Prediction is useful only when sensor evidence, diagnostic context, maintenance authority and post-repair verification connect in the real operating workflow.
An alert has a cost even when the machine does not fail
Predictive maintenance is often introduced through successful detections. Yet someone must examine each alert, confirm the operating context, decide whether to inspect, plan spare parts, obtain authorization and verify the result. A 2025 NIST systematic review notes that condition-monitoring adoption requires evaluation of engineering and financial benefits. An alert count alone is therefore a poor proxy for value delivered. For a buyer, the relevant question is not only “Did the system detect something?” but “What did this detection cause, and was the consequence justified?”
A simple workload example—explicitly hypothetical
Suppose a pilot generates 120 alerts in a month, with an average of 30 minutes of human review per alert. That is 60 technician or analyst hours before travel, planned inspections or repair work. If 30 alerts lead to useful action and 90 do not, the non-actionable alert investigation alone accounts for 45 review hours. This is a teaching illustration, not observed performance data from Siemens Senseye, SKF, I-care, Augury or Tractian. “Non-actionable” does not necessarily mean a model false positive: duplicate alerts, already-known issues and correct but low-priority findings may also be non-actionable.
Measure misses as carefully as alarms
A low alert volume may simply indicate a high threshold or limited sensing coverage. Define the set of failure modes that are observable with the chosen sensors, sampling rate and asset context. Track confirmed missed events and near misses against a transparent ground-truth process. Avoid comparing suppliers using a single precision or recall figure without the asset population, severity distribution, observation period and intervention rule. A repair avoided through early action and a failure that never would have occurred require different causal explanations.
Compare the entire route to a maintenance decision
Supplier product pages describe different combinations of sensing, analytics, service expertise and work-order integrations. For example, Siemens documents Senseye predictive maintenance while Tractian describes asset-health monitoring and CMMS integration. Those descriptions do not demonstrate comparative accuracy. Ask each provider how an alert is identified, prioritized, reviewed, converted into an approved task, linked to an asset record, and closed with post-maintenance evidence. Include who supplies reliability expertise, and whether work remains possible during a connectivity outage.
Build a fair pilot scorecard
For a defined asset group and observation period, record: alerts per 100 monitored asset-days; percentage requiring review; human minutes per alert; actionable findings; lead time before intervention; confirmed missed failure modes; avoided unplanned downtime with a documented baseline; and total cost of sensing, software, integration and labor. Segment these figures by machine type, operating regime and asset criticality. Publish both positive and adverse outcomes, including changes in routine inspection workload.
Keep the causal claim appropriately narrow
A condition alert may support a useful maintenance decision without proving that a catastrophic failure was prevented. Compare downtime and maintenance outcomes against comparable equipment and periods while recording confounders such as asset replacement, production load changes and scheduled overhaul. The most credible PdM purchasing decision is one that can explain the whole chain from measured condition to justified action and observed effect.
Practical takeaway
Before accepting an ROI slide, ask the supplier to show the alert-to-action record and the full operating cost of making that record usable. A detection is a technical event; maintenance value is an evidenced operating result.
Make the next vendor conversation more productive.
Use the free project fit check to generate an initial list of questions for your team.
Open planning tool ↗Primary sources and limitations
These links support the public product or research context. They do not substantiate comparative performance conclusions. Source content may change; verify directly with the supplier.