Definition
Prediction Coverage Breadth (PCB) measures the percentage of the total managed asset estate for which the AI lifecycle prediction model has sufficient telemetry to generate a confident refresh or lifecycle recommendation: distinguishing what the model can predict from what the model was expected to predict. This is a proposed metric. It doesn't appear in ITAM standards, CMDB coverage metrics, or AI model performance reporting. It's the question that should follow every AI model accuracy report: accurate for what fraction of the assets we actually manage?The
The Measurement Gap
The ITAM platform's quarterly review deck showed the AI lifecycle management model at 91% accuracy on hardware refresh predictions. The ITAM director presented this to IT leadership as evidence that the AI investment was paying off: the model correctly predicted failure or performance degradation for 91% of the devices it evaluated, substantially better than the old calendar-based schedule that had no predictive capability at all.
Three months after that presentation, a batch of lab machines started failing in clusters. High-end workstations used by engineers for simulation and rendering (averaging $3,200 each) had no telemetry in the endpoint management system because they ran a Linux distribution the Intune agent didn't support. The ITAM platform listed them as managed assets. The AI lifecycle model had no data on any of them.
When the failure wave hit, the organization replaced 34 lab machines on an emergency basis: premium pricing, expedited deployment, disruption to two product teams. Total unplanned cost: approximately $135,000. If those refreshes had been planned 6 months earlier through standard procurement channels, they would have cost around $89,000. The gap: $46,000 directly attributable to the model's blind spot on a device category nobody had checked for coverage.
The 91% accuracy figure was real. It just applied to the 68% of the asset estate the model actually covered.
This is the failure mode PCB measures: AI lifecycle management accuracy is reported for instrumented devices, but the size of the uninstrumented estate goes unmeasured. ITAM leadership believes the AI manages hardware refresh for the entire organization. The AI actually manages it for the subset of devices with compatible endpoint agents and sufficient telemetry history. Everything else runs on implicit calendar schedules, or runs until it fails.
Legacy discovery metrics track CMDB completeness: whether every physical device is listed. But completeness of the inventory and completeness of the AI model's prediction coverage are different things. A device can appear in the CMDB with full inventory accuracy and still have zero telemetry, which means the AI cannot make a confident lifecycle prediction for it. Standard discovery accuracy metrics don't distinguish between "we know this device exists" and "we know enough about this device's health to predict when it needs replacement."
The measurement gap is straightforward: there is no standard metric that tells you what percentage of the managed asset estate the AI lifecycle model can actually cover with reliable predictions. Model accuracy percentages are computed on the coverage that exists: they're not scaled to the estate the model was deployed to manage. When the ITAM team reports 91% accuracy, the right follow-up question is "91% of what?" Most ITAM teams don't have a clean answer.
Prediction Coverage Breadth makes that question answerable.
Formula
Prediction Coverage Breadth (PCB) measures the percentage of the total managed asset estate for which the AI lifecycle prediction model has sufficient telemetry to generate a confident refresh or lifecycle recommendation: distinguishing what the model can predict from what the model was expected to predict. This is a proposed metric. It doesn't appear in ITAM standards, CMDB coverage metrics, or AI model performance reporting. It's the question that should follow every AI model accuracy report: accurate for what fraction of the assets we actually manage?
The Formula
Plain language: Count the managed assets that fall within the AI lifecycle model's intended scope. For each one, check whether the model has enough data (sufficient telemetry history above the model's confidence threshold) to generate a prediction. Express as a percentage.
Notation:PCB = A_predicted / A_total
Where:
- A_predicted = count of assets for which the AI model generates a lifecycle prediction above its confidence threshold
- A_total = count of assets in the managed estate that fall within the model's intended scope (asset types the model was designed to cover)
Units: Percentage. Higher is better. 100% means the model covers every asset it was designed to cover.
Worked Example
Managed estate by asset type (within model scope):Corporate laptops: 1,200 total. Endpoint agent active and reporting: 1,140. Above model confidence threshold: 1,080.Desktop workstations: 340 total. Model coverage: 310. Above threshold: 285.Department-managed servers: 80 total. Model coverage: 71. Above threshold: 64.Lab and engineering workstations: 190 total. Model coverage: 52 (Intune agent incompatible with Linux for 138 devices). Above threshold: 44.Conference room equipment: 95 total. Model coverage: 12 (standalone devices, no management agent deployed). Above threshold: 9.A_total = 1,905. A_predicted = 1,080 + 285 + 64 + 44 + 9 = 1,482. PCB = 1,482 / 1,905 = 77.8%.The AI model the ITAM team thought was managing 1,905 assets is making confident predictions for 1,482 of them. The remaining 423 assets (22.2% of the estate) are operating on implicit calendar-based refresh because the model either lacks telemetry or lacks sufficient telemetry history to clear the confidence threshold. Knowing where that 22% lives is the first step toward managing it.
Edge Cases
Higher PCB is better, with the important qualification that the target should be defined by asset type and context. Corporate laptops in a standard managed environment should reach 90%+. Lab machines, conference room equipment, and IoT-adjacent assets may have structurally lower coverage, not because the organization failed, but because those device types are not compatible with standard endpoint telemetry approaches.
Edge Cases
Asset types the model was never designed to cover (printers, IP phones, IoT sensors) should be excluded from A_total, but documented separately as "uncovered by design" so the coverage boundary is explicit. Assets deployed fewer than 90 days ago should be excluded: insufficient operating history for the model to establish a performance baseline. CMDB ghost records (assets listed in inventory but physically retired or missing) should be flagged and excluded from both numerator and denominator; PCB computed against a dirty CMDB is meaningless.
Cross-References
Upstream: Discovery accuracy (legacy metric). You cannot predict lifecycle for devices you don't know exist. CMDB completeness is a prerequisite for accurate PCB. If 15% of the physical estate isn't in the CMDB, PCB's denominator is understated. Downstream: Refresh Decision Payback Accuracy (RDPA, GMI-IT-024). RDPA is only meaningful for covered assets: the refresh decisions the AI makes for uncovered devices are either not made at all (the device runs until failure) or made on a calendar basis without AI input. Tracking RDPA separately for covered vs. uncovered cohorts reveals how much of the AI's economic value is being left unrealized by coverage gaps.
Proposed formula, pending validation pilots.