A housing caseworker override of AI scores is never logged or analyzed. What should the auditor recommend?

In housing casework, human overrides of AI scores are a primary supervision signal: they show where the model conflicts with professional judgment, policy nuance, or fairness concerns. The auditor should recommend capturing and analyzing overrides—who overrode, why, outcome, and trends—so the municipality can improve the model, guidance, and training rather than treating overrides as invisible exceptions. Without that record, systematic model failure modes stay hidden while individual caseworkers absorb the friction alone.

Capturing override reasoning only in individual case files, without aggregating or analyzing it across cases, still leaves systemic model failure patterns invisible; the municipality has records but never turns them into model or policy improvement. Tracking override frequency in system audit logs without documenting reasons for regular supervisory review cycles fails because it records how often overrides happen but never captures why, leaving supervisors unable to tell sound judgment from systemic model failure. Monitoring model accuracy metrics separately from caseworker override decisions fails as incomplete supervision because it never connects model performance to human corrections, letting routine overrides and systemic bias stay hidden until they cause harm.

Exam caveat: override logging is a supervision control, not an attack on caseworker judgment—analysis improves both humans and models. Operational check: verify override events are recorded in the housing workflow, periodically reviewed for themes, and fed into model or policy remediation with clear ownership.