Indian manufacturing is in an unusual position. Capital is arriving: the government told the Lok Sabha in July 2026 that production-linked incentive schemes had drawn "actual investments of more than Rs 2.40 lakh crore" across 14 sectors by 31 March 2026, with exports of ₹15.2 lakh crore and "over 14.15 lakh direct and indirect jobs". New plants are being built for solar modules, pharmaceuticals, automobiles and electronics. At the same time, the installed base of older plants, where most of the country's output is still made, runs on equipment and software from several decades, with data that lives in paper logbooks, programmable logic controllers and spreadsheets in roughly equal measure.
Industrial AI for manufacturing in India therefore has two quite different jobs. In a greenfield plant it is a design choice, specified alongside the machines. In a brownfield plant it is a retrofit, and the retrofit is where the value and the difficulty both sit. This piece looks at the three use-case families that account for most deployed plant AI, vision, quality and maintenance, and proposes a readiness ladder that a plant head or a CIO can use to judge, before any vendor arrives, which rung a given site stands on and what the next rung costs.
It is written for the people who have to approve these projects: plant heads, operations directors, CIOs of manufacturing groups, and the investors and policymakers who are trying to understand why the same technology produces a step change in one plant and a stalled pilot in the next.
Industrial AI for manufacturing in India: where the sector stands
Manufacturing contributes around 17 per cent of India's GDP, according to IBEF's sector overview (last updated September 2026), which also reports manufacturing gross value added growth of 11.6 per cent in FY2025-26, up from 8.3 per cent the year before, and notes the National Manufacturing Mission's target of raising the sector's share of GDP to 25 per cent by 2035. The same source records record gross FDI inflows of ₹8,35,070.61 crore in FY2025- 26. Whether or not the 25 per cent target is met, the direction of capital is clear.
The digitisation base from which AI must start is thinner than the investment figures suggest. The Nasscom and Capgemini report "India Industry 4.0 Adoption: A Case to Mature Manufacturing Digitalization by 2025", released in March 2022, estimated that India spent between $5.5 billion and $6.5 billion on Industry 4.0 in FY21, nearly half of Indian manufacturers' annual technology spend, and that Industry 4.0 had risen from 5 per cent of manufacturing IT spend in 2011 to 20 per cent in 2021. The report named the gaps as scaling foundational technology readiness, moving "from proof-of-concept to ROI-driven, outcome-based deployment", and scaling adoption among MSMEs, which it said contribute 33 per cent of manufacturing output and 45 to 50 per cent of exports. It listed AI-based computer vision for quality inspection among the technologies to watch. Four years on, those three gaps describe the field accurately. PanScience Innovations reports that it published an Industry 4.0 Investor Roundtable report with the Nasscom Centre of Excellence in 2024; this article does not draw figures from it.
The international benchmark is also worth stating. The World Economic Forum's Global Lighthouse Network, co- founded with McKinsey, added 17 sites in January 2025 to reach 189 production facilities and value chains, including CEAT's Sriperumbudur tyre plant as a Factory Lighthouse and Hindustan Unilever's Tinsukia site as an End-to-End Value Chain Lighthouse. The Forum reported that the latest cohort achieved an average "53% boost in labour productivity and 26% reduction in conversion costs" attributed to digital tools including AI, machine learning and advanced analytics, alongside a 30 per cent average reduction in material waste. Lighthouses are, by construction, the top of the distribution. They show what is possible on a site that has already solved data, connectivity and people; they do not describe the median plant.
Three use-case families, and what each one needs from the plant
Deloitte's 2025 Smart Manufacturing and Operations Survey, fielded among 600 executives of large manufacturers with US headquarters or operations between August and September 2024, found that 29 per cent were using AI or machine learning at facility or network level and 23 per cent were piloting it, that 54 per cent reported a data standard through a unified data model and 45 per cent an architecture standard, and that respondents attributed "a 10% to 20% improvement in production output" and "a 7% to 20% improvement in employee productivity" to their smart manufacturing initiatives. The sample is not Indian, and the figures should be read as an indication of what a well- capitalised peer group looks like, not as a description of India. The pattern they reveal, though, travels: adoption sits in the low thirties, data standards are present in about half, and the gap between the two is where most industrial AI projects live.
Within that gap, three families of use case recur.
