EU AI Act: What Counts as a High-Risk AI System

What makes an AI system "high-risk" under the EU AI Act?

The EU AI Act uses a risk-based pyramid. Most AI — productivity tools, recommendation engines, general chatbots — sits at the base with minimal obligations. A smaller group of systems in categories where AI decisions significantly affect people's lives, rights, or safety is placed in the high-risk tier. High-risk status is not a judgment about how likely the system is to malfunction; it is a judgment about the severity of potential harm if it does.

There are two routes to high-risk classification under the EU AI Act. The first is being an AI system that functions as a safety component of a product already regulated by specific EU product-safety laws listed in Annex I — machinery, medical devices, aviation equipment, and similar. The second, and the one most software teams encounter, is falling into one of the eight use-case categories in Annex III. This guide focuses on Annex III.

A critical nuance: Annex III does not make entire industries high-risk. It targets specific functions within those industries. An AI tool that helps a recruiter write a job description is not high-risk. An AI system that ranks and filters job applicants for an employer — making or substantially influencing who advances — is. The question is always whether the system's specific function matches the specific use case described in the Annex.

The complete Annex III list: all eight categories explained

The table below sets out every category in Annex III of the EU AI Act, with a plain-English description of what the regulation is targeting in each one. These are the only eight categories. Do not infer additional categories from sector analogy — the list is exhaustive unless formally amended by the European Commission under Article 7.

Which of my AI systems might be high-risk? A decision framework

The following numbered steps form a structured decision tree. Work through them in order. A "yes" at any stage that sends you to a high-risk conclusion requires you to proceed with the full set of Annex III obligations; a definitive "no" at an early gate lets you stop and document your reasoning.

1. Is the system AI at all under the EU AI Act definition? The regulation covers machine-based systems that infer, from input, how to generate outputs such as predictions, content, recommendations, or decisions. Pure rule-based systems with no learned inference are not AI systems under the Act. If your system is purely deterministic rules, it is outside scope. If it involves any learned model component, continue.
2. Is the system used in the EU, or does it affect people in the EU? The EU AI Act applies regardless of where the provider or deployer is established, if the outputs are used within the EU or affect people located in the EU. If neither applies, the Act does not bind you — though other jurisdictions may have equivalent rules.
3. Is the system a safety component of an Annex I product? Check the list in Annex I of the EU AI Act: it includes machinery, medical devices, in-vitro diagnostics, aviation systems, motor vehicles, agricultural machinery, and marine equipment, among others. If your AI system is a safety component of one of those regulated products, it is high-risk regardless of what it does. Go directly to the obligations section below.
4. Does the primary function of the system match one of the eight Annex III categories? Work through each category in the table above. Focus on what the system actually does, not what industry it is sold into. Mapping by industry sector alone causes both false positives (non-high-risk tools in a listed sector) and false negatives (high-risk tools in unlisted sectors).
5. Does the system make or substantially influence individual decisions?Many Annex III categories are triggered not just by using AI in a listed domain but by the system making or substantially influencing consequential decisions about individual people. An AI that produces aggregate statistics for a recruiter's dashboard is in a different position from one that ranks and filters individual candidates. If your system outputs a ranked list, a score, or a recommended action about a specific person in a high-risk domain, it is most likely in scope.
6. Does any exemption apply? The EU AI Act contains narrowing provisions for research and development (systems not yet placed on the market), systems used purely for military or national security purposes, and AI intended exclusively for scientific research. If an exemption legitimately applies, document it precisely — regulators will ask.
7. If in doubt, assume high-risk and get legal review. The cost of misclassifying a high-risk system as low-risk is orders of magnitude higher than the cost of a legal review. The EU AI Act self-assessment calculator provides a structured scoping tool to guide this analysis and produce a documented classification record.

What obligations apply to high-risk AI systems?

Once a system is classified as high-risk under Annex III, Chapter III Section 2 of the EU AI Act lays out a mandatory set of technical and organisational requirements that must be met before the system can be placed on the EU market or put into service. These are not suggestions; they are legal prerequisites for deployment.

Risk management system

Article 9 requires providers to establish, implement, document, and maintain a risk management system throughout the entire lifecycle of the high-risk AI system. The system must identify and analyse known and foreseeable risks to health, safety, and fundamental rights; estimate and evaluate the risks that may emerge when the system is used as intended or in reasonably foreseeable misuse; and adopt risk mitigation measures. The risk management system is not a one-time exercise — it must be continuously updated as post-market data comes in.

Data and data governance

Article 10 sets requirements for the training, validation, and testing data used in high-risk systems. Data must be subject to governance practices covering data collection, processing, and preparation. Training data must be relevant, representative, free of errors to the extent possible, and complete in light of the intended purpose. Bias must be examined and addressed. These requirements apply to data used at the model level, not just at the application layer.

Technical documentation

Article 11 requires providers to draw up technical documentation before the system is placed on the market and to keep it up to date. The documentation must contain, at minimum, a general description of the system; a description of the training methodology, data, and training processes; information on validation and testing; and details of the monitoring, functioning, and control of the system. This documentation is what national authorities inspect — it must be detailed enough for a competent authority to assess compliance.

