Speaking at the recent Intelligent Manufacturing 500 conference, Debashish Roy opened by asking the room how many believed the shift from generative AI to agentic AI was actually yielding ROI. He then walked through CEAT’s decade-long journey, from connected machines in 2015 to a Lighthouse-recognized network of plants, and the use cases that came out of it: a root-cause finder that cuts analysis from days to hours, an AI control system for rubber mixing, and a model that predicts tyre tread design right the first time.
His message is that the hard parts of AI in manufacturing are not the algorithms. They are the data foundation, the change management, and a plant team that owns the outcome rather than waiting for IT to push solutions on it. The goal he has set is simple: get the first product right, every time.
Here is Debashish in his own words:
Is agentic AI paying off?
Before we start, a pertinent question, which we keep discussing, reading and hearing in all these conferences. How many of us in this room really feel that the movement from gen AI to agentic AI is yielding some ROI?
The reason I ask is that some weeks back, BCG came out with an article saying 50% of organisations have started seeing some ROI, which was not the case six months back. What I’m going to talk about today is largely on those grounds.
At CEAT, we believe in digital as the roadmap and digital as the way we manufacture. It’s no more a dashboard, it’s no more a model for us. It’s something we do every day, something we live every day.
Three shifts
There are three key shifts I want to talk about.
The first is the shift from predictive to prescriptive, and today to trying to become autonomous. Not in all areas, but depending on the use case, the severity and the implications it might have. We take those calls.
The second is the shift from digital teams pushing solutions to the business, to a pull from the business side. We are seeing a lot of that today.
The third is the role of AI. Today AI is not really a tool for us. It’s a capability that helps the plant sense, decide and then improve.
As we moved from Industry 4.0 to 5.0, what changed was interconnected systems, sustainability, and collaborative robots. That is the DNA we stand on. We have gone from generative AI and a lot of vision analytics, to predictive, and finally to trying to do agentic and autonomous. That is something we are really proud of.
The runway
It’s very important to understand what led to this foundation. It’s not a fortnight thing, and it’s not a one-year journey.
It started way back in 2015, when we laid the foundation of connected machines, the infrastructure, and the integration of various systems. In 2017, the first plant in Chennai got an MES system. In 2019, we started the transformation journey with the advent of the Internet of Things. All of this is the runway for us. To take a flight, we need a solid runway, and this foundation is what is helping us take the journey ahead.
In 2021, we laid the foundation of digital teams across all the plants, with data engineers, data scientists and business translators, living this journey day in and day out. In 2022, CEAT became the first tyre company to get Lighthouse status, for the Halol plant. We maintained that legacy, and the second plant, Chennai, also got Lighthouse status. We are also the only globally acclaimed tyre company to get a Deming Grand Award. Things started falling in place one after the other.
And today, in 2026, we are talking about agentic use cases and about making the plant autonomous, so that we get the first product right every time. That is our KPI and our KRA.
Five beliefs we live by
We are not trying to fit AI in. We look for problems where AI fits. There is always a business problem first, and tech comes later. We all believe tech is just 30% of the problem. Largely it is data and change management, which has the people element in it.
Digital and IT teams are enablers, and the plant owns the outcome. The pull has to come from the plant and business teams. We cannot keep pushing stuff and rolling out pilots, and the journey can’t just end there. It has to be outcome driven.
Pilots go through a proper process to scale. We scale up depending on viability and feasibility, and then cross-deploy across all the plants.
Foresighting matters. Today we have digital plants and we talk about digital twins. Foresighting is what moves the needle from a digitized factory to a connected factory and beyond. Today we are somewhere between predictive and prescriptive, and the goal is a completely autonomous plant, a living plant, which decides on itself and improves on itself.
Everything rests on a solid data foundation. Today we see a lot of good use cases and partner presentations, but when we take them to implementation, success depends on how solid the data is. Without the data lake, nothing works.
What this looks like in practice
Some of these use cases are in progress. Three have already been implemented, and we are seeing success.
TroubleTrace is a root cause finder. Using solid data and agents, we have reduced root cause analysis time from days to mere hours.
Boss.ai is largely for operators and engineers, to pinpoint a solution and bring down mean time to repair. In manufacturing, a lot of data lies in silos, which makes it very difficult to establish the relationship between an impact and what led to it. That is where we built this NLP-based, agentic solution, to help operators and engineers do root cause analysis and find where the problem exists, reducing downtime from days to hours.
The process quality model came out of a business challenge. Changing material from traditional carbon black to silica, which is more sustainable, increases the mixing cycle time, and mixing is a very critical part of tyre manufacturing. We worked on more than 50 parameters of pressure, temperature and material, and applied a gradient boosting model to predict the variation in mixing time. The result was close to a 50% reduction in cycle time variation, and an industry-first solution. It stabilized about six to eight months back and is rolled out across three plants right now.
An operations research model finds the best operating point across competing factors and constraints, helping us reduce total turnaround time by close to 15%.
The foresight factory takes the digital twin to the next level. Using simulation, we not only predict what is going to happen, but also get a fair understanding of the implications of the actions we simulate in a virtual space.
As we expand into new geographies, we need to customize products to customer requirements, and every time that requires a tyre tread design. That is critical because of the die swell arrangements, and it is time-consuming and resource-intensive. Every cycle needed three-plus trials per SKU, 15 days of trials per SKU, and more than 50 days of lead time. Every time the exercise failed, we had to scrap material.
We modelled all of it using AI and ML, based on 18 machine parameters and 15 material parameters, and defined a way to predict the tread design correctly the first time. That led to a 30% reduction in overall new product development time.
In conclusion, the algorithms are the easy part. The data, the people and the patience to build a runway are what let a plant take flight.
Also Read: