| Takeaway | Detail |
|---|---|
| Use the empirical P90, not mean MTTR or a flat 30-minute target | The 90th–95th percentile reflects the long right tail that a single average can hide in positive, variable repair times. |
| Check the log-normal fit against the empirical mean before trusting it | Log-normal fits often dominate repair-time data; if the log-normal estimate differs materially from the empirical mean, revisit the distribution. |
| Treat the 95th percentile as leaving 5% above the threshold | By definition, 95% of observations fall below the 95th percentile and 5% fall above it. |
| Pad only when missed-response cost exceeds holding cost | Set the buffer at empirical P90 first, then pad only if (P90 − median) × expected daily dispatch volume × cost per missed response exceeds the daily cost of holding. |
Set the dispatch buffer at the empirical P90 of your logged repair times, not at mean MTTR or a flat 30-minute response target.
Pad that buffer only if (P90 − median) × expected daily dispatch volume × cost per missed response exceeds the daily cost of holding.

Why the Mean MTTR Breaks Buffers
Repair time is a positive quantity, and in practice it carries a long right tail: most jobs finish near a typical value while a minority run much longer. According to the Maintainability Engineering treatment at eduKateSG (Nov. 1, 2026), that shape is driven by six delay factors—diagnosis, access, modularity, tools, spares, and verification. Each factor can add waiting, rework, or dependency time that is not captured by an “average” job, so the risk a buffer must cover lives in the upper tail rather than at the center.
A mean MTTR is a center-of-mass summary. It can describe overall repair performance, but it does not show how long the slow jobs take or how often they occur. The same eduKateSG source notes that a single average can hide a long right tail, which is why engineers use distributions and percentiles rather than means alone when sizing buffers.
Percentiles also keep the original measurement unit. As Wikipedia defines it, a percentile is a score greater than a given percentage of scores, and percentiles are expressed in the same unit as the input scores—not in percent. So P90 is a repair time in minutes: the value below which 90% of logged repair jobs fall, leaving the remaining 10% above it.
That is why a mean-based buffer breaks: it is tuned to the center while missed-response cost is generated by the tail. A buffer set at mean MTTR will be exceeded by every job above the mean, including outliers driven by diagnosis, access, spares, or verification delays. The practical check is to inspect the empirical repair-time distribution and compare the mean with the median; a wide gap signals that the tail, not the average, should govern the buffer.
Use the mean for availability or trend reporting if needed, but use a percentile for the time you promise to hold a dispatch slot. The buffer question is tail coverage, not center performance.

Fit Log-Normal, Check Empirical
Define the MTTR clock before any percentile work: decide whether time starts when the fault is reported, when the technician is en route, or when repair starts. PreventiveHQ's September 6 MTTR guidance stresses that the clock choice changes the average and any comparison, because each start point includes or excludes dispatch, travel, and access time. A percentile computed from one clock cannot be compared with a target written for another.
| Clock start | Includes before hands-on repair | Typical use |
|---|---|---|
| Fault reported | Dispatch, travel, access | Customer-facing response target |
| Technician en route | Travel and access | Dispatch policy review |
| Repair started | None | Pure maintainability study |
MetricGate's May 20 MTTR Calculator entry notes that log-normal fits often dominate repair-time data, but dominance is a hypothesis to test, not a default. Fit the log-normal, compute its mean, and compare that fitted mean with the empirical mean of your logged repair durations. If the two differ materially, reject the fit for buffer work and fall back to the empirical distribution.
Compute percentiles and confidence intervals from your own repair-duration log rather than vendor defaults or sector benchmarks. MetricGate's calculator also links MTTR to MTBF for availability context, which is useful for reporting, but the buffer percentile should come from your own durations and your own clock. Vendor defaults describe someone else's process, tooling, and spares position.
