Predictive maintenance: when does it pay off for an industrial SME?

Predictive maintenance is appealing because of what it promises: intervening just before a failure, neither too early nor too late. For an industrial SME, though, the useful question is not "which technology?" but "on which machines, for what avoided loss, and with what ability to respond?" We have seen sensor projects produce handsome dashboards without saving a franc, and more modest ones pay back in under two years. The difference rarely lies in the algorithm.

What you are actually buying

Predictive maintenance means continuously measuring the condition of equipment (vibration, temperature, current draw, pressure) and estimating from the drift in those signals when an intervention becomes necessary. It differs from corrective maintenance, which repairs after the failure, and from calendar-based preventive maintenance, which replaces parts at fixed intervals whether they are worn or not.

The economic gain does not come from measurement. It comes from moving from an unplanned failure, with its emergency call-outs, express parts and late deliveries, to a planned stop at a time of your choosing. An alert is therefore worth something only if the organization can do something with it. That is where projects are won or lost.

Four conditions for the numbers to hold

High, concentrated downtime cost

If every machine costs roughly the same when it stops, the effort is diluted. The favourable case is a bottleneck: a machine with no redundancy, whose stoppage blocks a line or a customer order. Where a machine has a buffer stock or a fallback station, a failure costs far less, and fine-grained monitoring rarely pays.

Progressive, detectable failure modes

Bearings, spindles, pumps, compressors and gearboxes usually degrade in a measurable way before they fail. A sudden electronic fault, an operator error or a material defect leaves no signal to work with. So the failure history has to be examined to see what share shows an observable period of degradation.

The ability to act in time

Knowing that a bearing will last three more weeks is useless if the part has a six-week lead time, if the technician is unavailable, or if production will not release the machine. Critical spares, joint planning between maintenance and production, and availability of the people who intervene weigh as much as the sensor. The routing and scheduling of field technicians belongs to the same chain.

Minimal history and recording discipline

Without traceability of failures (date, duration, cause, part replaced), the gain cannot be quantified and the relevance of alerts cannot be checked. Many SMEs discover at this stage that their maintenance system is only partly filled in. That is an organizational task, not a technology one, of the same nature as those we describe in our analysis of where support functions lose hours.

An order-of-magnitude calculation

Take an illustrative example, with explicit assumptions that each executive should replace with their own.

• Scope: 12 critical machines.
• Unplanned failures in that scope: 15 per year, averaging 8 hours, so 120 hours.
• Cost of one hour of downtime (lost margin, overtime, express shipping, penalties): CHF 3,000.
• Annual loss: 120 × 3,000 = CHF 360,000.
• Share of failures that can realistically be anticipated: 40%, so CHF 144,000 of loss concerned.
• Residual cost of a planned stop: 25% of that loss, or CHF 36,000.
• Gross annual gain at steady state: about CHF 108,000.

On the cost side: sensors, gateway and installation at CHF 3,000 to 5,000 per machine, about CHF 50,000 up front; software subscription of about CHF 15,000 a year; the time of a technician or supervisor to handle alerts, about 0.2 FTE or CHF 25,000 a year. Recurring costs: about CHF 40,000.

Assuming a first year at half effectiveness while the system ramps up, the cumulative position turns positive after roughly eighteen months. The recurring net gain is then close to CHF 68,000 a year.

Now change a single assumption: an hour of downtime costs only CHF 800, because the machines have buffers. The annual loss falls to CHF 96,000, the net gain to about CHF 29,000, which is less than the recurring costs. The project never pays back. Hourly downtime cost is the variable that decides, well before the choice of vendor.

What the calculation does not say

Three points deserve attention before signing.

• False positives. A system that alerts too often ends up ignored. Plan for a calibration period and a named person who qualifies alerts.
• Data ownership. When monitoring is offered by the machine builder, settle by contract who owns the data, who may reuse it and what happens when the subscription ends. Machine data is mostly non-personal, but once it is tied to an identifiable operator, the revised Swiss Federal Act on Data Protection (in force since 1 September 2023) may apply.
• The European framework. The EU AI Act, in force since 1 August 2024 with phased application, does not in general treat predictive maintenance as a high-risk use. A machine safety component is a different matter and should be discussed with the supplier.

A three-step path

Step 1: map, without sensors. Rank machines by criticality, reconstruct two to three years of failures, quantify the hourly cost of downtime and identify progressive failure modes. This takes a few weeks and often eliminates half the ideas. It also surfaces immediate gains: an adjusted preventive plan, critical spares stock, analysis of recurring causes.

Step 2: pilot on three to five assets for six months. Choose machines where the conditions above are met, and set in advance the success criterion (failures anticipated, downtime hours avoided) and the stopping criterion.

Step 3: extend only if the pilot has proved the gain. Extension assumes the teams have adopted the alert routine, which is as much a matter of organization and adoption as of technology.

Implications for leadership

In manufacturing, predictive maintenance is one tool among others within operations and processes, and rarely the first to reach for. Three questions help decide.

1. What does one hour of downtime cost on our five most critical machines, calculated with finance rather than guessed?
2. What share of our past failures showed early warning signs?
3. If an alert arrives tomorrow, who decides, with which part, and in which time slot?

If the answers are clear, a targeted pilot has a good chance of paying back. If not, the better investment is first recording discipline and coordination between maintenance and production. AI adds value only once those foundations are in place.

To see where your company stands, a WALVIS diagnostic can quantify the downtime cost of your critical machines and help you decide, within a few weeks, whether predictive maintenance deserves a pilot or whether other levers will pay back faster.