Factory AI needs more than a foundation model
A factory can look deceptively understandable to a general-purpose AI system because it contains so much information. Historians, maintenance records, quality data, operating procedures and production logs all appear to offer enough material for analysis. The problem is that manufacturing meaning is not held in data volume alone, but in the physical relationships between assets, recipes, operating states, process limits and the people accountable for action.
Prateek Kathpal, president of industrial at SymphonyAI, says that is where general-purpose models begin to run out of road. “A general-purpose model does not inherently understand how a manufacturing plant operates,” he says. “It does not know what a compressor, a heat exchanger or a production line is, or even what recipe you are making. Within a physical system, you are governed by the process, the engineering constraints, the operating envelopes and the specific failure modes, which might be very specific to the recipe.”
A temperature rise provides a simple example. The same movement in the data may be normal in one operating state, expected in one recipe and an early sign of failure in another. A model that cannot see those distinctions may still generate a fluent answer, but it cannot be trusted to make or support an operational recommendation without plant context.
The missing layer is context
Manufacturers do not usually lack data. Many already collect it from historians, MES, ERP, CMMS, quality systems, connected worker tools, operator logs and document libraries. What they often lack is a consistent representation of how those sources relate to the same physical operation.
“It is not about having a data problem. It is about having a data context problem,” Kathpal says. “The key information required to make an operational decision might be spread across multiple systems. The historian, MES, ERP, CMMS, connected worker system, operator logs and documents might all have different identifiers and different naming conventions. The key is to have a consistent representation of all that information.
“A good industrial model needs to understand the hierarchies and relationships of the physical operation and the KPIs. It needs to know what equipment a particular line belongs to, which area, which site, which sensors are available on that equipment, and how those assets combine with the operating state or recipe that is active at that point. It also needs to know the failure-mode knowledge, because that differs company by company.”
Useful data must be built
The hard work in industrial AI often begins before the model is trained. Plants grow through acquisitions, new lines, old assets, shifting equipment and inconsistent naming practices. Drawings may be on paper, P&IDs may not match the current plant, and much of the knowledge needed to interpret the system may still sit with experienced people.
“The hardest part is really making the data useful,” Kathpal says. “It is making the data understandable and contextualized across multiple systems and sites. Industrial data results in inconsistent tag naming, duplicate identifiers and multiple hierarchies within plants. Building a usable industrial model requires normalization, asset mapping, hierarchy consolidation or reconstruction, and ontology modeling. Those are the most difficult parts of the equation.”
In one global CPG deployment, chronic OE losses on a filler line and variable CIP cleaning cycle times were addressed by deploying an AI-powered asset twin for the filler line and a process twin for cleaning cycle optimization. The project unified PLC/SCADA, quality, maintenance and cleaning process data, enabled real-time operator guidance and delivered $830,000 in annualized savings, a three per cent OE improvement, reduced downtime and faster repairs in the first phase.
Anomaly is not a decision
In production, identifying abnormal behaviour is only useful if it leads to a decision someone can act on. A shift in performance may point to an emerging fault, a process drift, a recipe change or a temporary operating condition, and the system must help engineers understand which explanation is most likely. Otherwise, AI becomes another source of alerts, leaving the plant team to do the harder work of interpreting cause, urgency and response.
“The anomaly is useful, but anomaly is not a decision,” Kathpal says. “Industrial teams want to know what changed, why it changed and what they should do to fix it. That requires a combination of relationships, domain knowledge and physics to get to root cause analysis. The progression from detection to cause to recommendation, and then from recommendation to workflow, is where AI starts to create value.”
In production, explainability has to give engineers a route back through the recommendation, not just a confident answer. “The engineer in the plant needs to understand what triggered that recommendation, which variables contributed, what was considered, the failure modes, the engineering rules and the specific SOPs for that company,” Kathpal says. “When we make an AI decision, there is an explainability log showing how AI reached the decision. You can see what it looked at, how it moved from one step to the next and how it reached that decision.”
Production exposes the gap
Model performance must be judged after it has been connected to the plant, not before. “The typical breakdown is between the model and the operating environment,” Kathpal says. “A model may perform very well independently, but when you introduce it in the plant and connect it to different sensors, a lot of things play into the role. The frequency at which the data is coming in, sensor noise, downtime events, changes in production configuration and changes in recipe can all result in the model performing very differently in a real production environment.”
A glass manufacturing deployment demonstrates why this distinction matters. Furnace operations data was unified with energy usage and emissions data to create a digital twin, while deep learning models recommended operator setpoint adjustments and predicted temperatures, energy consumption and product quality in real time. The project maintained glass quality within acceptable limits after bottom temperature setpoints were lowered, producing $1 million in annual energy savings, a 3 percent yield increase and a 2 percent reduction in overall energy consumption.
The organizational environment matters as much as the technical one. A recommendation can be correct and still create no value if it reaches someone without the authority, time or role to act on it. In live operations, AI does not only change analysis; it changes the workflow around who investigates, approves and implements the response.
Recommendations must enter the workflow
That creates a different test for industrial AI: whether a recommendation can actually enter the work system. “The recommendation has to be part of workflow,” Kathpal says. “An agent can detect an event, gather the information and recommend an action, but then the maintenance workflow has to deliver the instructions to a connected worker, technician or engineer. The result also has to be captured so the system can continue to learn from the outcome. It becomes a closed loop, and that is what makes industrial AI much more valuable.”
Industrial copilots sit inside that same discipline. They can reduce time spent searching, correlating and interpreting information, but the benefit depends on plant data, governed documentation and clear limits on what the system can recommend or execute.
“The foundation model will become one component of the architecture rather than the differentiator by itself,” Kathpal says. “The real differentiation is the execution. It is not only about whether a system can generate better answers. What can the organization do with that? Are the job profiles set up for people to take those recommendations and act on them? If a system recommends something but it sits in the approval chain for three months, it adds no value.”
A factory is not just a data environment waiting to be queried. It is a physical and organizational system where every answer has to survive process limits, asset behavior, safety constraints and human accountability. General-purpose AI can contribute to that system, but it cannot understand the factory until the factory has been made understandable to it.
