The factory that thinks: Why intelligent manufacturing needs more humans: Gopi Thangavel, Group CIO, L&T

Will AI take over the factory? For Gopi Thangavel, Group CIO of Larsen & Toubro, the answer is a firm no, and he has the history to back it up.

Speaking at the Intelligent Manufacturing 500 conference, Gopi Thangavel drew on two unlikely sources: Eliyahu Goldratt’s The Goal, which shows why a single bottleneck can undo a factory full of perfect machines, and William Stanley Jevons’s 1865 The Coal Question, which explains why cheaper, more efficient technology tends to expand work rather than shrink it. From there he moved through the real world of manufacturing, from temperature-sensitive medicines losing potency on the road, to digital twins that let engineers fail safely before touching a ₹25 crore cylinder, to the shift from generative to agentic AI.

His conclusion is that the hardest parts of intelligent manufacturing are not technical at all. They are security, accountability and a workforce that is ready, and the “dark factory” with no lights and no people is a misreading of where this is heading.

Here is Gopi in his own words:

Every CIO in the room has lived this story : a stack that keeps shifting, upgrades that feel like heart attacks, and production that never gets to pause while we rebuild the engine in flight.

That is the real job of technology leadership in manufacturing today. And the question I keep hearing is the wrong one: Will AI take over the factory?

My answer is no. But to explain why, I need to start with a simpler question.

What manufacturing actually is
Strip away the industry jargon and manufacturing has one definition, whether you are in America, Africa or Asia: transforming resources into useful things, repeatably, in a way that is valuable and trusted. A shoelace, a car, a medicine, a tractor, an electrical grid. Same logic.

What is changing is the nature of the process. The physical process is becoming an information process. Quality, transformation, distribution, warehousing, logistics and value creation are all being instrumented, connected and made visible. That is what I mean by intelligent manufacturing: machines, data, people and products connected intelligently, with data you can rely on and with safety first, always.

We have been here before. The steam engine gave us the first revolution, electricity the second, IT the third. The intelligence revolution is the fourth, and I expect it to run for at least the next five to ten years. There is a nice historical echo: the first factories had to be built where water was, so water-rich land became expensive and strategic. Digital transformation removes that constraint. You can now build intelligent operations almost anywhere.

Lesson one: A perfect machine does not make a perfect factory
The first book I recommend to every manufacturing leader is Eliyahu Goldratt’s The Goal. Its core idea is deceptively simple. Imagine every process on your line runs at 80 units per hour, except painting, which manages 40. Your output is 40. It does not matter how superbly the other stations perform.

Optimizing each machine in isolation does not make the factory succeed. The bottleneck decides everything, and it keeps repeating until someone removes it.

This is exactly where AI earns its place. Not as a showpiece, but as a tool to find the constraint, predict it and clear it, continuously. If your AI program cannot answer “where is my bottleneck right now?”, it is decoration.

Lesson two: Cheaper intelligence means more of it, not less
The second book is William Stanley Jevons’s The Coal Question, published in 1865. When James Watt’s steam engine made coal use more efficient, everybody predicted coal was finished. Jevons observed the opposite. As efficiency improved, coal became cheaper to use and was adopted everywhere, including by people who could never have afforded it before. Total consumption went up, dramatically. That is the Jevons Paradox.

Today’s AI debate sounds uncannily similar. Tokens are expensive, GPUs are expensive, and many people predict that AI will simply eliminate jobs and switch the lights off. My view is that token costs will fall sharply, possibly within a year or two. When that happens, a premium AI subscription becomes an everyday commodity, and consumption explodes.

Two implications for CIOs. First, plan your cost model for abundance, not scarcity. Second, remember that new technology rarely retires the old on schedule. We keep buying new instruments, yet we still run legacy equipment and squeeze every drop of value from it. Intelligent manufacturing has to work with both.

