Definition
Ratio of AI-flagged anomalies resolved per period to anomalies generated, measuring whether investigation capacity is keeping pace with detection output or a silent backlog is accumulating.
The Measurement Gap
A Gulf Coast water utility deploys AMI across 78,000 meters, acoustic leak detection on 40 miles of transmission mains, and NLP-based analysis of customer complaint patterns. Within the first year, the combined analytics generate 5,820 anomaly flags: suspected leaks, unusual consumption patterns, potential meter failures, pressure deviations, complaint clusters suggesting infrastructure problems.
The utility has two analysts assigned to investigate anomaly flags. Working at capacity, they resolve approximately 280 anomalies per month: investigating, confirming or dismissing, and creating work orders for confirmed issues. The analytics platforms generate approximately 485 new anomalies per month. The queue grows by 205 unresolved flags every month. After 12 months, the backlog stands at over 2,400 uninvestigated anomalies.
Those 2,400 flags represent analytics output with zero operational value. The AI detected something. Nobody looked at it. The detection may as well not have happened. Among those 2,400 flags are real leaks losing real water, real meter failures losing real revenue, and real infrastructure problems that are getting worse while sitting in a queue nobody has time to process.
No existing water utility metric captures this dynamic. Utilities track anomalies generated (a vanity metric that measures AI output volume), anomalies investigated (a workload metric that measures staff activity), and resolution time for investigated anomalies (a speed metric for the fraction that gets attention). None answer the question that determines analytics ROI: is the organization keeping pace with the intelligence its AI systems produce?
Without ART, a utility cannot distinguish between an analytics program where every flag gets investigated within a week and one where 60% of flags are silently abandoned to a growing backlog. Both programs report the same anomaly generation rate and the same average resolution time, because resolution time is only calculated for anomalies that were actually resolved.
Formula
Layer 1: Plain Language
Count how many AI-flagged anomalies were resolved this period: investigated and either confirmed with corrective action taken, confirmed but deferred with a documented reason and review date, or dismissed as a false positive with a documented rationale. Count how many new anomalies were generated this period. Divide resolved by generated.
Layer 2: Mathematical Notation
ART = R / G
Queue_Depth = Σ (Gₜ − Rₜ) cumulative across all periods t
Where:
- R = anomalies resolved in period (count; source: anomaly tracking workflow)
- G = anomalies generated by AI systems in period (count; source: analytics platform logs)
- Queue_Depth = cumulative unresolved anomalies (count; derived)
"Resolved" requires one of three documented outcomes: confirmed and corrective action taken, confirmed and deferred with documented reason and review date, or dismissed as false positive with documented rationale. "Ignored" (no investigation, no documentation) is not resolved.
Proposed formula, pending validation pilots.