Vision. Cameras on a line, a model that classifies what it sees, and an action: reject, alert, log. Applications include surface and assembly defect detection, packaging and label verification, safety compliance (personal protective equipment, zone intrusion) and counting. Vision is the most retrofit-friendly family because the sensor is new and does not depend on the machine's own data. It is also the family where ground truth is cheapest to produce, because a human can look at an image and say what it shows. Its requirements are lighting, camera placement, a labelled image set that covers the real defect distribution including rare classes, and an operational decision about what happens when the model is uncertain.
Quality. Prediction and root-cause analysis of quality outcomes from process parameters: temperatures, pressures, speeds, raw-material lots, operator and shift. This family depends on the plant's own data, which must be time-aligned across machines, laboratory results and batch records. The hard part is almost never the model; it is establishing that the timestamp on the furnace log and the timestamp on the lab result refer to the same material. Quality AI pays when scrap, rework or customer returns are expensive relative to the cost of instrumentation.
Maintenance. Condition monitoring and failure prediction for rotating and process equipment, from vibration, current, temperature and acoustic signals. Predictive maintenance is the most-marketed industrial AI application and the most frequently disappointing, for a reason that is structural rather than technical: failures are rare, so labelled failure examples are scarce, and the first year of any deployment is largely spent learning what normal looks like. It pays where unplanned downtime is very costly, where spares have long lead times, or where failure has a safety consequence.
The table is a synthesis of practitioner experience and is illustrative. Its purpose is to make one point: the three families place very different demands on the plant, and a site that is ready for vision may be two rungs away from being ready for quality analytics.
The plant AI readiness ladder
Most AI readiness frameworks are written for enterprises with ERP systems and data warehouses. A plant is a different object. Its data is generated by machines, its network may be air-gapped by policy, its people work in shifts with varying literacy in software, and its processes are governed by standard operating procedures that were written for humans. The ladder below is an original framework with four dimensions and five rungs. The tests are deliberately observable: each can be answered by walking the floor and asking to see something, rather than by a questionnaire.
Four tests make the ladder usable.
The retrieval test (data). Ask for the furnace temperature, the operator and the lab result for one specified batch from three months ago. At rung 2 the answer takes a day and three people. At rung 3 it takes a query. At rung 4 the three values come back joined.
The gateway test (connectivity). Ask how a model would get data off the line and a result back onto it. If the answer involves a USB stick, the plant is at rung 2 whatever its IT department says. If it involves a documented gateway with a data classification policy, it is at rung 3 or above.
The owner test (people). Ask who would be called if a dashboard showed wrong numbers. If no name comes, the plant is below rung 3. If a name comes but the person is in corporate IT and has never been on this floor, the plant is at rung 3 with a people deficit.
The override test (process). Ask what an operator should do when a vision system rejects a part the operator believes is good. If the answer is "ignore it", the plant is at rung 3 and the model will be switched off within a quarter. If a written procedure exists with an escalation path, the plant is at rung 4.
The ladder's central claim is that the three use-case families have different minimum rungs. Vision can be deployed usefully at rung 2 in data terms, because it brings its own sensor, provided connectivity and process are at rung 3. Quality analytics needs rung 3 in data and rung 4 in people, because the joining of records is a human coordination problem before it is a software one. Predictive maintenance needs rung 3 in data and rung 4 in process, because its value depends on someone acting on a probabilistic warning in time, which is a procedure, not a model. A plant that attempts predictive maintenance from rung 2 will spend its budget on the climb and attribute the cost to the AI.
What this implies for sequencing and budgets
The sequencing conclusion is that most brownfield plants should start with vision, not because it is the highest-value family but because it is the one whose readiness requirements are closest to where the plant already stands, and because the ground truth it needs (labelled images) can be produced within weeks. A vision deployment that works forces the plant up the connectivity and process rungs; the gateway gets built, the override procedure gets written, and a team forms around the data. Those assets are then available for quality and maintenance work.
The budget conclusion is that the climb should be costed separately from the models. In the Deloitte sample, 40 per cent of respondents ranked data analytics as a first or second investment priority for the next two years, ahead of AI at 29 per cent; the ordering is telling. A manufacturing group evaluating an AI proposal should ask the vendor which rung the proposal assumes and what the plant will pay, in instrumentation, integration and people time, to reach it. If the answer is "nothing, the model handles it", the proposal is assuming rung 4 and the plant will find out otherwise after the purchase order. A separate PanScience Research piece sets out a fully loaded unit-cost model for one AI workflow; the plant ladder is the input that model needs on the cost side.
