The EU AI Act is no longer a horizon risk. For financial services firms with EU customers, EU-based operations, or EU-market products, the enforcement clock is running. High-risk AI system obligations under Annex III — which directly captures credit scoring, insurance risk assessment, employment screening, and essential services access — became fully enforceable on August 2, 2026. The conformity assessment process, technical documentation requirements, and ongoing monitoring obligations are not aspirational guidance. They are legal prerequisites for continued deployment.
This guide covers what financial services compliance teams and AI governance officers need to know to assess, document, and enforce EU AI Act obligations on their existing and planned AI deployments. It covers the prohibited practice categories, the high-risk classification framework, the specific technical obligations that Annex III classification triggers, and how purpose-built enforcement infrastructure maps to the Act's Articles 9 and 13 requirements.
The EU AI Act Structure: A Compliance Map
The EU AI Act organizes AI systems into four risk tiers, each with different compliance obligations. Understanding where your deployments land determines everything else in your compliance program.
Unacceptable risk (Article 5 — Prohibited Practices). These AI applications are banned outright in the EU market. Financial services firms need to ensure no deployment falls into these categories. The list includes social scoring systems that evaluate individuals based on social behavior to determine access to financial services — a direct concern for any institution considering broad behavioral scoring. It also includes real-time biometric identification in public spaces (not a typical FS use case), subliminal manipulation techniques, and exploitation of vulnerabilities of specific groups.
High risk (Annex III). The tier that most directly affects financial services. Systems listed in Annex III are high-risk by definition and trigger the full compliance apparatus under Articles 8-15. Financial services-relevant categories include: credit scoring and creditworthiness evaluation (point 5b), access to essential private services including insurance and banking (point 5b), employment and worker management AI (point 4), and AI-assisted access to public services (point 5a).
Limited risk. Systems with transparency obligations but no conformity assessment requirements — primarily chatbots and AI-generated content, where disclosure to users is the core obligation under Article 52.
Minimal risk. No specific obligations under the Act, though good governance practices apply.
Many financial institutions underestimate their high-risk AI footprint. Any LLM that informs a credit decision — including systems that summarize applicant information for underwriters, generate risk narratives, classify loan applications, or produce adverse action code recommendations — is likely high-risk under Annex III point 5(b). The test is not whether the AI makes the final decision; it is whether the AI output influences access to financial products or services for natural persons.
What Counts as High-Risk for Financial Services
The Annex III classification is broader than most legal teams initially assess. The critical criterion is whether the system's output is used to evaluate natural persons in connection with access to financial products, credit, insurance, or essential services. The AI does not need to be the sole decision-maker — systems that produce intermediate outputs used by humans to make those decisions are captured.
The following use cases are high-risk under Annex III for virtually all financial services deployments:
- Credit scoring and creditworthiness systems — Any model that produces a score, probability, classification, or narrative assessment used in evaluating whether to extend credit to a natural person. This includes LLM-based underwriting assistants that generate risk summaries, automated pre-qualification systems, and real-time credit limit adjustment models.
- Insurance risk assessment — Systems that evaluate risk profiles for natural persons in connection with life, health, property, or liability insurance underwriting. This explicitly includes AI systems that assist underwriters with risk classification or premium determination.
- Employment and HR AI in financial institutions — AI used in recruitment, CV screening, promotion assessment, or performance evaluation for financial services employees is high-risk under Annex III point 4. This has implications for AI-assisted hiring tools used within regulated firms.
- Access to essential services — Annex III point 5(b) covers AI that determines access to banking accounts, payment services, and other essential financial services for natural persons. Account opening AI, fraud-triggered account restriction systems, and KYC/AML classification systems that determine service access are in scope.
The Act uses "intended purpose" as the primary classification trigger (Article 3, point 12). An AI system is high-risk if it is intended to be used for a high-risk application, regardless of whether it is always used that way. A general-purpose LLM deployed by a financial institution specifically to assist with credit analysis is high-risk for that deployment context — even if the same model is used for low-risk summarization elsewhere. Each deployment context must be assessed independently.
