A vision-inspection system for a hand-tool line is not exotic technology. A camera, a light, a trained model, and a rejection arm. It catches the hairline forging crack that a tired inspector at the end of a shift does not. It runs at line speed. It does not have opinions about overtime.
It also costs more than most of the units on my road will spend on capital equipment this year.
That single sentence explains, better than any survey, why India's AI-in-manufacturing conversation has not reached Punjab. Not scepticism. Not illiteracy. Arithmetic.
And the arithmetic changes completely the moment you stop asking each factory to buy the system alone.
What the numbers actually say
The scale of the sector is not in dispute. According to the Economic Survey 2025-26, MSMEs contribute around 31.1 per cent of India's GDP, 35.4 per cent of manufacturing output, and 48.6 per cent of exports. The IndiaAI Mission carries a commitment of roughly $1.15 billion, and government estimates put AI's potential addition to the Indian economy at up to $1.7 trillion by 2035. The Union Budget for 2026-27 carries a ₹100 billion outlay for MSMEs.
The intent is not the problem. As the World Economic Forum put it earlier this year, India does not lack AI policy frameworks or technology intent what it lacks is translation at scale, the ability to convert that intent into measurable productivity on the shop floor. Execution, not ambition, is now the binding constraint. Adoption remains uneven and fragmented.
I would put it more bluntly. The unit of adoption is wrong.
The rational refusal
There is a tendency in policy conversations to treat MSME hesitancy as a cultural failing. Owners are described as conservative, digitally illiterate, resistant to change.
Research by PwC on MSME AI readiness records something more precise. One manufacturer told the researchers, in words that will be familiar to anyone who has sat in a Ludhiana factory office, that a short-term, survival-oriented mindset de-prioritises long-gestation digital projects. Others cited an inability to finance rising subscription fees and hidden integration costs, and the risk of being locked into a vendor.
Read that again, because it is not irrationality. It is an accurate risk assessment.
Consider what the last eighteen months looked like for an exporter in this cluster. A United States tariff regime that moved from 50 per cent to 18 per cent to a 10 per cent surcharge, with two of those regimes struck down in court. A European carbon border mechanism entering its financial phase in January 2026. A new EU steel quota system from July 2026. An owner who declines to sign a five-year software subscription against that backdrop is not backward. He is solvent.
The response cannot be to lecture him about mindset. The response has to be to change the size of the cheque.
Two hundred signatures
Here is the arithmetic that nobody has done publicly.
A vision-inspection cell, an edge-computing stack, a small data-engineering team, and a testing bay are, for one firm producing a few product families, a capital expense with an uncertain payback and a real risk of obsolescence.
The same assets, shared across a cluster of two hundred units producing spanners, wrenches, couplers, props, fasteners and forgings, are something else entirely. Utilisation rises from perhaps fifteen per cent to something close to continuous. The training data which is the actual scarce input, not the camera, accumulates across two hundred production lines rather than one. The model that learns to spot a forging defect on my line becomes better because it has also seen defects on my competitor's line. The engineer who knows how to maintain the stack is employable full-time by the cluster, and by nobody in it individually.
This is not a theory. It is how the WEF describes the only model that has worked: cluster-led deployment, because it enables peer learning, shared infrastructure and ecosystem partnerships. NASSCOM and CII are already running MSME cluster pilots through Centres of Excellence and smart manufacturing testbeds. A set of Indian technology firms Industry.AI, Flutura, Detect Technologies, Altizon Systems among them are deploying exactly the relevant use cases: visual quality inspection, predictive maintenance, compliance management, energy optimisation, much of it at the edge rather than in a distant cloud.
The pieces exist. The purchasing entity does not.
We have done this before, in this district
The most persuasive argument for a shared AI facility in Ludhiana and Jalandhar is that the cluster already built one, for the last technology it could not afford individually.
The Central Institute of Hand Tools in Jalandhar is a government technology centre supporting hand-tool MSMEs in Punjab. It provides common facility services forging, heat treatment, testing along with design assistance and quality improvement programmes. It exists because, decades ago, somebody understood that a hundred small forges could not each own a heat-treatment furnace and a metallurgical lab, but a hundred small forges together could.
That decision is why Jalandhar and Ludhiana today account for nearly 80 per cent of India's hand-tool exports, with twelve of the top fifteen hand-tool manufacturers based in these two cities. The Punjab State Industrial Development Corporation's industrial estate did not create that cluster by giving each firm a subsidy. It created it by giving all of them a furnace.
Heat treatment in 1975 is computer vision in 2026. The technology is different. The economics are identical.
The instrument already exists and nobody has used it for this
Here is the part that should embarrass all of us, myself included.
