Where this gets real: changeovers, cleaning, and the cost of running blind
Parts 1 and 2 explained that Industrial AI has lagged for structural reasons, and getting past them takes four things rather than a bigger model: a real-time signal worth acting on, a use case scoped to generalise, intelligence at the edge, and physics instead of a black box. Here is what that looks like in liquid food and beverage production, where the losses can be large and the economics are unforgiving.
The changeover problem
Consider a product changeover. When one liquid follows another inside a production line, there is a moment where they change over. Stop feeding the product tank too early and you push good product to the drain. Run it too long, chasing the last drop, and you pull water into the tank and dilute the batch. Drain loss on one side, dilution on the other, and the only way to win is a razor-sharp cutoff at exactly the right instant. Most plants manage that instant with timers and legacy sensors that cannot actually see the interface between the two liquids. They are guessing, conservatively, thousands of times a day, and paying for it on both sides.
The cleaning problem
Cleaning is the same kind of problem. A cleaning cycle has to remove product residue and then the cleaning chemicals themselves, and the plant has to be certain the line is clean before it runs again. The instruments most plants rely on cannot reliably see colourless residues like those from caustic or acid, so the cycle runs on a safety margin: longer rinses, more water, more energy than necessary, because stopping too early is unacceptable. The blindness is expensive every cycle.
Why these problems travel
Neither example is random. They are exactly the kind of tightly scoped, repeatable problems worth attacking. A changeover is a changeover whether it runs in Finland or France. A cleaning phase has the same physical structure across plants. These are the transferable archetypes that let a model be built once and travel, instead of being rebuilt at every site.
How Collo closes the loop
This is the gap Collo was built to close, and it maps onto the four principles.
The signal comes first, because without it nothing else matters. Collo's novel radio-frequency analyzer measures liquid properties in real-time, inline, while the process runs, including properties conventional instruments structurally cannot detect. It turns the live signal from a weak proxy into something close to lab-quality information, available in the moment rather than hours later from a lab bench.
The use cases are deliberately contained: changeovers, cleaning cycles, mixing endpoints. Small enough to solve reliably, common enough to travel.
The decision runs at the edge, right next to the process, taking in real-time data and feeding real-time parameters back to the local automation. The loop closes where the data is born, fast enough to act on the changeover or the cleaning phase as it happens, without hauling process data across the network and without making the control decision depend on an outside connection.
The models are grounded in the physics of the liquids
Not statistics alone. The philosophy is to use the simplest method that reliably solves each case and keep the reasoning rooted in physical behavior, so it holds up when conditions drift and so that a process engineer can see what it is doing and why. Above it sits an analytics layer that shows the process experts how the system is performing and where the next improvement is hiding, so the people who own the process stay in command of it.
The result
What industrial AI has promised for a decade and rarely delivered: not a dashboard with a suggestion on it, but a closed loop that acts, measurably, on losses legacy instruments could never see. Less product to drain. Less water and energy per cycle. Less running blind.
That is what the KPMG industrial manufacturing tech report points to: the appetite to deploy industrial AI at scale is real and it is now. The data problem underneath it is just as real. The plants that win the next few years will be the ones that fix the signal and scope the problem before they buy the model, because that order is the only one that works.
If you would like to see what that looks like on one of your own lines, that is a conversation we at Collo are always happy to have.
Blog series: Getting Industrial AI Right Is Hard by Jani Puroranta
Jani Puroranta is CEO of Collo, a deep tech company building Industrial AI and real-time sensing that enables liquid food and beverage producers to see and eliminate product, water, and energy losses as they happen. He has been working with physics-informed machine learning and industrial processes since 2017, already before the current AI wave.
