Where AI helps a store, and where it quietly adds risk.
The strongest signal that a COD order will come back is written on the shipping label, and your own delivery history for that pincode comes second. Which signals carry weight, which ones are noise or discriminatory, and why a false positive costs more than it looks.
7 min readBackground removal, alt text, auto-tagging and upscaling each break in a different place, and marketplace image rules are mechanical enough to catch every one. What each does reliably, where it fails, and what a product alt text needs to contain.
7 min readA language model fills any gap in a product row with a plausible material, dimension or certification, and each one becomes a return or a compliance notice. How to feed it structured attributes only, keep variants from collapsing into duplicates, and review per product line instead of per SKU.
7 min readA language-model API is another third-party dependency, except this one is non-deterministic, priced per token, and occasionally confidently wrong. Here's the checklist we'd apply before shipping one into a product that already works.
6 min readAI support chatbots get pitched as a blanket solution. In practice they're excellent at a narrow band of tickets and actively harmful outside it. Here's how we scope them so they help instead of adding a layer customers have to fight through.
6 min read