Definition
Measures the percentage of AI-generated root cause attributions from digital thread analysis that are confirmed correct by engineering investigation.
The Measurement Gap
Traditional root cause analysis in manufacturing is engineering-driven. A process engineer examines a defect, reviews process data, applies domain knowledge about physical mechanisms, and identifies the causal chain. The analysis is slow but grounded in physics. The engineer knows the difference between "the temperature was high when the defect occurred" and "the high temperature caused the defect through thermal stress on the coating."
AI-powered digital thread analysis offers speed and pattern detection across vastly more variables than a human can track simultaneously. But AI systems identify statistical associations, not physical mechanisms. When an AI reports that "parameter X is the root cause of defect Y," it means parameter X has the strongest statistical association with defect Y in the data. That association may reflect true causation, or it may reflect a shared underlying cause, a seasonal pattern, or a data artifact.
Without RCAA, organizations have no way to know how often the AI gets causation right versus merely identifying correlation. Teams that act on every AI attribution without validation risk implementing corrective actions on symptoms rather than causes. Teams that ignore AI attributions entirely lose the speed advantage. RCAA provides the trust calibration that lets organizations decide how much engineering validation each AI attribution requires before corrective action.
Formula
RCAA is calculated by sampling AI-generated root cause attributions and scoring them against engineering investigation findings:
RCAA = Attributions_confirmed_correct / Attributions_sampled x 100
Where:
- Attributions_confirmed_correct = AI root cause attributions confirmed as causally correct by engineering investigation (count)
- Attributions_sampled = total AI root cause attributions reviewed in the measurement period (count)
A score of 100% means every sampled AI attribution was causally correct. Lower scores indicate more correlation-not-causation errors.
Engineering confirmation requires domain expertise: for each sampled attribution, a process or quality engineer assesses whether the identified root cause is causally correct (the parameter caused the failure through a known physical mechanism), correlated but not causal (the parameter moved with the failure through a shared underlying cause), or incorrect (no meaningful relationship). Track these three categories separately because the risk profile differs. A "correlated but wrong cause" attribution may still point toward the right area of investigation, while a completely incorrect attribution leads engineering effort in the wrong direction entirely.
Minimum sample size is 20 attributions per measurement period. Track RCAA by attribution type (process parameter, material lot, equipment, environmental condition) because the AI may be accurate for some attribution categories and unreliable for others.
Proposed formula, pending validation pilots.