A material gap means the fitted curve is not reproducing the center of your data, so its tail percentiles are not trustworthy for deciding coverage. In that case, use empirical percentiles and their confidence intervals, and revisit the fit only after correcting the clock definition, data quality, or mix of repair types. The check is cheap; an unverified fit is not.
Buffer Options Compared
Use this comparison as a screening step before you write a dispatch rule. The question is not which buffer sounds disciplined; it is which one is tied to your logged repair times and which one leaves the long jobs outside the reserve. MetricGate's MTTR calculator guidance pairs empirical MTTR with percentiles and SLA comparison, so the table below uses that same empirical basis rather than a generic target.
| Option | Buffer basis | Practical appeal | Exposure | Verdict |
|---|---|---|---|---|
| Option A — Mean MTTR buffer | Logged mean MTTR | Cheapest on paper | Misses every tail job | Loses |
| Option B — Empirical P90 buffer with conditional pad | P90 from own repair-time log, pad only under separate cost test | Covers 90% of jobs at near-mean cost | Jobs above P90 remain exposed | WINNER — Option B |
| Option C — Flat 30-minute response target | Fixed convention | Simple to state | Defensible only if measured P90 ≤ 30 min | Verify before committing |
For Option A, the check is simple: if the buffer equals the logged mean, it is a mean-based buffer. It is the cheapest on paper because it consumes the smallest reserve, but it loses in this comparison because it misses every tail job. Do not rescue it by calling the mean a service level; keep it as a summary statistic, not the dispatch buffer.
For Option B, pull P90 from your own repair-time log and use that value as the starting reserve. Wikipedia's percentile definition says a P90 value is greater than approximately 90% of observations, which is why this option covers 90% of jobs without pretending the maximum is routine. MetricGate also lists percentiles alongside MTTR estimates, supporting the use of the empirical distribution as the control input. Add a pad only through the separate cost test, not as a reflex.
For Option C, verify before committing. Pull your own P90 and compare it with the 30-minute target. If measured P90 ≤ 30 min, the flat target is defensible for your fleet; if it is higher, the target is a convention, not a measured property of your fleet. Opsima's MTTR guide frames MTTR as actionable through a playbook, which supports checking the target rather than adopting it unchanged.
When you record the decision, write the buffer basis, the P90 value used, and whether any pad was applied. That record makes the next review mechanical: rerun P90 from new logs, recheck the 30-minute target only if you still use it, and reject any switch back to a mean-only buffer or an unverified flat target. The comparison should remain auditable from the log to the reserve.
The Pad-or-Dispatch Inequality
The pad-or-dispatch decision is a single inequality: compare tail exposure cost L = (P90 − median) × D × C with pad cost R = the daily cost of holding one reserve technician slot. If L is less than R, dispatch at the empirical P90 and absorb the tail; if L is greater than or equal to R, pad with a reserve slot. Treat this as a method formula and plug in your own logged repair times and costs rather than benchmark numbers.
| Term | Meaning |
|---|---|
| L = (P90 − median) × D × C | Tail exposure cost: minute gap × expected daily dispatches × cost per missed response |
| R = daily cost of one held slot | Pad cost: wage + vehicle + opportunity cost of idle time |
| Decision | L < R: dispatch at P90; L ≥ R: pad with a reserve slot |
On the left, (P90 − median) is the extra minutes in the tail beyond a typical repair, D is expected daily dispatch volume, and C is cost per missed response, including penalty, expedite, and goodwill. Multiplying the three gives the daily cost of letting the P90 buffer be exceeded without a reserve slot in place.
On the right, R is the daily cost of holding one reserve technician slot: wage plus vehicle plus the opportunity cost of idle time. This is the price of padding the schedule so a tail event does not become a missed response.
Compare the two on the same time basis and the same cost basis. The inequality is directional, not a universal rule: a site with low missed-response cost and high idle-slot cost will often dispatch at P90, while a site with the opposite profile will often pad. Because both sides depend on local logs and local cost accounting, use your own figures; do not substitute industry benchmarks for either side.
