High-risk under the EU AI Act, Annex III (employment and worker management), because the system materially influences shortlisting decisions about people. Also mapped to: UK (ICO guidance and the principles-based regime), US (NIST AI Risk Management Framework and New York City Local Law 144 bias-audit duty), and the equivalent expectations in the UAE, Canada and Australia.
| Attribute | Detail (illustrative) |
|---|---|
| Model type | Gradient-boosted ranking model over structured CV features, with a text encoder for free text. |
| Training data | 180,000 historic applications (2019 to 2025), de-identified, with documented lineage and lawful basis. |
| Human oversight | A recruiter reviews every shortlist; overrides are logged; no automatic rejection. |
| Transparency | Candidates are told AI assists screening; an explanation is available on request. |
| Metric | Result | Threshold | Status |
|---|---|---|---|
| Accuracy (top-20 precision) | 0.81 | at least 0.75 | Pass |
| Demographic parity ratio (gender) | 0.86 | at least 0.80 | Pass |
| Equal-opportunity gap (ethnicity) | 0.11 | 0.10 or less | Watch |
| Data-governance record completeness | 72% | 100% | Gap |
| Risk | Mitigation | Owner |
|---|---|---|
| Historic hiring bias amplified by the training data | Quarterly fairness testing, reweighting, protected-group monitoring | People Ops |
| Recruiters over-relying on the ranking | Mandatory review step, override prompts, periodic audit of overrides | Talent lead |
| Incomplete data-governance evidence | Close documentation gaps to lift Confidence above 90 | Data owner |
Assessed for the stated role families and regions only. Performance on roles or populations outside the training distribution is not assured. Re-assessment is required after any material change to the model.
Conditional Cleared for continued use provided the two evidence gaps above are closed within 60 days. Next scheduled review: December 2026, or on any material change, whichever is sooner.