Abstract
Every measure of a successful readiness or planning programme is a reduction in visible failure. Fewer schedule breaks, fewer emergency purchases, fewer aborted jobs, fewer clarification requests. Success therefore presents as an absence, which is the hardest thing to defend in a budget cycle and the easiest to attribute to luck. This paper argues that the measurement problem is the primary reason planning is chronically underfunded, and proposes counting the caught failure rather than the prevented one.
The best day you will never hear about
I once tried to think of a great story about a down day that went perfectly, and I could not come up with one.
The good ones do not get remembered. We changed the bearing, everything was staged, the lockout was done, everybody knew, it went smooth, we handed it back early and went home. The end.
Nobody tells that one at the bar.
The oilfield was often like that. You were handed a box with every part you could possibly need, and it was rarely wrong, so the work just went. That was not luck. It was a system somebody had built properly, and the reward for building it properly was that nothing memorable ever happened.
A perfectly planned day is boring, and boring is the scoreboard. The goal of designing maintenance is to make it unremarkable, and the positive proof is not a war story. It is the absence of one.
Now try to put that in a budget submission.
The attribution trap
A plant that prevents six unplanned outages cannot prove it prevented them. The counterfactual is unavailable, the outages did not happen, and the person responsible has no artifact to point at.
Meanwhile the man who recovered a line at three in the morning has a visible, dramatic, gratitude-generating artifact, and he will be promoted.
This is not a cultural failing that exhortation fixes. It is a structural bias toward measurable heroics, and it reproduces in every organization that does not deliberately instrument against it.
It has a second-order effect that is worse and less often noticed. If the only legible achievements are recoveries, then the people who are good at prevention learn that their work is invisible, and some of them stop doing it. Not consciously. They simply drift toward the work that gets seen.
Measure the near miss, not the outcome
The way out is to stop trying to measure prevented failures and start counting caught ones.
A readiness review that stops a job produces an artifact at the moment it stops it, and that artifact is countable. It requires no counterfactual and no claim about what would have happened.
Four counts are enough:
- Jobs stopped before release, and which gate stopped them.
- Discovery tasks raised and closed, meaning unknowns converted to knowns on purpose.
- Assumptions carried into the field and named, rather than dissolved into the plan.
- Parts kits found incomplete before the crew mobilized.
None of those claims a prevented outage. Each is a documented instance of the system doing the job it was funded to do, produced as a byproduct of doing it.
The research gap is instructive here and it is not a coincidence. Crespo-Márquez and Gómez Fernández, reviewing agentic AI across the maintenance management framework, find empirical evidence concentrated in technical phases while managerial and strategic phases such as KPI alignment, asset criticality and lifecycle costing remain largely unexplored. The managerial half of maintenance is under-researched for the same reason it is underfunded: its outputs are absences, and absences are hard to publish about too.
The cheapest detection system I ever built, and it counted nothing
Water bottling plant, high-speed fillers. One of them piled up and I was working the repair. Because it is food production you account for everything, every bolt, washer and tool, so nothing gets left inside to contaminate product.
I pointed at an empty bolt hole and told the production guy helping me that we needed a bolt matching the others. He pulled one out of his pocket and said he had one right here.
Where did that come from? He had found it on the floor earlier, sweeping up.
Then he said: maybe this is the one that is supposed to be there.
Yeah. No shit.
Out of that I built a standing rule. Any bolt, spring, or anything out of the ordinary gets carried to the shop and tagged with where it was found. Not swept up, not pocketed. A bolt on the floor is the machine telling you something is backing out, and on high-speed gear a loose thing grenades fast.
That rule caught a stack of problems before they were problems, and it told me where to point PMs. I widened it to odd sounds and odd smells, reported to me immediately. It was my first PM programme before I had heard the term.
Now the measurement point. What did that rule produce, on paper? Nothing. No downtime avoided that anyone could name, no failure prevented that anyone could point to. A tagged bolt in a shop is not a KPI. Every single one of those catches was invisible to every system in the building, and the entire value of the programme existed only in things that did not happen.
Under the four counts above, it would have produced a number every week.
The thing worth having is not the saving
There is a related trap on the other side of a success.
I rebuilt the PM routing in a wood room once. The CMMS was generating individual work orders that marched guys back and forth in whatever order the software produced, ignoring the physical layout of the building. I mapped every bearing, gearbox, chain and motor into one continuous route following the plant's natural progression, built iteratively over three or four passes. I found equipment I did not know existed on the later passes.
The wood room went from four men walking PMs all week to one man in a day. The other departments adopted the same approach and each landed on the same shape.
The saving was real, and banking it in headcount would have been the obvious move and the wrong one. The thing worth having was that the plant could then all but guarantee everything worth looking at got looked at once a week.
Efficiency was the vehicle. Daily coverage and early detection were the destination. A plant that converts every readiness gain into a smaller crew has spent the gain and kept the exposure, and it will report the headcount reduction as the success because that is the number that is legible.
The twenty-year measurement
There is one measure that does work, and nobody can use it in a budget cycle because it takes two decades to read.
I went back to a plant I had worked at almost twenty years earlier. The company had been bought by a large corporation, and a number of the processes I built there had been adopted across that corporation because they worked. Turnover had been running close to eighty percent when I started. Twenty years later a good deal of the staff who had been there when I left were still there.
Retention is a maintenance metric, not an HR one. Eighty percent turnover to a stable twenty-year crew is the same curve as unplanned downtime falling, and it is caused by many of the same things: work that can actually be done, parts that are actually there, and a system that does not make people fight it to do their jobs.
I should also be honest that none of it got me rehired, and being right about the systems did not convert into a reference or a callback. Good measurement design and good career outcomes are unrelated variables. Anyone telling you otherwise is selling a seminar.
