Prediction accuracy is not the hard part anymore. On a semiconductor line running six-second sampling, Decisyon models call failure events with better than 90% live accuracy and hand back between 24 minutes and 11 hours of warning. The question that decides whether a predictive maintenance program pays for itself is what happens inside that window. A prediction nobody acts on is an expensive alarm. Score yourself on the ten questions below, then see what the window is worth on your assets. A reliability engineer at a packaging plant has a dashboard that works. It flagged the number-two filler eleven days ago. The vibration signature was clear, the model was confident, and the alert went to a shared inbox that four people can see. The filler came down on a Thursday. The alert was correct. Nobody opened a work order. This is the failure mode that almost never appears in a vendor evaluation, because evaluations are built around whether the prediction is right. The prediction was right. What was missing was everything between the alert and a technician with a part in their hand.
The gap is not in the model
Most plants that are disappointed by predictive maintenance are not disappointed by the accuracy. They are disappointed by the return, and the return leaks in three places that have nothing to do with data science. The warning arrives and no work order opens, because opening one requires a person to decide to. The technician who takes the call gets a fault code but not the procedure, so the first forty minutes go to finding out what to do rather than doing it. And when the same asset class fails at another plant in the network six months later, nobody there can find out that this plant already solved it. Each of those is a handoff. None of them is a modeling problem. All of them are why a program that predicts accurately can still fail to change the downtime number.
The audit
Ten questions. Answer each one yes or no, about the program you actually run rather than the one in the implementation plan. If you would need to check with someone before answering, that is a no.
Signal: do you actually know?
- You can name which of your critical assets are instrumented and which are not, without asking anyone.
- The condition data comes off the machine directly, not off a round sheet or an operator's entry.
- You could rank your top five failure modes by cost rather than by frequency.
Response: does the warning become work?
- When a prediction fires at 2 a.m., a work order opens without a person deciding to open it.
- The technician who receives the alert also receives the procedure, not just the fault code.
- You know how many of last quarter's predictions were acted on before the warning window closed. If you would have to estimate, that is a no.
- A prediction that nobody acknowledges escalates on its own.
Memory: does the plant get smarter?
- When the same asset class fails at another plant, this plant finds out without someone remembering to send an email.
- The corrective action from the last occurrence is findable in under a minute by whoever is on shift tonight.
- If your most experienced reliability engineer left tomorrow, the plant would still know which assets to watch and why.
What your score means
8 to 10. Your program is converting warnings into work. Further gains will come from widening coverage to more assets, not from tightening the model. 5 to 7. The most common result. You have a good signal and a weak response. The models are earning their keep and the organization around them is not, which shows up as a downtime number that improves less than the accuracy number would suggest. 0 to 4. You have monitoring, not predictive maintenance. The distinction is not the sensors. It is whether anything happens when the alert fires and nobody is looking at the screen. The pattern worth noticing: most plants score well on signal and badly on response. Signal is what the vendor sold and what the pilot proved. Response is what the plant has to build, and it is where the return actually lives.
What the tiers look like side by side
| Capability | Run to fail | Condition monitoring | Predictive with execution |
|---|---|---|---|
| Live condition data off the asset | No | Yes | Yes |
| Advance warning before failure | No | Minutes | Hours |
| Work order opens automatically | No | No | Yes |
| Procedure routed with the alert | No | No | Yes |
| Unacknowledged prediction escalates | No | No | Yes |
| Prior fix findable at the line | No | No | Yes |
| Fix shared across plants | No | No | Yes |
| The jump most plants think they are making is column one to column two. The jump that changes the downtime number is column two to column three, and it is mostly not a sensor purchase. |
What it looks like when it holds
A global electronics and semiconductor manufacturer runs six-second sensor sampling into live failure models, gets better than 90% prediction accuracy with 24 minutes to 11 hours of advance warning, and reports effectively zero unplanned machine downtime at the pilot plant. Read the full breakdown. Schneider Electric runs the execution layer across more than 200 plants and over 110,000 users, which is what makes a fix found at one site available to the next. Read the full breakdown. The semiconductor result is the interesting one, because near-zero unplanned downtime did not come from the accuracy alone. It came from the accuracy plus a response that fired inside the window every time.
Four questions to ask any vendor in this category
Skip the anomaly scores and the failure curves in the demo. Every vendor's charts look good.
- Show me, live, what happens when a prediction fires at 2 a.m. and nobody is at a screen.
- Does the technician get the repair procedure with the alert, or a fault code and a search box?
- How do I report what percentage of predictions were acted on inside the warning window?
- If a different plant already fixed this failure mode, how does the technician on shift tonight find out? If the answers are an email, a fault code, you would have to build that, and they would have to ask around, you are buying monitoring.
Where Decisyon fits
Smart Gateway connects to the assets you already run, including legacy equipment, and standardizes what they emit. Predictive Maintenance turns the signal into a prescribed action rather than a score. The response layer is where the warning window actually gets used. Fix Finder reads incident and corrective-action history across every connected plant and routes the fix that already worked to whoever is standing at the asset now. Deep Find puts the SOP, the manual and the work instruction in the same place, in plain language, so the first forty minutes of a repair are not spent locating the procedure. Both are agents in the Decisyon AI Workforce. What changes is not the accuracy of the prediction. It is whether the eleven hours you were given got used. See what that window is worth on your assets → · Read the customer results →