Intelligence in the real world: Four short stories

Pharma: the journey of a tablet. A chemical reaction that works in the lab does not automatically behave the same way at industrial scale, and it is just as fragile in transit. When we traced where temperature-sensitive products were losing effectiveness, the findings were uncomfortable. A 15-minute idle time on the floor had stretched to three hours. A 20-minute shift to the warehouse took four. Trucks meant to travel at 40 km/h were doing 80 to 100, and drivers sometimes switched off cooling to rest.

Temperatures swung up and down, and the product’s chemistry changed with them.

We fitted vehicles with GPS integrated with blockchain. Now an alert fires if the vehicle exceeds the speed limit, stops unexpectedly or switches off the engine, and a shipment that sits too long at a distribution center gets flagged. The beneficiary is not the balance sheet. It is the patient who receives a medicine that actually works.

Oil and gas: sensors as the nervous system. When a pipeline is built, it is pressure-tested to extremes. The value lies in how much sensing you have, how well you predict behavior under pressure, and how vibration and communication between systems are controlled. Global majors such as Shell have shown what a sensor-rich, predictive approach can do.

Heavy engineering: you cannot experiment on the real thing. A single large cylinder can cost around ₹25 crore. Nobody can afford to learn on it live. Digital twins let interns and experienced engineers alike experiment, fail safely and build best practice before anyone touches the physical asset.

Automotive and electronics: data is the new differentiator. Even a 100-gram reduction in a car component changes efficiency and design. Tesla is often described as a data company that happens to make cars, and Indian manufacturers such as Ashok Leyland, Tata and Maruti are sharpening their focus on manufacturing data as well. Whether the product is a vehicle, a laptop or a phone, every minute of delay is calculated, and every minute of unplanned downtime is a cost you can see.

From generative to agentic: why the distinction matters
The World Economic Forum’s “lighthouse” network now counts roughly 230 of the world’s most advanced factories, and about 62% of them have already put AI to work. The maturity curve is steep, and the important shift is from generative to agentic AI.

Here is the simplest way I explain it. A new joiner arrives at your plant. Generative AI can answer, “How many holidays do I have?” Useful, but it only responds. Yet that new employee also needs a laptop,      Wi-Fi, a user ID, gateway access and, in a manufacturing environment, access to machinery and OT platforms. You cannot build a separate bot for every one of those steps. Agentic AI orchestrates the whole thing in one shot, with 20 or even 200 agents working in parallel.

On the shop floor, the same maturity shows up in the morning routine: detect a breakdown instantly, fix it the same day, and eventually predict that a given machine always goes down on Thursday at 10 a.m.

The next frontier is physical AI: drones inspecting facilities, robotic arms, and “cobots” that work alongside people. The question is no longer only can we, but whether we can do it safely, reliably and economically, with a predictable cost per transaction, per unit manufactured and per safety incident.

Safety is where AI quietly delivers some of its best value. On a construction site, vision systems can spot open pits, missing helmets and risky zones. Combined with temperature sensing, they can alert supervisors the moment conditions drift past a threshold, so the problem is fixed before it becomes an incident.

The hard part is not the technology
Three things decide whether intelligent manufacturing succeeds, and none of them is a model.

Security and accountability. IT and OT security are non-negotiable, and AI must always be accountable to a person who owns what it does. “We don’t know why it happened” is not an acceptable answer on a factory floor.

Workforce readiness. Too many organizations buy tools and under-invest in training. Capability has to become culture, not a one-time rollout.

Safe environments by design. Intelligence should be used first to remove risk from the places people work.

And for CIOs, a reminder on priorities. We used to talk about people, process and technology. In manufacturing, the order that works is product, process, people, supported by machines, digital twins and one information system that can serve the entire group.

The future is not dark
Many people have misread where this is heading. They picture a dark factory: no lights, no people, AI running everything. Jevons teaches us to be skeptical. Cheaper, more capable technology does not shrink the world of work. It enlarges the world of activity.

The future I am working toward is intelligent, sustainable, human-centered manufacturing. Quality matters. Safety matters. Sustainability matters. And people matter more than any of them.

The steam engine did not end coal. AI will not end the factory floor. It will, if we lead it well, make that floor smarter, safer and more human than it has ever been.