Definition
Ratio of confirmed-actionable access anomaly alerts to total anomaly alerts generated, tracked by alert confidence tier.
The Measurement Gap
Measurement Gap and Decision
The Measurement Gap
Traditional access control generates binary alerts: badge read at unusual time, access at unusual door. These alerts are high-volume and low-specificity. AI-enhanced systems add confidence scoring and pattern matching, promising to surface the alerts that matter while suppressing the noise. But without AACI, security management has no way to verify whether the confidence tiers are calibrated. A "high-confidence" alert that is actionable only 30% of the time is not high-confidence: it is misleading labeling that erodes trust faster than having no confidence scores at all.
The additional gap is uninvestigated alerts. When AACI drops and fatigue sets in, the uninvestigated rate rises. A rising uninvestigated rate is itself a leading indicator of security exposure, not because the system missed something, but because the humans stopped looking.
The Decision It Supports
AACI tells security leadership whether the anomaly detection system is trustworthy at each confidence tier. If the high-confidence tier shows AACI above 70%, security teams can be directed to investigate every high-confidence alert. If it shows 25%, the confidence model needs retraining before anyone should rely on it. AACI also reveals whether the system improves over time as it ingests more data, or whether it stagnates at initial calibration levels.
Formula
Plain Language
For each anomaly alert, track whether the investigation outcome was confirmed actionable, benign, or uninvestigated. Compute the actionable rate within each confidence tier separately.
Formal Notation
$$AACI(t) = \frac{Confirmed_actionable(t)}{Total_alerts(t)} \times 100$$
Variable Definitions
| Variable | Definition | Unit |
|---|---|---|
| t | Alert confidence tier (High ≥80%, Medium 50-80%, Low <50%) | Tier |
| Confirmed_actionable(t) | Alerts in tier t confirmed as genuine security concerns requiring response | Count |
| Total_alerts(t) | All alerts generated in tier t during the measurement period | Count |
Units: Percentage per tier. Report AACI for each tier independently, a composite obscures whether the confidence model is meaningful. Also track the uninvestigated rate per tier as a companion metric.
Proposed formula, pending validation pilots.
At a Glance
Decision Supported
Is the access anomaly system generating alerts security teams can trust, or noise that trains the team to dismiss alerts?