Prediction is only valuable when the factory can act

An early warning is only useful if it arrives while the factory still has choices. A developing fault may be visible days before failure, yet production commitments, maintenance workload and spare-parts availability can still leave the plant with no practical way to intervene. Predictive maintenance has improved the timing of the warning, but the operational value depends on what the organization can do with that time.

Jared Pfeiffer, Product Manager for Operations Agent at Augury, has seen machine-health decisions from several sides of the plant, having worked as a maintenance planner and manager as well as a production leader and plant manager. That experience gives him a practical view of what happens after a prediction reaches the factory floor. “The industry has gotten much better at identifying that something is going wrong,” he says. “The harder problem is deciding what to do about it. A prediction might tell you that a bearing, gearbox or motor is developing a problem, but the team still must determine why it is happening, how quickly it will become serious, what the consequence of failure would be and when they can safely intervene.”

The alert is not the decision

Machine-health data enters an operating environment already shaped by production commitments and constraints that maintenance does not control. A diagnosis may be technically accurate while the decision around it still depends on what the line is producing and what would happen if the asset were taken out of service. That broader operating picture determines whether an alert becomes useful intelligence or simply another item competing for attention.

“If I am coming at this as a maintenance person, I do not just want to look at one piece of equipment with blinders on,” Pfeiffer says. “I also need information from operations about what product is running on the line, what the deadline is and what happens to production or shipments if that machine goes down. The goal moves from asking whether there is a problem to saying we believe the problem is here, here is the evidence and here are the appropriate next steps.”

Maintenance teams rarely have the capacity to respond immediately to every developing condition across a plant. When several assets demand attention at the same time, choosing where to intervene becomes a question of how limited maintenance resources should be deployed while production continues around them. The apparent urgency of the fault is therefore only the starting point for that decision.

“A problem on a redundant asset with a spare available is very different from the same problem on a constrained production line running a high-value product with no replacement,” Pfeiffer explains. “You must consider asset criticality, how fast the condition is deteriorating, production requirements, quality and safety consequences and what resources are available. A reliability engineer, a maintenance manager and a production leader will all look at the same problem differently, so the opportunity is to give each of them the operational context to make a coordinated decision.”

A diagnosis only creates value when it can be converted into work quickly enough to change the outcome. That means moving from technical evidence to an approved intervention, with the right people, parts and timing aligned around it. Pfeiffer argues that this handoff is where many otherwise sophisticated maintenance programs still lose time.

“One of the biggest sources of friction in manufacturing is that we work across different systems,” he says. “Someone must switch between the CMMS, ERP, supply chain and MES and build the story themselves. The objective is not to replace those systems. It is to connect the intelligence across them so somebody investigating an issue can understand the problem and move toward action within the workflows they already use.”

Maintenance and production need the same picture

Maintenance and production are often accountable for different outcomes, even when they are dealing with the same piece of equipment. One team is responsible for protecting reliability while the other is measured on output and customer commitments, so the same developing fault can create competing pressures inside the plant. Pfeiffer’s experience in both functions has shown him how easily those priorities can pull decisions in different directions.

“Neither perspective is wrong,” he says. “Maintenance may want to repair the equipment immediately because they understand the downstream impact, while production knows that stopping now could jeopardize an order. If we can show how quickly the condition is changing, the risk of continuing to run, what production is scheduled and what resources are required, the conversation shifts from maintenance versus production to what is the best outcome for the plant.”

Even a well-timed decision is difficult to execute without the people, skills and parts required to complete the work. Augury’s 2026 State of Production Health research found workforce constraints and upskilling were the most frequently cited manufacturing challenge among 501 respondents, at 43 percent, with unplanned downtime close behind at 40 percent. Poor data quality was also identified by 47 percent as a primary barrier to AI adoption, showing how heavily better decisions still depend on the information available to people and systems.

Workforce shortages become particularly difficult when the expertise needed to resolve a problem is concentrated in a small number of experienced people. A plant may have technicians available, but not necessarily someone who has seen that particular failure before or knows the quirks of an asset that has been running for years. That dependence becomes more exposed as experienced employees retire or move on.

“A significant amount of manufacturing knowledge is not documented,” he says. “It lives in the heads of operators, technicians and engineers. The goal is not AI instead of expertise. It is AI that can use that expertise, preserve it and make it available to more people.”

At ICL, a global materials manufacturer with production operations across five continents, earlier diagnosis created enough time for a more controlled intervention on critical equipment at one of its engineered materials facilities. Augury’s monitoring identified a developing fault early enough for the team to confirm the issue and schedule a repair without disrupting production, avoiding an estimated $400,000 in downtime and production loss. Across the first 10 months of deployment, the company reported more than $1 million in avoided downtime, repair costs and production losses, while knowledge from incidents was documented so other sites could reuse it.

Closing the loop changes the measure of success

Production can start losing performance well before an asset reaches the point of failure. Throughput may fall, energy use may rise or product quality may begin to drift even though the equipment is still running and no obvious breakdown has occurred. For manufacturers, that creates a harder diagnostic problem because the loss is visible in the process before the machine itself appears to have failed.

“We have to ask more than whether the machine is healthy,” he says. “If a pump is degrading, it may still be running, but it could be affecting flow, pressure, temperature or another process variable. That can show up as lost throughput, increased consumption or process-related issues before the machine fails. When we bring machine, process and operational data together, manufacturers can start addressing problems that traditionally sit in the gaps between maintenance, operations and process engineering.”

As industrial AI moves beyond diagnosis into recommending and coordinating work, the stakes around its decisions rise sharply. The technology is no longer simply describing what is happening; in some cases it is beginning to influence what people do next. That shift makes the question of authority much more important.

“Manufacturing is different because there are real physical consequences to decisions,” Pfeiffer says. “There will be tasks where an agent can explore information, draw conclusions and recommend next steps, but there are also decisions where you need a human in control. Those boundaries need to be defined with the customer, with clear permissions, human approval and auditability around how the decision was reached.”

Closing a work order should not end the evidence trail that began with the original machine-health signal. If a repair was recommended because vibration changed, the factory should be able to establish whether vibration returned to normal, whether the process recovered and whether the fault reappeared weeks later, then connect that outcome back to the original diagnosis. Avoided downtime, recovered production, cost and yield ultimately matter more than the technical accuracy of the prediction alone. “A successful prediction tells you the AI was right,” Pfeiffer says. “A successful intervention tells you the plant got better.”