EU AI Act Art 6: What It Requires and Who It Applies To

What is EU AI Act Art 6?

Article 6 of Regulation (EU) 2024/1689, known as the EU Artificial Intelligence Act, is the classification gateway that determines whether an AI system must comply with the regulation's full high-risk requirements. The European Parliament and Council adopted the regulation in June 2024, with the final text published in the Official Journal of the EU on July 12, 2024. It entered into force on August 1, 2024.

The classification runs on two tracks. Track one, Article 6(1), applies when an AI system is a safety component embedded in a product covered by existing EU harmonization legislation listed in Annex I (medical devices, machinery, aviation equipment, railway systems) or is that product itself, and the product requires third-party conformity assessment. Track two, Article 6(2), applies when an AI system performs a function listed in Annex III regardless of the product context. Annex III is where most of the financial services exposure sits.

The regulation was a direct response to AI deployment in consequential decisions affecting people's access to credit, employment, and essential services with no documented methodology, no bias testing, and no obligation to explain outcomes. The European Commission's impact assessment identified unchecked AI deployment across credit, insurance, and employment as a systemic risk affecting hundreds of millions of EU citizens.

Timeline matters. Prohibited AI systems became unlawful February 2, 2025. General-purpose AI model obligations applied from August 2, 2025. The high-risk requirements triggered by Article 6 apply from August 2, 2026. Annex I product-embedded systems get until August 2, 2027. That window looks long. It isn't, once you factor in the internal change management required to build compliant risk management systems, data governance processes, and audit-ready technical documentation from scratch.

Who does EU AI Act Art 6 apply to?

The regulation distinguishes between providers (organizations that develop or place AI systems on the market) and deployers (organizations that use AI systems in a professional context). A bank that licenses a credit scoring model from a fintech is a deployer. The fintech is the provider. Both carry compliance obligations, with providers bearing the heavier burden for documentation and conformity assessments.

Entities in scope for financial services include:

• Retail and commercial banks using AI to evaluate the creditworthiness of natural persons or to establish credit scores, including models embedded in loan origination and card underwriting platforms (Annex III, point 5b)
• Insurance firms using AI for risk assessment, pricing decisions, or eligibility determinations on life and health insurance products (Annex III, point 5b)
• Banks, fintechs, and investment firms using AI in HR and employment decisions affecting workers, including AI-assisted CV screening, performance monitoring, or promotion recommendations (Annex III, point 4)
• Banks and payment institutions using AI for real-time biometric identification of account holders (Annex III, point 1; note that real-time remote biometric surveillance in public spaces faces outright prohibition under Article 5)
• AI model vendors and fintechs that supply high-risk AI systems to EU-based financial institutions, regardless of where the vendor is located

There's no minimum size threshold. A regional building society with £800 million in assets using an AI mortgage affordability model has the same Article 6 obligations as a large universal bank.

Geographic reach is broader than most compliance teams expect. The AI Act applies when an AI system's output affects EU residents. A US bank using AI to evaluate credit applications from EU consumers through a mobile app is in scope. So is a Singapore-based payment platform using AI to verify EU customers. The question is where the person affected is located, not where the AI system is operated.

Deployers have lighter obligations than providers, but don't mistake "lighter" for "manageable without preparation." Fundamental rights impact assessments, human oversight systems, and usage logs aren't trivial requirements. That's not light.

What does EU AI Act Art 6 require?

Article 6 is the classification rule. The actual obligations sit in Articles 8 through 15. Here's what those require in practice:

1. Continuous risk management system (Article 9). Not a one-time assessment. A documented, iterative process that identifies, analyzes, and mitigates risks throughout the system's lifecycle. The system must be updated when the AI is modified, when new use cases emerge, or when post-deployment monitoring reveals problems. "We did a risk assessment at deployment" doesn't satisfy this.

2. Data governance and management practices (Article 10). Training, validation, and test datasets must be subject to documented governance practices covering relevance, representativeness, freedom from material error, and completeness. Data bias must be identified and corrected before deployment. The EBA's June 2024 discussion paper on AI in banking flagged credit scoring model bias as a supervisory priority for the first compliance cycle.

3. Technical documentation before deployment (Article 11 and Annex IV). Providers must produce detailed documentation before placing a system on the market. Annex IV specifies thirteen content categories: general description, design choices and logic, development process, training data characteristics, performance metrics and benchmarks, known limitations, human oversight measures, cybersecurity controls, and more. This documentation must be kept current as the system changes.

