EU AI Act Risk Pyramid: What Each Risk Tier Means for Your Product
The EU AI Act’s Annex III high-risk obligations apply from August 2, 2026. For AI systems embedded in Annex I CE-marked products, the deadline is August 2, 2027. Engineers designing products with AI components today are making architecture decisions that must hold up against those requirements.
Regulation (EU) 2024/1689 assigns every AI system to one of four tiers: minimal risk, limited risk, high risk, or prohibited. The tier determines your documentation obligations, conformity assessment route, and post-market requirements.
This article explains what each tier demands from your product development process.
To determine which tier your system belongs to, use the EU AI Act Classification Framework, which covers Article 5, Article 6(1), Article 6(2), and Article 50 gate by gate.
What does minimal risk actually require from a product?
Minimal risk is the default outcome for AI systems that do not trigger Article 5, Article 6, or Article 50. Under Regulation (EU) 2024/1689, minimal risk systems carry no mandatory requirements.
No technical documentation, no conformity assessment, no EU database registration, and no CE marking obligation applies at this tier.
The Commission encourages voluntary codes of conduct for minimal risk systems. These are not enforceable and carry no legal consequence under the current regulation.
One structural risk applies regardless of current classification: the Commission can expand the Annex III list through delegated acts, without full legislative procedure.
A system classified as minimal risk today can fall into a new Annex III category after a delegated act is issued.
Documenting the classification rationale at product approval makes any future re-evaluation faster and more defensible.
What does the limited risk tier require under Article 50?
Limited risk systems are not high-risk, but they carry transparency obligations under Article 50 of Regulation (EU) 2024/1689. These apply regardless of the product’s CE marking status or other regulatory classification.
Three system characteristics trigger Article 50 obligations:
- Conversational AI systems that interact with people must disclose that the user is interacting with an AI, unless this is obvious from context
- AI systems that generate synthetic audio, video, images, or text must label the output as AI-generated in a machine-readable format
- Emotion recognition systems and biometric categorisation systems must inform individuals who are subject to those processes before exposure
None of these obligations requires a conformity assessment, Annex IV technical documentation, or EU database registration. Each requires a disclosure mechanism, and synthetic content additionally requires machine-readable labelling.
What does a high-risk classification mean for your product engineering?
High risk imposes the most substantive engineering obligations in the regulation. The classification can arise from two distinct pathways under Article 6, each with different conformity assessment consequences.
Article 6(1): AI in CE-marked products
Article 6(1) applies when your AI system acts as a safety component of a product already covered by EU harmonised legislation listed in Annex I of the AI Act. The product type determines the high-risk classification, not the AI system in isolation.
The Annex I product families most directly relevant to hardware engineers are the Machinery Regulation (EU) 2023/1230, Medical Devices Regulation (EU) 2017/745, In Vitro Diagnostics Regulation (EU) 2017/746, and Radio Equipment Directive 2014/53/EU.
[VERIFY: confirm the complete Annex I list against the published text of Regulation (EU) 2024/1689 before finalising technical documentation.]
An AI-based vision system that triggers an emergency stop on an industrial robot is a safety component of a Machinery Regulation product. An AI module logging production output on the same robot is not.
The distinction turns on whether the AI contributes to the product’s safety function, not its technical complexity.
Under Article 6(1), the conformity assessment follows the existing process of the applicable Annex I legislation, extended to cover the AI system.
For a Machinery Regulation product, the AI component is included in the EC declaration of conformity and technical documentation. For an MDR product, the notified body assessment extends to include the AI system.
Article 6(2): Annex III use cases
Article 6(2) applies when the use case of your AI system is listed in Annex III, regardless of the product it sits inside. Two Annex III categories are most frequently encountered by hardware and embedded engineers.
The first is biometric identification and categorisation: systems that identify or categorise natural persons using biometric data. An embedded face recognition module used for access control is within scope. A face detection system that adjusts display brightness based on user proximity is not.
The second is critical infrastructure management: systems operating or controlling road, rail, water, gas, heating, electricity, or digital infrastructure. An AI-based load balancing controller for a power distribution product falls here. A report-generation module for the same network does not.
All eight Annex III categories and their scope are summarised below. The first two receive detailed treatment above because they arise most frequently in embedded and hardware product contexts.
| Annex III Category | Scope |
|---|---|
| 1. Biometric identification and categorisation | Systems identifying or categorising natural persons through biometric data, including face recognition for access control |
| 2. Critical infrastructure management | Systems operating or managing road, rail, water, gas, heating, electricity, and digital infrastructure |
| 3. Education and vocational training | Systems determining access to educational institutions or evaluating learner performance |
| 4. Employment and worker management | Systems used in recruitment, task allocation, performance monitoring, or termination decisions |
| 5. Essential private and public services | Systems determining eligibility for public benefits, credit scoring, insurance risk assessment, or emergency service dispatch |
| 6. Law enforcement | Systems used for crime prediction, evidence evaluation, or profiling of individuals |
| 7. Migration and border control | Systems supporting asylum processing, border checks, visa applications, or migration monitoring |
| 8. Justice and democratic processes | Systems influencing electoral outcomes, assisting judicial decisions, or used in alternative dispute resolution |
Providers of systems in any Annex III category should additionally check Article 6(3), which provides a self-classification exemption for systems that demonstrably pose no significant risk of harm despite falling within a listed category.
