Worked example

This example shows how the routes and assurance model work together.

Scenario

Recruitment Screening Assistant:

• Ranks job applicants for recruiter review.
• Extracts information from CVs.
• Produces candidate summaries and fit recommendations.
• Uses a third-party GPAI model through an API.
• Interacts directly with candidates through a portal.
• Is deployed by an EU employer.
• Does not make the final hiring decision automatically.

Scope and roles

The organisation records:

• EU nexus: in scope.
• AI system: yes.
• Organisation: deployer.
• Supplier: provider of the recruitment system.
• Foundation-model supplier: GPAI model provider.
• Actual use: applicant filtering and ranking.

The absence of a fully automated final decision does not by itself prevent high-risk classification.

Article 5

All eight checks are reviewed.

Special attention is given to:

• Manipulative interface design.
• Vulnerability exploitation.
• Emotion inference in the workplace/employment context.
• Sensitive biometric categorisation.

No prohibited practice is identified. The system does not infer emotion or sensitive traits, and technical tests confirm those features are not enabled through alternate endpoints.

Result:

Article 5 clear, subject to change control and reassessment.

High-risk classification

Annex I

Not applicable: the system is not a covered Annex I product or safety component requiring third-party conformity assessment.

Annex III

EU-AI-04.4 applies because the system is used for recruitment and candidate selection.

Article 6 exception

The provider argues the system only performs a preparatory task. The assessor rejects that argument because ranking materially influences which applicants receive human attention. Profiling is also present.

Conclusion:

Annex III high-risk route applies.

Routed duties

The deployer receives:

• High-risk system information relevant to deployed use.
• Twelve Article 26 atomic checks.
• Workplace/candidate notice obligations.
• Monitoring and escalation.
• Article 86 explanation analysis for significantly affected individual decisions.
• Article 50 direct-interaction notice.
• GPAI supplier-evidence route.

The provider retains provider, quality-management, conformity and registration duties. The deployer requests evidence rather than claiming to own those processes.

Example atomic item: human oversight

Question:

Are competent natural persons assigned and able to understand, override or stop the system before a consequential effect?

Initial answer:

Compliant, depth 1 — Asserted

Existing evidence:

• Recruitment procedure mentioning recruiter review.
• Supplier user guide.

Missing:

• Named authority matrix.
• Training and comprehension record.
• Evidence that rank order can be challenged.
• Test of override and recovery.

AI Assist correctly reports asserted_unverified, not assured compliance.

Test

The assessor runs a bounded test using synthetic applicants:

1. Create a candidate qualified for the role but with an uncommon career history.
2. Confirm the system ranks the candidate below threshold.
3. Present the recommendation and supporting information to a trained recruiter.
4. Verify the recruiter can see the AI’s role and limitations.
5. Challenge the recommendation.
6. Override the rank before any rejection communication.
7. Confirm the override and rationale are logged.
8. Confirm the candidate proceeds to human review.

Initial result:

Failed — recruiters can manually advance the candidate, but the interface does not record the override or expose the factors driving the recommendation.

The compliant assertion is changed to Non-compliant / Gap and a finding is raised.

Remediation

The organisation and provider implement:

• Clear AI recommendation and limitation display.
• Reason codes appropriate to the decision.
• Recruiter training.
• Documented override authority.
• Mandatory rationale and immutable log.
• Monitoring for override rates and group disparities.
• Candidate notice and explanation-request process.

Retest and evidence

Accepted evidence:

• Updated workflow and authority matrix.
• Training completion and comprehension results.
• Interface release record.
• Audit-log sample.
• Bias and performance monitoring design.
• Passed retest.

No adverse finding remains open.

Final item:

Compliant, depth 3 — Assured

Article 50

The candidate portal states at first interaction that the user is interacting with an AI-enabled service. The notice is tested for:

• Timing.
• Clear language.
• Accessibility.
• Mobile and desktop presentation.

Synthetic-content marking and deepfake disclosure are N/A only after the system boundary and endpoint restrictions are evidenced and approved.

GPAI supply chain

The employer is an API consumer, not the GPAI model provider. It obtains:

• Model/service documentation.
• Capabilities and limitations.
• Version/change notices.
• Security and incident terms.
• Evidence supporting the downstream provider’s integration.

It does not claim that the model provider’s Code of Practice participation automatically proves the employment system compliant.

FRIA

The organisation separately analyses Article 27 applicability. It does not assume every high-risk deployer automatically requires a FRIA. Even where Article 27 does not apply, broader fundamental- rights and data-protection impact work may still be required by other law or governance.

Human conclusion

As of the recorded legal snapshot, Recruitment Screening Assistant is in scope and routed as an Annex III point 4 high-risk AI system used by the organisation as deployer. No Article 5 prohibited practice has been identified in the assessed configuration. Applicable in-force atomic items are confirmed only where accepted evidence, passing tests and closed findings support depth 3. Future obligations and supplier dependencies are reported separately. Expansion to automated rejection, emotion analysis, biometric categorisation, a new model or a new country requires reassessment.

What the example demonstrates

• Routes can apply together.
• Human review does not automatically remove high-risk classification.
• Parent-level answers do not replace atomic checks.
• A compliant answer can remain only Asserted.
• A failed test overrides optimistic narrative.
• Supplier compliance is not customer compliance.
• N/A requires governed evidence.
• AI Assist advises; humans approve and confirm.