4. Automatic event logging (Article 12). High-risk AI systems must automatically generate event logs that capture sufficient data to trace decisions back to the deployment context. Deployers must retain these logs for at least 6 months, or longer when sector-specific rules require it. Banking prudential rules typically require longer retention periods for credit decision records.

5. Transparency and instructions for use (Article 13). Natural persons subject to AI-driven decisions must be told they're interacting with an AI system. Deployers must receive clear instructions of use from providers, covering the system's intended purpose, performance boundaries, and known failure modes.

6. Human oversight by design (Article 14). Systems must be built so that humans can monitor outputs in real time, understand the logic behind decisions, override or correct the system, and halt it entirely. The kill switch requirement is real and auditable. Documenting the existence of a kill switch isn't enough; regulators will ask whether it's been tested.

7. Accuracy, robustness, and cybersecurity (Article 15). Providers must define and document the accuracy levels their systems achieve and maintain them post-deployment. Systems must remain resilient to adversarial inputs. Cybersecurity controls must match the risk profile.

8. Conformity assessment (Article 43). For most Annex III use cases, including credit scoring, this is a self-assessment by the provider. For remote biometric identification systems, an accredited notified body must conduct it. Either way, the assessment must precede market placement.

9. EU database registration (Article 49). Providers must register high-risk AI systems in the EU AI Act database maintained by the European AI Office before deployment. Some deployers using Annex III systems in public interest contexts must also register.

What evidence do regulators expect?

National market surveillance authorities and sectoral supervisors (the EBA, ECB, and national prudential regulators) can request documentation at any time. Here's what an Article 6 audit for a bank typically covers:

Technical documentation package:

• Annex IV documentation: system purpose, description, design choices, training data sources and governance, performance benchmarks, and known limitations
• Risk management records demonstrating the continuous process required by Article 9, including records of identified risks and mitigation decisions taken over time
• Data bias testing reports for training and validation datasets
• Conformity assessment records (self-assessment report or notified body certificate, depending on the system type)

Operational records:

• Automatic event logs for the deployment period, retained for at least 6 months; banking prudential requirements typically extend this to several years for credit decisions
• Evidence that human oversight procedures are operational, not just written, including records of actual interventions, overrides, and escalations
• Post-deployment performance monitoring reports confirming accuracy has remained within declared bounds

Governance evidence:

• Internal AI governance policy covering approval workflows, periodic review cycles, and escalation procedures for unexpected outputs
• Training records for staff responsible for supervising AI system outputs
• Fundamental rights impact assessments for systems affecting EU residents' access to financial products or services

Vendor and third-party management:

• Contracts with AI providers that allocate Article 6 obligations and give deployers access to Annex IV technical documentation
• Due diligence records on third-party AI systems integrated into credit underwriting or fraud detection processes

The European AI Office has indicated that supervisory attention in the first compliance cycle will focus on financial services, healthcare, and employment, precisely because these are the sectors where Annex III misclassification risk is highest.

Common failure modes

Most Article 6 failures won't come from institutions that ignored the law. They'll come from institutions that thought they'd done enough.

• Misclassifying systems as low-risk. Treating a credit scoring model that drives 95% of loan decisions as a low-risk tool because "a human approves the final decision." If the human override rate is effectively zero, that framing won't survive a supervisory review. The CJEU's December 2023 ruling in SCHUFA (Case C-634/21) already found that automated credit scoring constitutes a form of automated individual decision-making under GDPR Article 22. Regulators will apply similar scrutiny under the AI Act.

• Deploying without Annex IV documentation. Buying a scoring model from a fintech vendor without requiring Article 11-compliant technical documentation. Deployers are responsible for ensuring this exists. "The vendor didn't provide it" doesn't hold up.

• Paper human oversight. Writing kill switch procedures into policy without testing whether they actually work. Examiners will ask for records of AI decision overrides. A zero-override rate over 12 months is a red flag, not a sign of smooth operations.

• Log retention gaps. Retaining decision logs for 90 days when prudential rules demand more. Six months is the AI Act's floor. Banking supervisors typically require 5 to 7 years for credit decision records.

• Vendor contracts without AI Act provisions. Third-party AI agreements drafted before 2024 routinely don't allocate Article 6 obligations. At audit time, this creates a dispute over who owns the conformity assessment. That dispute typically goes badly for the deployer.

