AI on the Indian Factory Floor: What Actually Works

1. Visual quality control

Cameras + a small vision model catch defects faster and more consistently than line operators. Payback: 6-9 months in textiles, auto components, packaging.

2. Demand forecasting

Forecasting at SKU × region × week beats the planner's spreadsheet by 15-25% MAPE in most categories. The win is in inventory turns, not forecast accuracy itself.

3. Predictive maintenance

Vibration + temperature sensors on critical motors. Don't boil the ocean - start with the 10 machines whose downtime hurts most.

4. Energy optimisation

ML on plant SCADA data routinely finds 5-12% energy savings. With Indian power tariffs, that pays back fast.

5. Document automation

GRNs, invoices, lab reports - extracting structured data with vision LLMs has gone from 70% to 95% accuracy in 18 months.

What to skip

Full-blown "digital twins" for SME plants. Beautiful demo, multi-year payback, rarely worth it before the basics.