Predictive vs preventive maintenance

There's a vendor pitch we've heard variations of a dozen times: "Stop wasting money on preventive maintenance. Use AI to predict failures before they happen."

It sounds great. It's also misleading. Predictive maintenance is real and powerful, but for most Malaysian SMB factories the right first move is not AI-driven prediction — it's getting basic preventive maintenance properly done, with data to back it up. This is the version we tell clients on discovery calls.

The three flavours of maintenance

To make this make sense, here are the three actual modes a factory can run on:

Reactive maintenance. "Run it till it breaks, then fix it." This is where most SMB factories without a system actually live, even when they think they're doing preventive. Cheapest in calendar terms, most expensive in lost-production and emergency-repair terms.

Preventive maintenance (PM). "Service the machine on a fixed schedule — every X hours of operation, or every Y weeks, or every Z units produced." Calendar-driven or counter-driven. Time-tested, well understood, what every equipment manufacturer's manual recommends.

Predictive maintenance (PdM). "Watch the machine's actual condition and predict when it's about to fail." Vibration analysis, temperature trends, current draw, oil chemistry, acoustic signatures. Increasingly: ML models trained on sensor and failure history.

These aren't mutually exclusive. A grown-up maintenance program uses all three: reactive for non-critical stuff, preventive as the baseline for critical assets, predictive as the upgrade path for the assets where prediction actually pays off.

What each one costs

Honest cost framing for a Malaysian SMB factory:

Reactive (no system): "free" in software terms but expensive in production terms. The hidden cost is unplanned downtime — and in food, FMCG, and high-throughput packaging, an unplanned line stop during peak production can cost more in one event than a year of CMMS subscription.

Preventive (with a CMMS): software typically RM 5,000–RM 30,000 per year for SMB-sized CMMS, plus the labour cost of actually doing the PMs on schedule. Implementation maybe RM 20,000–RM 50,000 if you want it integrated into your OEE / production data and customised to your assets. Pays off the first time you avoid a major unplanned stop on a critical asset.

Predictive (with sensors + ML): materially more expensive on the build side. Vibration sensors, temperature sensors, current sensors per critical asset (RM 1,500–RM 8,000 per asset depending on what you're measuring). ML models trained on your data — typically RM 50,000–RM 150,000+ for a real predictive maintenance build that actually predicts something. Plus tuning time after launch.

When predictive maintenance actually earns its keep

We're an AI-heavy team and we like building predictive maintenance models. We're also honest about when they don't pay off. Real signals that PdM is worth the investment:

• The asset is expensive when it fails. A bearing replacement on an injection moulding machine that idles a 24/7 line is a different financial event than a bearing on a low-utilisation packing machine.
• Failures have a measurable run-up. Not all failures do. Vibration-based prediction works on rotating equipment because vibration profiles drift before catastrophic failure. Sudden electrical failures on motors with no precursor pattern — there's nothing to predict.
• You have at least 12–18 months of operational data, including some real failures. ML models need failure examples to learn from. If your equipment is too reliable to have failed yet, the model has nothing to learn.
• Maintenance windows are scarce. If you can pull a machine for PM whenever you want, scheduled PM works fine. PdM matters when downtime is expensive and getting a planned window is hard.

If you don't hit at least three of these, predictive maintenance is mostly a science project that doesn't pay back fast enough.

What we usually recommend for SMBs

The practical recipe for a Malaysian SMB factory in 2026:

1. Get OEE measurement working first. You can't run a real maintenance program without knowing what's actually breaking and how often. (We wrote a pricing breakdown and an overview of OEE vs MES vs CMMS that go deeper.)
2. Add a CMMS layer wired into OEE downtime data. PM schedules per critical asset, downtime-driven work orders, spare-parts inventory. This alone moves most SMB factories from "reactive disguised as preventive" to "actually preventive". Big jump.
3. Run the basic preventive program for 6–12 months. Collect data: which assets fail despite PM, which assets fail in patterns the PM schedule misses, which assets are over-serviced.
4. Add predictive monitoring for the 1–3 assets where the case is obvious. This is rarely "all assets" — it's usually a small number of high-value rotating assets where vibration or thermal monitoring pays back fast.
5. Layer ML on top once you have failure data and sensor data both flowing. Start with anomaly detection (statistical, no ML required) and only graduate to true ML prediction when the case is overwhelming.

A common mistake

The most common mistake we see is SMBs going straight from reactive maintenance to a vendor's "AI-driven predictive maintenance platform" without doing the boring preventive layer first.

What happens: the predictive system flags issues. Maintenance team has no process to act on the flags. Either the flags get ignored, or the team scrambles to fix things that weren't actually about to break, generating false-positive fatigue. Six months later the system is muted and the same reactive maintenance pattern is back.

The reason preventive maintenance comes first is process, not just data. PdM only works when there's already a maintenance team that responds to scheduled work orders. Without that, PdM is a notification system going into a void.

A second common mistake

The reverse mistake is also common: factories that have invested heavily in preventive maintenance assume predictive is just incremental and skip it forever. For a few specific high-value assets where vibration / thermal / electrical signatures actually drift before failure, predictive measurably reduces both downtime and over-servicing. Worth doing on those — just not on everything.

What this looks like at the AI layer

For Malaysian SMB factories that have OEE + CMMS already running, the predictive maintenance build we typically recommend:

• Anomaly detection first. Statistical models on sensor streams (current draw, vibration, temperature, cycle time). No ML required. Catches drifts and step-changes that humans miss. Cheap to build, immediate value.
• ML failure prediction second. Only on assets where (a) failures have been historically observed, (b) sensor data shows a run-up pattern, (c) the cost of failure justifies the model build cost. Usually 2–5 assets, not 50.
• Integration with the CMMS. Predicted failures auto-create work orders in the CMMS, with confidence ranges. Maintenance team treats them like any other scheduled work, not as a separate alerting channel that competes for attention.

Sized this way, predictive maintenance becomes a high-value layer on top of working preventive — not a replacement for it.

How to choose

If you're an SMB factory owner thinking about maintenance software:

• No CMMS yet? Start with CMMS + OEE. Skip the predictive pitch for now.
• CMMS running but reactive in practice? Fix the process first. Software won't fix a process problem.
• Solid PM program, asking what's next? That's when predictive earns its keep — but on a small set of carefully chosen assets, not everything.

We do free discovery calls. Drop us a line if you'd like to walk through your maintenance setup and what would actually move the needle.