Factory Automation for Indian MSMEs: What to Automate and What to Leave Alone

Automation conversations in Indian MSME plants usually start with a vendor demonstration and a labour-cost calculation. Both are the wrong starting point, and the resulting projects are why a lot of expensive equipment sits idle in Indian factories.

The right starting point is a measured loss list.

Step zero: measure before you spend

Across the plants we have audited, true overall equipment effectiveness before any intervention has run 31–68%. The median is around 55%. That means roughly a third of the capacity these businesses were about to buy already exists inside the plant they own.

Four weeks of shift-by-shift logging of availability, performance and quality losses costs nothing but discipline, and it reorders every subsequent decision. On one Nashik engagement it deferred ₹7 crore of planned capex indefinitely — the four interventions that came out of the loss list took line output from 2.8 to 4.6 tonnes an hour on the same equipment.

Automate for these three reasons

1. Consistency. Where product quality depends on operator judgement and variation costs you rejections or customer complaints, automation pays reliably. This is the strongest case in food processing.
2. Data. Instrumentation that captures weight, temperature, count and downtime automatically is the cheapest automation available and usually the highest return, because it makes every other decision better.
3. Genuine bottleneck relief. Where one station demonstrably constrains the whole line and cannot be relieved by better scheduling or changeover practice.

Do not automate for these reasons

• Labour cost alone. At Indian wage levels, straight labour-substitution payback is often five years or more, and vendor calculations usually assume a utilisation you will not reach.
• A process you have not standardised. Automation locks in your current practice, including its defects, and makes them harder to change.
• Because a competitor did it. Their bottleneck is not yours.
• A seasonal peak. Equipment bought for eight weeks of the year is very expensive capacity.
• Because a scheme will subsidise it. A subsidy on the wrong asset is still the wrong asset, and you will operate it for twenty years.

A sequence that works for a ₹10–60 crore plant

1. Instrument for data — weighbridge integration, line counters, temperature logging, downtime capture. Typically ₹2–6 lakh, payback measured in weeks through better decisions.
2. Standardise and document the process — SOPs, changeover routines, quality plan. Costs time, not capital.
3. Attack changeover and maintenance losses — usually the largest items on the loss list and mostly addressable without capital.
4. Automate the genuine bottleneck station, chosen from measured data rather than intuition.
5. Automate packing and coding, where consistency and traceability requirements are rising fastest.
6. Consider process automation and recipe control, which pays above roughly ₹25 crore of revenue where batch consistency is commercially critical.

Where AI genuinely helps, and where it does not

There are narrow, real applications at MSME scale: vision-based defect detection on a packing line, extracting figures from supplier quality certificates, summarising tender documents, classifying and routing inbound enquiries, and drafting RFQ responses. All of these sit behind human review and all of them save real hours.

What we will not do, and advise against: putting an unreviewed model anywhere near a quality release decision, a price quotation or a statutory filing. The cost of a wrong answer in those places is not proportional to the labour saved.

Automation does not fix a process. It makes whatever process you have permanent, and more expensive to change.

How to evaluate a proposal

Ask the vendor for the assumed utilisation and the assumed labour redeployment, then replace both with your own measured numbers. Most proposals lose their payback at that point, and the ones that survive are worth doing.

Also ask what happens to the rest of the line. Automating one station frequently moves the bottleneck rather than removing it, and a proposal that does not model the whole line has not been thought through.