Definition
Estimated reduction in claims frequency or severity attributable to AI-triggered loss control interventions, measured by comparing outcomes for intervention-received policyholders against a matched control group.
The Measurement Gap
Carriers use AI models to identify high-risk policyholders and trigger loss control interventions: inspections, engineering recommendations, premium credits for mitigation measures, or targeted safety communications. These programs have budgets, staffing, and performance metrics. But the fundamental question remains unanswered: do the interventions actually reduce losses?
The problem is attribution. A carrier sends loss control recommendations to 1,000 policyholders and observes that their claims frequency is lower the following year. But without a proper control group, this observation proves nothing. Perhaps those 1,000 policyholders would have had lower frequency anyway because of regression to the mean. Perhaps the model selected policyholders who were already improving. Perhaps the intervention had zero effect and the improvement was coincidental.
Legacy loss control metrics report activity volumes: inspections completed, recommendations delivered, compliance rates. They do not measure whether the activities caused the desired outcome. A loss control department can report impressive activity numbers while having no demonstrable impact on loss experience.
LCIE closes this gap by implementing a matched-control comparison design. Each intervention-received policyholder is matched with a comparable policyholder who did not receive the intervention. The difference in outcomes between the two groups, if statistically significant, is the intervention's causal effect. This transforms loss control evaluation from activity measurement to outcome measurement.
Decision Supported:
Are AI-triggered loss control interventions actually reducing losses, or is the model recommending interventions to policyholders who were low-risk regardless, and should loss control resources be reallocated to higher-impact targeting?
Formula
Plain Language: Match each policyholder who received an AI-triggered loss control intervention with a comparable policyholder who did not receive the intervention. Compare claims frequency between the two groups over 12 months. The difference is the intervention's effect.
Formula:
LCIE = (Frequency_control - Frequency_intervention) / Frequency_control x 100
Variables:
- Frequency_intervention: claims frequency for policyholders who received AI-triggered loss control intervention in the 12 months following intervention (claims per policy-year)
- Frequency_control: claims frequency for matched control group policyholders over the same period (claims per policy-year)
- Matching dimensions: risk score band, property type, geography, coverage type, policy vintage
- Positive LCIE means the intervention reduced claims (effective). Zero or negative means the intervention had no effect or targeted the wrong population.
Units:
Percentage reduction in claims frequency attributable to the intervention.
Data Requirements:
| Data Element | Source System | Classification |
|---|---|---|
| Loss control intervention records (who received, what intervention, when) | Loss control management system / underwriting notes | NATIVE |
| Risk scores at time of intervention targeting | Risk scoring model output | NATIVE |
| Claims frequency and severity post-intervention | Claims management system | NATIVE |
| Policyholder characteristics for matching | Policy administration system + underwriting data | NATIVE |
| Intervention compliance (did the policyholder implement the recommendation?) | Loss control follow-up records | MANUAL / NOT CAPTURED |
Proposed formula, pending validation pilots.