- A predictive maintenance program is an operating change, not a software purchase. The tooling is the smallest decision in it.
- Start from the assets whose failure costs the most, not the assets with the best data. The data gap is solvable; a program aimed at cheap equipment never earns its budget.
- Most of the effort, and most of the cost, sits in making existing sensor and maintenance history usable together. Price that separately from the modeling.
- Pick the measure before go-live. A program without an agreed definition of success gets judged on whichever number looks worst at review time.
Somewhere between reading how predictive maintenance works and asking who should build it sits the question that actually decides the budget: is this worth doing here, on which equipment, and what would it return.
If you are still working out what the technology does, start with the guide to predictive maintenance for industrial operations. If you have decided and are comparing suppliers, the roundup of predictive maintenance companies covers nine of them.
What a Predictive Maintenance Program Actually Is
The word that matters is program. Buying a monitoring platform is a procurement event; running a predictive maintenance program changes how maintenance decisions get made, who makes them, and what evidence they rest on.
That distinction shows up in the failure mode. Plants rarely fail to install sensors. They fail to change the work order that follows, so the readings accumulate in a system nobody is accountable for, and the annual review finds a subscription with no attributable saving behind it.
A program has four things a purchase does not: a defined asset scope, an owner who is accountable for acting on what it produces, an agreed measure, and a route by which a prediction becomes scheduled work. If any of the four is missing at the start, it does not get added later.
Which Assets Justify Predictive Maintenance
This is the decision that determines everything downstream, and the instinct most operations follow is the wrong one.
The instinct is to start with the equipment that already has good instrumentation, because it looks easiest. That produces a technically successful pilot on assets whose failure nobody minds, and no business case for extending it.
Start instead from consequence. For each candidate asset, ask three questions:
- What does an unplanned stop actually cost? Not the repair — the production it defers, the crew standing idle, the contractual penalty, the restart. On process plant this figure is usually an order of magnitude above the repair.
- How much warning would change the outcome? Some failures are worth predicting because two weeks’ notice turns an emergency into planned work. Others fail with no useful lead time, or are cheap to swap on failure, and belong on run-to-failure however sophisticated your tooling.
- Does it fail in ways that leave a signal? Rotating equipment — pumps, compressors, motors, fans — degrades along measurable paths. That is why nearly every program in heavy industry starts there.
Five assets is the right number for a first phase. Enough to prove the mechanism, few enough that a failed phase is a lesson rather than a write-off.
The Data a Predictive Maintenance Strategy Depends On
The uncomfortable part of any predictive maintenance strategy is that the modeling is rarely the constraint. The constraint is whether three things can be read together for the same asset:
- Condition readings over enough history to show what healthy looked like, from a historian or from retrofitted sensors.
- Operating context — load, throughput, mode — because a reading is only abnormal relative to what the machine was being asked to do.
- Maintenance history — what was done, when, and what was found — because that is where the model learns what a failure looks like.
Most operations hold all three and cannot join them, usually because the same pump is identified differently in each system. Resolving that is unglamorous, invisible to the board, and routinely the majority of the work.
We would rather say that plainly than discover it in month two. On an oil and gas program, the data infrastructure was the entire first engagement, with no prediction attempted until it was in place. On a separate production-pumps program, sensor data structuring and integration came before any model was built — and that program reached 90% prediction accuracy, cut maintenance costs by 40%, and reduced unplanned downtime by 65%. Those figures belong to that operation and its asset class.
The practical test before you commission anything: can you pull two years of readings for your five candidate assets, alongside every work order raised against them? If yes, you are scoping a modeling project. If no, the first project is the data, and it should be priced as its own piece of work.
Building the Predictive Maintenance Business Case
A predictive maintenance business case that survives scrutiny counts avoided events, not vendor averages.
What to count
- Deferred production avoided. Usually the largest number, and the one operations can estimate credibly from history.
- Unplanned interventions converted to planned. Cheaper labor, no premium parts, no shutdown around them.
- Preventive work not done. Servicing on condition rather than on a calendar removes unnecessary interventions.
- Extended asset life, where you can evidence it.
What not to count
Published predictive maintenance ROI percentages from anyone’s marketing. They were measured on someone else’s equipment, and a good number of them turn out to be category averages rather than any customer’s result. A finance director who spots one borrowed figure will discount the whole case.
Build it from your own failure history instead. Take the last two years of unplanned stops on the candidate assets, price them, and ask what proportion had a detectable lead-up. That proportion is your realistic ceiling, and it is defensible because it comes from your own records.
Where Predictive Maintenance Programs Stall
- No owner. The most common cause. Sensors have a budget holder; predictions need someone accountable for acting.
- Alerts nobody trusts. Once ignored, a system is finished socially long before anyone cancels it.
- Data discovered late. Found in month two rather than week three, it looks like failure rather than the scope it always was.
- A pilot with no exit. Without an agreed measure, a first phase runs until someone loses patience.
- Success nobody can prove. If the baseline was not recorded before go-live, the saving cannot be demonstrated afterwards.
How to Implement Predictive Maintenance: The First 90 Days
A phased shape that can be stopped cheaply at any point, which matters more than speed:
- Weeks 1–2 — scope. Confirm the five assets and their failure history. Agree the measure now, in writing.
- Weeks 3–6 — data. Join readings, operating context, and maintenance records for those assets. This is where a program either becomes possible or reveals what has to be fixed first.
- Weeks 7–10 — model and validate. Build against known failures, and have the engineers who will use the output check its calls before anyone relies on them. Their skepticism at this stage is the cheapest quality control available.
- Weeks 11–12 — wire it into the work. A prediction that does not become a work order is a screenshot. Route it into the system your planners already use.
- Then decide. Extend, adjust the asset list, or stop. All three are acceptable outcomes of a first phase; only “continue indefinitely without a measure” is not.
How Crunch-IS Runs a Predictive Maintenance Program
We work that sequence in the order it is written, which mostly means the unglamorous part comes first. Where condition readings, operating context, and maintenance history cannot be joined for the candidate assets, that join is scoped and priced as its own engagement before any model is discussed. On one oil and gas program, it was the entire first phase, with no prediction attempted until it was in place.
Two programs, two asset classes, and the figures belong to the operations they were measured on:
- The production-pumps program described earlier — 90% prediction accuracy, maintenance costs down 40%, and unplanned downtime down 65%, on rare high-value equipment for a European producer.
- An AI-powered water main failure prediction platform for a UK utility — 42% fewer unexpected failures, with a further 28% reduction in emergency maintenance costs projected rather than measured.
Neither figure transfers. They are what those asset classes returned under those conditions, and the only number that will predict yours is the one already sitting in your own failure history.
