Predictive maintenance in an SME: when it pays and when it's money down the drain

What it really is (without the word AI)

Predictive maintenance is simple to explain: you put sensors on a machine (vibration, temperature, current), a model learns how that machine vibrates when healthy, and it warns you when the pattern changes - weeks before the failure. No magic: bearings that are going to fail start singing long before they fail.

What the vendor doesn't mention in the first meeting: the model needs months of data, someone who understands the alarms, and a machine where the failure is expensive enough to justify it all.

The maths to do before signing anything

A serious monitoring point (sensor + platform + installation) costs 1,000-3,000€ per year per machine. The question is not "does it work?" - it's "what does the failure it prevents cost me?".

• Cost of one failure: downtime hours x margin per hour of what you stop making + urgent repair + customer penalties. On a compressor that stops the whole plant, one failure can cost 5,000-20,000€. On a machine with backup, 500€.
• Frequency: if that machine fails once every 3 years, expected benefit is low even if the failure is expensive.
• Detectability: progressive mechanical failures (bearings, misalignment, wear) are detected well. Sudden electrical failures or human error are not.

Quick rule: if (failure cost x avoidable failures per year) exceeds 3 times the annual monitoring cost, go for it. If not, don't.

Where it DOES pay in an SME

• The bottleneck machine that stops the factory: compressors, boilers, furnaces, the continuous line with no alternative.
• Machines whose failure is catastrophic in cost: even if rare, the repair is five figures (large motors, extruders).
• Failures with a clear signature: vibration on rotating equipment, temperature on furnaces, current draw on pumps.

Where it's money down the drain

• A varied fleet of 15 small machines: monitoring everything costs more than all the failures combined. A preventive calendar wins here.
• Machines that fail randomly: knocks, operator error, material quality. No sensor predicts that.
• Buying it before having preventive: without a basic maintenance plan running, predictive is strapping a rocket to a cart. AI amplifies what you have; with no maintenance discipline, it amplifies chaos.

The right ladder

1. Ordered corrective: log every failure with hours and cost. Without this data you can't even run the maths on whether predictive pays.
2. Serious preventive: calendar by criticality. With this alone, most SMEs cut downtime 30-50% in the first year. Cost: one afternoon and an Excel.
3. Condition-based (predictive's poor cousin): periodic manual measurements - a handheld vibration meter, a 300€ thermal camera, the trained ear of your maintenance person written on a sheet. 70% of the benefit for 10% of the cost.
4. Predictive with sensors: only on the 1-3 machines where the maths above clearly works.

Almost every SME should be on rungs 2 and 3. Vendors sell rung 4 because that's the one they charge for.