By: Dr Mandar Kulkarni, Vice President and General Manager, AI Centre of Excellence, Findability Sciences
Walk into any modern manufacturing plant in India and you will see automation everywhere. PLCs run the sequences, SCADA screens show thousands of tags, and robots weld and palletise without tiring. Indian manufacturing has invested heavily in this layer over the last two decades, and it has paid back in speed, consistency and safety. Yet the questions plant heads ask me have not changed much. Why did yield drop on the night shift? Which bearing will fail next month? What setpoint should the operator choose when raw material quality changes mid-shift? Automation cannot answer these. It carries out instructions; it cannot tell whether the instruction was the right one. Intelligent manufacturing is about closing that gap.
From fixed rules to systems that learn
Traditional automation works on fixed logic: if temperature crosses a limit, open the valve. The logic is written once and stays the same until an engineer changes it. Intelligent manufacturing adds a layer above the PLC and the DCS that learns from the plant’s own history. It predicts a problem before the alarm goes off, recommends a setpoint instead of just displaying a trend, and reads production, quality, maintenance and energy data together.
What two sugar mills taught us
At a large sugar producer we work with in Latin America, the first step was not a model. The mill agreed on a benchmark for what it should be able to achieve, measured how far it was from that benchmark, and made its finance team the owner of the number. Only then did the models go in. In the 2025 harvest, validated recovery rose from 10 to 10.3 per cent, roughly three more kilograms of sugar from every tonne of cane, and the mill recorded its lowest lost time ever.
At a sugar mill in India, we worked through a full crushing season on three areas: the crushing plan, anomaly detection at the shredder, and process optimisation. The shredder shows the difference between automation and intelligence clearly. The control system raises an alarm once a limit is crossed. The anomaly model watches the pattern underneath and flags abnormal behaviour earlier, so the team can intervene at a time of its choosing. Over that season, these early warnings helped the team cut unplanned shredder stoppages by around 20 per cent.
Why most AI pilots stall
Most AI pilots in manufacturing stall, and the algorithm is rarely the reason. They stall because they are run as demonstrations. A data scientist hand-cleans three months of data, the model looks good in a review meeting, and nobody builds the pipes that would feed it every day. I would go further: a pilot on hand-prepared data proves very little about a running plant, and I advise plant heads not to fund one.
In a real plant, the data is scattered. Machine data sits in the historian, quality results in the lab register, maintenance history in the ERP, and a great deal of practical knowledge in the heads of senior operators.
At Findability Sciences, we address this with a framework we call ICUPP™: Infrastructure, Collection, Unification, Processing and Presentation. It begins with Infrastructure: the right connectivity, storage and compute, on-premises or in the cloud, as the plant requires. Collection starts with a data census, where we list every source, from sensors and SCADA to lab sheets and logbooks, and bring it in. Unification joins these sources on a common timeline and context, creating what we call the corporate memory of the plant. Only then comes Processing, where predictive models, optimisation and computer vision are applied to the unified data. Presentation delivers the result where the decision is taken: on the operator’s screen, on the manager’s phone, or back into the control system.
The order matters. Many companies jump straight to Processing and then wonder why the results do not hold. Once the corporate memory is in place, each new use case comes faster because the foundation is reused.
Intelligent workflows keep people at the centre
The last mile is the workflow: who gets the insight, in what form, and what happens next. At the Indian mill, advisories went to the shift engineer as a plain-language message on the control room screen, with the reason attached. The system did the reading and the reasoning; the engineer made the decision. I stress this because the fear on the shop floor is real. Intelligent manufacturing is not about removing the operator. It is about giving every shift the judgement of a 25-year veteran, and capturing that judgement before it retires. Plants that involve operators from day one, and show them why a recommendation was made, see adoption. Where the system is pushed as a black box, operators quietly stop using it.
Where to begin
For manufacturers ready to make this move, my advice is practical.
•Start with a business problem, not a technology. Pick one loss that hurts: unplanned downtime, yield, energy or rejection.
• Do the data census early. Know what you have, what is missing and what is still on paper.
• Build for scale from day one. Design the first use case so it can be repeated on the next line or the next plant. This is the thinking behind the Factory Model that our founder and CEO, Anand Mahurkar, sets out in his book AI Factory: The Executive’s Blueprint for Turning Data into Intelligence. A plant should produce intelligence the way it produces goods: every day, repeatably and with quality checks.
• Measure in the plant’s own language: recovery, OEE, steam per tonne, rupees saved. If the benefit does not show up in these numbers, it is not real.
Automation gave Indian manufacturing its muscle. Intelligence will give it judgement. The manufacturers who combine the two, with their people at the centre, will set the pace in the coming decade.