Key takeaways
See our roundup of the platforms this should be measured against.
The standard PdM pitch is structured around a small number of high-impact failures. A motor seizes, a bearing destroys itself, a gearbox fails catastrophically. The avoided cost of even one of these events, production downtime, secondary damage, emergency parts, runs into hundreds of thousands of currency-units. PdM detects the early signal, the maintenance team intervenes, the catastrophic event never happens. Multiply by three or four assumed avoidances per year, the case pays back the sensors and software within 18 months.
The pitch is internally consistent and almost never accurate as written. Three structural errors hide inside it.
Most PdM business cases implicitly compare against a baseline where the asset would have run to failure. That is rarely the actual baseline. The real baseline is whatever the plant's current PM and condition-based program would have caught.
A well-tuned PM program catches many of the failure modes PdM is sold to prevent. Bearing failures are caught by lubrication-cycle checks. Belt failures are caught by visual inspection. Motor seizures are often preceded by current-draw signatures that an alert technician notices on routine rounds. The honest PdM ROI subtracts the gains the existing program is already producing, not the catastrophic events it would not have prevented. Department of Energy figures capture exactly this gap: predictive maintenance tends to save on the order of 8-12% over a well-run preventive programme, but 30-40% over purely reactive maintenance, so a case benchmarked against "no maintenance" looks far stronger than one benchmarked against the PM programme the plant already runs.
Plants that run this subtraction usually find the PdM payback period stretches well beyond the pitched 12-18 months, often into multiple years. Still positive on some asset classes, but no longer a clear-cut decision. The piece on the preventive maintenance schedule covers how a well-tuned PM program performs against PdM-style failure modes.
PdM models produce alerts. Some alerts are real; some are not. The false-positive rate matters because every false alert consumes maintenance time investigating a non-failure and erodes operator trust in subsequent alerts.
Vendor business cases often assume high model accuracy. The achieved accuracy on real installed systems is typically much lower in the first year and improves only after extensive tuning on the plant's own assets. The gap matters: even a modest false-positive rate can erode the projected savings, one widely-cited McKinsey programme saw a 10% false-positive rate wipe out its gains. When false positives accompany a meaningful share of alerts, the investigation cost of chasing them can exceed the labour saved on the true positives, especially in the first year.
A more honest case models false-positive cost explicitly: false-positive rate × average investigation cost × annual alert volume. The number can consume a large share of the projected ROI, often enough to turn a slam-dunk case into a marginal one. The article on root cause analysis covers the investigation overhead in adjacent terms.
PdM vendors talk about model accuracy as a single number ("our system is 92% accurate"). Real accuracy varies dramatically by asset class. A rotating-machinery vibration model is well-developed and high-accuracy. A gearbox-temperature model is reasonable. A reciprocating-asset model is less mature. A custom or low-volume asset class may have no working model at all.
The business case usually applies a single accuracy figure to a heterogeneous asset population. The asset classes where the model is well-developed will perform; the classes where it is not will produce mostly noise. A plant-wide rollout averages a productive subset with a noisy subset and delivers below the pitched number.
The honest case scopes per asset class: which assets the available model actually performs on. That subset is often only a fraction of the total assets the pitch covers. The deployment-cost-per-protected-asset becomes substantially higher than the vendor's plant-wide number suggested. The piece on manufacturing KPIs covers the per-class metric framing this depends on.
Targeted PdM, deployed on specific asset classes with well-developed models and high failure cost, can pay back cleanly. The criteria:
Plants that deploy against these criteria, targeted, on instrumented critical assets with mature models, typically see a clean payback on the targeted subset. Plants that deploy plant-wide because the vendor offered a discount usually see no clear payback at all.
If you are evaluating a PdM proposal, the four questions:
Answers that come back specific and defensible point to a project worth doing. Answers that come back vague point to a project that will disappoint the operator team in year two. The article on the work order management system covers how the targeted PdM alerts get integrated into the existing work-order flow.
Fabrico does not pitch a plant-wide PdM solution. The platform's role is to be the operational data layer that any PdM project, vendor's or in-house, lands on top of. Failure modes, asset hierarchy, work-order integration, the false-positive investigation loop, these all need a stable foundation before PdM can deliver value. To see how a targeted PdM project would fit against your current operational data, book a demo.
Usually not. Vendors apply benchmark cases from their best-case deployments. The mismatch comes from buyers' situations being more average than the benchmark. Read every case study as "best case for the vendor's most favourable customer" and adjust expectations accordingly.
"AI" labels apply to models that range from genuinely sophisticated to repackaged threshold rules. The honest test is the same: ask for false-positive rates on the asset classes you actually have, not on a generic benchmark.
Yes, on the right scope. Targeted PdM on rotating equipment with mature vibration models is a defensible investment for most mid-market plants. Plant-wide PdM rollouts usually are not.
One quarter, three assets, with explicit success criteria defined before the pilot starts. Shorter pilots do not collect enough failure events to measure model accuracy; longer pilots delay the deployment decision past the point where vendor pricing pressure matters.
Committing to plant-wide deployment before validating model performance on the plant's own asset population. The vendor's accuracy claims are usually true for some asset classes and false for others, and the only way to know which is which is to test on your assets, not theirs.
Login
Blog
