Past the Pilot: Intelligent Manufacturing Summit Puts Readiness Ahead of Technology

At the second Intelligent Manufacturing Summit, industry leaders agreed the AI is real and ready. The open questions are data, people, money and MSMEs.

Speaker after speaker in Chennai made the same point. The technology is not what’s holding Indian manufacturing back. The second edition of the Intelligent Manufacturing Summit, organised by CII, ran four sessions: generative AI and foundation models, autonomous and self-healing factories, sustainability and traceability, and automation for MSMEs. Topics changed from session to session, but the diagnosis stayed the same. Factories have more data and more tools than ever, yet few manage to turn either into results they can measure.

The stakes

The opening addresses set the scale. Manufacturing is about 14–15% of India’s GDP, and the government aims to raise that to about 25% as the economy more than doubles. One speaker put the implication bluntly: manufacturing output faces a roughly four-fold change. Much of that growth is expected from advanced sectors such as defence, semiconductors, electronics, drones and medtech, alongside traditional manufacturing.

A paper-industry leader added the cost-pressure view. Mill realizations were ₹95,000–96,000 per tonne just after COVID. Today they are below ₹70,000, a fall of more than ₹25,000 per tonne. In a sector with excess capacity and no growth, he said, the only lever left is efficiency. His plants already collect continuous data instead of keeping it in silos. AI is now being tested to choose the raw-material mix that hits a target paper grade, matching thickness, shade and opacity, a job that used to be done by trial and error.

Theme 1: The pilot trap

The sharpest numbers of the day described a funnel. The session moderator cited figures for generative AI adoption in manufacturing:

About four in five manufacturers adopt generative AI and foundation models.
Fewer than three in ten turn a pilot or demo into something deployable.
About one in twenty scale it across the business.
Fewer still translate it into value that shows up in the P&L.

Panelists listed why pilots stall, and most of the reasons had nothing to do with the model:

Data that exists but can’t be used. Having data and having data ready for AI are different things.
No business owner. A pilot can succeed technically and still go nowhere if no business leader commits to putting it into production.
The last 20%. A demo that works 80% of the time impresses. In production, the remaining 20%, including hallucinations, carries real risk.
Integration. A standalone chatbot can be built quickly. Connecting it to existing plant and enterprise systems is the hard part.

The panel also gave a practical set of success metrics. These include reduced unplanned downtime, and precision and recall in defect detection. That means tracking false negatives (defects that escape) separately from false positives (good parts rejected), with extra weight on critical defects that could trigger product recalls. The final metric was adoption: how often operators accept AI recommendations, and how often they override them.

One panelist argued that incremental gains rarely win investment. Leadership does not get excited by a 5–10% improvement, he said, but a 10x improvement in cost, speed or quality does get funded.

On technique, the session described how rare defects are handled. Large foundation models run in the cloud. Smaller “teacher” and “student” models, trained on specific low-frequency defects, run at the edge where the data is created. This setup was presented as a way to tackle rare-defect detection while keeping security concerns in check.

Theme 2: Data is not context

Talks on factory platforms kept returning to one phrase: having data is not the same as having context. Plants run ERP systems, PLCs, SCADA and manual inspection sheets, but these rarely connect. One presenter said planning teams often learn about shop-floor disruptions three to six hours late, sometimes a full day. By then they are re-planning against a situation that has already changed. The result is missed deliveries, low capacity use and no single source of truth.

The answer proposed from the stage was a maturity ladder, climbed one step at a time:

Digitize: move from paper to digital records.
Integrate: connect business applications.
Connect: link people, processes and machines.
Add intelligence: predictive maintenance and demand forecasting.
Go agentic: AI acts, but only with human approval.

Several speakers warned against trying to build intelligence for the whole factory at once. The advice was to pick one function, such as production planning, and build its context first.

Here the summit showed a real tension. The MSME panel argued that today’s generative AI can work with messy data from spreadsheets and messaging apps and still connect the dots, so data quality is less of a barrier than before. The generative AI panel said the opposite: curating data so a model can use it is the main problem, and models are easy to obtain. Both positions were stated with confidence, and no one reconciled them.

Theme 3: Augment people, don’t replace them

The special address took a firm line: no technology should be adopted for its own sake. Each tool must solve a real business problem, such as catching a quality defect, predicting a breakdown, saving energy or cutting waste. The speaker argued that India’s biggest strength is its people, so AI should augment workers rather than replace them. It should also capture senior staff’s experience in a form younger workers can learn from.

