Define the system
Start by naming the AI feature, users, decision impact, data categories, vendors, and the team member responsible for maintaining the review record.
Risk guide
What Is a High-Risk AI System?
A high-risk AI signal usually appears when an AI system affects safety, rights, access, eligibility, employment, education, public services, or other consequential decisions.
Common high-risk signals
| Area | Examples to review carefully |
|---|---|
| Employment | CV screening, worker scoring, task allocation, promotion or dismissal support. |
| Education | Admissions, exam scoring, student assessment, learning-path decisions. |
| Credit and essential services | Creditworthiness, insurance eligibility, benefits, housing access. |
| Biometrics | Identification, categorization, emotion recognition, access control. |
| Healthcare and safety | Clinical triage, patient recommendations, safety-critical monitoring. |
| Public sector and law | Law enforcement, migration, border control, courts, democratic processes. |
What to do if you see a signal
- Do not rely on a marketing label like "AI assistant" or "automation." Look at the real decision impact.
- Document the intended purpose and user group.
- Describe human oversight and escalation.
- Get specialist review before launch if the system affects rights or access.
Official sources
Last reviewed: 2026-07-03.
- EU AI Act Article 6: high-risk classification
- EU AI Act Annex III: high-risk areas
- European Commission high-risk guidelines
practical AI compliance self-assessment
Practical notes for What Is a High-Risk AI System?
What Is a High-Risk AI System? | EU AI Act Guide is maintained for founders, product managers, compliance owners, agencies, and small teams building AI workflows who need AI governance workflow. The goal is to help visitors complete a real task and leave with an AI inventory, risk note, disclosure draft, vendor question set, policy outline, or review workflow report, not only read a generic summary.
Separate triage from advice
Use the generated output as first-pass operational triage. Legal, medical, hiring, credit, education, biometric, and public-sector uses still need specialist review.
Keep evidence
Save the output, assumptions, date, source links, and reviewer notes so the team can explain why a risk level, disclosure, or vendor question set was chosen.
Before relying on this page
- Review high-impact use cases manually.
- Keep policies and disclosures aligned with the real product behavior.
- Re-run the workflow when vendors, data, or user impact changes.
Review record
How to use What Is a High-Risk AI System? in an AI compliance file
A compliance tool is useful when it leaves a traceable record. The output should name the AI system, explain the assumptions, and show what the team still needs to verify with product, legal, security, or vendor owners.
Describe the real system
Record the feature name, user group, decision impact, data sources, vendor dependencies, and the human owner. A generic "chatbot" label is rarely enough. A hiring assistant, support summarizer, medical triage bot, and product recommender can have very different risk profiles even if they all use language models.
Separate signal from conclusion
Treat this page as first-pass triage. It can flag high-risk areas, transparency duties, missing evidence, and questions to ask a vendor. It should not be treated as legal approval, clinical advice, hiring approval, credit approval, or permission to launch without human review.
Save evidence and changes
Save the generated result with the date, reviewer, source links, and unresolved questions. Update the record when the model, data, users, product flow, vendor, or region changes. This keeps the site useful for actual operators rather than only being a static explanation page.
Evidence checklist
- Document what the AI system does and what it does not do.
- Record whether the system influences employment, education, credit, public benefits, healthcare, biometric identification, safety, or other high-impact outcomes.
- Keep vendor documentation, model notes, data descriptions, user notices, human oversight notes, and monitoring plans together.
- Re-run the review when the product behavior changes, not only when the law changes.
- Use specialist review for high-impact or regulated workflows before relying on any generated text.