Cost Anomaly Investigation Rate (CAIR)

GMI-IT-018IT OperationsIntegration-Required

Definition

Cost Anomaly Investigation Rate (CAIR) is the proposed metric for this: the percentage of AI-flagged cost anomaly alerts that received a documented investigation (reviewed, triaged, and either resolved or explicitly dismissed with a stated reason) within the defined response SLA.CAIR closes the loop that anomaly detection systems deliberately leave open. Detection systems measure their own output. CAIR measures whether the organizational response function is keeping pace with that output: whether the human-AI feedback loop on financial anomalies is actually functioning or has become decorative.

The Measurement Gap

Your spend anomaly detection system is working. It generated 340 alerts last quarter. Each one represented a deviation from the expected spending pattern, a spike in data transfer costs, a sudden increase in compute in an account that usually runs quiet, an EC2 instance family the team doesn't normally use showing up in billing.

Your FinOps team investigated 41 of them.

The other 299 were dismissed, not acted on, or lost in the queue. Most of those 299 probably were legitimate anomalies: campaign launches, developer testing, a team spinning up a new environment without notifying FinOps. But 6 of them were a misconfigured auto-scaler that added 140 instances to a development account over 17 days before anyone noticed. The cost: $28,000. The time to detect: 17 days. The investigation status in the anomaly system: "pending," never opened.

The anomaly detection system did its job. It flagged the auto-scaler event on day one. The alert sat in the queue alongside 56 other alerts that week, competing for attention from a team that had learned (correctly, given past experience) that most alerts were either expected activity or low-priority spend that would self-correct. The team's triage heuristic was reasonable. The outcome was a $28,000 undetected cost event.

This is what alert fatigue looks like in FinOps: the AI anomaly detector generates alerts faster than the team can investigate them. Early in deployment, the team tries to keep up. Over time, investigation rates drop. The detection system appears to be working (alert count is high, the dashboard shows active monitoring) but the actual investigation rate has drifted to 12%, and the team only knows about the anomalies they chose to look at.

Legacy metrics don't reveal this because they measure detection, not response. Alert volume counts how many anomalies the system flagged: it doesn't measure what happened next. RI utilization, cost per unit, and spend vs. budget are all trailing indicators that eventually reflect uninvestigated anomalies, but weeks or months after the event, diluted across the broader spend base. There is no metric that tells you, right now, whether the anomaly detection system is actually influencing the organization's financial decisions or whether it's running as expensive background noise.

The measurement gap for CAIR is this: there is no metric that tracks whether AI-flagged cost anomaly alerts are actually being investigated. The detection side of the system has full instrumentation: every alert is timestamped, categorized, and stored. The response side is a black box. You don't know whether your FinOps team is keeping up, falling behind, or has effectively stopped engaging with the alert queue for anything but the most obvious events. Without CAIR, you manage the detection system and assume the investigation is happening. That assumption is where the $28,000 surprises come from.

Formula

Cost Anomaly Investigation Rate (CAIR) is the proposed metric for this: the percentage of AI-flagged cost anomaly alerts that received a documented investigation (reviewed, triaged, and either resolved or explicitly dismissed with a stated reason) within the defined response SLA.CAIR closes the loop that anomaly detection systems deliberately leave open. Detection systems measure their own output. CAIR measures whether the organizational response function is keeping pace with that output: whether the human-AI feedback loop on financial anomalies is actually functioning or has become decorative.

Formula: plain language:

Count all cost anomaly alerts generated by the AI system in a measurement period. Count how many of those alerts received a documented response (an investigation opened in the ticketing system, a root cause identified, or an explicit dismissal with a stated reason) within the response SLA. Divide the second number by the first. The result is CAIR: the percentage of alerts that were actually engaged with.

Formula: notation:

CAIR = A_investigated / A_total

Where:

  • A_investigated = anomaly alerts that received a documented investigation within the response SLA window
  • A_total = total anomaly alerts generated in the measurement period
  • Response SLA (proposed starting points): 48 hours for alerts with estimated impact >$1,000; 5 business days for alerts with estimated impact <$1,000Worked Example In Q3, the FinOps platform generated 280 cost anomaly alerts. Reviewing investigation records: 190 alerts had a documented investigation record (ticket opened, Slack thread linked, or dismissal noted in the platform) within the SLA window. 90 did not. CAIR = 190 / 280 = 67.9%. Segmenting by impact: of the 45 alerts with estimated impact >$1,000, 38 were investigated within 48 hours, a high-severity CAIR of 84.4%. Of the 235 alerts with estimated impact <$1,000, 152 were investigated within 5 business days, a low-severity CAIR of 64.7%. The aggregate number looks acceptable. The high-severity CAIR reveals that 7 significant alerts went uninvestigated. Those 7 are the ones that matter.

Sign convention:

Higher is better. 100% means every alert generated a documented response within SLA. Lower values indicate either alert fatigue (too many alerts relative to investigation capacity) or missing investigation documentation (alerts investigated informally but not recorded). Both are problems worth addressing: for different reasons and with different remedies.

Edge Cases

Auto-resolved anomalies (spend spikes that self-corrected before the SLA window closed) count as investigated only if someone reviewed the auto-resolution and confirmed it was expected behavior. Recurring anomalies from the same source: each occurrence counts separately. If the same misconfigured service generates a new alert each day, each day's alert should be investigated or the source should be resolved. Low-severity alerts with no estimated impact should be segmented separately or excluded from the primary CAIR calculation: mixing $50 alerts with $50,000 alerts in a single average obscures the metric's most important signal.

Cross-references:

Upstream: Analyst Trust Calibration Index (ATCI, GMI-IT-010). ATCI measures whether security analysts trust AI-generated alerts appropriately. CAIR applies the same trust-calibration dynamic to FinOps. When analysts lose confidence in the signal quality of cost anomaly alerts because the false positive rate is high, CAIR declines as a direct consequence. The mechanism is identical across domains: eroded trust reduces engagement with the alert queue. Lateral: Stranded Commitment Waste Rate (SCWR, GMI-IT-015). Some uninvestigated anomalies are commitment-related waste events that SCWR would capture eventually. High SCWR combined with low CAIR often reflects the same root cause: anomalous spending patterns that the detection system flagged and the investigation function didn't engage with. Downstream: Cost-per-unit and spend-vs-budget (legacy metrics). These will eventually reflect uninvestigated anomalies when the spend impact is large enough to move aggregate numbers, but with a lag of weeks to months and without identifying the source. CAIR is the early warning that those lags are accumulating.

Proposed formula, pending validation pilots.

At a Glance

Category
AI-Era IT Operations
← Compression Quality Regression Depth (CQRD)Cross-Client Pattern Detection Rate (CPDR) →

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