Many years ago, on the shop floor of a plant manufacturing heavy engineering equipment, we watched a robotic welding cell come alive. The trials went exactly as promised; clean welds, consistent cycle times, an engineer beaming beside the control panel. The machine worked.What took the next 18 months was something else. Qualifying the wire feed across three shifts. Training operators who had welded by hand for 20 years to trust a system they could not see inside. Building a maintenance regime that did not exist yet.By the time that cell ran reliably across volumes, the technology had become the smallest part of the story.A near-identical lesson came from the digital side. When a Dealer Business Management System was deployed across a national dealer network of a large automobile company, provisioning took weeks; operational alignment took nearly two years standardising workflows, cleansing field data, converting users from manual habits. The technology was again the smallest part of the story.We think about both examples when we sit through AI demonstrations today. The question is no longer whether AI works. The demonstrations have settled that. What remains unsettled is whether organisations can deploy it reliably and at scale.The first phase of AI was a demonstration phase and a remarkable one. Models that write serviceable code, draft contracts, summarise board papers and pass examinations were, a few years ago, improbable. But it has stopped being the interesting question.Inside many organisations today, almost every function can show a pilot. Far fewer can show AI embedded in how the organisation runs, week after week, without a champion keeping it alive.The gap between a pilot and a system is not a technology gap. It is an industrialisation gap. And it is exactly the gap manufacturing spent a century learning to close.Why pilots are easyPilots succeed for reasons that have little to do with the technology. Data is curated by the team that wants the pilot to work. Users are the most capable people in the building. The problem is narrowed until it fits inside what the model can do. None of this is dishonest. It is how any controlled trial behaves.Scaling removes every one of those conditions at once.The data that looked clean in one plant turns out inconsistent across 12. The workflow built for an early-adopter team must survive contact with a function that never asked for it. Accountability, informal during the pilot, must now be designed. Who owns the decision when it goes wrong?India’s own transport policy offers a precise illustration. In 2015, Rule 118 mandated Speed Limiting Devices on commercial vehicles, sealed by the transport authority to prevent tampering.The device worked exactly as specified. But a sealed component is not a monitored system: without live tracking behind it, seals were routinely broken and devices bypassed. Compliance on the road stayed inconsistent for years. The lesson was not that the device failed. It was that a component, however well engineered, cannot industrialise itself.A factory floor, an enterprise network, a national regulatory mandate: three domains, one lesson. Scale is also financial. Running millions of decisions every day costs far more than running a few thousand test queries. The gap catches most organisations by surprise.The answer is leaner AI purpose-built for specific tasks, not large models for everything. Scale is not only an operational challenge. It is a financial one.Infra over intelligenceThis is why the centre of gravity in serious AI conversations is moving from the model to everything around it. Data architecture, governance, security and the unglamorous plumbing that determines whether intelligence survives contact with a real organisation.Across sectors, enterprises are expected to shift from buying off-the-shelf AI to building it in-house, recognising that the advantage lies in integration depth, not model access.A capable model does not redeem a weak operating system. It accelerates it surfacing problems the organisation had perhaps quietly agreed to ignore.Manufacturing leaders recognise this instinctively. A new machine tool does not fix poor scheduling discipline, it simply produces the same defects faster, at higher cost.AI follows the same logic. Organisations treating it as a software upgrade rather than an operating-system question are most likely to be disappointed and most surprised, since the model performed exactly as advertised.Industrialisation also changes who does the work, not only how. The most valuable people on a shop floor are rarely pure specialists. They connect a process problem to a materials answer and a cost constraint in one sentence. AI is producing the same demand at the desk.The engineers and managers capturing disproportionate value are not those using AI cleverly in isolation, but those who hold a model’s output against business judgment, regulatory constraint and operational reality at once. That is a systems skill, not a technical one and scarcest in most organisations. Capability is migrating from performing tasks to integrating systems and the gap is widening faster than most systems acknowledge.Structural arbitrageThe opportunity for India here is large.The global conversation remains fixated on who builds the largest model which is an expensive race, with entry costs in the billions. The industrialisation race rewards different capabilities and India has been building them for two decades without recognising that as preparation.Consider what India has already achieved at scale. Aadhaar authenticates a billion identities in real time, across infrastructure of sharply uneven quality. UPI processes more transactions than most nations settle in a year. These are not technology achievements. They are industrialisation achievements; embedding digital systems into a fragmented multi-lingual economy and making them work reliably at volume. That capability is what the AI race will reward.The more durable ambition is not to build the most powerful model. It is to become the world’s most credible laboratory for deploying AI inside real economic systems — manufacturing lines, hospital networks, logistics corridors, agricultural supply chains — reliably enough so that the world studies the deployment, not the demonstration.Viksit Bharat 2047 will be built less on who invents AI first than on who industrialises it first; at scale, across real systems, without losing reliability. Invention proves what a machine can do. Industrialisation decides who benefits from it.Sondhi is Independent Director: Global–India; former MD & CEO, Ashok Leyland and JCB India; Sundararaman is Chief Scientist and Head of Wipro Research. Views are personalPublished on July 10, 2026