There are three ways to decide when a machine gets serviced. Run it until it breaks, that's reactive. Service it every 500 hours whether it needs it or not, that's preventive. Or measure it and service it when the measurement says to, that's predictive. The trigger is the whole definition. Everything else, the sensors, the dashboards, the vendor's AI claims, is implementation detail around that one design decision.
Predictive vs preventive maintenance covers the economic trade-off between the last two in full. This page covers what predictive actually is and what it takes to run one.
The sensing to work-order loop
A predictive maintenance system is four stages, and it only produces value if all four are actually connected.
Sensing. Vibration, thermal, acoustic, or current-draw sensors on a piece of equipment produce a continuous signal. This is the cheapest part of the system and the part vendors show you first.
Modeling. A model, trained on that asset's own failure history and often on a baseline from similar assets, scores how far the current signal has drifted from healthy. This is where "failure history" stops being a nice-to-have and becomes the thing the whole system is built on. No history, no model, or a model trained on too little data to trust.
Alerting. When the deviation score crosses a threshold, the system raises an alert, ideally with enough specificity that a reliability engineer can act on it, "bearing wear signature on Line 4 Motor 2," not just "anomaly detected."
Work order. The alert has to become a scheduled intervention, not a notification someone dismisses. In a plant with a CMMS, this is a direct handoff: the alert opens a work order, a technician gets dispatched before the failure, and the outcome, parts replaced, hours saved, gets logged back into the same maintenance history the model was trained on. That closed loop, alert in, history out, is what separates a working predictive program from a dashboard nobody looks at.
What it actually requires
Failure history. A model needs to have seen failures, or close approximations of them, to learn what one looks like before it happens. This is why predictive maintenance is not a good first project for a plant with a thin maintenance record. It's also why a well-run CMMS is worth more than it looks like on the surface, its history is exactly this input.
Sensor coverage on the right assets. Instrumenting every machine in a plant is rarely worth it. Predictive maintenance pays off on assets where failure is expensive or dangerous and where the failure mode is physically measurable, bearings, motors, rotating equipment. A $200 pump that's cheap to replace on failure doesn't need a sensor.
Someone who acts on the alert. The single most common way predictive maintenance programs fail isn't a bad model, it's an alert that goes to an inbox nobody checks. The workflow has to route to a person with the authority and the parts on hand to act before the model's prediction window closes.
Two deployments, worked
Fiberon, a Fortune Brands Innovations consumer-goods plant, deployed Augury's machine-health monitoring and over eight months saved $274,000 and avoided 178 hours of downtime, a 2.5x ROI. Do the arithmetic on that: $274,000 over 178 hours works out to roughly $1,540 saved per hour of downtime avoided, and annualized, the eight-month figure runs at roughly $411,000 a year. That's the actual economic case for predictive maintenance in one plant: not a percentage on a slide, a dollar figure per hour of downtime that didn't happen.
DuPont ran Augury at proof-of-concept sites and reported 7x ROI in under a year, with 100% prediction accuracy at those sites. Read that accuracy figure for what it is, a proof-of-concept result at a small number of sites, not a claim that the model never misses at scale. The honest version of that story is still a strong one: fast payback, concentrated on the assets where it was tried first.
What the benchmark actually says
Our Predictive Maintenance benchmark covers 62 deployments. Reported annual savings sit at a median of $200,000, ranging from $20,000 to $45 million depending on scale, across 8 deployments that report a figure. Productivity and OEE gains sit at a 9% median across 5 deployments. By vendor, the deployments concentrate heavily: Rockwell Automation (48), Augury (9), C3.ai (2), and Tulip (1).
The benchmark deliberately publishes no downtime median, which is worth explaining because downtime is the outcome this category is sold on. Only five deployments in the corpus report a parseable downtime figure, and three of those are the identical round number 80%. A median of that sample would describe round-number clustering in vendor case studies, not the results a plant should expect. The individual outcomes are real and worth reading — Fiberon avoided 178 hours of downtime in eight months — but the average of five figures, three of them identical, is not a number worth printing.
Where to go next
- For the trigger-level distinction and a rule for which assets deserve which approach, see predictive vs preventive maintenance.
- Predictive maintenance is built on top of the maintenance history a CMMS accumulates. What is a CMMS covers what that record actually contains.
- The full list of documented deployments, filterable by industry and vendor, is on the Predictive Maintenance use-case hub.
- Vendors active in this category are ranked with their evidence on the manufacturing operations management software hub.
- For the sensing technology underneath these deployments, see IoT and sensors.
AI Roadmap — Eight workflows in your plant where AI would pay back this quarter. We know which. 12 pages, in 72 hours.