The factory data problem is no longer volume but context
Factories are producing more data than ever, but the ability to collect it has moved ahead of the ability to use it. Sensors and connected machines can expose almost every part of a production process, yet visibility alone does not tell a manufacturer why performance has changed or what should happen next. The harder problem is turning a stream of measurements into enough context to support a better decision.
Tim Long, Global Head of Manufacturing at Snowflake, argues that the industry is beginning to move beyond the idea that more data automatically creates more intelligence. In his view, the harder step is deciding which operational problem deserves attention and preserving enough context around the relevant information for people to act. That concern now reaches beyond the plant itself as manufacturers look for ways to combine shopfloor and enterprise information without losing the meaning of either.
“Most of the industry is still figuring out how to use the latest technologies to improve shopfloor performance,” Long says. “What I have observed over the last two years is manufacturers seeing the role of the cloud in improving shop floor performance. IT/OT convergence has been discussed as a concept for a long time. It is now becoming a common priority, where manufacturers bring OT data to the cloud and join it with IT data to unlock performance.”
Context is the missing layer
As manufacturers bring production data out of individual machines and plant systems so it can be combined with information from elsewhere in the business, they create the opportunity to analyze operations in a much wider context. The difficulty is that operational systems often describe assets in ways that make sense locally but become harder to interpret once the data is separated from the machine or process that produced it. A technically successful integration can therefore move the information perfectly while losing enough of its original context to make subsequent analytics or AI far less useful.
“Sometimes you simply get a tag name and a value, and you do not know exactly what that tag relates to, which asset it is tied to or what units of measure are associated with it,” Long says. “One of the biggest challenges for manufacturers is making sure they obtain the full context of that data and bring it with them so that when they land the data, it is useful.”
Trying to reuse what one factory has learned elsewhere exposes another weakness in industrial data. Different sites may have evolved their own naming conventions and technology stacks, so apparently similar production information can mean different things from one plant to another. Long believes technology has largely solved earlier concerns around the volume and variety of industrial data, leaving standardization and the preservation of operational context as more important barriers to scale.
Bringing information together is therefore valuable only if the relationships between shopfloor activity and the business outcomes it affects remain visible after the move. Without that structure, a larger pool of data can simply create a larger search problem. Manufacturers may have more information available while still leaving operational teams to work out which part of it matters.
Start with the decision not the data
Data programs become expensive when collection turns into the objective rather than a means to improve performance. A factory can spend months building pipelines and organizing information before anyone has defined the operational decision the project is supposed to improve. Long sees that sequence as one reason manufacturers can accumulate substantial data estates without producing an equally substantial return.
“We start with the business outcome they are trying to achieve because that narrows the focus on which data sources are relevant,” he says. “If the problem is equipment uptime, we might go down the predictive maintenance path and ask why there is unplanned downtime. Once you know where the business problem is, you have much better specificity on which data is relevant. Start with the minimally viable solution and then layer in complexity as it pays off.”
Once the relevant data has been narrowed around a specific business problem, the next test is whether the resulting analysis helps operational teams decide what deserves attention first. Maintenance resources are finite, so a warning about one asset needs to be understood in relation to the consequences of losing it rather than treated as an isolated probability. That means combining the likelihood of failure with the operational impact of the machine, which is why Long sees predictive maintenance as a useful example of where shop-floor data alone is not enough.
“Predicting a failure on a machine with a probability is one thing, but what is the actual business impact if that machine goes down?” Long says. “Different machines have different impacts within the factory. Understanding both the probability and the severity helps prioritize which actions you take, and that requires being able to link shop-floor data with enterprise and supply chain information so you can understand how a decision on the floor affects your ability to ship product to customers.”
Deciding what information matters is only part of the architecture because manufacturers must also determine where analysis can run without creating unacceptable latency or reliability risk. Use cases involving an immediate safety decision or the risk of scrapping product may need inference close to the machine, while predictive maintenance can tolerate the delay involved in running analysis in the cloud. Long describes manufacturing as a hybrid environment in which the location of the computation should follow the operational requirement rather than a preference for either edge or cloud.
AI is changing the interface to factory data
Manufacturers have spent years building dashboards to make performance visible, but the next change may be less about adding another screen and more about changing how people interrogate the information behind it. Long does not see the proliferation of dashboards as evidence that manufacturers have collected too much data. He sees it as evidence that people are trying to answer more questions, while the interface they have traditionally used to find those answers is becoming less efficient.
“The value of the dashboard is limited in the same way that we used to go to the internet and search for something to get an answer,” he says. “Today, people increasingly go straight to a large language model and ask the question. I think you are seeing the same change in manufacturing. Rather than go to a dashboard and explore until they find the answer, people can ask the question directly.”
Dashboards are unlikely to disappear simply because conversational analysis becomes easier. Long sees a stronger model emerging in which a formal dashboard continues to provide the agreed scorecard while an AI agent works alongside it, with access to the underlying data and enough context to investigate questions that the dashboard designer could never anticipate. Instead of clicking through several layers of OEE metrics to find the source of deterioration, an operator could ask what is driving the change and receive an analysis built from the detailed records behind the headline measure.
That ability to interrogate the detail behind a headline metric becomes more valuable when the analysis reaches beyond live machine data. Sensor readings may show that a temperature or vibration pattern is unusual, but they do not necessarily explain what caused a similar problem before, or which intervention restored the equipment most effectively. Historical maintenance records and the experience captured in them can provide that missing context, particularly as experienced technicians retire.
At semiconductor manufacturer Wolfspeed, Long says shop-floor data is being combined with historical maintenance records so AI can compare a current problem with previous failures and recommend the most efficient troubleshooting path. The company has reported reducing the time needed to resolve downtime events from hours to minutes, while building safety guardrails into the agent so recommendations do not compromise established procedures. The same approach can feed analysis back into the maintenance system, raising a ticket with the relevant evidence and directing it to the team responsible for the response.
“Sensor data over time may be predictive of failures, but the historical lessons learned are also incredibly useful,” he says. “If that knowledge is captured in the maintenance system, AI can learn from it and advise based on what has worked in the past. The agent can also connect back to the maintenance system, raise a ticket and include the analysis used to detect the risk, so you begin to close the loop between finding the issue and acting on it.”
For manufacturers disappointed by the returns from years of data collection, that closed loop may be a more useful measure of maturity than the size of the data estate. Long’s advice is to resist beginning with the information already available and asking what might be done with it. The better starting point is the business problem, followed by the operational and enterprise context needed to solve it, because the answer to a manufacturing problem increasingly sits across systems rather than inside one of them.