The control that got deleted for working
I want to put the opposite ending beside that one, because it is the more common one and I lived it.
I was a supervisor at a mill with a wall of bolt bins. The tops were flat, and every flat spot in a plant turns into a collection area. Guys would pull out a handful of bolts, nuts and washers, count out what they needed, and dump the remainder on top rather than find the right bin again. I spent an enormous amount of time every day sorting that pile back into the bins.
I raised it in tailgate meetings as a small thing we needed to fix. Every single time I brought it up it got worse. Dirty coveralls started appearing on the pile.
So I engineered it. I had an angled top put on the bins so there was no flat spot to use. That made everything worse. Guys went out of their way not to put hardware back, wedging it into the joints, sliding it under the ramp, or throwing it on the floor. Now I was picking hardware off the floor instead, and doing it more often than before, because loose hardware underfoot is a safety hazard.
I put a box there and asked them to at least use the box. They used the floor.
What finally worked was none of that. I wrote a weekly work order to clean and organize the bins, and I rotated it so a different man had it every week. It came right almost immediately, because the crew started policing each other. Nobody knew whose turn was coming, so making a mess meant possibly making your own job.
Then I was demoted, and my replacement decided the weekly work order was unnecessary and removed it. Everything went straight back on the floor, worse than before. He tried to reinstate the same work order later and it never took hold.
There are three measurement problems in that, and they are the ones this paper is about.
The control was deleted precisely because it was working. For as long as it ran, the area stayed clean and nothing bad happened, so a recurring cleanup work order read as a line of waste on a schedule. That is this paper's argument with an ending attached. The absence that proves a control is working is indistinguishable, on paper, from the absence that proves it was never needed. Nothing in any report told the difference, so the difference got decided by whoever was looking at the schedule that month.
The failure returning was the only measurement anybody ever got. It took a removal and a few weeks to produce evidence that no report had produced in the entire time the control ran. That is an expensive way to learn something, and it is the default way plants learn it.
And the fix did not survive me. The successor had the identical work order in his hand and it would not take. That proves the work order was never the whole mechanism. Whatever standing I had with that crew was an ingredient nobody had written down, including me. A control whose real dependency is who installed it will look structural right up until the day it is tested, and the test is somebody else trying to run it.
That last one sits badly next to the twenty-year story above, and both are true. Some of what I built outlived me by two decades. That one did not last a month.
And it did not stop there, which is the part I find hardest to argue with.
It ended with the operations guys sweeping the hardware up and throwing it in the garbage, and the plant bringing in a third party contractor to manage the bins.
Sit with the shape of that. A problem that had been solved, for free, by one man doing an hour of tidying on a rotating schedule, finished as a purchased service with a continuous inventory loss running underneath it. And the contractor does not correct the behaviour that caused it. It absorbs it, which removes the last reason anybody had to put anything back.
Here is why the expensive answer won, and it is the whole thesis of this paper stated as a purchase order.
A contractor invoice is legible and a rotating work order is not. The invoice is a number on a budget line. Somebody can point at it, defend it, forecast it, and report it upward as action taken. The work order produced nothing on paper for the entire period it was working, because what it produced was the absence of a mess. When the time came to choose, the organization did not weigh a free control against a paid one. It weighed a visible thing against an invisible one, and it picked the visible thing, and it experienced that as diligence.
The hardware going into the garbage never appeared anywhere either. Nothing in a plant meters its own bin, so a continuous stores loss ran alongside a contractor bill and only one of them was ever a number.
This is the failure mode the four counts exist to prevent. Not because counting catches would have saved those bolts, but because a control that generates a number every week is a control somebody has to argue against explicitly, in writing, with a reason. The bolt bin work order could be deleted silently by one person who looked at a schedule and saw a line that seemed to be for nothing. That is the whole distance between a control that survives and one that does not, and it has nothing to do with whether the control works.
What to stop reporting
Two metrics actively damage the case.
Percentage readiness, because averaging manufactures confidence and hides the single red gate that stops the job. Ninety percent ready is not a state, it is a rounding of "not ready" into something that reads like progress.
Schedule compliance measured against a schedule that was itself negotiated, because it rewards the fiction rather than the work. A plant hitting ninety-five percent compliance against durations everyone knows are invented has measured its own willingness to pad.
Replace both with the four counts and a single stop-go state. The report gets shorter, harder to argue with, and considerably less comfortable, which is the point.
Conclusion
Good maintenance is chronically underfunded because it is chronically unmeasurable in the terms budgets are written in. That is fixable, but not by proving prevented failures, because that proof does not exist and chasing it makes you sound defensive.
It is fixed by counting the catches, publishing them as they occur, and refusing to average them into a number that makes a stopped job look mostly finished.
And by accepting the thing that makes this work genuinely hard to sell: if you do it properly, the best day you will ever have is one nobody remembers, and one you cannot tell a story about afterwards.
References
- Crespo-Márquez & Gómez Fernández, “Agentic AI for maintenance management: a process-centric review and a staged framework for industrial adoption,” Computers & Industrial Engineering, 2026. https://www.sciencedirect.com/science/article/pii/S0360835226004869
- Misra, Vyas, Gutta & Mercangöz, “ADMITBench: A Safety-Governed Reference Framework for Evaluating the Admissibility of Industrial LLM Advisories,” 2026. https://arxiv.org/abs/2608.03866
- Prometheus Group, “From Hours to Minutes: How AI Is Transforming Maintenance Planning and Scheduling,” June 26, 2026. https://www.prometheusgroup.com/resources/posts/from-hours-to-minutes-how-ai-is-transforming-maintenance-planning-and-scheduling
- Liu et al., “ConMem: Contribution-Aware Memory for Long-Horizon Manufacturing Inspection Logs,” 2026. https://arxiv.org/abs/2607.28126