• Late EU database registration. Registration in the EU AI Act database is a pre-deployment requirement, not a post-deployment administrative step. Deploying first and registering later is non-compliant from day one. Many providers are still unaware of this.

Penalties for non-compliance

The AI Act establishes a three-tier fine structure. National market surveillance authorities enforce Article 6 violations; the European AI Office holds oversight authority.

Tier 1, prohibited AI violations (Article 5): Up to €35 million or 7% of total worldwide annual turnover, whichever is higher. This covers systems that cross into prohibited territory: real-time remote biometric surveillance in public spaces, systems that manipulate behavior through subliminal techniques, or social scoring by public authorities.

Tier 2, high-risk AI non-compliance (Articles 6 through 51): Up to €15 million or 3% of total worldwide annual turnover. This is the tier that catches banks failing to maintain Annex IV documentation, failing to register in the EU database, or deploying without a completed conformity assessment. For a bank with €10 billion in annual revenue, the 3% ceiling is a €300 million exposure.

Tier 3, incorrect information to authorities: Up to €7.5 million or 1.5% of worldwide annual turnover for providing incorrect, incomplete, or misleading information to national authorities or notified bodies.

For SMEs, the lower of the two figures applies. The regulation states this explicitly.

Banks face additional consequences beyond AI Act fines. The EBA and national prudential supervisors can treat AI governance failures as operational risk management deficiencies under CRD VI, leading to capital add-ons. GDPR fines run in parallel if the system processes personal data without adequate safeguards. These aren't alternative penalties; they stack.

There are no formal Article 6 enforcement cases yet (most requirements don't apply until August 2026), but the European Commission's GDPR enforcement track record is a reliable indicator. The Irish Data Protection Commission's €1.2 billion fine against Meta in May 2023 for systematic data transfer violations shows EU regulators will issue headline-level fines for non-compliance at scale.

Related regulations and frameworks

Article 6 doesn't operate in isolation. Banks face overlapping obligations from multiple directions.

GDPR (Regulation (EU) 2016/679). GDPR Article 22 already restricts fully automated individual decisions with legal or similarly significant effects. The AI Act doesn't replace this; both apply simultaneously. Banks using AI for credit decisions must satisfy GDPR Article 22's legal basis requirements and the full Article 6 high-risk regime at the same time. Recital 59 of the AI Act confirms this.

DORA (Regulation (EU) 2022/2554). The Digital Operational Resilience Act's ICT risk management requirements, which applied from January 2025, overlap with Article 15's cybersecurity and robustness obligations. Banks must map AI systems against both frameworks to avoid duplicated but divergent controls.

SR 11-7 (US Federal Reserve / OCC). For banks with US operations, SR 11-7 model risk management guidance has required model inventory, validation, and documentation since 2011. The frameworks aren't identical, but a bank with mature SR 11-7 compliance has a real head start on Annex IV documentation. The logic is the same; the specifics differ.

FATF Recommendation 15 on new technologies. FATF requires member jurisdictions to identify and assess ML/TF risks from new technologies in financial services. For banks using AI in transaction monitoring, FATF R15 creates a risk assessment obligation that maps directly onto Article 9's continuous risk management requirement.

UK FCA AI governance. UK banks aren't subject to the EU AI Act post-Brexit, but FCA supervisory expectations for AI are converging with the EU approach. Banks with EU subsidiary operations must comply with Article 6 for those entities regardless of their UK regulatory status. Running two governance frameworks in parallel is the operational reality for most large UK-headquartered banks.

EU AML framework. The new EU Anti-Money Laundering Regulation (2024/1624) and the upcoming AMLA supervisory authority will expect banks to document the methodology behind AI systems used in transaction monitoring and customer risk scoring. That documentation obligation runs parallel to Annex IV, and the same documentation can typically serve both.

How FluxForce supports EU AI Act Art 6 compliance

FluxForce AI agents generate full decision audit trails for every risk signal processed, satisfying Article 12's automatic logging requirement. For credit risk and fraud detection deployments, agents surface the contributing factors behind each decision in plain language, supporting Article 13's transparency obligations and making human oversight (Article 14) operationally viable rather than theoretical. FluxForce's regulatory compliance automation platform includes configurable kill switches and real-time monitoring dashboards built around the Article 14 human oversight specification. Request a demo to see how the platform maps to your specific Article 6 obligations.