Predictive vs preventive maintenance: the real cost.

Three ways to pay for maintenance

Every maintenance strategy is a decision about when you pay. Reactive (run-to-failure) pays nothing until the machine stops. Then it pays for the repair, the collateral damage, the expedited parts, and every hour of lost production at once. It remains the default in much of industry: the DOE estimates that over half of maintenance resources at an average facility are still spent reactively.

Preventive maintenance pays on a schedule. Components are serviced or replaced at fixed intervals of calendar time or run hours, sized to the expected life of the part. It converts chaos into a budget line, which is why it became the industrial standard. Its weakness is that the schedule, not the machine, decides what gets done.

Predictive maintenance pays for information first. Sensors, historians, and models watch the actual condition of each asset: vibration, temperature, current draw, process behavior. Work is called for only when the evidence says a failure is developing. Done well, it eliminates both the surprise failures of reactive and the unnecessary interventions of preventive.

Cost comparison: reactive vs preventive vs predictive

The strategies differ less in what you pay for and more in when and how predictably you pay. This is the comparison that matters when you build the business case:

Savings ranges are US Department of Energy industry figures for functioning programs; your numbers depend on asset criticality, downtime cost, and how much of your maintenance is currently reactive.

The hidden costs of preventive maintenance

Preventive maintenance looks cheap because its waste is invisible: it shows up as normal, budgeted work rather than as an incident. Three effects drive the gap the DOE measured:

• Servicing healthy equipment. Calendar intervals are sized to the worst case, so most interventions happen long before the part needed them. The labor, parts, and lost production of every early replacement is pure cost.
• Random failures don't read the calendar. The landmark United Airlines reliability study by Nowlan and Heap found that only about 11% of failure modes were age-related. The rest failed at random or from induced causes, and a schedule can only prevent the age-related minority.
• Maintenance-induced failures. Every intrusive intervention, from opening to disassembly to reassembly, is a chance to introduce a defect into a machine that was running fine. Some fraction of tomorrow's breakdowns are caused by today's PMs.

None of this makes preventive maintenance wrong. It makes it a blunt instrument: highly effective on parts that genuinely wear with age, expensive everywhere else.

What a predictive program actually costs

Predictive maintenance moves cost forward. Before it saves anything, you pay for four things:

• Instrumentation: vibration, temperature, current, acoustic, or process sensors on the assets you want covered. Many plants already have more of this than they use.
• Data infrastructure: a historian or equivalent pipeline that gets sensor and process data somewhere a model can see it, reliably.
• Analytics: the models that turn signals into failure warnings, whether rule-based condition monitoring, machine learning, or physics-informed approaches.
• People and process: someone has to trust the alert, plan the work, and close the loop. A prediction nobody acts on saves nothing.

Against that investment, the DOE's industry averages for a functioning program: roughly 10× return on investment, 25–30% lower maintenance costs, 70–75% fewer breakdowns, and 35–45% less unplanned downtime. Deloitte's analysis lands in the same direction: 10–20% more equipment uptime and 20–50% less time spent planning maintenance.

The denominator matters as much as the program cost. Siemens' True Cost of Downtime research estimates the world's 500 largest companies lose roughly $1.5 trillion a year to unplanned downtime, which is about 11% of their revenues. When one hour of stoppage costs six or seven figures, the instrumentation bill stops being the interesting number.

When preventive is still the right call

An honest cost comparison has to say where predictive loses. Preventive maintenance remains the better economic answer when:

• The asset is cheap and non-critical. If a failure costs less than the monitoring, run it to failure or keep the simple PM.
• The failure mode is genuinely age-related. Filters, belts, lubricants, and wear parts with predictable life are exactly what calendars are for.
• Regulation or warranty mandates the interval. Statutory inspections and OEM warranty terms don't care about your condition data.
• You can't yet act on predictions. Without the planning discipline to use an early warning, predictive spend buys alerts, not savings.

In practice every mature program is a hybrid: predictive on the critical, expensive, randomly-failing assets; preventive on true wear items and mandated tasks; run-to-failure on things that don't matter.

How to decide, asset by asset

The business case is built bottom-up, not plant-wide. A sequence that works:

1. Rank assets by criticality: what stops production, what threatens safety, what has no installed spare.
2. Price an hour of downtime for each critical asset, including lost production, restart costs, and contractual penalties. This number anchors everything.
3. Sort failure modes: age-related wear stays on the PM schedule; random, condition-detectable failures are predictive candidates.
4. Pilot on your worst actors, the handful of assets with the most unplanned events. They carry the fastest payback and the most convincing before-and-after story.
5. Measure avoided events, not alerts. Track unplanned downtime hours and maintenance spend per asset before and after; that delta is the program's ROI.

Where physics-informed AI changes the math

Two line items dominate the cost side of most predictive programs: false alarms, which burn planner trust and technician hours, and the long data-collection runway that pure machine-learning models need before they're useful. Most plants don't have years of labeled failure history for their critical assets.

Physics-informed models attack both. Because the model reasons from the physics of the equipment rather than only from historical examples, it needs far less failure data to become useful, and its warnings come with an explanation an engineer can check against reality before spending a shift on it. That's the difference between an alert and a work order, and it's the part of the cost equation Vermilion was built for.