The hard part of industrial AI begins when the model works
Manufacturers are getting better at building models that detect anomalies, predict quality problems or identify the conditions associated with stronger process performance. Yet a technically successful model can still have almost no effect on the factory if its output remains on a dashboard, inside a data science environment or outside the way production teams work. The difficult transition is from proving that an insight is valid to making it part of the decisions taken every shift.
Jason Dietrich, CRO at TwinThread, describes that transition as operationalization. The term is less fashionable than predictive or agentic AI, but it addresses the point at which much of the potential value is either realized or lost. “You have done all the hard work of getting the data, identifying an outcome you are targeting and delivering a model that drives towards that outcome,” he explains. “Now it is about getting the organization to adopt those changes. Operationalization means it becomes part of the standard operating procedures; part of how an operator works day to day and part of how management reviews how we perform versus how we performed in the past. The hard part is making it part of the muscle memory of how an operational organization operates.”
Once industrial AI moves from validation into production, the challenge shifts from proving technical accuracy to proving that the technology can change performance consistently. Operators must trust the recommendations enough to act on them, while management needs confidence that improvements in yield, quality or other measures are real and repeatable. Dietrich sees this transition as one of the points where adoption often slows, because a model may have demonstrated that it works without yet proving that the wider organization can embed it successfully into day-to-day operations.
A prediction must change what happens next
Industrial AI is often presented as a succession of technologies, but their value becomes easier to understand when they are treated as parts of the same decision process. Predictive AI provides an indication of what is likely to happen; prescriptive capability turns that prediction into a recommended response, while agentic systems can begin executing the workflow created by that decision.
“Predictive AI is saying that, based on the way you are currently operating, you are going to miss a specification,” Dietrich adds. “Prescriptive is saying, in order not to miss that, these are the things we recommend you do right now. Generative AI gives you the ability to quickly search, query and understand the problem behind the problem. Then agentic AI is really the workflow engine behind those things that makes it easier for us to execute.”
For a process application, execution could eventually mean writing an approved set-point adjustment directly to a control system. On an asset model, an anticipated failure could trigger interaction with the maintenance system and creation of a work order. The important progression is from knowing that something may happen towards changing the outcome before it does.
Operators cannot reach that point by absorbing another stream of indiscriminate alarms. Plants already contain systems competing for attention, and recommendations become useful only when teams can distinguish the action that matters from the background noise. “The first thing is to eliminate the noise,” Dietrich says. “Make it very clear from a priority standpoint which recommendations have the most value to the organization, which have the highest priority, and which need to be acted upon right now. Then it is back to that muscle memory. It is having a consistent set of operations leaders enforcing that this is part of the normal operating procedure going forward.”
Nestlé’s Waverly factory in Iowa provides a practical example. Its legacy agglomeration process for Nesquik and Ovaltine relied on samples taken every 30 to 60 minutes, leaving operators with lagging information about moisture and density. A TwinThread model used real-time production data to predict those characteristics and provide set-point recommendations while production was still running. During an eight-hour trial, Nestlé reported more consistent powder and a ten percent product saving, equivalent to one kilogram saved for every ten one-kilogram jars produced.
The significance lies less in predicting moisture or density than in making the prediction useful before the next physical sample arrived. Analytics became part of the production decision rather than an explanation of what had already happened.
Industrial AI needs operational context
Manufacturers have spent decades accumulating historians, MES platforms, laboratory systems, maintenance applications and automation data, but those investments do not automatically create an environment in which AI can work effectively. Dietrich describes the familiar problem as continuing “islands of automation” and “islands of data”, with information divided between systems that were designed for different purposes.
“There are situations where those disparate data sources do not communicate well together,” he continues. “It is difficult for a person to understand my historical data versus my MES data versus my LIMS data versus my real-time things coming from OPC and my alerts, because they are on different screens. The data is not contextualized or harmonized in a way around the idea of a digital twin.”
TwinThread uses an operational digital twin to organize that information around the process, asset, line or site. The twin is therefore more than a visualization of equipment: it provides the context needed to associate data with what is physically happening and creates a foundation on which predictive and prescriptive models can operate.
Manufacturers are more likely to build momentum when the first AI use case is tied to a problem the plant already recognises and values. Dietrich recommends starting with the current operating reality, identifying an outcome that matters commercially or operationally, and involving the people who will have to use the model from the outset. Operators, process engineers, supervisors and plant management all need to understand what is being improved and why, because adoption becomes much easier when the use case is solving a problem they already experience rather than introducing technology in search of a purpose.
Scale then depends on carrying what has been learned to similar equipment without rebuilding everything from scratch. TwinThread uses classes that allow a proven model and outcome to be reused while accommodating differences in local tags and data connections.
“If I want to optimize a spray dryer, the data inputs and the historian might be different, and the tags might be different, but the idea of what I am trying to optimize is common,” Dietrich says. “I can build a collective class that can then be deployed to any spray dryer in the enterprise, knowing that the wiring of the data will be different from site to site, but the outcome we’re trying to accomplish and the model itself have been proven out.”
A deployment with Hill’s Pet Nutrition illustrates the same principle at workflow level. TwinThread says model recommendations were integrated into startup and in-run processes across multiple plants, with the project improving quality capability and reducing material losses. The company also reports that less experienced operators were able to deliver stronger quality performance, demonstrating why repeatability depends on embedding the guidance into how people work rather than merely copying a model between sites.
Human oversight changes as the loop closes
The relationship between operator and model becomes more important as AI moves closer to direct process control. Human involvement does not necessarily mean continuing to make every adjustment manually; the more useful question is where judgement adds value and where constant intervention simply leaves operators chasing process variation.
“We are not trying to replace people. We are trying to make those people more productive,” Dietrich says. “Can we have those people doing things that are more proactive versus reactive? Can we have those people innovating and solving other problems that they do not have time to solve today because they are constantly dealing with fires on the plant floor?”
Dietrich compares the eventual relationship with an aircraft autopilot. Operators should not have to continuously chase the desired condition when AI can monitor the process and identify when intervention is required. People remain responsible for the operation, but their attention moves towards exceptions, improvement and decisions where experience matters most.
The attraction of generative AI is its ability to make complex operational information easier to access, but that usefulness depends heavily on the quality and boundaries of the information it is allowed to use. TwinThread initially applied its Advisor capability to search and interrogate operational data and is now extending it towards creating digital threads and supporting model development. Dietrich is more cautious when external information is introduced, because hallucinations or inconsistent responses from general-purpose models can create uncertainty that is difficult to tolerate in a production environment.
Manufacturers do not need to reach closed-loop autonomy for industrial AI to be successful. For some, stabilizing a process and reducing variation may be the immediate objective; elsewhere the priority could be yield, throughput or quality. The common measure is whether the technology changes the outcome the organization originally set out to improve and becomes part of normal operations.
“Data is data. Model outputs are model outputs,” Dietrich concludes. “If you do not turn data into information, if you do not turn model outputs into recommendations and improvement, it is really the people who operationalize it and make it part of their new operating rhythm who are going to win.”
Better models will continue to emerge, but model sophistication is an incomplete measure of industrial AI maturity. The more meaningful test is whether an insight can move reliably into a recommendation and then into repeatable operational action. That is the point at which AI stops demonstrating what it can do and starts changing how manufacturing is done.
