Scope
This guide helps your security and governance team assess cloud AI service compliance with the EU AI Act. It covers provider responsibilities, customer obligations under the shared responsibility model, and practical steps to maintain compliance when deploying AI systems on AWS infrastructure.
This is not legal advice. Your organization's obligations depend on your specific use cases, risk tier, and role in the AI supply chain.
Key Concepts
Provider vs. Deployer
The EU AI Act distinguishes between providers (those who develop or substantially modify an AI system and place it on the market) and deployers (those who use an AI system under their authority). As a cloud customer, you're typically a deployer when using managed AI services, but you may become a provider when building custom models.
Shared Responsibility Model
AWS provides compliant infrastructure and tools. You're responsible for configuring controls, documenting your use case, and ensuring your application meets EU AI Act requirements for your risk tier.
Risk-Based Framework
The Act categorizes AI systems into prohibited, high-risk, limited-risk, and minimal-risk. Your compliance obligations scale with risk. For example, a chatbot requires transparency disclosures, while a credit scoring system requires Technical Documentation (Annex IV), human oversight, and ongoing Post-Market Monitoring.
General-Purpose AI Models
Separate requirements apply to General-Purpose AI Model providers. If you fine-tune a foundation model substantially, you may trigger provider obligations. If you use a pre-trained model via API, you're typically a deployer.
Timeline and Enforcement
The EU AI Act entered into force on August 1, 2024. Obligations apply in phases:
- February 1, 2025: Prohibited AI Practices and AI literacy requirements enforceable
- Next 14 months: Remaining obligations become applicable gradually
Even non-EU organizations face compliance requirements if their AI systems are used in the EU or produce outputs used in the EU.
Provider Capabilities: What AWS Delivers
AWS was the first major cloud service provider to announce ISO/IEC 42001 accredited certification for AI services, covering Amazon Bedrock, Amazon Q Business, Amazon Textract, and Amazon Transcribe. This certification demonstrates alignment with the international standard for AI Management Systems, including Plan-Do-Check-Act (PDCA) processes and Annex A Controls.
AI Service Cards provide transparency documentation for each managed service. You'll find intended use cases, model limitations, responsible AI design choices, and performance optimization guidance. Use these as inputs when drafting your own Technical Documentation (Annex IV) for high-risk systems.
Frontier Model Safety Framework addresses severe risks unique to large-scale models. This framework informs how AWS evaluates and mitigates risks before making foundation models available through Amazon Bedrock.
Amazon Bedrock Guardrails let you implement content filters, denied topics, word filters, and sensitive information redaction. Configure these to enforce your responsible AI policies and meet transparency obligations.
Your Compliance Checklist
1. Determine Your Role and Risk Tier
Start by mapping your AI system to the EU AI Act's risk categories. High-risk systems include those used for:
- Credit scoring or creditworthiness assessment
- Employment decisions (CV screening, performance evaluation)
- Access to education or training
- Law enforcement (predictive policing, emotion recognition)
- Critical infrastructure management
If you're using AWS services to build a high-risk system, you're the provider and must meet the full set of requirements in Articles 8-15.
2. Avoid Prohibited Practices
Review Article 5 prohibitions before deployment. The European Commission's Guidelines on Prohibited AI Practices clarify these restrictions. AWS services aren't designed for prohibited uses, and the AWS Acceptable Use Policy and Responsible AI Policy reinforce this, but you must verify your specific application doesn't:
- Deploy subliminal manipulation techniques
- Exploit vulnerabilities of specific groups
- Evaluate or classify individuals based on social behavior or personal traits (social scoring)
- Conduct real-time remote biometric identification in public spaces (with narrow exceptions)
3. Build Technical Documentation
For high-risk systems, prepare Technical Documentation (Annex IV) covering:
- General description of the AI system and its intended purpose
- Detailed description of system elements and development process
- Data governance and training methodologies
- Validation evidence and performance metrics
- Human oversight measures
- Cybersecurity and resilience measures
AWS AI Service Cards provide baseline information for managed services. Extend this with your application-specific architecture, data sources, and control implementations.
