01
Start with job relatedness
If you cannot defend why a competency predicts success, do not score it with a model. AI accelerates structure; it does not invent validity.
02
Monitor selection rates by process stage
Track who advances after each AI-assisted step. Version assessments so you can reconstruct what candidates experienced. Pair quantitative monitoring with rubric quality reviews.
03
Keep humans in the loop for adverse actions
Silent auto-reject increases risk. Require human review, capture override rationale, and offer accommodation paths.
