Most writing on AI in manufacturing describes a plant that does not exist in India below ₹1,000 Cr: a sensor on every machine, a data lake, a camera on every line. The plants AgentJi walks into run on Tally, Excel, a boiler register and four WhatsApp groups. This note is about what AI actually runs in such a plant, what each piece saves, and what AgentJi tells promoters not to buy yet.
What “AI” means on a plant floor
Not robots. Not a control room. In a mid-size Indian plant, AI in manufacturing means software agents that read what the plant already produces (vouchers, registers, meter photos, weighbridge slips, WhatsApp messages), match it against what should be there, and tell a named person when it is not.
The reading is the new part. Until recently, software needed clean, typed data. Now it can read a handwritten boiler log or a photo of a weighbridge slip well enough to work from. Everything downstream, reconciling, costing, reporting, is arithmetic that was always possible and never done, because the numbers were never in one place.
Where the data actually is
On paper and in spreadsheets. The Manufacturing Leadership Council, the digital arm of the US National Association of Manufacturers, found in 2024 that 70% of manufacturers surveyed still collect data manually. That is a US survey; the Indian figure is not in print, and AgentJi has yet to visit a plant where it would be lower. McKinsey Global Institute’s 2017 study put 64% of time spent collecting data and 69% of time spent processing it as technically automatable with the technology of that year, before software could read a register.
The practical consequence: in a ₹100 Cr plant, two to four people in accounts and planning spend most of their week moving numbers between systems that do not talk. That is where AI in manufacturing starts, because that is where the hours are and where the errors are made. Barchard and Pace, in a 2011 study in Computers in Human Behavior, measured manual single-entry error at about 1% of values. One wrong digit in every hundred, across every register in the plant, every day, and nobody knows which one.
The four agents that actually go live
AgentJi has settled on four, in the order that pays.
Reporting over Tally and Excel
The first agent reads Tally, the dispatch register and the production sheet every morning and sends one message: yesterday’s production, dispatch, collections and cash. Nobody exports anything. This is what AgentJi calls AI ERP and reporting, and it goes live in 30 days because it changes nothing in the systems it reads.
What it saves: the MIS hours, and the month-end. APQC’s 2017 benchmark, as reported by Numeric, puts the median monthly close at 6.4 calendar days across roughly 2,300 organisations. Plants on Tally and Excel are usually at the slow end of that range, and the close shrinks to a morning once reconciliation is done daily instead of monthly.
Plant KPIs from the meters you have
The second agent reads the boiler log, the energy meter and the shift register, by photo, by meter export, or by a small gateway where one exists, and reports steam per tonne, power per tonne, downtime by cause and yield per shift. No new SCADA. This is plant KPI monitoring.
What it saves: energy first. A plant that sees steam per tonne by shift finds the shift that runs the boiler badly within two weeks. Deloitte and MAPI’s 2019 smart factory study, which surveyed more than 600 US manufacturing executives, reported average three-year gains of around 10% in output, capacity utilisation and labour productivity from smart factory initiatives. AgentJi targets 3 to 6% off energy cost per tonne in the first year, in writing, because that is what reading the existing meters reliably delivers.
Costing per batch, and the metal that goes missing
The third agent costs every batch from raw material intake to finished goods, and a route-and-weight agent watches every inbound truck: weighbridge slip against purchase order against GPS route. This is costing, yield and loss prevention.
What it saves: the batches that lose money, which the average margin hides, and raw material that leaves between the supplier’s weighbridge and yours. AgentJi’s target is raw-material loss below 0.5% of intake. No Indian manufacturing shrinkage figure exists in print, so AgentJi does not quote one; the target is set against the plant’s own intake and consumption, measured in the first 30 days.
Controls an auditor will sign
The fourth agent checks every voucher, bank line and stock movement overnight and keeps a controls register any auditor can read. For a family business that wants an SME IPO or an outside investor, this is the one that matters: the merchant banker wants three years of books that reconcile, and SEBI’s eligibility rules assume them. This is AI audit and IPO readiness.
What it saves: internal audit effort, and the weeks spent answering auditor queries from memory.
What does not run yet
Promoters are shown three things at every conference that AgentJi declines to build in a plant still on registers.
Predictive maintenance. It needs years of sensor history on the machine that fails. A plant with no sensors has no history. Put the KPI agent on first; the downtime log it builds is the history.
Computer vision for quality. It needs thousands of labelled defect images per product. Possible, expensive, and rarely the biggest loss in the plant.
Demand forecasting. Three years of lumpy, dealer-driven sales in Excel do not forecast anything. Fix the sales register first.
None of these are wrong. They are later. AgentJi says so in the roadmap and ranks them below the four above by rupees saved per month of effort.
How to judge it
Not on the demo. On what is running. The promoter should be able to answer three questions every Monday: which agents are live, what did they do last week in counts and rupees, and what does the vendor need from the plant. AgentJi sends that page every week and publishes engagement notes 90 days after go-live, in the client’s numbers.
The commercial test is as plain. An implementation fee per agent, a monthly fee to keep it running, and a saving agreed in writing before anything is built. Not met in 90 days, the vendor keeps working at no charge for up to 90 more days until it is. A vendor who will not sign that is selling capability, not completion.
What it needs from the plant
Three things, and the third is the one that fails. Access: a Tally login, the WhatsApp groups, and permission to photograph the registers. A named person: one clerk or supervisor who answers the agent’s exceptions each morning, because an exception nobody owns is a loss nobody catches. The promoter’s attention: ten minutes on Monday to read the brief and answer what AgentJi needs. Plants where the third goes missing after week six are the plants where the agent stops paying, and AgentJi says so in the first week rather than the twelfth.
Where to start
Pick the register that causes the most arguments at month end. In most plants it is the bank reconciliation or the stock count. Put the first agent on that. Thirty days later there is a number on the promoter’s phone that was not there before, and the next agent is easier to choose.