Carmen Sook, director of customer solutions at Bestmix Software, told attendees during Petfood Essentials at Petfood Forum 2026 in Kansas City, Missouri, on April 27 that pet food manufacturers are sitting on a goldmine of untapped production data — and that AI can be the key to unlocking it.
During her session, "From nutritionist to operator: AI as your trusted copilot," Sook said quality and efficiency in pet food manufacturing still depend too heavily on individual operators, even with advances in automation and formulation software.
Three systemic gaps drive the problem, she said:
- Operator judgment varies shift to shift, so quality depends on who is running the line.
- Operators adjust settings in real time but do not see QC results for about an hour, creating a sustained blind spot.
- Formulation, production settings and QC results sit in separate systems with no feedback loop connecting them.
"AI-driven decision support can give every operator access to the same level of insight as the most experienced person on the floor, reducing moisture variability, cutting startup waste and removing quality outcomes from the hands of any one individual," Sook said.
Sook pointed to extrusion as an early deployment area where AI analytics are already showing results. A moisture miss in extrusion, she noted, can mean 30 to 60 minutes of product held in the dryer that may ultimately need to be discarded.
Predictive AI models can flag that outcome earlier in the process, shifting operators from reactive troubleshooting to proactive decision-making. "When production drifts, the consequences extend beyond moisture," she explained. "Kibble texture changes, coating absorption shifts and palatability changes, even when the original formula was sound.
Alltech case study
Sook detailed a deployed moisture prediction model built from four years of production data collected by Alltech operators. As part of their normal QC routine, operators had been logging approximately 27 production parameters per hour into Bestmix quality control, including temperatures, steam, screw speed and pressure, along with finished product measurements such as moisture and density. Recipe and ingredient data for each run was simultaneously stored in Bestmix recipe management.
Bestmix data scientists used those inputs — what recipe was running, what the operator set on the extruder, and what the finished moisture actually turned out to be — to build a model with one goal: predict finished moisture before the product hit the dryer. The model was tested against historical runs, with no changes made on the floor.
The results were significant. Operators were getting moisture within tolerance about 60% of the time on their own judgment. The model hit the target 87% of the time — roughly nine out of 10 runs — a 50% improvement in accuracy. Startup stabilization was 43% faster. Projected annual savings across two extrusion lines came to approximately $567,000 at a 6x ROI.
Sook broke down where those savings come from. Reduced rework accounted for about $157,000 — when finished moisture falls outside of tolerance, product is recycled back through the line, consuming energy, labor and time, and the model cut rework by roughly a third. Tighter formulations represented the largest bucket, at nearly $400,000, or about 70% of total savings. Faster startups contributed about $16,000 from the 43% reduction in dial-in time. Energy savings from running the dryer less aggressively and yield recovery from running closer to target were not included in the projections, Sook noted, meaning actual savings could be higher.
Alltech has since connected its manufacturing execution system to the Bestmix QC system, moving from one manually logged data point per hour to automated data collection every six minutes across all lines. Real-time operator dashboards are currently in development. Sook said the increased data density and elimination of manual entry errors should improve model accuracy further.
"That conversation only happens when you have the data to show regulatory and quality that you can consistently hit it," she said.
What's deployed, what's being built
Sook said the Alltech case study represents the first deployed example of connecting those data streams in a way that is practical, measurable and repeatable.
Additional capabilities currently in development include real-time operator dashboards and a formulation starting-point prediction tool that would give formulators a data-informed baseline for new recipes rather than a blank screen. Sook said Bestmix is actively seeking a partner to build the formulation prediction capability.
Looking ahead, Sook described a roadmap that extends AI beyond extrusion to close the loop between production and formulation — using real-time line data to inform how the next recipe is written. Over the next five years, she envisions AI embedded across the full value chain, from helping formulators generate smarter starting recipes based on historical data to accelerating quality and compliance decisions.
"The companies that get there first are creating their AI foundation right now," she said.
Four steps processors can take this week
Sook closed with four actions she said require no budget approval:
- Ask whether the plant is already collecting extrusion parameters somewhere — in a QC system, a spreadsheet or a SCADA system
- Pull target moisture and actual finished moisture on one product and look at the gap
- If no one is tracking production settings, start on one line with one operator logging settings hourly; six months of data is enough to begin seeing patterns
- Try an AI tool on one task done every day.
"Your data doesn't have to be perfect," Sook said. "It just has to exist."
It's also important to remember that AI is designed to be a tool to help your operators, not replace them.
"The operator sees the prediction; they make the call," Sook said. "The model does not touch the equipment. It just gives the operator better information than they had before. This does not replace your operator's expertise — it enhances it."
