A temperature trace rises before a machine stops. To an analyst, it looks like the beginning of a predictive-maintenance story. To the technician standing beside the machine, it may be a known change after cleaning, a sensor moved during a repair, or a warning that matters only when another condition is present. The chart has information. The people who operate the line have context.
I would not treat either source as sufficient on its own. Manufacturing AI earns its place when it helps somebody make a better decision during a real shift, under the constraints that shift actually faces.
Start with the intervention, not the sensor feed
It is tempting to begin with whatever data a plant has collected. There may be years of sensor readings and maintenance records, so the question becomes, “What can we predict?” I would reverse it: what action is the plant trying to take earlier or more consistently?
For a maintenance team, the action may be planning an inspection before an avoidable stoppage. For quality, it may be separating a batch for human review before defects travel downstream. For a production planner, it may be deciding whether a changeover is feasible with the staff and materials available. These decisions have different time windows and different costs of a false alarm.
Ask the team to describe the current decision. Who notices the issue? What else must they check? Who can authorize a stop or a change? If the proposed signal arrives after that choice has already been made, the model can be technically accurate and operationally irrelevant.
Make the data argument in the plant
Data collected for operating equipment is not automatically suitable for evaluating a new decision. A maintenance log may record the repair but not when the fault first became visible. Operators may use different labels for the same condition. A sensor may have been replaced halfway through the history. None of this is a criticism of the plant; the information was created for a different purpose.
I would ask our delivery team to review a handful of historical events with operators, maintenance and engineering together. Put the timeline on the table: the reading, the shift note, the alarm, the inspection, the action and the outcome. Where accounts differ, preserve the disagreement until it is understood. A clean-looking training set built by hiding uncertainty can teach a model the wrong lesson.
This exercise also reveals whether the plant has enough examples of the condition it hopes to predict. If not, a simpler rule, better instrumentation or a disciplined logging practice may be the right first investment. There is no shame in improving the evidence before asking an algorithm to interpret it.
Design an alert someone can use
Suppose a model flags possible trouble during a busy shift. The operator needs more than a red icon. What changed? How certain is the system? What should be checked first? How does the team record a false alarm or a genuine issue? What happens when the person with authority to act is unavailable?
I would expect our team to trial the alert alongside the current process, with clear responsibility for reviewing it. Watch what people do, not just whether the prediction matches a historical label. If alarms are dismissed because they arrive too often, or because acting on them requires a shutdown nobody can authorize, the design has failed even if the model score looks respectable.
The best result may be an advisory tool rather than automatic control. An operator can compare the signal with sound, material condition and recent maintenance work in a way the data may not capture. Their ability to challenge the system should be part of its design and evaluation.
What this means for leaders
I would make the plant team's place on the pilot's decision group a condition from the beginning. Give them time to question the data, define useful lead time and identify the cost of acting or waiting. Evaluate a change in planned work, avoidable disruption and trust in the alert, with appropriate attention to safety and quality.
If an AI proposal cannot explain what the next shift will do differently, it is still a research exercise. Useful manufacturing intelligence begins with the people who will carry the decision.
Put this insight to work.
If you are weighing a similar decision, our team can help you clarify the problem, test your options and choose a practical next step.