4. Implement Transparency Controls
All AI systems that interact directly with people require Disclosure of AI Interaction. Configure your application to inform users they're engaging with an AI system.
For AI-Generated Content Labelling, use metadata tagging or watermarking where technically feasible. Amazon Bedrock supports custom metadata fields you can populate during generation.
5. Configure Guardrails and Oversight
Use Amazon Bedrock Guardrails to:
- Block harmful content categories (hate speech, violence, sexual content)
- Deny specific topics relevant to your domain (medical advice for non-healthcare applications)
- Filter personally identifiable information
- Implement custom word filters
Document your guardrail configurations as part of your risk management procedures. Test them against Adversarial Simulation scenarios to verify effectiveness.
6. Establish Post-Market Monitoring
High-risk systems require ongoing Post-Market Monitoring to detect performance drift, bias emergence, or safety incidents. Implement:
- Logging of system inputs, outputs, and decisions
- Regular performance metric reviews against validation baselines
- Incident response procedures for adverse events
- User feedback collection mechanisms
AWS CloudWatch and CloudTrail provide infrastructure-level monitoring. Extend this with application-level metrics specific to your AI system's risk profile.
Common Pitfalls
Assuming cloud compliance equals application compliance
AWS's ISO/IEC 42001 certification covers infrastructure and managed services. Your custom application requires separate compliance assessment and documentation.
Underestimating documentation burden
68% of European businesses struggle to understand their responsibilities under the EU AI Act. Don't wait until an enforcement action to build Technical Documentation (Annex IV). Start documenting during development, not after deployment.
Ignoring General-Purpose AI Model fine-tuning implications
Substantial modification of a foundation model may shift you from deployer to provider. If you're fine-tuning on domain-specific data and releasing the model, consult the General-Purpose AI Code of Practice for transparency and systemic risk obligations.
Treating transparency as a checkbox
Disclosure of AI Interaction isn't just a banner message. Users need meaningful information about how the AI system makes decisions and what limitations apply. Your Instructions for Use should be specific and actionable.
Missing extraterritorial applicability
Non-EU organizations face EU AI Act obligations if their systems are used in the EU. Geographic location of your infrastructure doesn't determine compliance scope; the location of your users and the effects of your AI system do.
Quick Reference
| Your Activity | Your Role | Key Obligations | AWS Tools |
|---|---|---|---|
| Using Amazon Bedrock via API | Deployer | Prohibited practice review, transparency disclosures | AI Service Cards, Guardrails |
| Fine-tuning Bedrock model minimally | Deployer | Same as above, plus use restrictions documentation | Model Cards, Guardrails |
| Fine-tuning and releasing model | Provider | Full Technical Documentation (Annex IV), conformity assessment | Model Cards, Responsible AI Guide |
| Building custom high-risk system | Provider | Annex IV, risk management, human oversight, Post-Market Monitoring | CloudWatch, CloudTrail, Guardrails |
| Providing General-Purpose AI Model | GPAI Provider | Transparency obligations, systemic risk assessment if applicable | Frontier Model Safety Framework |
Next Steps
- Conduct an AI System Impact Assessment using ISO/IEC 42005 methodology to identify affected systems and risk tiers.
- Review AWS AI Service Cards for services you currently use.
- Map your existing model governance framework to EU AI Act requirements.
- Identify documentation gaps for high-risk systems.
- Configure Amazon Bedrock Guardrails aligned with your responsible AI policies.
- Establish Post-Market Monitoring procedures with defined metrics and review cadence.
The EU AI Act represents a shared journey between cloud providers and customers. AWS delivers compliant infrastructure and transparency tooling. You're responsible for configuring these tools correctly, documenting your specific use case, and maintaining ongoing compliance as your AI systems evolve.