Use the same logged repair-time clock for both P90 and median, and use the same operating period for both cost estimates. If some cost components are not yet tracked, build the inequality with the components you can attribute and revisit it as the cost model improves.
What the Evidence Does Not Establish
No grounding source in this packet supplies a sector-level P90 repair-time benchmark. Opsima (May 18, 2026) states that industry benchmarks by sector exist, but the specific figures are unavailable here, so check that table directly before citing any number and treat any borrowed figure as unverified until you confirm the sector, sample, and repair-time definition match your operation.
The "30-minute response" target is a service convention, not an evidence-backed constant. Treat it as a hypothesis: compare your empirical P90 to 30 minutes. If your P90 is above it, the target fails for your logged process; if it is below, that only shows your current logs clear the convention, not that 30 minutes is a validated standard.
Percentile buffers also assume stationarity. A fleet mix change, a new equipment class, or new hires can invalidate the percentile until you re-estimate it from fresh logs. Do not carry an old P90 across a material process change, and do not pad a stale percentile as if it were current evidence.
| Claim or limit | What the evidence supports | What you must verify |
|---|---|---|
| Cross-industry P90 benchmark | No source here provides one; opsima (May 18, 2026) claims sector benchmarks but figures are unavailable | Open the opsima table and confirm sector, sample, and definition before citing |
| 30-minute response standard | Service convention only | Compare it with your empirical P90 and record pass or fail |
| Stationarity of the percentile | Assumed, not proven | Re-estimate after fleet, equipment, or staffing changes |
Practical check: before using any external threshold, record its source, date, sector, and repair-time definition. If any element does not match your logs, discard the number and set the buffer from your own empirical P90.
Therefore, treat external benchmarks as hypotheses and set every threshold from your own repair-time logs.
60-Job Log
This worked example uses one 60-job repair log and is meant to be recalculated, not copied: 60 completed jobs, median repair time 42 minutes, P90 78 minutes, 12 dispatches per day, $180 missed-response cost, and $340 per day to hold one reserve slot. Every figure is a method input from this log, so replace each value with your own before using the result.
| Item | Value |
|---|---|
| Completed jobs | 60 |
| Median repair time | 42 min |
| P90 repair time | 78 min |
| Dispatches per day | 12 |
| Missed-response cost | $180 |
| Reserve slot cost | $340/day |
Checkpoint 2 is the pad decision. Using the log above, (P90 − median) = 78 − 42 = 36 minutes, so L = 36 × 12 × $180 = $77,760 per day of tail exposure, which exceeds R = $340 per day to hold one reserve slot. Because L ≥ R, pad one reserve slot. The cost of being wrong by padding unnecessarily is about $340 per day; the cost of being wrong by dispatching bare is the tail exposure the same arithmetic places at $77,760 per day.
The decision rule is comparative, not automatic. If your own log has a smaller P90-minus-median gap, lower dispatch volume, or lower missed-response cost, the inequality can flip and bare dispatch can be the cheaper error. Re-run the arithmetic when the log changes, and do not carry the $77,760 figure into another site's decision.
With 60 completed jobs, treat the 78-minute P90 as a working estimate rather than a settled population value, and refresh it as the log grows. The buffer should follow the empirical P90 of your own repair times, with padding justified only by the inequality above.
Worked Example: Run the Numbers
For this example, the winner is the empirical-P90 base buffer: it is anchored to the operation's own upper-tail experience rather than to a mean-based buffer or to the flat 30-minute target. Padding is not automatic; it is governed by the pad-or-dispatch inequality computed above, where the $77,760 tail exposure exceeded the $340 daily reserve cost.
Before using the result, check the fit: if the log-normal estimate differs materially from the empirical mean, do not trust the fitted percentiles and use the empirical P90 directly, consistent with MetricGate's May 20, 2026 MTTR guidance. Keep the MTTR clock definition unchanged across the median, P90, and any comparison, because PreventiveHQ's September 6, 2026 MTTR guidance stresses that clock choice changes the average.
