Maintenance strategy gets pitched as a hierarchy, reactive at the bottom, preventive in the middle, predictive at the top, with the implication that every shop should be climbing toward predictive as fast as budget allows. That’s not quite right. Each strategy has a real cost profile and a real set of conditions where it’s the correct choice, including reactive, which is unfairly treated as the strategy nobody should ever choose. The right answer for most plants is a deliberate mix, not a single winner.
Reactive: fix it when it breaks
Reactive maintenance is exactly what it sounds like: run the equipment until it fails, then fix it. It has a bad reputation because it’s often the default by accident, not by decision, a plant with no PM schedule and no monitoring runs reactive by neglect, and that version is expensive, unplanned failures at the worst possible time, on the critical path, taking down whatever depends on that machine with it.
As a deliberate choice, though, reactive is the right call for low-cost, low-consequence equipment. A cheap fan, a redundant pump with a backup already in place, a component that’s simple and fast to swap and doesn’t take anything else down when it fails. Spending PM labor hours inspecting a $40 part that takes ten minutes to replace and never stops production when it fails is wasted effort. The mistake isn’t running some equipment reactive, it’s running everything reactive by default because nobody ever sat down and picked a strategy per asset.
Preventive: fix it on a schedule, whether it needs it or not
Preventive maintenance services or replaces a component on a fixed cadence, hours run, cycles completed, or calendar time elapsed, regardless of its actual condition. A filter changed every 500 hours gets changed at 500 hours whether it’s still clean or already clogged, because the schedule doesn’t know which.
That’s both preventive’s strength and its cost. The strength: it’s predictable, budgetable, and it catches the failures that follow a fairly consistent wear curve, a bearing, a belt, a filter, things that degrade on a knowable timeline. The cost: some of that maintenance is wasted, performed on a component that had life left in it, and it doesn’t catch a failure that happens off the wear curve’s usual pattern, a part that fails early for a reason the calendar never accounted for.
Predictive: fix it when the evidence says to
Predictive maintenance uses a condition signal, vibration, temperature, a trend in cycle time or stop frequency, to trigger work close to when a component actually needs it, instead of on a fixed calendar. Done well, it captures most of preventive’s failure-prevention benefit while cutting the wasted labor on parts that had life left, and it catches some failures preventive’s fixed schedule would have missed because they didn’t follow the assumed wear curve.
The real cost of predictive is what gets skipped in the pitch: it needs a real signal, consistently collected, and a track record long enough to know what normal looks like before it can tell you what abnormal looks like. Full condition-based monitoring, vibration analysis, oil analysis, thermal imaging, is a real program with real sensor cost and real expertise to interpret it, not a checkbox a plant flips on. A lighter version of predictive, using stop patterns and reliability trends a plant is likely already capturing rather than new dedicated condition sensors, is a realistic middle ground for most shops, not a substitute for full condition monitoring on the handful of assets where it actually pays for itself, but a genuine step up from a pure calendar schedule.
That lighter version is what a reliability metric like MTBF (Mean Time Between Failures) is actually for. It’s not a vibration sensor, but a machine whose MTBF is visibly trending down is giving you a real early signal, evidence something is drifting, worth investigating before it turns into an unplanned stop, without installing anything new. Alongside MTTR (Mean Time To Repair), which measures how long a stop takes to close once it starts, MTBF is the kind of data a plant already generates just by tracking stops and repairs consistently, no extra hardware, that starts to behave like a predictive signal once there’s enough history behind it. A statistical failure-risk estimate built on that same history, an asset’s own pattern of past stops, not a generic industry curve, is a reasonable next step up from just watching MTBF trend by eye, and it’s still fundamentally different from true condition-based sensing, it’s pattern recognition on stop history, not a physical measurement of the machine’s actual wear.
The real decision: which strategy for which asset
The strategy question isn’t preventive versus predictive versus reactive as a plant-wide policy. It’s a decision made per asset, weighing two things: how expensive is a failure, in downtime and consequence, not just parts cost, and how predictable is the failure mode.
A machine whose failure stops the whole line, and whose bearings show a clear vibration signature before they go, is exactly where investing in predictive pays back fastest. A backup pump that’s cheap to replace and fails unpredictably anyway is exactly where predictive would be spending real money to learn very little. Most plants have all four quadrants on their floor at once, and the right program treats them differently instead of applying one policy to everything.
Start where the data already exists
The realistic sequence for most shops isn’t preventive, then predictive, as separate projects with separate budgets. It’s building the maintenance repair loop first, a real queue that tracks every stop through acknowledge, start repair, and resolve, with an actual resolution note, the same discipline behind good downtime reason codes, because that queue is what generates the MTTR and MTBF history predictive work depends on later. A plant with no consistent repair logging has no baseline to compare a predictive signal against, so it has nothing to notice as abnormal. Preventive tasks on a due list, separate from that stop-driven queue, cover the calendar-based half of the picture. Neither replaces a dedicated CMMS with parts tracking and routing if a plant’s complexity has grown into needing one, but both are the real starting point before anyone spends money on vibration sensors.
Quick recap
- Reactive is a legitimate, deliberate choice for cheap, low-consequence equipment, chosen per asset instead of defaulted into everywhere
- Preventive trades some wasted labor on early service for predictability against a known wear curve
- Predictive needs a real, consistently collected signal and enough history to know what normal looks like, it is a real program, not a checkbox
- MTBF and MTTR, built from consistent stop and repair logging, are a lightweight predictive signal most shops already have the raw data for
- Pick the strategy per asset, weighing failure cost against how predictable the failure mode actually is
- Build a real repair queue first, it is the data foundation everything predictive is compared against later