6: Compile and Sign EU Declaration of Conformity

The EU Declaration of Conformity (DoC) is the legal document where the provider declares that the AI system complies with EU AI Act requirements. It is mandatory and must be signed by the organization's legal representative.

Key elements of the DoC:

  • System identification (name, version, serial number)
  • Declaration of conformity with legal requirements
  • Reference to Notified Body (if engaged)
  • Signature date and authorized signatory
  • Compliance with harmonized standards (e.g., ISO/IEC 42001)

EU Declaration of Conformity template:


# EU DECLARATION OF CONFORMITY

Product: [System Name] v[Version]
Manufacturer: [Company Name], [Address]
Responsible person: [Name, Title]

This is to certify that the AI system described above is in conformity with:
- Regulation (EU) 2024/1689 (EU AI Act)
- Articles 15-30 (Transparency, Data Governance, Risk Management, Human Oversight)
- Annex III (High-Risk Categories): [List applicable categories]
- Aligned standards:
  * ISO/IEC 42001:2023 (AI Management System)
  * ISO 31000:2018 (Risk Management)

The AI system has undergone the following conformity assessment procedures:
1. ☐ Internal audit (date: ___)
2. ☐ Third-party assessment by Notified Body (cert number: ___)

Identified risks have been mitigated per Risk Management System Document (ref: ___)

Post-market monitoring plan is in place and will be executed as per document (ref: ___)

This declaration is issued under the sole responsibility of the manufacturer.

Date: _______________
Signature: _______________
Name and Title: _______________
  

Step 7: Apply CE Marking

Once conformity assessment is complete and the EU Declaration of Conformity is signed, the provider must apply CE marking to the system or its documentation. For software AI systems, CE marking is not physical but digital — must appear in technical documentation, packaging, and product pages.

CE marking specifications for AI:

  • CE logo (circle + "CE")
  • Notified Body number (if applicable, e.g., "CE 0123")
  • Reference to EU Declaration of Conformity (public document)
  • Application date

Where to apply CE marking for software AI systems:

  • On product documentation (datasheets, user manuals)
  • On website product page (CE logo visible)
  • On EU Declaration of Conformity document (public accessibility)
  • On compliance certificate (if issued by Notified Body)

Step 8: Register in NEBA Database (National Authorities)

For high-risk AI systems, the provider must register the system in the NEBA (National Authorities) database of the country where the system is distributed. This registration enables monitoring by national authorities and facilitates cross-border coordination on enforcement.

NEBA registration procedure:

  • Identify national competent authority (each EU member state has one)
  • Complete registration form (requires: system name, provider info, risk category, Notified Body cert if present)
  • Submit EU Declaration of Conformity and risk assessment summary
  • Receive registration number (e.g., "NEBA-IT-12345")
  • Update EU Declaration with NEBA number

Competent authorities by country (example Italy):

In Italy, the designated competent authority is the Directorate General for Digitalization, Research and Innovation (DGDRI) under the Ministry of Foreign Affairs and International Cooperation. The registration procedure will be available by 2025 through a centralized European portal.

Step 9: Implement Post-Market Monitoring

Post-market monitoring is a continuous obligation of the provider. It means tracking real-world system performance, identifying anomalies, and taking corrective action if actual risk exceeds expectations.

What to monitor:

  • Performance Metrics: accuracy, precision, recall, AUC on real data (not just test set)
  • Bias & Fairness: disparate impact across demographic groups, demographic parity, equalized odds
  • Data Drift: does real-world data distribution diverge from training data?
  • Model Drift: does performance degrade over time?
  • Incident Reports: reports of malfunction, discrimination, privacy breach
  • Human Audit: sample review of critical AI decisions to verify human oversight works

Minimalist monitoring architecture:


Post-Market Monitoring System

Components:
  1. Data Collection
     - Log all predictions (input, output, confidence, human override)
     - Collect ground truth labels (actual outcome)
     - Capture metadata (timestamp, user, region, model version)

  2. Metrics Calculation (daily/weekly)
     - Performance: accuracy, precision, recall, ROC-AUC
     - Fairness: demographic parity (diff in approval rate between groups)
     - Drift: KL divergence between live data and training data distribution

  3. Alert Thresholds
     - Accuracy drops below 92% → alert compliance team
     - Demographic disparity > 5% → alert compliance team
     - Data drift KL divergence > 0.1 → retrain model

