99.96% accuracy and six plants show how JSW Cement is scaling AI

The real test of industrial AI begins after the pilot. A model can demonstrate technical accuracy in a controlled environment, but manufacturing organisations have to answer a harder question before taking it elsewhere. Can the intervention work repeatedly, within an operating plant, and produce an outcome that the business can measure?

That is the approach Raghu Vokuda, Chief Digital and Information Officer at JSW Cement, describes in discussing the company’s AI programme. Rather than starting with a collection of technology solutions, he says the organisation first sought to establish where AI could have a measurable impact across manufacturing.

“I suggested that the approach, the pointed solution approach, is not something that we should do as a group,” Vokuda says. “We should have some kind of a clear cut roadmap, a blueprint.”

The distinction is important in an industry where AI projects can remain confined to proofs of concept. At JSW Cement, the roadmap was built around manufacturing areas including quality, maintenance, energy, process optimisation, safety and sustainability. The intention was to identify specific interventions and the objectives they were expected to address before implementation.

From one plant to six

One of the more tangible examples is AI based packer automation. The problem involved discrepancies between the bags dispensed through packers and those subsequently loaded, making reconciliation difficult and leaving unaccounted bags that the company described as pilferages.

JSW Cement introduced a vision based AI system into the process. Vokuda says the intervention has achieved more than 99.96% accuracy. “With the packer automation, with the AI intervention, the accuracy is more than 99.96,” he says.

The more significant development is the subsequent expansion. The system was first implemented at one plant and, within a year and a half, had reached six plants. Implementation at a seventh plant was underway at the time of the interaction, Vokuda says.

The figures indicate that the company has moved at least one AI application beyond experimentation. They do not, however, establish a uniform return across the plants, and Vokuda declined to provide specific financial or operational numbers for other interventions. That distinction matters when assessing the maturity of an enterprise AI programme.

Prediction has to lead to a decision

AI deployment at JSW Cement extends beyond computer vision. Vokuda cites predictive quality, predictive maintenance and predictive energy models, alongside energy forecasting and optimisation.

Predictive quality is being used to anticipate cement strength at different points in the testing cycle. “We are talking about predicting the C3S strength of cement for day one, seven, and 28 and so on, well ahead of time,” he says.

The significance of such models lies less in the prediction itself than in the time it creates for intervention. A prediction becomes operationally useful only when manufacturing teams can act on it.

That is also why Vokuda says the AI programme was structured around clear KPIs. The approach was intended to give business teams visibility into what would be monitored and what outcomes the models were expected to produce. “Integrate, automate, and build intelligence,” is how he describes the strategy.

Once a model becomes stable and the organisation develops confidence in its outputs, the next step can be closed loop operation, he says. This represents a gradual progression from analytics to decision support and potentially greater automation.

Using process twins to test the factory

Digital twin technology forms another part of the architecture, although Vokuda distinguishes between an asset twin and a process twin.

JSW Cement has not pursued the former in the manner described in the interaction. Instead, its AI models use process twins to simulate manufacturing processes and examine possible interventions without changing the live production environment.

“Process twin, yes, all the AI models that we are talking about are based on the kind of process twins that we have developed,” Vokuda says.

The approach allows teams to examine what could happen if a process parameter were changed, including potential effects on variation, before making an intervention on the production line. It therefore provides a controlled environment for testing models that could otherwise carry operational consequences.

Beyond the financial case

Vokuda’s definition of intelligent manufacturing also extends beyond efficiency and financial returns. Quality, maintenance and energy remain core manufacturing concerns, but he places safety and environmental considerations alongside them.

“We are not talking just about the financial gains and impacts on the bottom line,” he says. “We are also very keen on the safety of our employees.”

The final element is the workforce. Vokuda describes a possible smart worker application that would give shop-floor employees access to relevant intelligence through their mobile devices.

The broader implication is that industrial AI is gradually shifting from a technology project to an operating model question. The challenge is not simply whether a manufacturer can build an AI model, but whether it can establish the governance, metrics and operational confidence required to reproduce the result.

JSW Cement’s experience, as described by Vokuda, suggests that scaling may depend less on how many AI experiments an enterprise can initiate and more on whether it can identify a small number of interventions that demonstrate enough evidence to justify replication.