Definition
Refresh Decision Payback Accuracy (RDPA) measures the percentage of AI-recommended hardware refreshes where the total cost of replacement was lower than the estimated cost of continuing to operate the old device: verified retrospectively using actual post-refresh and post-retention cost data. This is a proposed metric. It does not appear in ITAM standards bodies, ITIL frameworks, or vendor dashboards. What it does appear as is the question every ITAM director should be asking after implementing AI-driven refresh: did those recommendations pay off?The
The Measurement Gap
Here's how the story usually goes: The ITAM platform's AI refresh engine scores a fleet of laptops by predicted failure risk, performance degradation, and support ticket frequency. A batch of 85 machines shows elevated risk scores. The recommendation comes through: refresh now. IT procurement orders replacements. Budget is consumed. The old machines are wiped and sent to asset disposal.
Six months later, someone pulls the support data on that cohort. Half of those machines were two and a half years old: comfortably within their four-year lifecycle. The refresh model flagged them because a firmware update caused a temporary spike in performance metrics that correlated with failure in the training data. Actual failure risk on 2.5-year-old machines in that hardware generation: low. The model had no way to tell the difference between "this device is declining" and "this device just had a rough quarter." The organization spent $127,500 on replacements it didn't need that year.
What made this invisible: hardware refresh cycle adherence. It showed 94%, an excellent score. The AI recommended refreshes, and IT fulfilled them. From the metric's perspective, the system worked. Nobody measured whether those specific recommendations were economically justified.
This is the failure mode at the core of AI-driven hardware refresh: the model optimizes for device health and failure prediction, but the decision it generates (refresh now or extend) is fundamentally economic. Is the cost of continuing to operate this device greater or less than the cost of replacing it? Legacy ITAM metrics weren't built to answer that question.
The traditional metrics in this space track three things: refresh cycle adherence (did you refresh on schedule?), incident rate by asset age (are old devices causing more tickets?), and total cost of ownership (what did you spend on this device over its life?). None of them tell you whether a specific refresh decision (this device, this week) was economically correct.
Incident rate by asset age can be a starting point, but it's backward-looking. It tells you the average relationship between age and failure across your fleet, not whether this particular device's current health trajectory justifies premature replacement. The AI model can also be wrong in the other direction: keeping a device that's generating repeated support tickets because it hasn't crossed the refresh threshold yet, when replacement would have been cheaper months ago.
The measurement gap is straightforward but nobody plugs it: there is no metric that tells you whether an AI-recommended hardware refresh was economically optimal compared to extending the device's life. You know how many refreshes happened. You know cycle adherence. You don't know if the underlying recommendations were sound, and you won't know until the evidence accumulates across hundreds of decisions, which is exactly when a measurement framework would tell you you've had a problem.
Refresh Decision Payback Accuracy is the metric that closes this gap.
Formula
Refresh Decision Payback Accuracy (RDPA) measures the percentage of AI-recommended hardware refreshes where the total cost of replacement was lower than the estimated cost of continuing to operate the old device: verified retrospectively using actual post-refresh and post-retention cost data. This is a proposed metric. It does not appear in ITAM standards bodies, ITIL frameworks, or vendor dashboards. What it does appear as is the question every ITAM director should be asking after implementing AI-driven refresh: did those recommendations pay off?
The Formula
Plain language: For each device the AI recommended for refresh, compare what the refresh actually cost (new device + setup + migration + disposal) to what it would have cost to keep the old device running for its remaining planned lifecycle (projected support tickets × average resolution cost + productivity impact from degraded performance + risk-adjusted probability of unplanned failure × replacement cost). Score each recommendation as economically correct (replace < retain) or incorrect (replace > retain).
Notation:RDPA = (1/N) × Σ I(C_replace,i < C_retain,i)
Where:
- N = number of AI-recommended refreshes evaluated
- C_replace,i = total replacement cost for device i (acquisition + setup + migration + disposal)
- C_retain,i = estimated remaining operating cost for device i if retained (projected support costs + performance-adjusted productivity impact + risk-adjusted failure cost)
- I(·) = indicator function (1 if replacement was cheaper, 0 if not)
Units: Percentage. Higher is better. 100% means every recommendation was economically correct: the AI never told you to replace a device that would have been cheaper to keep.
Worked Example
Device: A 2.5-year-old laptop flagged for refresh. Replacement cost: New device ($1,400) + imaging and setup ($150) + data migration ($50) + disposal ($20) = $1,620.Estimated retain cost: The device generated 3 support tickets in the past year at an average resolution cost of $80/ticket = $240/year projected support. Multiplied by the 1.5 years remaining in planned lifecycle = $360. Add performance degradation impact: benchmark shows 15% productivity reduction on heavy workloads for this generation at this age; user is a developer, so estimated cost = 15% × $85/hour loaded rate × 2 hours/day × 250 days × 1.5 years = $9,563. Add risk-adjusted unplanned failure: 8% failure probability in the remaining lifecycle × $1,620 replacement cost = $130.Total estimated retain cost: $360 + $9,563 + $130 = $10,053.Result: Replace ($1,620) << Retain ($10,053). This recommendation scores as correct: replacement was economically justified. Repeat this calculation across 200 AI-recommended refreshes. If 160 score as correct, RDPA = 80%.
Edge Cases
A higher RDPA is always better. There is no case where lower RDPA is preferable. If the AI recommends replacing devices that would be cheaper to keep, the model is wrong and it costs money. Edge cases to handle explicitly: devices refreshed without an AI recommendation (manual decision by IT) should be excluded from RDPA and tracked separately. Devices where the recommendation was overridden and the device was kept. This is valuable data. Their actual remaining operating cost is now observable, and you can score the AI's recommendation in hindsight. Small samples (fewer than 10 refreshes in a category) make RDPA noisy: aggregate across quarters before drawing conclusions.
Cross-References
Upstream: Prediction Coverage Breadth (PCB, GMI-IT-027). If the AI model doesn't have sufficient telemetry on a device, its refresh recommendation is based on partial information: RDPA for low-coverage devices will be worse than for well-instrumented ones. Downstream: TCO per device (legacy metric). RDPA explains why TCO is what it is: whether the lifetime cost reflects good timing decisions or premature replacements.
Proposed formula, pending validation pilots.