A high-risk AI system is an AI system that the EU AI Act classifies as posing a significant risk to health, safety or fundamental rights, either because it is a safety component of a regulated product or because it is used in a listed sensitive area such as hiring, credit or law enforcement. In AI regulation, this category carries the heaviest obligations short of an outright ban, including risk management, data governance, human oversight and a conformity assessment before the system reaches the market.
The classification concerns the intended purpose, not the underlying technique. The same machine learning model can be unregulated when it recommends films and high-risk when it ranks job applicants, because the second use can shape a person’s livelihood.
How the AI Act classifies a system as high-risk
The AI Act uses two routes into the high-risk category. The first covers AI systems that are a safety component of a product, or are themselves a product, governed by EU product legislation listed in Annex I, such as machinery, medical devices, toys or lifts, where that legislation requires a third-party conformity assessment. The second covers stand-alone systems used in the areas listed in Annex III.
Annex III areas are defined by use case, and the Commission can update the list through delegated acts as technology and evidence develop.
| Area | Typical high-risk use |
|---|---|
| Biometrics | Remote identification, categorization, emotion recognition |
| Critical infrastructure | Safety components in energy, water, traffic, digital infrastructure |
| Education | Admission, grading, exam proctoring |
| Employment | Recruitment screening, promotion, task allocation, monitoring |
| Essential services | Credit scoring, public benefits, health and life insurance pricing |
| Law enforcement | Evidence reliability, risk assessment of individuals |
| Migration and borders | Visa and asylum assessment, risk screening |
| Justice and democracy | Supporting judicial decisions, influencing elections |
An Annex III system can escape the category through a filter for low-impact uses: it is not high-risk when it performs a narrow procedural task, improves the result of completed human work, detects patterns without replacing human assessment or performs a preparatory task. The filter never applies when the system profiles natural persons, and a provider relying on it must document its assessment.
Requirements a high-risk AI system must meet
The Act sets seven core requirements that apply throughout the lifecycle of the system. Providers usually meet them by following harmonized standards, which create a presumption of conformity once they are published.
| Requirement | What it demands in practice |
|---|---|
| Risk management system | Continuous identification and mitigation of foreseeable risks |
| Data and data governance | Relevant, representative training data checked for bias |
| Technical documentation | Design, training and testing records for authorities |
| Record-keeping | Automatic event logs that allow traceability |
| Transparency to deployers | Instructions on purpose, limits and accuracy levels |
| Human oversight | Design that lets people monitor, override or stop the system |
| Accuracy, robustness and cybersecurity | Declared performance and resistance to errors and attacks |
These requirements are outcome-based rather than prescriptive. The Act does not fix an accuracy figure; it requires the provider to declare appropriate metrics, justify them for the intended purpose and keep performance within those limits.
Duties of providers and deployers
Obligations fall mainly on the AI provider, the organization that develops the system and places it on the market under its own name, but the AI deployer that uses it professionally has its own duties. A deployer that substantially modifies a system or rebrands it can itself become the provider.
| Duty | Provider | Deployer |
|---|---|---|
| Quality management system | Required | Not required |
| Conformity assessment and CE marking | Required before market entry | Verify marking exists |
| Registration in the EU database | Required | Required for public bodies |
| Human oversight | Design it into the system | Assign competent staff |
| Logs | Enable automatic logging | Keep logs for at least six months |
| Fundamental rights impact assessment | Not required | Public bodies and certain service providers |
| Serious incident reporting | Report to market surveillance authority | Inform the provider promptly |
For most Annex III systems, the conformity assessment is an internal control procedure carried out by the provider. Some biometric systems and products under Annex I legislation require the involvement of a notified body.
High-risk compared with other risk tiers
The AI Act follows a risk-based pyramid, and high-risk systems sit just below the prohibited practices. Getting the tier right is the first and most consequential compliance decision.
| Tier | Examples | Regulatory consequence |
|---|---|---|
| Prohibited practices | Social scoring, manipulative techniques | Banned outright |
| High-risk | Recruitment screening, credit scoring | Full requirements and conformity assessment |
| Transparency obligations | Chatbots, deepfakes | Disclosure to users |
| Minimal risk | Spam filters, game AI | No specific obligations |
General-purpose AI models are regulated through a separate set of rules. A system built on such a model becomes high-risk only when it is deployed for a high-risk purpose.
Timelines, penalties and common mistakes
The Act originally scheduled the high-risk obligations to apply from August 2026 for Annex III systems and August 2027 for Annex I products. The EU has since proposed tying these dates to the availability of standards and support tools, so organizations should confirm the applicable dates against the latest amendments. Breaches of high-risk obligations can be fined up to 15 million euros or 3 percent of worldwide annual turnover, whichever is higher.
- Classifying by technology instead of intended purpose.
- Assuming a vendor carries all duties, while the deployer’s oversight and logging obligations remain.
- Overusing the low-impact filter without documented reasoning.
- Treating compliance as a one-time audit, although monitoring continues after deployment.
A plain-language summary
A high-risk AI system is software that helps make decisions that can seriously affect people’s lives, such as who gets a job interview, a loan or a place at a university. Because mistakes in these areas can be unfair or harmful, the law asks the makers to test, document and supervise such systems carefully.
It is similar to how cars or medical devices must pass checks before sale. The higher the stakes for people, the stricter the rules.
An example from practice
A software company sells a tool that ranks job applications for large employers in several EU countries. Because recruitment appears in Annex III and the tool profiles applicants, it is a high-risk AI system, and the company acts as its provider.
The company sets up a quality management system, tests the ranking model for disparities between applicant groups, documents accuracy levels and adds an interface where recruiters can see why a candidate was ranked and override the result. Each employer, as deployer, assigns trained recruiters to review rankings, informs staff representatives and keeps logs. The outcome is a product that can carry CE marking and be sold across the EU market.