What high-risk requirements cover
High-risk AI systems must meet the requirements in Articles 9 through 15 before being placed on the EU market.
The six requirement areas, with their governing articles, are as follows:
- Risk management system (Article 9): a documented, iterative process for identifying, estimating, and mitigating AI-specific risks across the product lifecycle
- Data governance (Article 10): training, validation, and test datasets must be relevant, representative, and reviewed for errors that could produce discriminatory outcomes
- Technical documentation (Article 11, Annex IV): a technical file covering system architecture, training methodology, performance metrics, and instructions for use
- Logging (Article 12): automatic logging of system operation to the extent technically feasible, retained for the required period
- Transparency for deployers (Article 13): instructions of use enabling the deployer to understand system capabilities, limitations, and correct operation
- Human oversight (Article 14): design features enabling a human operator to monitor, intervene, or halt the system during operation
Article 15 adds requirements for accuracy, robustness, and cybersecurity, including resilience against adversarial inputs.
Before market placement, the provider must complete a conformity assessment, register the system in the EU database, issue a declaration of conformity, and apply CE marking [VERIFY: confirm CE marking for standalone Annex III systems under Article 48].
For most Annex III systems, the conformity assessment route is internal control (self-assessment). Remote biometric identification systems require assessment by a third-party notified body [VERIFY: confirm the exact notified body scope in Article 43].
Post-market monitoring under Article 72 is mandatory and ongoing for the product lifecycle. For the QMS requirements that accompany high-risk AI certification, see the EU AI Act QMS guide for high-risk compliance.
What practices does Article 5 prohibit outright?
Article 5 lists AI practices that are prohibited with no conformity route and no market access. It became applicable on February 2, 2025.
The prohibitions most relevant to hardware engineers building deployed products include the following:
- Real-time remote biometric identification of natural persons in publicly accessible spaces, subject to narrow law enforcement exceptions
- Biometric categorisation systems that infer race, political opinion, trade union membership, religious belief, sexual orientation, or health status from biometric data
- Emotion recognition systems deployed in workplaces or educational institutions
- AI systems that manipulate persons through subliminal techniques or by exploiting vulnerabilities related to age, disability, or social circumstances
The prohibition on real-time remote biometric identification is the one most frequently encountered by hardware engineers building camera-based systems for public deployment. The prohibition applies to the use of the system, not only its technical capability.
Article 5 is already in force. Products implementing prohibited practices are not grandfathered, regardless of when they entered the market.
How do the application timelines affect engineering decisions today?
The EU AI Act applied its obligations in phases. Each tier has a distinct application date with direct implications for product development scheduling [VERIFY: confirm dates against the published text of Regulation (EU) 2024/1689].
The application dates by tier are as follows:
- Prohibited practices (Article 5): applicable from February 2, 2025
- GPAI model obligations (Chapter V): applicable from August 2, 2025
- High-risk AI systems under Annex III: applicable from August 2, 2026
- High-risk AI systems in Annex I CE-marked products: applicable from August 2, 2027
The Annex III deadline of August 2026 means products currently in development and carrying Annex III-scope AI systems need Article 9 risk management and Article 11 technical documentation integrated into the design process now.
Retrofitting compliance documentation onto a completed hardware design is significantly more costly than building it in from the schematic phase.
For engineers on Machinery Regulation products, prEN 50742 and its intersection with AI Act requirements provides context on overlapping obligations.
Frequently Asked Questions
How do I know which EU AI Act risk tier applies to my product? Work through Article 5, Article 6(1), Article 6(2), and Article 50 in sequence. The EU AI Act Classification Framework covers each gate step by step.
What does high-risk mean under the EU AI Act? Articles 9 through 15 apply: risk management, data governance, technical documentation, logging, transparency, human oversight, and cybersecurity. A conformity assessment, EU database registration, and CE marking are required before market placement.
Does the EU AI Act apply to AI embedded in a CE-marked product? Yes, under Article 6(1). If the AI system acts as a safety component of an Annex I product, it is automatically high-risk. The conformity assessment of the applicable product legislation is extended to cover the AI system.
Can a minimal risk AI system become high-risk later? Yes. The Commission can expand Annex III through delegated acts. A minimal risk classification today can change. Document the classification rationale at product approval to reduce the cost of future re-evaluation.
When must high-risk AI systems comply with the EU AI Act? Annex III high-risk systems must comply from August 2, 2026. Annex I high-risk systems embedded in CE-marked products must comply from August 2, 2027 [VERIFY: confirm dates against Regulation (EU) 2024/1689].
Conclusion
Minimal risk carries no current mandatory requirements but is not permanently settled. The Annex III list can be expanded by delegated act, and a current minimal risk classification can change.
High risk triggers six requirement areas under Articles 9 through 15, a conformity assessment, EU database registration, and ongoing post-market monitoring.
Article 5 prohibitions have applied since February 2, 2025, with no grandfathering for products already on the market.
The Annex III deadline of August 2, 2026 means the documentation and risk management decisions being made in current product development cycles are the ones that will need to be defended first.
