A machine that never complains is not a machine that is fine. It is a machine nobody is listening to. The whole premise of predictive maintenance is that equipment talks constantly, in vibration, temperature and current draw, and that the expensive failures are the ones where the conversation was happening and nobody had a microphone in the room.
Key takeaways
- DOE benchmark: predictive saves 8% to 12% versus preventive alone, and preventive saves 12% to 18% versus reactive.
- A typical facility still runs 40% to 60% reactive; best-in-class sites run under 10%.
- Siemens put unplanned downtime at $1.4 trillion a year across the world’s 500 largest firms in its 2024 report.
- The median cost of an hour of unplanned stoppage is around $125,000, and about $2.3 million in automotive.
Three maintenance regimes competing for the same budget
Every maintenance strategy is a bet about when to spend money. Reactive maintenance runs equipment to failure and pays afterwards, in overtime, expedited parts and secondary damage. Preventive maintenance replaces or services on a fixed interval, which prevents a good share of failures and also throws away useful component life. Predictive maintenance monitors condition and intervenes when the data says the asset has started to degrade, not before and not after.
The DOE guide is unusually blunt about what each of those choices costs, and its comparison table remains the cleanest public benchmark we know of for arguing a business case internally.
| Regime | Typical facility | Best-in-class target | Stated saving |
|---|---|---|---|
| Reactive | 40% to 60% | Under 10% | Baseline |
| Preventive | 30% to 50% | 25% to 35% | 12% to 18% over reactive |
| Predictive | 10% to 15% | 45% to 55% | 8% to 12% over preventive |
The gap between the second and third columns is the whole argument. Most plants are not choosing between predictive and preventive. They are still paying for a reactive base load they inherited.
What an hour of silence now costs
Siemens quantified the other side of the ledger in its True Cost of Downtime report published in 2024. Unscheduled downtime was costing the world’s 500 largest companies around $1.4 trillion a year, equivalent to roughly 11% of revenue, against $864 billion and about 8% in 2019 and 2020. The median hourly cost of a stoppage came in near $125,000, with automotive assembly at roughly $2.3 million an hour.
One detail in that report matters more than the headline. The average plant took 81 minutes to restart after a stoppage, against 49 minutes five years earlier. Thinner maintenance headcount and slower emergency parts sourcing have made recovery itself more expensive, which changes the arithmetic of prevention independently of how often things break.
Machines do not get bored, they degrade quietly
The metaphor in this article’s title is affectionate, and it is also slightly misleading, so it is worth being precise. Equipment does not go idle and lose interest. It accumulates wear that produces measurable signatures long before anything stops turning: bearing defect frequencies in vibration spectra, rising winding temperature under constant load, particle counts and viscosity drift in oil analysis, changes in motor current draw that show up before heat does.
Predictive maintenance is the discipline of catching those signatures and converting them into a scheduled intervention. The forecast is not mystical. It is a trend line crossing a threshold that somebody defined, and the quality of the whole programme rests on how well that threshold was chosen for that specific asset in that specific duty cycle.
Where these programmes actually stall
In our reading of how these deployments go wrong, three failure modes come up far more often than any technical limitation of the models.
No usable failure history. Supervised models need examples of the thing you want to predict. A plant with excellent uptime and poor record-keeping has plenty of data and almost no labels, which is why pilots on critical assets so often produce confident nonsense in the first year.
Alarms with no owner. A condition alert that lands in a shared inbox and competes with four hundred other notifications is not a maintenance system, it is a liability. The organisational question, who stops the line and on whose authority, has to be answered before the first sensor is fitted.
Instrumenting the wrong assets. Criticality analysis is unglamorous and it is the step that decides the return. Monitoring a redundant pump with a spare on the shelf produces beautiful dashboards and no saving. This is where the case for smart sensors in quality control and the case for condition monitoring genuinely diverge: one is about the product, the other is about the asset, and they rarely justify the same hardware budget.
How we would scope a first deployment
- Rank assets by consequence, not by age. What does an hour of this machine being down cost, and is there a bypass? Everything else follows from that number.
- Pick failure modes, not machines. “Bearing degradation on the main drive” is a scope. “The extruder” is a wish.
- Establish a baseline before the algorithm. Several months of condition data under known-good operation is worth more than a more sophisticated model trained on nothing in particular.
- Write the response procedure first. Threshold, notification, authority to act, and the spare part that needs to be on the shelf when the alert fires.
- Measure the programme on avoided stoppages and schedule adherence, not on model accuracy. Accuracy is an input. Uptime is the output anyone will fund.
Frequently misunderstood
Does predictive maintenance replace preventive maintenance?
No. The DOE benchmark for a best-in-class site still allocates 25% to 35% to preventive work. Some tasks, lubrication and statutory inspections among them, stay on a calendar regardless of what the sensors say.
Can it eliminate unplanned failures?
It cannot. Random failures, installation defects and operator error are not preceded by a degradation trend, so no amount of monitoring will forecast them. The honest claim is a reduction in a specific class of failures, not their disappearance.
Do you need machine learning to start?
Not at first. Route-based vibration analysis, thermography and oil sampling on a defined schedule are decades-old techniques that deliver a large share of the benefit. Continuous monitoring and modelling raise the ceiling; they are not the entry ticket.
What is the realistic payback window?
It depends entirely on the hourly cost of the assets covered, which is why step one is ranking by consequence. A site where an hour of stoppage costs six figures and a site where it costs four are running the same technology on completely different economics, and vendor payback figures rarely say which one they were measured on.
The version of this we find defensible: predictive maintenance is not a forecasting trick, it is a shift in when a decision gets made. The machines were never bored. They were simply running unobserved, which for a long time was the cheapest option available and, at $125,000 a median hour, no longer is.
Want the strategic case rather than the deployment detail?
Why condition-based work has become the reference standard across industrial operations.
Sources: US Department of Energy Federal Energy Management Program, Operations & Maintenance Best Practices Guide, Release 3.0, prepared by Pacific Northwest National Laboratory (maintenance approach comparison and savings ranges); Siemens, The True Cost of Downtime (2024), as reported by Siemens and by the Institute for Supply Management. Updated August 2026.