The Article 5 Prohibited Practices: Financial Services Exposure
Before addressing the high-risk compliance program, financial services firms must eliminate any Article 5 exposures. The Act prohibits specific AI practices regardless of business justification, and fines for violations reach €35 million or 7% of global annual turnover.
The most relevant Article 5 prohibitions for financial services are:
Social scoring (Article 5(1)(c)). The Act prohibits AI systems that evaluate or classify natural persons based on social behavior over time and that produce a social score used to determine access to financial services. This is a direct prohibition on broad behavioral scoring systems that aggregate social media behavior, spending patterns across unrelated categories, or social network analysis to produce credit or insurance access decisions. Traditional credit scoring based on financial behavior remains permissible; broader social scoring is not.
Exploitation of vulnerabilities (Article 5(1)(b)). AI systems that exploit specific vulnerabilities of groups of persons — including age, disability, or social or economic situation — to materially distort their behavior in a way that harms them are prohibited. This has implications for financial institutions that use AI in customer communications, particularly in collections, retention, or product upselling contexts targeting vulnerable populations.
Manipulation techniques (Article 5(1)(a)). AI using subliminal techniques beyond governance that materially distort behavior to cause harm is prohibited. For customer-facing AI in financial services, this requires careful review of any system designed to influence financial product adoption or retention decisions through methods that bypass conscious deliberation.
Article 9: The Risk Management System Requirement
For high-risk AI systems, Article 9 is the operational core of the EU AI Act. It requires providers and deployers to establish, implement, document, and maintain a risk management system throughout the AI system's lifecycle. This is not a one-time assessment — it is an ongoing operational program.
What Article 9 Requires
Article 9 specifies four components of the risk management system:
Risk identification and analysis (Article 9(2)(a) and (b)). The system must identify and analyze known and reasonably foreseeable risks of the high-risk AI system to health, safety, or fundamental rights. For financial services AI, this includes: discriminatory output risks (disparate impact on protected classes), accuracy degradation risks in out-of-distribution inputs, risks from adversarial manipulation of inputs, data quality risks, and risks arising from the interaction between the AI and human decision-makers.
Risk estimation and evaluation (Article 9(2)(c)). The system must estimate and evaluate risks that may emerge when the AI system is used in accordance with its intended purpose and under reasonably foreseeable conditions of misuse. For LLM-based financial AI, this requires explicit analysis of prompt injection risks, the impact of model updates on decision behavior, and edge case behavior under extreme market conditions or unusual applicant profiles.
Risk management measures (Article 9(2)(d)). Appropriate risk management measures must be adopted in response to identified risks. The Act requires that residual risk be evaluated after measures are applied, and that risk management measures include testing procedures.
Testing (Article 9(6)). High-risk AI systems must be tested to identify appropriate risk management measures. Testing must be prior to placing on the market, conducted against defined metrics, and performed on a population representative of the intended deployment context.
Article 9(5) requires that risk management measures take into account the effects the high-risk AI system may have on persons or groups of persons, and that the measures are appropriate to the specific circumstances of deployment. Critically, risk management must address the interaction between the AI and the human oversight mechanisms — meaning the risk assessment must evaluate not just what the AI does in isolation, but how human decision-makers actually interact with its outputs in practice. Systems designed for "human in the loop" oversight must demonstrate that the oversight mechanism is genuinely effective, not nominal.
Documentation Requirements Under Article 9
The risk management system documentation must be maintained and updated throughout the lifecycle. This means financial services firms need living documentation processes, not a one-time assessment delivered for deployment approval. The documentation must cover:
- The risk identification and analysis methodology used
- The identified risks and their assessed severity and likelihood
- The risk management measures adopted and evidence of their effectiveness
- Test procedures, test metrics, and test results
- Residual risk assessment after mitigation measures
- Monitoring procedures for post-deployment risk tracking
Article 13: Transparency Obligations
Article 13 requires that high-risk AI systems be designed and developed in a way that ensures their operation is sufficiently transparent to enable deployers to interpret the system's output and use it appropriately. This is not merely a documentation requirement — it is a technical design requirement.