India does not need a new scheme for this. The Micro and Small Enterprises Cluster Development Programme has, for years, financed exactly this kind of shared asset. Under its hard-intervention component, the Government of India grant covers up to 70 per cent of the project cost for a Common Facility Centre containing machinery and equipment for critical processes, research and development, and testing rising to 90 per cent for clusters where more than half the units are micro, women-owned, or SC/ST. More than a thousand interventions have been taken up under the programme across twenty-nine states.
The scheme is demand-driven. State governments forward proposals reflecting the common needs of an existing cluster. Which means the reason no AI Common Facility Centre exists for the Punjab hand-tools cluster is not that Delhi refused one.
It is that Ludhiana never asked.
What the facility should actually contain
If the ask is to be taken seriously, it has to be specific. Four components, in order of payback:
A shared visual-inspection and defect-classification service. Units send sample batches; the centre runs inspection; defect data returns to the unit and, anonymised, into the shared model. PwC's research notes that for digitally mature MSMEs equipped with sensors and IoT, AI-driven quality control and defect reduction addresses one of the key barriers to integrating MSMEs into global value chains. Defect rate is not a productivity metric in an export business. It is the qualification criterion.
A compliance and traceability data layer. This is the component nobody is discussing and the one I would fund first. The EU's carbon border mechanism penalises exporters who cannot document their emissions, applying punitive default values to those who cannot produce verified figures. From October 2026, EU steel imports must declare the country where the steel was melted and poured. US Section 232 requires the same declaration on steel derivatives. Every one of those obligations is a data problem before it is an engineering problem. A cluster-level system that captures per-batch energy, origin and process data once, properly converts a compliance burden into a commercial asset. Two hundred firms cannot each build this. One centre can build it for all of them.
A shared OT security function. At the India AI Impact Summit in February, Palo Alto Networks' India head warned specifically about cybersecurity risk in operational-technology environments, and the need to build modern network-security principles into how this workforce is trained. We are about to connect several thousand machines in a border state's industrial cluster to the internet. Somebody should be responsible for that. No individual unit will be.
A skills bay. Rockwell Automation's India managing director made the point that real AI value emerges only when it is integrated into industrial automation systems at scale. Integration is a human skill. The centre should train line supervisors, not data scientists.
Why the window is now, and narrow
Three things make this the moment.
First, the money is allocated. The ₹100 billion MSME outlay and the IndiaAI Mission's commitment are both live, and MSE-CDP is a standing instrument awaiting proposals.
Second, and more uncomfortably: when MeitY and the IndiaAI Mission commissioned a roadmap for AI adoption across more than 350 manufacturing MSME units earlier this year, the sectors chosen were textiles, pharmaceuticals and medical devices, and electronics.
Engineering goods and metals fabrication were not among them. The sector that accounts for the largest share of India's merchandise exports is absent from the country's flagship study on AI adoption in manufacturing MSMEs. That is not a conspiracy. It is what happens when a cluster does not organise itself to be counted.
Third, the pressure is already here. Buyers in the EU are asking for emissions data. Buyers in the US are asking for melt-and-pour origin. Neither of them will wait for us to hold a seminar.
The ask
To the Government of Punjab, the Ministry of MSME, EEPC India, and the hand-tool and scaffolding associations of Ludhiana and Jalandhar, a single, costed, unglamorous proposal:
Commission a diagnostic study for an AI and Traceability Common Facility Centre for the Punjab hand-tools and scaffolding cluster, and route it through MSE-CDP within this financial year.
Anchor it at, or alongside, CIHT Jalandhar, which already holds the cluster's trust and its testing infrastructure. Structure the special purpose vehicle so that the smallest forge on the street has the same access as the largest exporter. Fund the compliance data layer first, because that is where the export orders are being decided this year, not in 2030.
And publish the utilisation numbers, quarterly, so that the next cluster in the next state can copy what worked and skip what did not.
The point
The debate about AI in Indian manufacturing has been conducted, so far, as a debate about willingness. Are MSMEs ready? Do they understand? Will they adopt?
Wrong questions. A hand-tool exporter in Ludhiana understands perfectly well what a defect rate costs him, because a German buyer has already told him. He is not unwilling. He is under-capitalised, and he is being asked to solve a shared problem privately.
Fifty years ago this cluster faced the same choice about heat treatment, and chose correctly. It built one furnace and shared it, and turned two Punjab cities into eighty per cent of a national export industry.
The furnace of this decade is a model, a camera, and a database. It should be built the same way, in the same place, by the same people, for the same reason.
We do not need to be convinced. We need to be counted, and we need to be allowed to buy this together. (By Gaurav Singal, CEO, Eastman IMPEX)
