Definition
Percentage of AI-assessed claims where the computer vision model's attribution of damage to the current loss event versus pre-existing conditions was incorrect, measured by comparison against expert review.
The Measurement Gap
Computer vision models can identify damage types and estimate repair costs from photographs. What they struggle with is temporal attribution: determining whether visible damage was caused by the current loss event or existed before the event occurred. A roof with hail damage may also have pre-existing wear, previous storm damage, or deferred maintenance. The model must separate current from pre-existing to produce an accurate payment.
This temporal attribution challenge is unique to insurance. In manufacturing or infrastructure inspection, the question is "what is the condition now." In insurance claims, the question is "what changed because of this specific event." The model must answer both the detection question (what damage exists) and the attribution question (when did it occur).
Legacy damage estimation metrics report accuracy of the total damage estimate. They do not separately measure the accuracy of the temporal attribution component. A model can produce an accurate total estimate while systematically misattributing pre-existing damage to the current loss (overpayment) or current damage to pre-existing conditions (underpayment and potential bad faith exposure).
PEDMR closes this gap by sampling claims where the AI made a temporal attribution determination and comparing against expert review using baseline imagery and claim history. It decomposes misattribution into two directions: over-attribution (pre-existing damage counted as current loss, creating leakage) and under-attribution (current damage counted as pre-existing, creating customer harm and litigation risk).
Decision Supported:
Is the computer vision model correctly separating current-loss damage from pre-existing conditions, and should baseline imagery requirements be strengthened before relying on AI for damage attribution?
Formula
Plain Language: Sample claims where the AI made a damage attribution determination (new versus pre-existing). Have an expert reviewer independently assess the attribution using the same imagery plus additional evidence. Count misattributions and decompose by direction.
Formula:
PEDMR = Misattributed_claims / Total_attribution_sampled x 100
Decompose into:
Over_attribution_rate = New_damage_attributed_to_loss_but_pre_existing / Total_sampled x 100
Under_attribution_rate = Pre_existing_attributed_but_actually_new / Total_sampled x 100
Variables:
- Misattributed_claims: claims where the AI's attribution (new vs. pre-existing) disagreed with the expert reviewer
- Over_attribution: AI attributed pre-existing damage to the current loss (leads to overpayment)
- Under_attribution: AI attributed new damage to pre-existing (leads to underpayment/denial)
- Total_attribution_sampled: all sampled claims where the AI made an attribution determination
Units: Percentage. Track overall and directional rates.
Data Requirements:
| Data Element | Source System | Classification |
|---|---|---|
| AI damage attribution determination (new vs. pre-existing, with confidence score) | Computer vision platform | NATIVE |
| Expert attribution review | Structured audit program | MANUAL |
| Baseline / pre-loss imagery | Aerial imagery providers (Nearmap, EagleView), Google Earth, pre-inspection photos | INTEGRATION |
| Prior inspection reports for the property | Loss control / underwriting inspection records | NATIVE (if inspections exist) / NOT CAPTURED (if not) |
| Claim history for the property | Claims management system | NATIVE |
Proposed formula, pending validation pilots.