Apollo Tyres: AI that cuts production scrap

An instability window at every recipe change

Every recipe change on an extrusion line opens a window of instability, and that's where AI that cuts production scrap does its best work.

For a few minutes the compound comes out of specification and ends up in rework or scrap. In a high-variety factory that window reopens dozens of times a day.

Apollo Tyres chose to close it with a control system built in-house. The Indian group, headquartered in Gurgaon with plants in India and Europe, deployed a predictive model on the edge and an agent that corrects the line while it produces. The case is discussed on the AWS for Industries blog alongside Harsh Vardhan, Global Head of Apollo Tyres' Digital Innovation Hub.

The interesting choice lies in the verb: built, never bought.

The original idea: closing the loop on a safe lever

The original idea is almost trivial in form and rare in practice: closing the loop.

Classical systems work in feed forward: they set parameters at startup and hope the material follows. Apollo Tyres flipped the logic. The controller measures the tread and sidewall profile at each measurement cycle, compares the value against specification, and automatically corrects line speed as soon as it sees a deviation.

The predictive model estimates optimal startup speed at the moment of recipe change. It works on over 40 critical variables within a set of over 200 process features.

The refinement lies in the choice of actuator. Among hundreds of observed variables, the system moves one: line speed. A tight loop on a safe lever makes behavior predictable, verifiable, and acceptable to those on the shop floor.

The numbers, and where they come from

Four numbers are published, and it's worth separating them.

• Industry context: 30-35 setup changes per day in a high-variety plant
• Structural loss: 5-6% of production volume wasted in startup rework
• Result: stabilization time down by approximately 40%, startup rework down by approximately 26%
• Maturity: approximately 44% of AI and machine learning initiatives reach production

The four figures come from the same source, the post published by AWS for Industries and co-signed by the company^[1]: results declared by the protagonist, with the mechanism described in full.

The metric that matters to a plant director is the second one, because it brings the denominator. The 5-6% of volume lost is the pie; the 26% is the slice recovered from part of that pie. A serious return is calculated there, never on the isolated percentage.

There's also the gap between the two result figures. Stabilization time drops more than rework (40% versus 26%), a signal that the closed loop shortens the critical phase more than it eliminates out-of-spec material. The game remains open in the final stretch.

The friction: humans remain above the loop

This is where the part comes in that makes the case truly useful: the friction.

A closed loop on an extrusion line is a system that touches the product autonomously. Apollo Tyres accepted it with two explicit constraints: safety guardrails and Human on the Loop supervision.

The person remains above the loop, sees the corrections, and can stop them. It's a voluntary reduction in scope, never a step backward. The alternative, fully autonomous control over dozens of parameters, would have multiplied failure modes and made quality validation burdensome.

Those piloting similar projects know the typical breaking point: the shop floor shuts down the system at the first strange correction. A tight loop with humans above it survives the first month. This, more than the percentages, is the decision that explains why the solution remains operational.

The most honest figure is that 44%

The most instructive number in the story is the share of initiatives that reach production.

Approximately 44% means that a significant portion of the group's AI projects stay upstream of the line: prototypes, trials, things running in the lab that stop before the factory. Marketing communication would have kept that figure out of the text.

Seeing it written changes how you read the entire case. The closed loop on extrusion is the survivor of a portfolio, never an isolated stroke of luck.

For decision-makers, the practical lesson lies in the ratio. You need a pipeline of attempts to achieve industrial success, and you need a common data platform that makes the tenth attempt economical. Apollo Tyres built the cloud-native data platform first, then placed the models on top: the order of these two moves explains the percentage.

In-house beats off-the-shelf

The case reinforces a thesis that recurs in successful industrial deployments: those who know the process build better than those who sell the solution.

Apollo Tyres' APC is born inside the group's Digital Innovation Hub. The cloud provider supplies the bricks (the MLOps pipeline runs on Amazon SageMaker and Amazon Bedrock), domain intelligence remains with the company.

The practical difference is the correction cycle. An internal team changes the deviation threshold the same day the shop manager raises it; an external vendor opens a ticket.

The same direction emerges elsewhere. Forrester argues that private AI will beat public AI for B2B marketing use cases^[2], for similar reasons: proprietary data, specific context, control of the model lifecycle. The pattern crosses industries.

What this result is really worth

A methodological clarification is needed to close the circle with rigor.

The two key percentages come from text co-signed by the company and its cloud provider. It's a result declared by the protagonist, pending independent verification.

Three pieces of information are also missing that a demanding reader would want: the time window of the before-and-after comparison, the number of lines covered, and the absolute economic value of the recovery. The two figures remain credible in order of magnitude and for the mechanism described, consistent with process physics. They should be read for what they are: company results, measured in-house.

The distinction applies to any industrial case. A number verified by third parties carries more weight; a number declared with the mechanism explained nonetheless carries more weight than enthusiastic language.

What you can take away

Three things, transferable beyond tire manufacturing.

First: find the small, recurring loss, the one the income statement hides. Five points of volume wasted in startup are worth more than many projects with brilliant names.

Second: close the loop on one lever at a time. One actuator, one measurement, one guardrail, one person watching from above.

Third: keep the model close to those who know the process, and buy only the infrastructure from outside.

An open question remains, good for any organization. Which of your processes generates systematic scrap at every transition, and which variable could you measure at each cycle to correct it automatically?