High-risk exposure
Annex III high-risk AI systems explained
Annex III focuses on sensitive use cases rather than particular model brands. Classification depends on intended purpose, deployment context and the effect on people.
Prepared by EU AI Fit editorial team · Published 24 August 2026 · Source review 31 August 2026 · Page updated 24 September 2026
Areas that require close attention
Potential categories include biometrics, critical infrastructure, education, employment, access to essential services, law enforcement, migration and the administration of justice or democratic processes.
Avoid classification by keyword
A tool used near a listed sector is not automatically high-risk, and an innocent product label does not make it minimal-risk. Document the decision influenced, the people affected, the significance of the outcome and the role of human review.
Prepare for a defensible review
Retain the intended-purpose statement, workflow diagram, input and output examples, human-oversight design, provider documents and rationale for inclusion or exclusion. Escalate borderline cases to a qualified adviser.
A human in the process is not an automatic exemption
Compare a recruitment tool that formats interview notes with one that ranks applicants for shortlisting. Both operate in HR, but their purpose and influence differ. Record what the person actually sees and whether the output changes who is considered. For an Annex III use, any Article 6(3) exception needs a reasoned assessment; profiling of people requires particular attention. A human clicking ‘accept’ is not enough reasoning.
Keep classification separate from commencement and readiness
Maintain three distinct questions: what category does this use fall into, when do the relevant provisions apply, and what preparation remains? An implementation date does not change the system's purpose or remove other applicable obligations. Likewise, a negative Annex III screen does not close the separate regulated-product, prohibited-practice or transparency analysis.
High-risk classification review prompts
Use this sequence to prepare a review, not to generate a pass/fail conclusion. Check the current official provisions and amendments before relying on an exception.
Download editable worksheet (CSV)No email or account required. Includes response, evidence, owner, review and follow-up columns for your team. This is a working aid, not an official form or proof of compliance.
On smaller screens, scroll the table horizontally or download the worksheet.
| Step | Evidence question | Record before concluding |
|---|---|---|
| Define the use | Which decision or activity does the output support? | Purpose, actual workflow and affected people |
| Test the listed category | Which precise Annex III point might apply? | Matching facts and facts that do not match |
| Assess influence | Can it materially influence the decision or outcome? | Interface, operating rules and realistic human review |
| Review exceptions | Is an Article 6(3) route genuinely supported, and is profiling involved? | Qualified rationale, uncertainty and required documentation |
| Check other routes | Is there a regulated-product, prohibited or transparency issue? | Separate assessments, not a single ‘low risk’ label |
| Approve and revisit | Who can decide, and what change would invalidate the reasoning? | Reviewer, date, source version and triggers |
Recommended next step
Use EU AI Fit's exposure check to identify triggers, then complete a source-linked classification in the workspace.
Run the free exposure checkRelated practical guides
Put the guidance into practice