Definition
The fraction of technician decisions (both overrides and acceptances of AI recommendations) that were validated as correct by subsequent failure-report outcomes or by absence of the predicted failure within the prognostic window.
The Measurement Gap
AI-era maintenance creates a four-quadrant decision space (correct acceptance, correct override, incorrect acceptance, incorrect override). Override rate collapses all four into a single number and treats every override as a deviation. A senior technician who correctly overrides false alarms is penalized in the same dashboard as a junior who blindly accepts bad recommendations. Legacy compliance metrics (PM Compliance, schedule compliance, AI adoption rate) share the same defect: they treat following the recommendation as success regardless of outcome quality. F1, precision, and recall evaluate the model alone, with no view of the human decision quality that determines whether model output produces value.
Decision Supported:
Whether to invest in coaching versus model improvement versus workflow redesign, and which technicians are exercising expert judgment versus which need additional support.
Formula
Plain language: For every AI recommendation, the technician either accepted or overrode it. Each decision turned out to be right or wrong. OA is the share of decisions that turned out to be right, regardless of whether the technician followed the AI or overruled it.
Notation:
OA = (Correct Overrides + Correct Acceptances) / Total AI Recommendations
Variables:
| Variable | Definition | Unit |
|---|---|---|
| Correct Override (True Negative) | Technician overrode; asset did not fail as predicted within the prognostic window | Count |
| Correct Acceptance (True Positive) | Technician accepted; failure report confirmed predicted failure mode | Count |
| Incorrect Override (Miss) | Technician overrode; asset failed as predicted within the window | Count |
| Incorrect Acceptance (False Alarm) | Technician accepted and completed work; nothing was found wrong | Count |
| Total AI Recommendations | All AI-triggered work orders in the period | Count |
Stratify by technician, asset class, and AI confidence band to expose where expertise is strong and where training is needed.
Proposed formula, pending validation pilots.