Decision Rules for Buffer Setting
The protocol is one if/then chain: set the dispatch buffer at the empirical P90 of your logged repair times, where P90 is the repair time exceeded by about 10% of jobs, add a pad only when the pad-or-dispatch test says reserve capacity is cheaper than missed responses, and re-fit on a defined trigger so percentile choice, padding, and maintenance stay bound together.
Run the fit check before quoting any percentile. MetricGate's May 20, 2026 MTTR guidance says log-normal fits often dominate for repair-time data; if the fitted mean differs materially from the empirical mean, discard the fitted percentiles and use empirical order statistics until the fit is repaired. Treat the gap as material when it could move the buffer or the published target; do not blend fitted and empirical values.
For the response promise, use the 30-minute target only when the buffer supports it: if empirical P90 ≤ 30 minutes, dispatch at P90 with no pad and publish the 30-minute response target. If P90 > 30 minutes, do not promise 30 minutes; publish P90, or the padded buffer if the pad test adds one. The external commitment then tracks the observed tail, not a flat target.
Re-fit at the next log closeout after the fit is repaired, and at any later closeout where job mix, data quality, or the repair-time definition changes; if the mean gap reopens, switch back to empirical order statistics immediately. Keep the published target tied to the controlling value: empirical P90 when the fit fails, fitted P90 only while the fit passes. With 60 completed jobs, treat the 78-minute P90 as a working estimate and refresh it as the log grows.
Operationally, a pad is reserve capacity added to the buffer, and publish means the response target you communicate internally or externally. If a pad is added, move the published target to the padded value; if no pad is added, leave the target at P90 or at the 30-minute figure allowed by the first rule.
Frequently Asked Questions
What should I set as the dispatch buffer before considering any pad?
Set the dispatch buffer at the empirical P90 of your logged repair times, not at mean MTTR or a flat 30-minute response target.
When is it justified to pad the empirical P90 buffer?
Pad that buffer only if (P90 − median) × expected daily dispatch volume × cost per missed response exceeds the daily cost of holding.
What does choosing the 95th percentile imply about observations above the threshold?
By definition, 95% of observations fall below the 95th percentile and 5% fall above it.
Why does mean MTTR break buffers for repair-time data?
Repair time is a positive quantity, and in practice it carries a long right tail: most jobs finish near a typical value while a minority run much longer.
When should a log-normal fit for repair times be questioned?
Log-normal fits often dominate repair-time data; if the log-normal estimate differs materially from the empirical mean, revisit the distribution.
Which percentile band reflects the long right tail hidden by a single average?
The 90th–95th percentile reflects the long right tail that a single average can hide in positive, variable repair times.
Quick answers
| How should you set the dispatch buffer? | Set the dispatch buffer at the empirical P90 of your logged repair times, not at mean MTTR or a flat 30-minute response target. |
| Why does mean MTTR break buffers? | Repair time is a positive quantity, and in practice it carries a long right tail: most jobs finish near a typical value while a minority run much longer. |
| How can you verify a log-normal fit for repair-time data? | Check the log-normal fit against the empirical mean before trusting it, and if the log-normal estimate differs materially from the empirical mean, revisit the distribution. |
| What does the 95th percentile represent? | By definition, 95% of observations fall below the 95th percentile and 5% fall above it. |
| When should you pad the dispatch buffer? | Pad only when missed-response cost exceeds holding cost, and only if (P90 − median) × expected daily dispatch volume × cost per missed response exceeds the daily cost of holding. |
Also worth reading: Reassign or hold: computing a per-dispatch technician break-even: Reassign or hold: computing a · Warranty checks for field service: 28-second verify vs blind dispatch: Warranty checks for field service: · On-device redaction for field-service photos: 120ms vs service level, redact or skip: On-device redaction for field-service photos: