Loss Prevention Intervention Equity Index (LPIEI)

GMI-RET-012Retail OperationsInfrastructure-Dependent

Definition

Ratio of CV loss-prevention false positive rates across customer segments, measuring whether shoplifting detection interventions fall equitably or concentrate on specific populations.

The Measurement Gap

CV-based shoplifting detection systems flag suspicious customer behavior and trigger security interventions, but most deployments do not measure whether the system flags certain customer segments disproportionately relative to actual theft rates. Bias in pattern recognition is well documented in adjacent domains, and training data for retail LP models inherits whatever biases existed in historical intervention data. If security staff historically intervened more frequently with certain groups, the model learns to replicate that pattern. Most LP platforms report aggregate detection counts and false positive rates but do not segment false positives by customer characteristics. The measurement gap is both technical and organizational: technically, the segmentation analysis is rarely performed; organizationally, retailers are reluctant to measure demographic patterns in security interventions because the findings carry legal and reputational implications. But not measuring carries greater risk. A retailer that does not know whether its AI security system is biased cannot fix the bias and cannot demonstrate due diligence.

Decision Supported:

Is the loss prevention AI flagging certain customer segments disproportionately relative to actual theft rates, and should the detection model be retrained or intervention protocols changed?

Formula

Formula:

LPIEI = max(FPR_segment) / min(FPR_segment)

Variables:

  • FPR_segment = False_positives_segment / Total_flags_segment for each demographic or behavioral segment
  • False_positive = security intervention triggered by AI where no theft was confirmed
  • Demographic segment = groupings identifiable in CV behavioral analysis (age, gender, and behavioral proxies). The recommended approach uses behavioral proxy segments (shopping duration, number of items handled, time of day) rather than attempting demographic classification directly.

Units:

Ratio. LPIEI = 1.0 means perfectly equitable false positive distribution across all segments. LPIEI = 3.0 means the most-flagged segment has three times the false positive rate of the least-flagged segment.

Edge Cases and Exclusions:

Low theft volume creates noisy false positive rates. Use 90-day rolling windows for stability. Compute LPIEI per store because a system-level average may hide store-level bias driven by local model tuning or store-specific training data. Require a clear definition of "false positive" (flagged, security engaged, no apprehension and no evidence of theft). This is the most sensitive metric in this catalog. The finding itself, if shared externally, carries brand risk. But not measuring it carries greater risk.

Proposed formula, pending validation pilots.

At a Glance

Unit
Ratio
Trend Direction
Toward 1.0; equitable false positives
Reporting Frequency
Quarterly
Category
In-Store AI Operational Effectiveness
← Supplier Lead Time Trust Index (SLTTI)Shelf-Gap Economic Priority Score (SGEPS) →

Assess Your Organization

The AI-GPS Assessment evaluates your AI governance maturity and recommends KPIs like this one for your implementation roadmap.

Take the Free AI-GPS Assessment