  4. Incident Investigation
     - Any flagged incident logged with root cause analysis
     - Determine if risk was adequately mitigated
     - If not, escalate to legal/compliance

  5. Periodic Review (quarterly)
     - Formal review of monitoring data
     - Update risk assessment if needed
     - Publish monitoring report (retained for authority inspection)

  6. Model Update & Versioning
     - Version tracking for all model updates
     - Retraining justified by monitoring data
     - Updated risk assessment before deploying new version
  

Step 10: Serious Incident Management

A serious incident is an event that has caused or could cause significant harm (physical, economic, legal) to a person or society. The EU AI Act requires providers and deployers to immediately report serious incidents to national competent authorities.

Criteria for "serious incident":

  • Physical harm to a person (injury, illness)
  • Significant economic harm (>€10,000 or 1% annual revenue)
  • Massive rights violation (systematic discrimination, privacy breach affecting >1000 people)
  • Critical infrastructure compromise (energy, transport, healthcare)

Incident management process:


Serious Incident Management Protocol

Detection (when):
  - Customer report / internal monitoring alert
  - Potential serious incident identified

Immediate Response (hours):
  - Stop or reduce scope of affected system (if possible)
  - Containment: isolate affected users/data
  - Document: timeline, people affected, damage estimate
  - Root cause hypothesis: AI bug, training data bias, configuration error?

Notification (within 24-72 hours):
  - Internal escalation to compliance/legal/CEO
  - Notification to affected users (if personal data involved)
  - Notification to national competent authority (as per regulation)
  - Notification to Notified Body (if involved in original assessment)

Investigation (within 2 weeks):
  - Detailed root cause analysis
  - Determine if risk was adequately mitigated
  - Evaluate if risk assessment was insufficient
  - Collect evidence (logs, test data, audit trails)

Corrective Action (within 30-60 days):
  - Implement fix (model retrain, threshold adjustment, human review process)
  - Regression testing to verify fix doesn't create new risks
  - If needed, update risk assessment and Declaration of Conformity

Reporting:
  - Submit incident report to competent authority with corrective action plan
  - Public transparency: publish incident summary (if affects significant user population)
  - Retain documentation for 7 years (inspection by authority)

Follow-up:
  - Monitor system for recurrence
  - Verify corrective action effectiveness
  - Close incident after 90-day stability
  

Complete Operational Checklist for Conformity Assessment

Below is the unified checklist covering all 10 steps, usable as an internal compliance project:


# Conformity Assessment Project Checklist

## Phase 1: Risk Classification & Planning
- ☐ System risk classification completed (high-risk? limited-risk? prohibited?)
- ☐ Applicable requirements identified (transparency, data governance, human oversight, risk management)
- ☐ Compliance officer assigned
- ☐ Budget & timeline approved
- ☐ Steering committee established (dev, legal, compliance)

## Phase 2: Risk Management System & Documentation
- ☐ Risk assessment performed (NIST AI RMF or equivalent)
- ☐ 9 required documents prepared:
  - ☐ System Description Document
  - ☐ Risk Assessment Report
  - ☐ Data Sheet (bias analysis, data quality)
  - ☐ Model Card (performance metrics)
  - ☐ Risk Management System Document
  - ☐ Human Oversight Plan
  - ☐ Audit Trail & Logging specification
  - ☐ Quality Assurance & Testing Report
  - ☐ Post-Market Monitoring Plan
- ☐ All documentation reviewed by legal counsel
- ☐ Documentation translated to English (for international contexts)

## Phase 3: Internal Audit
- ☐ Independent assessor (not development team) assigned
- ☐ Internal assessment checklist completed
- ☐ All findings documented (pass/conditional pass/fail)
- ☐ Remediation plan for any failures
- ☐ Final sign-off from assessor and management

## Phase 4: Third-Party Assessment (Optional but Recommended)
- ☐ Notified Body identified (NANDO database)
- ☐ Quotation & timeline received
- ☐ NDA signed
- ☐ Notified Body assessment completed
- ☐ EU Type-Examination Certificate received (if applicable)
- ☐ Assessment findings remediated

## Phase 5: Formal Compliance Declaration
- ☐ EU Declaration of Conformity drafted
- ☐ Legal review of Declaration
- ☐ Authorized signatory identified and trained
- ☐ Declaration signed and dated
- ☐ Digital signature/stamp applied (if required)