He framed AI’s role through the Japanese ideas of muri, mura and muda, which the speaker glossed as excess, unevenness and waste. He paired this with a four-stage approach to quality: don’t accept a defect, don’t make one, don’t pass one on, and finally make defects impossible. AI fits the last stage. A camera that compares each assembly with the ideal and signals OK or not OK can stop a defect before it happens.

The autonomous-operations session offered a counterpoint based on geography. A factory in India can still run with minimal digitalization, one panelist said, but a new plant in Europe, where labour costs are far higher, must be highly automated with minimal human involvement. He added that the dark factory is “not far away.” The summit did not settle where India’s balance between automation and people should lie, and the honest answer may differ by product and market.

Theme 4: Sustainability as a business case

The sustainability panel spoke about money as much as the environment. One panelist cited studies indicating that 80% of a product’s environmental impact is decided at the design stage. He listed the design levers:

Material: recycled content, sustainable options such as jute and bamboo, green steel and aluminium, and lighter materials.
Product design: fewer parts, parts made from a single material, and bolts instead of adhesives so products can be taken apart.
Process: compact plant layouts designed virtually to reduce material movement.
Use: connected products updated over the air, which extends their life.

Panelists also drew a line between “recyclable” and actually “recycled.”

The economic argument rests on life-cycle cost rather than part cost. In vehicle programmes, one panelist said, life-cycle carbon is now locked in before ROI is calculated, driven by the rising cost of non-compliance. Green products can also earn a premium, and carbon credits can be traded, turning sustainability into a revenue line.

An equipment manufacturer added scale to the argument. Pumping accounts for up to 20% of global electricity use, she said, and even small efficiency gains at scale could save up to 4% of electricity. Pumps are often bought on purchase price rather than lifetime energy cost.

The panel was also candid about the limits. A product take-back and recycling programme, piloted in 2024–25, is growing, but customers are “not ready yet” because the incentive is unclear. A fully recycled pump does not exist yet. Consumers’ appetite for new products also works against repair and reuse.

On traceability, a five-layer model was presented: IoT sensing, AI-driven insight, digital twins, context, and customer-facing digital product records. The speaker added that technology covers perhaps 80–90% of the need, and supplier maturity is the gap. In automotive assembly, roughly 65% of revenue flows to the supply chain, so tier-1 and tier-2 suppliers decide the outcome. The advice was for OEMs to act as collaborators, not auditors. That means training suppliers, sharing costs and offering IoT on an operating-cost model so suppliers avoid capital spending.

Theme 5: Making automation work for MSMEs

The closing panel named a paradox: the companies that gain most from automation are often the least able to absorb its implementation risk. It judged every initiative on three tests: relevance, economics and adoption.

The panel’s practical guidance, in brief:

Start with the real problem. Separate a real problem from a perceived problem or a symptom, using root-cause analysis backed by data. A “three-day delivery delay” may turn out to be a supplier who never received the order.
Begin small, and measure. Pick the top three bottlenecks. Record performance before and after any intervention, and give the project a clear owner.
Respect legacy machines. Use bolt-on IoT that doesn’t touch running PLCs. Expect two or three cycles before the data is reliable, and plan for clean power, which shop floors often lack.
Redesign, don’t patch. For high-impact workflows, rebuild the process with technology at its core, then switch off the old one.
Test retrofits honestly. Ask whether a retrofit solves the problem or just pushes the cost into next year.
Pool resources. One panelist suggested MSMEs build shared platforms through industry bodies, because off-the-shelf tools often fit 80% of needs while charging for 100%.

One speaker summed it up: the technology has rarely failed. What fails is readiness of process, people and culture. Adoption also depends on visible commitment. Shop-floor teams adopt new tools when they see senior leaders using them too.

The ecosystem agenda

The organisers outlined their own capacity-building work: an AI immersion programme with IIT Madras that drew 85 industry participants in its first, oversubscribed edition. The target is 500 participants across southern India within six months. An agentic AI developer certificate and MSME-focused workshops are also running; MSMEs make up half of CII’s member base. The theme address closed with a call for responsible AI, built on governance, transparency, data security and workforce empowerment, and a direct challenge to the audience: move beyond pilots and proofs of concept.

Analysis

The summit’s strongest consensus was not about any one area or technology. It was that factories fail at AI through weak problem definition, unready data and low adoption, not because the models fall short. Two questions remain open: whether generative AI truly lowers the data-quality barrier, and who funds the first step for MSMEs and the incentives for circular products. How those are answered will decide whether the pilot funnel widens.