Article 13(3) specifies what the instructions for use must include. For financial services deployers, the critical elements are:
- The intended purpose of the AI system and the specific circumstances in which it is intended to be used
- The level of accuracy, robustness, and cybersecurity of the system against which it has been tested and validated
- Any known or foreseeable circumstances related to the use of the high-risk AI system that may lead to risks to health, safety, or fundamental rights
- The human oversight measures under Article 14, including the technical measures to facilitate interpretation of the outputs by deployers
- Computational and hardware resources needed, expected lifetime of the AI system, and maintenance and care measures for proper functioning
For LLM-based financial AI, the transparency obligation has a specific implication: deployers must be able to understand the basis of the system's output well enough to use it appropriately. Systems that produce opaque outputs without explanation metadata, confidence indicators, or reasoning traces may not satisfy Article 13 for high-stakes financial decisions.
Technical Documentation: Annex IV
Annex IV specifies 15 categories of technical documentation that must be prepared and maintained for high-risk AI systems. This documentation must be available to national market surveillance authorities on request and forms the basis for the conformity assessment process. For financial services AI, maintaining Annex IV documentation requires collaboration across legal, technology, data science, and compliance functions.
| Annex IV Category | Financial Services Implication | Documentation Owner |
|---|---|---|
| General description of AI system and its intended purpose | Must specify exact use case — "credit risk" is insufficient; must describe the decision workflow | Business + Legal |
| Detailed description of elements and development process | Architecture, training data sources, training methodology, validation approach | Data Science + Technology |
| Monitoring, functioning, and control of AI system | Ongoing performance metrics, drift detection thresholds, escalation procedures | Model Risk + Technology |
| Risk management system (Article 9) | Full Article 9 documentation as described above | Compliance + Model Risk |
| Data governance and management practices | Training data provenance, bias analysis, data quality documentation | Data Governance |
| Detailed description of the accuracy, robustness and cybersecurity measures | Performance benchmarks, adversarial robustness testing, security controls | Technology + Security |
| Testing procedures and evaluation metrics | Pre-deployment testing protocols, bias testing results, fairness metric documentation | Model Risk + Compliance |
Conformity Assessment for Financial Services AI
For most Annex III high-risk AI systems — including credit scoring, insurance risk assessment, and access-to-services AI — the EU AI Act uses a self-assessment conformity process rather than mandatory third-party notified body assessment. This may sound less burdensome, but the self-assessment requirement is substantive and generates significant liability if done inadequately.
The conformity assessment process requires the provider to:
- Verify that the AI system complies with all requirements in Articles 8-15
- Prepare the technical documentation specified in Annex IV
- Implement a quality management system meeting Article 17 requirements
- Register the high-risk AI system in the EU database established under Article 71
- Draw up an EU Declaration of Conformity
- Affix the CE marking (where applicable) or include it in the Declaration of Conformity
The EU database registration requirement (Article 71) is frequently overlooked. High-risk AI systems in the financial services categories must be registered before deployment in the EU market. The registration creates a public record of the system, its intended purpose, and the provider's contact information. National competent authorities have access to the full technical documentation on request.
The EU AI Act distinguishes between providers (those who develop and place high-risk AI systems on the market) and deployers (those who use those systems in their own context for a natural person). Financial services firms are often both: they are deployers of AI provided by technology vendors, and providers of their own custom-built AI systems. Where a financial institution takes a foundation model or third-party AI product and customizes it for their specific high-risk use case, they may assume provider obligations under Article 25. This determination requires careful legal analysis of the customization depth and the resulting change in the AI system's risk profile.
