Ask a maintenance program how many of last quarter’s predictions held. Most cannot answer, and the reason is rarely that the program is badly run. The question was never made answerable in the first place. The recommendation was made, the work order was raised, the work was done, and the asset did not fail, which everyone agrees is a good outcome and nobody can attribute to anything in particular. Would it have failed anyway? Nobody wrote down when.
How do you measure whether predictive maintenance is working?
By recording the predicted failure window and the confidence in that prediction before the work is scheduled, then recording the outcome once the window closes. Three results are possible: the failure arrived inside the window, it arrived outside it, or the work prevented it. Each gets recorded, and a miss is worth as much as a hit.
That is the whole discipline. It sounds trivial and almost nobody does it, because the moment of recommendation is busy and the moment of review is months later.
Why can’t most programs answer the question?
Because the record was never created at the point where it would have been cheap to create.
Afterwards, the program is judged on recollection and on a general availability figure. Availability improved this year. Was that the predictive model, the new parts supplier, the mild winter, or the two units that were retired? A number that cannot separate those will not survive a serious operations review, and it should not.
A prediction nobody recorded is a guess that happened to be written down.
What two trends carry most of the signal?
Two trends carry most of the signal, and both become visible well before a failure arrives. Time between failures falling means failures are becoming more frequent rather than more unlucky. Repair time rising means something has changed in parts availability, in the skills of the people doing the work, or in how quickly a fault can be diagnosed.
Time between failures falling
Failures becoming more frequent is a leading indicator that an asset is entering a failure pattern rather than having bad luck. The distinction matters because the responses are different, and because a run of bad luck is what every maintenance organization tells itself it is looking at.
Repair time rising
Something has changed in parts availability, in the skills of the people doing the work, or in how quickly the fault can be diagnosed. Rising repair time is rarely about the asset and almost always about the system around it.
Both trends are visible only if the history has been resolved to the asset rather than scattered across work orders under three different identifiers. That is the prerequisite, and it is why this is a data problem before it is a modeling problem.
What has to be recorded before the work is scheduled?
Three things, captured at the moment the recommendation is made rather than reconstructed afterwards: the predicted failure window stated as a range with dates, the confidence in that prediction, and the expected outcome of the work that answers it.
Because those were captured at the point of recommendation, the comparison becomes available the day the window closes. No reconstruction, no argument about what was meant, no reliance on anyone’s memory of a meeting in March.
Why is asset work easier to prove than most business commitments?
This is worth saying plainly rather than implying that every commitment can be closed out this cleanly.
Many business commitments land over quarters, tangled with other commitments and with whatever the market did in the meantime. Attribution genuinely decays. Eighteen months after a system consolidation, no platform can isolate the contribution of one initiative from the dozen running alongside it, and a number that pretends otherwise will not survive a serious review.
A failure prediction is not like that. The window was stated in advance, the condition inputs are known, and the asset either failed inside that window or it did not. The answer arrives when the window closes, and it is separable from everything else happening in the operation.
That is what allows a maintenance program to close its loop honestly while a portfolio commitment cannot. Both are worth tracking. Only one produces a clean result, and an honest platform keeps the two apart.
How does a maintenance decision become a financial position?
By writing availability, maintenance cost, and downtime into the same governed structure the rest of the business reads from.
A deferred overhaul recorded in a maintenance system is a maintenance matter. The same deferral, carried into the operating model as a signal with a quantified effect on expense and cycle time, against the outcome that was expected and with an accountable owner attached, is a financial position with a number on it.
Maintenance signals are among the classes where that comparison can be read directly, because the window is short and the inputs are few. Most P&L lines are not so cooperative. This one is.
What about the agents?
The same discipline is what makes extending an agent’s authority defensible.
An agent acting on unresolved, unscored asset data produces confident recommendations with unknown foundations. Give it scored data and keep a record of what it recommended and what happened, and the authority question becomes evidential rather than political. A Dataiku and Harris Poll survey of data leaders found 95% could not fully trace an AI decision from input data through model output. In an asset context that is not an abstraction. It is somebody signing off on a deferral they cannot defend.
The measure worth watching
The measure is not the accuracy of the first prediction. It is whether the organization can state, at the end of a quarter, how many predictions it made, how many held, and what it did differently as a result.
Run it for one quarter on one asset class before extending the model’s authority anywhere else. Record the window and the confidence before the work is scheduled. Record the outcome either way. The miss is the only thing that improves the next prediction.
The discipline behind the number
The PolyPhaze white paper Best decision. Best action. covers the twelve agents across five phases, compliance evidence assembled from lineage, and how the same record is configured per asset class. Download the full predictive maintenance ebook for the complete method.
Frequently asked questions
How do you measure predictive maintenance accuracy?
Record the predicted failure window and the confidence in it before the work is scheduled, then record the outcome once the window closes. Accuracy is the proportion of predictions where the failure arrived inside the stated window or where the work demonstrably prevented it.
Is a missed prediction useful?
Yes. A miss is the only input that improves the next prediction, provided the window and the confidence were recorded in advance so the miss can be diagnosed rather than argued about.
Why is predictive maintenance easier to prove than other AI use cases?
Because the window is stated in advance, the condition inputs are few and known, and the outcome is separable from other activity in the business. Most commitments land over quarters alongside a dozen others, which makes attribution decay. A failure window closes on a date.