The Pringles AI project in brief:
- Pringles says its AI-supported system increased capacity by 10% without requiring another fryer or production line.
- A digital twin predicts how variations in potatoes, flour and other ingredients will affect dough behaviour, allowing operators to adjust the process earlier.
- The $4m-$5m investment reportedly delivered a return above 40%, but its wider viability will depend on each factory’s data, infrastructure and existing automation.
Increasing factory capacity usually means installing faster machinery, extending a production line or building another facility. Pringles claims it has found another 10% inside equipment already operating at its Kutno plant in Poland.
The manufacturer hasn’t installed another fryer or rebuilt the line. Instead, it’s invested between $4m and $5m in sensors, machine learning and a digital twin that predicts how its dough will behave before variations cause sticking, breakages, waste or lost output. The project has generated a reported return of more than 40%, according to The Wall Street Journal.
That makes Kutno more than an experiment in producing a consistent Pringle. It’s a test of whether AI can help food manufacturers extract greater capacity from existing assets while reducing waste, energy use and disruption caused by variable agricultural ingredients.
Siemens says the technology has increased capacity by 10%, reduced waste by 13% and improved energy efficiency by 7%. However, these reported gains must be weighed against the data, infrastructure and investment required to achieve them – and whether the economics would work in other food factories.
Finding output without buying another line
Kellanova spent four years developing the system with Siemens before its acquisition by Mars was completed in December 2025. At the centre of the project is a real-time digital twin of the Pringles dough moving through the Kutno line.
Digital twins are commonly associated with virtual representations of machinery, production lines or entire factories. The Pringles model instead follows the behaviour of the food itself, using live production data to predict whether the finished crisp will meet its target shape, texture and consistency.
A machine can be calibrated to repeat the same movement precisely, but potatoes, flour and other agricultural ingredients will still vary from one delivery to the next. More than 200 parameters can influence a Pringle, according to Kellanova senior engineering director Jan Laenen. These include flour particle size, dough humidity, protein levels, potato origin and seasonal variations in raw materials.
Even potatoes sourced from the same supplier and region don’t necessarily behave identically throughout the year. A small change in the flour or starch mixture can affect how the dough absorbs water, travels through the rollers, releases from processing surfaces and responds to frying.
“We want to make a perfect chip… the same crunchiness, the same taste and feel when you open the can,” Laenen said.
Experienced doughmakers traditionally managed those variations through touch, sight and repeated sampling. They might stretch a piece of dough, inspect freshly fried crisps or weigh the product before adjusting water, oil or machinery settings. This relied heavily on individual experience and generally responded to a variation after it had started affecting production.
“I want to move from the art of food making to the science of food making,” Kutno plant director Ronny Matthijs said in Siemens’ account of the project.
Sensors, lasers and cameras now collect information from the line, capturing around 200 data points every millisecond. An AI model inside the factory analyses variables including temperature, humidity and protein levels, while the digital twin simulates how the dough is likely to behave.
If the model detects an anomaly, it can recommend adjustments such as changing the amount of water or oil in the formulation. Problems that can’t be resolved locally can be referred to a larger cloud-based model with access to historical production data.
This isn’t generative AI of the ChatGPT variety. It’s conventional machine learning trained to identify patterns between raw materials, processing conditions and product outcomes. Pringles hasn’t handed autonomous control of the line to an algorithm, either. The system recommends changes, but operators decide whether to make them.
Could AI loosen ingredient specifications?
Capacity is only one potential benefit. The tech could also change how manufacturers manage variability in crops and other biological raw materials.
Producers generally try to control that variability by specifying particular varieties and imposing tightly defined supplier requirements. Ingredients outside those parameters may be more difficult to process consistently, even when they remain suitable for food production.
An AI model that understands how different raw materials behave could allow a line to compensate by adjusting processing conditions. Laenen believes Pringles could eventually handle two or three potato varieties and multiple varieties of rice or corn while maintaining the same finished product.
Greater flexibility could give procurement teams more options when crop quality, prices or availability change. It may also allow manufacturers to use ingredients that would previously have fallen outside narrow processing specifications, although Pringles hasn’t demonstrated those benefits at scale.
The system could also help retain knowledge held by experienced operators by recording relationships between ingredients, environmental conditions and machinery settings. The operator’s role would move from detecting every variation manually to interpreting recommendations and deciding when the model can be trusted.
Food producers have good reason to retain human oversight. An incorrect recommendation could affect product quality, allergen management, food safety or regulatory compliance. The Institute of Food Technologists (IFT) says AI can help identify risks earlier, but stresses that reliable data, clear accountability and human judgement remain essential.
Can the economics travel?
The Kutno system is operating on one production line. Expansion into Belgium is expected to follow, with deployment in the US planned for 2027. Those factories will provide a stronger test of whether the model can adapt to different equipment, suppliers, operating conditions and climates.
Questions remain about the cost of achieving the reported gains. Neither Kellanova nor Siemens has disclosed how much of the $4m-$5m investment covered sensors, software, cloud computing, engineering or integration with existing equipment. Continuing costs for model maintenance, cybersecurity and workforce training also remain unclear.
The 10% capacity increase needs context, too. Without baseline production volumes, reject rates or independently verified measurements, producers can’t directly compare Kutno’s performance with their own lines. The figures have been supplied by the companies involved rather than established through an independent assessment.
Results may also depend heavily on a factory’s existing infrastructure. A highly automated plant with extensive sensors and clean historical data is a much easier starting point than an older facility running disconnected machinery and paper-based records.
Connecting production equipment to wider data systems introduces further risks. Poor sensor readings, incomplete data or a model trained on unrepresentative conditions could produce misleading recommendations. Greater connectivity can also increase cybersecurity exposure, particularly when legacy equipment wasn’t designed to exchange data continuously.
The business case won’t rest on whether AI can produce a more perfect Pringle or generate another factory dashboard. It will rest on whether manufacturers can use the technology to delay an expensive line expansion, process variable ingredients more reliably and extract more saleable product from the assets they already own.