## Phase 6: Marking & Registration
- ☐ CE marking applied to product documentation & website
- ☐ System registered in NEBA (national database)
- ☐ NEBA registration number recorded in Declaration
- ☐ Declaration of Conformity published (public accessibility)

## Phase 7: Post-Market Monitoring Setup
- ☐ Monitoring infrastructure deployed (logging, metrics collection)
- ☐ Monitoring dashboard created (daily performance overview)
- ☐ Alert thresholds configured
- ☐ Incident investigation procedure documented
- ☐ Monitoring reports generated and retained

## Phase 8: Governance & Audit Trail
- ☐ All compliance documentation stored in secure, version-controlled system
- ☐ Access logs maintained for all compliance files
- ☐ 7-year retention policy implemented
- ☐ Annual compliance review scheduled
- ☐ Authority inspection readiness verified

## Ongoing Obligations (Every Release)
- ☐ Risk assessment updated for any major model/data changes
- ☐ Monitoring data reviewed (performance, bias, drift)
- ☐ Incident management protocol triggered if serious incident detected
- ☐ Declaration of Conformity updated if risk profile changes
- ☐ Notified Body re-assessment requested (if material changes)
  

Frequently Asked Questions (FAQ)

Q: If our AI system is developed by a vendor and integrated internally, who is responsible for conformity assessment?

A: It depends on the role. If the vendor is the original producer (develops model, training data, deployment architecture), the vendor is the provider and responsible for conformity assessment. The company integrating the system is the deployer and responsible for verifying the vendor provided conformity documentation and EU declaration. If the company materially modifies the system (retrain with own data, adjust output), it becomes a co-provider with shared responsibility for conformity.

Q: How long does the conformity assessment process take for a high-risk system?

A: Typically 3-6 months for internal evaluation (depends on system complexity and existing documentation quality). If engaging a Notified Body, add 4-12 additional weeks. Priority should be parallelizing tasks (prepare documentation while doing internal audit).

Q: If our system uses only hardcoded rules and no machine learning, is it subject to the EU AI Act?

A: It is not an AI system per the EU AI Act definition (which requires autonomous learning capability or pattern recognition). However, if the system uses NLP, computer vision, or other AI techniques, it may be subject even without ML. Consult Annex I of the regulation for precise definition of "AI system".

Q: What are the penalties for non-compliance with the EU AI Act?

A: Administrative fines of €10-30 million or up to 4% of global annual revenue (tier 1). Possible ban from European market. Civil liability for damages caused by non-compliance. Confiscation of profits derived from the non-compliant system.

Q: How do we ensure our risk management system is "sufficient" per the law?

A: There is no numerical definition of "sufficient." The law references harmonized standards (ISO/IEC 42001, ISO 31000). Best practice is to document your risk identification process (e.g., NIST AI RMF), demonstrate each identified risk has proportionate mitigation, and that post-market monitoring can detect risk realization. A Notified Body or external audit can validate sufficiency.

Q: If we retrain our AI system with new data, must we repeat conformity assessment?

A: It depends on the scope of retraining. If it is minor fine-tuning (parameter updates on 5-10% new data), a complete reassessment is not necessary — only update the monitoring plan and re-evaluate the risk assessment. If it is substantial retraining (new architecture, +50% new data, significantly different performance), internal re-evaluation or Notified Body re-assessment is advised. Rule of thumb: if the risk profile changes, update the conformity assessment.

Conclusion

Conformity assessment under the EU AI Act is a rigorous but achievable process. The 10 steps described (risk classification, risk management, documentation, internal audit, Notified Body engagement, conformity declaration, CE marking, NEBA registration, post-market monitoring, incident management) provide a clear path to compliance. The key is investing in accurate documentation and robust governance from early development stages — retroactive compliance is ineffective and costly. For organizations with high-risk systems, third-party external audit (Notified Body) increases credibility and mitigates legal risk. Conformity assessment is not a sprint but a structural commitment by the organization toward transparency and accountability in AI deployment in Europe.

Commenti

Caricamento commenti...

Sign in to comment

Discuss this article in the community

Join the Telegram group to discuss with other developers, ask questions and share your experiences.

Did you like this article?

Explore other content on the blog or discover my projects