The operations director of a private industrial-water operator receives a commercial proposal from a sensing company. The package is concrete: vibration, current and temperature instrumentation for the forty critical pumps in the treatment-plant portfolio, cloud analytics that claim to detect degradation three to six weeks before failure, and a combined tariff of one hundred eighty thousand euros upfront plus twenty-four thousand annually in subscription. The sales rep points to a published case with a forty percent corrective reduction and calls the model "the future of maintenance." The operations director has to decide. Preventive maintenance on the same pumps currently costs around eight thousand euros per asset per year, fully loaded. The math has to work before signing. Before the meeting with finance, a defensible reading must be produced.
Predictive maintenance sits today at the boundary between commercial discourse and real operational value. For some asset classes and some operational profiles it produces measurable return on a reasonable horizon; for others it adds cost, complexity and operational noise without significant benefit. The question is not whether to adopt predictive as a general maintenance philosophy, but which specific assets, at what moment, with what realistic expectation of corrective reduction. Four concrete conditions decide whether the jump pays off in a given operational context.
What predictive maintenance actually is
Before the conditions, the technical definition helps to defuse the hype. Predictive maintenance is the continuous or high-frequency observation of an asset's behavior through sensing, with pattern detection by threshold, by trend analysis or by statistical model of failure signature. It differs from traditional preventive maintenance in that the latter acts by calendar, and from simple condition-based maintenance in that the latter applies point readings without pattern analysis. The predictive layer adds sensor cost, connectivity cost to the central platform, data infrastructure cost and alert interpretation cost. That cost is not marginal, and that is why the comparison with the expected corrective saving must be done per asset class, not as a global maintenance-policy decision.
The four conditions that support the case
The first condition is a high cost of unplanned failure per asset. It has three components that must be summed before comparing against the sensing cost: downtime cost valued at the business margin during out-of-service hours, direct repair cost under urgent corrective versus planned intervention, and cascade cost if the failure affects downstream assets or SLA-bound tenants. When the sum of the three is substantially larger than the annualized cost of the sensing package, the first condition holds; when it is of the same order or smaller, it does not.
The second condition is availability of enough failure history to train models or calibrate thresholds seriously. A predictive model without prior behavioral data on the asset, or without access to a well-mapped equivalent industrial corpus, is a fixed threshold with an elegant name. If the asset class has two or three years of structured traceability with enough failure events to characterize patterns, the second condition holds; if not, the instrumentation runs but the analytics are speculative.
The third condition is per-asset economics. The instrumentation package has a sensible minimum threshold: if the asset's unit value and its annualized failure cost are modest, the package cannot amortize. As an order of magnitude, assets with replacement value below the twenty-five to fifty thousand euro range typically do not justify individual predictive sensing, unless criticality is exceptional. Above that range, the economic sensitivity shifts and the case holds more often.
The fourth condition is the operational maturity to act on predictive alerts. A degradation warning three weeks before failure does not generate savings by itself: it generates savings when the organization has a plannable maintenance window, has the necessary spare on hand, and has the intervention-on-alert process defined, staffed and practiced. Without operational maturity, alerts accumulate without translating into action, and the predictive system becomes a source of noise the organization quietly begins to ignore.
The three failure modes when the jump is premature
When the decision is taken without evaluating the four conditions, three predictable failure modes appear within months. The first is instrumentation on assets whose expected failure cost is below the annualized cost of the sensing package; the system runs technically but produces no return. The second is predictive alerts that no one acts on: they are emitted at the expected cadence but the intervention chain is not prepared to handle them. The third is false-positive fatigue: poorly calibrated models or thresholds generate alerts most of which do not anticipate real failure, and the maintenance team learns to ignore them, with the consequence that the true positives are lost as well.
Operational integration as a precondition
What Maptainer contributes is the foundation that decision rests on: a georeferenced asset register with criticality, corrective and preventive history linked to the specific asset and the responsible contractor, accessible through an API documented in OpenAPI and syncable with external systems through REST APIs. The sensing platform and the failure model live outside; what they get from Maptainer is a clean, traceable history, which is the raw material of any model. This foundation does not remove the need for the four conditions to hold, but it is what allows the result, when they do hold, to translate operationally into decisions rather than into parallel dashboards.
The granular decision per asset class
Predictive maintenance is not a maintenance philosophy adopted for the whole portfolio. It is an investment made on specific asset classes where the four conditions are met. The asset manager who says "no predictive" on a class where the conditions do not hold is doing the same defensible technical work as the one who says "yes predictive" on the class where they do. The conversation with finance holds up far better when the proposal identifies two or three classes where sensing pays off with detailed economics and explicitly rules out the rest of the portfolio with the same discipline. That selectivity is what separates an operator that invests in predictive with judgment from one that buys it from the sales rep who calls.