Automatic record-keeping and logging (Article 12)

High-risk AI systems must be designed to automatically record events over the system's lifetime. The logs must enable identification of risk situations and substantial modifications, and must support post-market monitoring by the provider and monitoring by the deployer during operation. For biometric systems, additional specific fields must be logged. Logs must be retained for at least six months. For a full breakdown of the logging obligation, see our guide to EU AI Act Article 12 logging requirements.

Transparency and instructions for deployers

Article 13 requires that high-risk AI systems are designed and developed to be sufficiently transparent so that deployers can interpret the system's outputs and use them correctly. Providers must supply instructions for use that cover the system's identity, capabilities and limitations, performance metrics, intended deployment contexts, and the human oversight measures that should be applied.

Human oversight

Article 14 requires high-risk AI systems to be designed so that human oversight is possible and effective. The system must allow the humans responsible for its use to understand its capabilities and limitations, detect and address failures, override or disregard outputs, and interrupt or stop the system. Human oversight must be built in, not bolted on after deployment.

Accuracy, robustness, and cybersecurity

Article 15 requires high-risk AI systems to be designed to achieve an appropriate level of accuracy, robustness, and cybersecurity. They must perform consistently throughout their lifecycle, must be resilient against attempts by third parties to exploit vulnerabilities, and must include technical redundancy and failsafe mechanisms where appropriate for the risk.

Conformity assessment and CE marking

Before placing a high-risk AI system on the market, providers must carry out a conformity assessment demonstrating that the system meets all the above requirements. For most Annex III systems, providers may self-assess. For certain biometric systems, a third-party notified body assessment is required. After a successful assessment, the provider affixes the CE mark and registers the system in the EU database of high-risk AI systems.

Registration in the EU database

High-risk systems must be registered in the EU-level public database before market placement. The database is maintained by the European AI Office and is intended to give authorities, deployers, and the public visibility into which high-risk systems are in use across the EU.

How to determine if your company's AI is in scope

In practice, the scoping exercise involves three parallel workstreams: a technical inventory, a legal mapping, and an organisational assessment.

Build an AI inventory first

You cannot classify what you have not found. Start by identifying every AI system or AI-powered feature your company develops, deploys, or procures — including third-party tools that integrate AI into your workflows. Many companies discover that AI is more pervasive than their initial estimate: embedded in HR tools, in customer-scoring APIs, in fraud-detection services, and in document-processing pipelines.

Map each system to Annex III

For each item in your inventory, work through the decision steps above. Document the mapping explicitly: which Annex III category, if any, is triggered; what the basis for the conclusion is; and who signed off on the analysis. Undocumented classification decisions create audit risk — a regulator who asks why you treated a system as low-risk should find a clear written answer, not a verbal recollection.

Identify which role you occupy

Are you the provider (the entity that developed the system and placed it on the market), the deployer (the entity that puts it into service within its own organisation), or both? The role determines which obligations fall on you. A company that builds an AI recruitment tool and sells it to employers is a provider. An employer that buys that tool and uses it is a deployer. A company that buys a foundation model and builds its own credit-scoring product on top of it may be reclassified as a provider for that downstream product.

Use the self-assessment calculator as a starting point

Scoping analysis is not a one-time event — it should be repeated whenever you launch a new product, substantially modify an existing system, or enter a new market. The EU AI Act self-assessment calculator provides a structured, question-driven framework that walks through the scoping decision and produces a documented output you can keep on file. It does not replace legal advice for complex cases, but it is the right starting point for every system in your inventory.

Penalties for non-compliance

The EU AI Act's enforcement regime (Article 99) sets tiered penalties. The bands are "whichever is higher" of a fixed euro ceiling or a percentage of global annual turnover — except for SMEs and start-ups, for whom it is the lower of the two.

Non-compliance with Annex III obligations tends to cascade: a missing risk management system means inadequate testing, which means inadequate technical documentation, which means a conformity assessment that cannot be completed. Regulators investigating one gap typically find several. Treating compliance as an interconnected system — not a checklist of independent boxes — is the only approach that survives a real audit.

Next steps if your AI system is in scope

If you have determined that one or more of your AI systems is high-risk under Annex III, the practical starting point is a gap analysis: measure your current documentation, data governance, logging, and oversight practices against each Article 9–15 obligation, identify what is missing, and sequence remediation from the obligations that require the longest lead time (risk management systems and training-data governance) to those you can implement more quickly (logging, transparency documentation).

Article 12 logging is frequently the first obligation teams tackle because it is technically concrete — it involves building or adapting infrastructure — and because it also serves the post-market monitoring obligation in Article 9. If your logging is not yet structured around risk identification and modification tracking, AuditProof can help you build Article 12-oriented audit trails and convert your AI process records into evidence that meets the EU AI Act's documentation standard. For initial scoping and classification, begin with the EU AI Act self-assessment calculator.