Definition
Measures the false positive rate of AI theft/anomaly detection separately by customer segment, revealing whether the model disproportionately flags specific populations (new solar, EV, low-income, multi-family).
The Measurement Gap
Traditional revenue protection tracks total investigations, confirmed theft cases, and recovered revenue. These fleet-level aggregates completely obscure whether specific customer populations are carrying a disproportionate share of false investigation burden. A fleet-level 40% false positive rate looks the same whether it's uniformly distributed or whether solar customers are being falsely flagged at 70% while non-solar customers are flagged at 25%.
The specific failure of legacy tracking is the absence of a segment dimension. AI anomaly detection models are trained on historical usage patterns. New solar installations, EV chargers, and time-of-use shifting all create consumption signatures that look anomalous to a model trained on pre-DER usage data. The model flags them as potential theft, not because they're stealing, but because their legitimate usage doesn't match the model's training distribution. Without RPSBR, nobody sees this pattern until a regulator or advocacy group files a complaint.
The operational decision this KPI enables: which segments require model recalibration, threshold adjustment, or an additional pre-investigation filter to reduce harassment of legitimate customers? This is a model governance decision, not a field operations decision, and it can only be made if the segment-level false positive rate is tracked.
What goes wrong without it: the utility installs 10,000 smart meters and deploys AI anomaly detection, solar adoption increases by 40%, and the revenue protection team starts initiating hundreds of false investigations against solar customers who simply have unusual consumption signatures. The regulatory complaint comes 18 months later. RPSBR makes this visible in the first 90 days. AI and AMI together enable this metric immediately: investigation outcomes already exist in the case management system, and customer segments exist in CIS.
Formula
RPSBR(s) = FalsePositive_investigations(s) / Total_AI_investigations(s)
Companion: Segment Disparity Ratio:
SDR = max(RPSBR(s)) / min(RPSBR(s))
Variables:
- s: Customer segment identifier (new solar install, EV charger, low-income/LIHEAP enrolled, high-usage commercial, standard residential). Source: CIS with AMI-derived segment flags
- FalsePositive_investigations(s): Investigations in segment s concluded as "no theft / legitimate usage". Source: Revenue protection case management system
- Total_AI_investigations(s): All AI-flag-initiated investigations in segment s in the measurement period. Source: Revenue protection case management system, AI anomaly detection log
- SDR: Ratio of the highest-RPSBR segment to the lowest-RPSBR segment. Source: Derived
Calculation Frequency: Quarterly, with annual trend review
Output: RPSBR per segment, 0–100%. Fleet average below 30% is typical for anomaly detection systems. SDR above 3.0 means the worst-targeted segment is falsely flagged at triple the rate of the best, a pattern that is difficult to defend in regulatory proceedings. SDR above 3.0 combined with RPSBR above 50% in the highest-flagged segment constitutes a material reputational and regulatory risk.
Proposed formula, pending validation pilots.
At a Glance
Decision Supported
Is the revenue protection AI creating segment-specific customer relationship damage, and which segments need model recalibration or investigation process changes?