Article 14: Human Oversight Requirements
Article 14 requires that high-risk AI systems be designed and developed in a way that enables effective human oversight when they are in use. This is an affirmative design requirement, not just a procedural policy requirement. The system must be technically designed to facilitate human oversight — it cannot merely be a policy that humans review outputs.
Article 14(4) specifies three dimensions of the human oversight requirement:
Ability to understand (Article 14(4)(a)). Persons responsible for oversight must be able to understand the AI system's capabilities and limitations — specifically its performance characteristics and failure modes.
Ability to recognize anomalies (Article 14(4)(b)). Oversight persons must be able to detect anomalies, dysfunctions, and unexpected outputs. This implies that the AI system must generate output in a form that makes anomalous behavior detectable — not merely that oversight personnel are trained to look for problems.
Ability to override (Article 14(4)(c)–(e)). Human oversight persons must be able to disregard, override, or reverse the output of the AI system in accordance with the intended purpose. They must also be able to intervene or interrupt the system through a "stop" button or similar procedure.
How CoreGuard Maps to EU AI Act Obligations
Purpose-built AI enforcement infrastructure like CoreGuard addresses several of the EU AI Act's most technically demanding requirements — not by replacing the compliance program, but by providing the technical evidence layer that the Act requires.
The Enforcement Timeline: What Is Already Live
Building an EU AI Act Compliance Program for Financial Services
The compliance program for EU AI Act obligations in financial services requires five concurrent workstreams. These are not sequential — they must proceed in parallel given the enforcement timeline.
1. AI System Inventory and Classification
The first step is a complete inventory of AI systems used in EU-affecting operations. For each system, the compliance team must determine: Is this system in scope under the Act? If so, what risk tier applies? For systems identified as high-risk, the full Article 8-15 compliance program applies. This inventory must be maintained as a living document — new deployments, significant modifications to existing systems, and changes in intended use all require re-classification.
2. Provider / Deployer Determination
For each high-risk AI system, the institution must determine whether it is acting as a provider, a deployer, or both. This determination affects which obligations apply. Third-party AI vendors typically supply conformity assessment documentation and technical documentation — but financial institutions that customize or fine-tune vendor models, or that build their own high-risk systems, become providers and assume the full provider obligation set.
3. Article 9 Risk Management Program
For each high-risk system, establish the Article 9 risk management documentation. This should be integrated with existing model risk management programs (SR 11-7 for US-regulated institutions) rather than built in parallel. The risk identification, estimation, and mitigation documentation required by Article 9 aligns substantially with the conceptual soundness and ongoing monitoring components of SR 11-7 validation.
4. Technical Controls and Human Oversight Infrastructure
Article 14's human oversight requirement is a technical design mandate. For LLM-based financial AI, this requires: interpretable output formats that enable anomaly detection, override mechanisms that function at the speed of human review, audit trails that capture the basis of AI outputs, and governance controls that ensure the oversight mechanism is practically effective. Pre-execution enforcement infrastructure — systems that apply and document policy compliance before AI output reaches decision-makers — provides the technical substrate for Article 14 compliance.
5. EU Database Registration and Declaration of Conformity
High-risk AI systems must be registered in the EU AI Act database before deployment. The registration requires information about the system's intended purpose, the risk management system, and the technical documentation availability. The EU Declaration of Conformity must be prepared and signed by a person with the authority to legally bind the organization.
Financial services firms with mature SR 11-7 model risk management programs should not build a parallel EU AI Act compliance track. The conceptual soundness (Article 9 risk identification), ongoing monitoring (Article 9 risk management measures), and outcomes analysis (Article 9 residual risk evaluation) components of SR 11-7 map directly to EU AI Act Article 9 requirements. A unified AI governance program that satisfies both frameworks simultaneously is more efficient and creates a more coherent operational process than maintaining separate compliance programs for US and EU regulatory obligations.