The MSME point from the Nasscom and Capgemini report also bears on budgets. A large group can afford to climb one plant to rung 4 and copy the architecture across sites; the Deloitte finding that 45 per cent of large manufacturers report an architecture standard is what rung 5 looks like at group level. A single-plant MSME cannot amortise that climb. For that segment, the realistic path is vision at rung 2 to 3 with a vendor who brings the gateway and the procedure as part of the product, and a deliberate decision not to attempt quality or maintenance analytics until the data owner exists.
A practice example
PanScience Innovations reports that its venture PredCo works in industrial AI and predictive maintenance for manufacturing plants. The company-reported description is used here only to illustrate the sequencing argument: a venture serving brownfield plants has to carry readiness work, from sensor placement and gateway configuration to the writing of response procedures with the maintenance team, as part of its offer, because the plants it serves are rarely at the rung the model assumes. That has consequences for how such a venture is staffed and priced, which a separate PanScience Research piece on venture building covers; this article notes only that the ladder is as much a product-design input for the vendor as a diagnostic for the buyer.
What the policy environment adds and leaves
PLI capital is creating rung-4 plants by construction, since a new solar module or electronics line is specified with historians, networks and edge compute from the start. The policy instrument that is missing is one for the brownfield climb: the retrofit of connectivity, the structuring of maintenance records and the training of plant staff to act on model outputs. The Nasscom and Capgemini report identified MSME scaling as a gap in 2022; the ladder suggests why. The cost of the climb is largely people and process, and those are harder to subsidise than hardware.
For the plant head, the practical message is simpler. Before asking what AI can do for the plant, ask which rung the plant is on, in each of the four dimensions, using tests that can be run on the floor this week. The answer decides which of the three use-case families is available now, which is a year away, and what the year will cost. Industrial AI for manufacturing in India will be built plant by plant, and the plants that move fastest will be the ones that know where they are standing before they start to climb.
Methodology and sources
This article combines published figures with an original framework. PLI figures (investment of more than ₹2.40 lakh crore, exports of ₹15.2 lakh crore, 14.15 lakh jobs, 14 sectors, as of 31 March 2026) are from the written reply of the Minister of State for Commerce and Industry in the Lok Sabha, reported by IANS via Morung Express on 21 July 2026 (https://morungexpress.com/pli-schemes-attract-rs-24-lakh-crore-investment-create-1415-lakh-jobs-govt). Manufacturing's share of GDP (around 17 per cent), GVA growth, FDI and the 25 per cent by 2035 target are from IBEF's manufacturing sector overview, last updated September 2026 (https://www.ibef.org/industry/manufacturing- sector-india). Industry 4.0 spend and adoption gaps are from the Capgemini press release for the Nasscom and Capgemini report "India Industry 4.0 Adoption: A Case to Mature Manufacturing Digitalization by 2025", 1 March 2022 (https://www.capgemini.com/in-en/news/press-releases/industry-40-adoption-in-india-can-help-manufacturing- sector-meet-national-growth-targets-and-contribute-25-to-gdp-by-fy26/). Global Lighthouse Network figures and the Indian sites named are from the World Economic Forum release of 14 January 2025 (https://www.weforum.org/media/global-lighthouse-network-2025-world-economic-forum-recognizes-companies- transforming-manufacturing-through-innovation). Adoption, data-standard and benefit figures are from Deloitte's 2025 Smart Manufacturing and Operations Survey, published 1 May 2025, based on 600 executives of large manufacturers with US headquarters or operations surveyed in August and September 2024 (https://www.deloitte.com/us/en/insights/industry/manufacturing/2025-smart-manufacturing-survey.html and https://www.deloitte.com/us/en/about/press-room/deloitte-2025-smart-manufacturing-survey.html); the sample is not Indian and is used as a peer benchmark only. The use-case table and the plant AI readiness ladder are original frameworks and are illustrative; their tests are proposed, not validated against a sample. References to PredCo and to the 2024 Industry 4.0 report with the Nasscom Centre of Excellence are company-reported by PanScience Innovations (https://panscience.xyz) and carry no client, deployment or outcome claims.