Four things industrial AI has to get right before it earns a setpoint
Part 1 argued that industrial AI lagged for structural reasons, not technical ones. Getting past them takes four unglamorous things, none of which is a bigger model.
A real-time signal worth acting on
First, the data feeding the model has to be real-time and good enough to act on. This is what has held everything back. You could build an excellent model on lab data, but in production it runs off whatever the process offers, and a flow rate or a temperature is a weak proxy for what is actually happening inside the pipe. The fix is not a cleverer algorithm on a poor signal, it is a better signal. Measurement that approaches lab quality but happens live, in the line, while the process runs. Get that, and much of the modelling difficulty evaporates, because the model finally sees the thing it is meant to reason about.
A use case scoped to generalise
Second, the use case has to be small enough to generalise. This is where most industrial AI quietly dies. The instinct is to model the whole process end to end, with all its delays, feedback loops, and chained sub-steps. That ambition fails every time, on accuracy, cost, or time to market. The discipline that works is the opposite: break the process into steps small enough that the same model can travel between plants: a liquid changeover, a cleaning phase, a mixing endpoint. A tightly scoped problem behaves consistently enough across sites to be solved once and reused. An end-to-end model behaves differently in every plant and is never finished.
A decision made at the edge
Third, the closed-loop decision has to live at the edge, right next to the asset. The principle is simple: do not move process data across the industrial network to make a control decision. Keep the decision where the data is born. That buys you three things at once. Latency stays near zero, because there is no round trip to a server room or a data center before a valve moves or a phase ends. The network load stays low, because you are not streaming high-rate process data across the plant to be reasoned about somewhere else. And the integration burden stays small, because a decision made locally does not have to thread through layers of plant IT, OT security, and outside connections that each have a good reason to be denied. The cloud still has its place for analytics, benchmarking, and learning across a fleet of assets. It does not, however, belong inside the control loop.
A model that isn't a black box
Fourth, the model cannot be a black box. A process engineer will not, and should not, hand control to something nobody can explain. "The neural network said so!" is no answer when a batch goes wrong. The models that earn trust are rooted in the physics of the process, not pattern-matching alone. That matters for more than comfort: a model grounded in real physical behavior holds up when conditions drift outside its training range, because the physics still applies. A pure statistical model extrapolates and hopes. On a plant floor, hope is not a control strategy.
None of these four is about model sophistication. The honest position is to use the simplest method that reliably solves the problem. Sometimes that is machine learning. Often it is signal processing or a well-grounded statistical classifier an engineer can inspect. What wins in industrial AI is not the fanciest algorithm. It is the combination of a signal worth trusting, a problem scoped to generalise, intelligence placed where it can act, and physics that keeps it honest.
Walk the four
So when you weigh a vendor or an internal initiative, walk the four:
1) Is the live signal good enough to act on?
2) Is the use case scoped to travel between plants?
3) Does the decision happen at the edge, without hauling data across the network?
4) Can someone explain why it does what it does?
Clear all four and it is worth your time. Clear none and it is a pilot headed for the drawer, next to the others.
The next piece puts the four to work, in the parts of liquid production where the losses can be large and the measurement gap has been wide.
→ Closing the AI loop in liquid production
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.
