A mid-sized manufacturer runs on three things: an ERP that everyone half-uses, a large number of spreadsheets, and two or three people who know where everything is. The third one is the real system. When that person is on leave, the company slows down, and everybody has quietly accepted this.
That is the actual opportunity for AI in Indian manufacturing, and it has almost nothing to do with the things AI is usually sold on.
What gets pitched, and why it mostly does not land
The standard pitch is predictive maintenance and computer vision on the line. Both are real, both work, and both are wrong for most companies reading this.
Predictive maintenance needs sensor history you probably do not have. If your machines are not instrumented, step one is a capex project, not an AI project, and the payback conversation looks very different. Vision-based quality inspection is genuinely excellent, but it is a per-line installation with fixtures, lighting and calibration. It is a factory project with an AI component, not the other way round.
The thing that costs you money every day is that nobody can find out what was quoted to this customer in 2023 without calling Suresh.
The unglamorous version that actually pays
A private AI sits on the material you already have. Not sensors. Your ERP tables, your purchase orders, your quotations, your drawings folder, your quality reports, your supplier correspondence. Fourteen years of it, in inconsistent formats, with three spellings of every customer name.
What it does is let a person ask a question in a sentence and get an answer with the source attached. Here is what that looks like in practice.
Quotation memory
A customer calls for a repeat order. What did we quote them last time, what was the material rate then, what did we finally settle at, and did we take a hit on that job. Today that is forty minutes and two phone calls. It should be one question.
This is the single highest-value use case we see, because it directly protects margin. Companies lose money quoting from memory far more often than they lose it on the shop floor.
Receivables that answer back
Not a report. A question: which dealers in the south are over sixty days, how much, and what did they say last time we chased them. The AI can read the ledger and the follow-up emails together, which no report does.
Drawings and specification lookup
Every manufacturer has a folder structure that made sense to the person who created it in 2014. Finding the right revision of the right drawing for a part number you half remember is a daily tax. A private AI that has been pointed at that folder turns it into a question.
Supplier and material history
Which of our suppliers for this grade has ever had a rejection, what was the reason, when. This is the sort of thing that lives in the quality team's memory and evaporates when they change jobs.
Compliance and audit preparation
An ISO or customer audit is largely an exercise in producing evidence that already exists somewhere. Halving the time it takes to assemble that is worth real money in a week where everybody is doing it instead of their job.
Why the data has to stay yours
Manufacturing data is the most sensitive commercial information a company has, and it is sensitive in a way people underrate. It is not just personal data. Your quotation history is a map of your margins. Your BOMs are your cost structure. Your customer list with volumes is what a competitor would pay for.
If you are supplying to an OEM, you have likely signed an agreement that restricts where their drawings and specifications may be processed. Uploading them to a public AI service to summarise is, in many of those contracts, a breach. Nobody in the room usually knows this, because the person who signed the agreement is not the person pasting the drawing into a chat window.
This is the practical argument for a private setup, and it is a contractual one before it is a philosophical one. The test to apply is simple: does your data leave your control, is it retained anywhere after the answer is given, and is it ever used to improve someone else's model. Get those three answers in writing. We wrote a longer checklist in how to check a private AI claim.
A two-hour exercise before you spend anything
List the ten questions your team chases most often. Against each one, write down which system or folder holds the answer today.
Questions with a clear source are AI projects. Questions with no source are record-keeping problems, and no software recovers a quotation that was only ever verbal.
The three things that decide whether this works
Your data is messier than you think, and that is fine
Everyone worries their data is too dirty. It usually is dirty, and it usually does not matter as much as feared, because the AI can be taught that ABC Ltd, A.B.C. Limited and ABC are the same party. What does matter is whether the data exists at all. If quotations were verbal and never recorded, no software recovers that.
A useful two-hour exercise before spending anything: pick your ten most-asked questions and, for each, write down which system or folder the answer lives in. If a question has no source, it is not an AI problem yet.
Somebody on the floor has to want it
Projects that are bought at the top and handed down do not survive. Projects that start because the accounts head is tired of chasing the same number every Monday do. Find that person first. We wrote about this failure mode in most AI projects do not fail, they go unused.
It has to show its working
A factory runs on numbers people are accountable for. An answer without a source will be checked manually, and once it is checked manually twice, the tool is dead. Every reply needs the voucher number, the file, the row.
Where to start, concretely
Pick one question that costs somebody time every week. Quotation history is usually the right first one, because the value is obvious and the data is contained.
Run it on a sample of your real data before committing to anything. A demo on somebody else's dataset tells you nothing about whether it survives your naming conventions. Any vendor unwilling to do this under NDA has told you something useful.
Then judge it on week six, not week one. Accuracy in the demo is table stakes. Whether the purchase team still opens it in six weeks is the only number that predicts whether this was worth doing.
That first question is roughly a ₹2.5 lakh piece of work, not a ₹50 lakh transformation programme. Which is the point. Melon exists to be the small version that earns the right to the big version.
