Predictive maintenance: when it pays, when it doesn't

Predictive maintenance pays when three things are true: the assets are expensive to lose, the failure modes are known and somewhat gradual, and the data to detect them exists. When any of those is missing, the economics collapse, and pushing through with a predictive maintenance build anyway is how businesses end up with an expensive monitoring system that doesn't change any maintenance decisions.

The industry pitches predictive maintenance as a universally applicable improvement on time-based maintenance. It isn't. Plenty of businesses should stick with well-tuned time-based maintenance or simple condition monitoring, because the economics of prediction don't clear the bar. Here's the honest framework.

The economic test

The math is simple. Predictive maintenance pays when:

(Downtime cost × avoidable failures × detection rate) − (monitoring cost + false positive cost) > 0

For a typical piece of industrial equipment, that's a real calculation. High-value assets with expensive downtime (mining haul trucks, refinery pumps, manufacturing line motors) clear the threshold easily. Low-value assets with cheap failures (light bulbs, consumables) don't, no matter how accurate the model.

Three rough asset categories and whether predictive maintenance typically pays:

The category most businesses misjudge is the middle one. They default to either treating everything like category 1 (overinvesting) or everything like category 3 (underinvesting). Neither is right.

When sensor data isn't required

A common myth: predictive maintenance needs IoT sensors. Sometimes yes, often no. A well-maintained CMMS (computerised maintenance management system) with accurate work-order history, failure mode codes, and runtime hours is frequently enough to predict failures on common industrial equipment. The signals look like:

• Maintenance interval creeping. Time between minor services getting shorter on a specific asset, compared to its fleet average.
• Work-order cost trending up. Same asset, rising service costs over the last 6-12 months.
• Specific fault code patterns. Certain combinations of minor faults precede major failures by weeks or months.
• Runtime vs age vs failure rate curves. Well understood for most industrial equipment.

Putting IoT sensors on a piece of equipment adds data but also adds cost, maintenance burden, and failure modes of its own (the sensor itself can fail). Before investing in sensors, check whether CMMS-only predictions clear the accuracy bar. Often they do.

When sensor data is required

For assets where failure is genuinely fast-progression, or where the CMMS data is poor quality, sensors are the right investment. The honest indicators:

• Rotating machinery with bearing failure modes (vibration sensors).
• Electrical equipment with thermal signatures preceding failure (thermal imaging, periodic or continuous).
• Hydraulic systems with particulate contamination leading to failure (oil analysis, continuous where economical).
• Critical pumps and compressors where even hours of downtime are expensive.

Sensor deployment on a mid-fleet (50-200 assets) is a meaningful capital outlay for hardware alone, plus an ongoing operating cost for the data infrastructure and analytics that sit behind it. That's a real bar to clear. The downtime savings need to be significant for the math to work.

Predictive maintenance on a fleet of assets that don't individually generate enough downtime cost to justify the monitoring investment is a solution looking for a problem. It happens more often than the industry admits, because the technology is interesting and the business case is rarely interrogated hard enough.

The scope trap

Businesses that get predictive maintenance right almost always start narrow. Three to five failure modes on one asset class, with a clear payback calculation. They ship, measure, scale. Businesses that get it wrong try to predict every failure on every asset simultaneously. The build takes three years, the accuracy is mediocre on everything, and nobody can point to a specific decision that changed because of the system.

A good first predictive maintenance engagement scopes to:

• One asset class (e.g. the 40 haul trucks, not all 400 pieces of equipment on the site).
• Three to five known failure modes on that class.
• A clear downtime cost number per failure and an accuracy target.
• A 12-week build and a 12-week monitored pilot before scoping expansion.

Typical ROI when the conditions are right

For mining, manufacturing, energy, and logistics businesses with the right asset profile:

• Year 1 ROI: 2x to 4x on the total engagement cost, driven mostly by avoided downtime on the first 3-5 failure modes.
• Year 2-3 ROI: 4x to 8x as the model improves and scope extends to adjacent failure modes.
• Year 4+: Diminishing returns. Most of the easy wins are captured by year three, and ongoing maintenance of the model becomes the operating cost.

If a vendor promises 10x year-1 ROI, either the baseline is unusually broken (possible) or they're optimising their numbers. The honest range is 2x to 4x, and businesses that are happy with that ship on time and scale. Businesses that demand 10x upfront tend to not ship at all.

Scope the first engagement around the biggest failure, not the biggest asset

Counterintuitively, the highest-ROI predictive maintenance engagements aren't always on the most expensive assets. They're on the asset class where one specific failure mode accounts for a disproportionate share of unplanned downtime. Sometimes that's the haul trucks. Often it's something less glamorous: the conveyor belts, the compressor feeding the whole site, the critical pump whose failure stops three production lines.

Find that one failure mode, predict it reliably, save a known amount of downtime. Use the win to fund the next three. That's the pattern that compounds.

When to not do predictive maintenance at all

• Asset fleet is under 20-30 comparable units. Not enough data to train on.
• Downtime cost per failure is under $10k. The monitoring investment doesn't clear the economic bar.
• Failure modes are genuinely random (e.g. external impact damage, operator error). Prediction doesn't work on random.
• CMMS data is so bad that six months of data cleanup would be needed before any model could train. Fix the CMMS first, come back to prediction in year two.

The test: can you draw a line from "predicted failure" to "different maintenance decision" to "avoided downtime cost"? If yes, predictive maintenance pays. If not, don't start.