The AI ROI test is no longer about adoption
Adoption was the easy part. Now manufacturers have to prove the intelligence pays for itself.
A manufacturer can potentially save the cost of 150 employees with AI and still lose the ROI battle if the technology consumes ₹1 crore in tokens. That is the paradox confronting manufacturers as artificial intelligence moves from experimentation into production.
AI is no longer being judged simply by what it can automate. The harder test is whether the value it creates survives the cost of creating, governing and scaling that intelligence. Across Bosch India, Berger Paints and Havells India, that calculation is beginning to reshape how technology leaders approach intelligent manufacturing.
When AI savings have to survive AI costs
At Bosch India, Shankar Rao, CIO and CDO, says the focus is on use cases with a clear impact on the top or bottom line, particularly engineering, quality, maintenance and supply chain.
One example is design review based failure mode analysis, or DRBFM, traditionally dependent on engineers with at least five years of experience on a particular product. Bosch has applied AI to historical knowledge so that an engineer with around six months of experience can undertake much of the initial analysis, with a certified engineer validating the result.
The economics are tangible. A process that could earlier take up to 45 days can now be reduced to a few hours or a maximum of one to two days for validation. Across parts of the engineering V-cycle, Rao says Bosch has achieved a 40% reduction.
But the more consequential question is what that efficiency costs.
“If you’re deploying AI and it reduces headcount by, say, 150, but increases your token consumption by one crore,” Rao says, “the economics have to be examined holistically.”
That is pushing Bosch towards use-case-level token metering and greater discipline around technology selection. “Just because you have a hammer, not everything should look like a nail,” he adds. Not every manufacturing problem requires GenAI. Traditional AI/ML, RPA or simpler software can sometimes deliver the outcome at a lower cost.
The ROI is also visible beyond engineering. Bosch is using AI for predictive maintenance, where unplanned downtime provides a direct before-and-after metric, and for quality, where changes can be measured against established quality indicators. In supply chain, AI is being used to consolidate partial truck loads, optimise routes and support negotiations with around 30 logistics service providers.
The lesson is becoming clear. AI ROI has to be measured where the business actually feels the gain.
ROI does not arrive on one clock
At Berger Paints, Partha Protim Mondal, CIO, argues that the return cannot be reduced to incremental revenue.
“AI should be viewed as a strategic transformation investment rather than a traditional capital expenditure measured solely by direct revenue generation.”
Berger’s AI value scorecard spans four dimensions: revenue growth, productivity improvement, customer experience and risk reduction. The latter can be particularly important in manufacturing, where avoiding cybersecurity incidents, supply-chain disruption, forecasting errors or downtime can create economic value without appearing as new revenue.
The timeline also varies. Mondal estimates 6–18 months for productivity benefits, while transformative revenue growth and competitive differentiation can take 18–36 months.
That longer payback period is significant because AI can initially increase technology spending. GPU infrastructure, cloud consumption, platform licensing, data modernisation, security, governance and specialised talent all add to the cost base.
The AI business case, therefore, increasingly resembles a portfolio rather than a single IT project. Proven use cases need to demonstrate measurable outcomes before they are scaled, while experimentation needs to be funded with an acceptance that not every pilot will survive.
The return may extend beyond the factory
At Havells India, Dipesh Shah, Executive President and CTO, sees another dimension to the equation. As manufacturers invest in new capacity, the question is not merely how much they produce, but how intelligently those assets connect with engineering and product development.
Havells has approved an additional ₹255 crore for its greenfield Tumakuru plant, while expanding capacity at Ghiloth, and operates 17 manufacturing facilities. As manufacturing capital carries higher stakes, the return from intelligence increasingly depends on how effectively those assets can be utilised, adapted and connected to the product lifecycle.
“We are increasingly looking at manufacturing as an extension of the R&D ecosystem,” Shah says.
AI is being embedded across design exploration, simulation and validation, while real-world usage data from connected products can feed back into engineering. Shah calls AI “an enabler of engineering excellence,” making the product-development ecosystem more agile, intelligent, efficient and consumer-centric.
That is why he does not see intelligent manufacturing and intelligent products as competing sources of advantage. “I don’t see these as two separate advantages,” he says.
The return, then, can extend beyond lower production costs to faster design, shorter validation cycles, better quality and a more responsive product-development cycle. Over the next three years, Shah says Havells will assess intelligent manufacturing through productivity, quality, agility and uptime, with the larger measure being how effectively those capabilities improve the overall product-development cycle.
The scale test
But proving ROI in one successful use case is not the same as proving it across an enterprise.
Bosch has around 70 manufacturing sites and seven development and application centres in India, with digital maturity varying considerably between plants. Some older machines remain unconnected, making data the first prerequisite for more sophisticated intelligence.
Digital twins illustrate the challenge. Bosch has deployed the technology, but scaling it economically remains difficult because the value of a pilot does not automatically translate across a manufacturing network.
“The real value does not come from pilot. Real value comes from scale,” Rao says.
That may be the defining test of manufacturing’s AI era. A pilot proves that intelligence can work. Scale proves that it can create value repeatedly.
The manufacturers that gain the most from AI may therefore not be those deploying the most models. They will be those capable of answering three harder questions before they scale: what value will this use case create, what will that intelligence really cost, and does the economics improve when it moves from one successful pilot to the enterprise?
The real AI advantage in manufacturing may ultimately lie not in putting intelligence everywhere, but in having the economic intelligence to know where it belongs.
The numbers back that up. In Deloitte’s 2026 AI in Manufacturing survey, high implementation cost was the single most-cited barrier to scaling AI, named by 43% of respondents — ahead of talent gaps, change resistance and data quality. Bosch, Berger and Havells are, in effect, running ahead of an industry-wide problem, not around it.