We do AI consulting work, which means we have a commercial interest in you believing AI is useful. It also means we have spent enough time putting it into real stores to know which parts of it are worth paying for and which parts are a subscription you will cancel in four months.
The useful test is not whether a tool is impressive. It is whether it removes a cost you are currently paying, in a task where a wrong answer is cheap to catch. Everything below sorts on that.
Where it pays for itself today
Support, on the channels Indian customers actually use
The highest-return AI work we do in eCommerce is not customer-facing cleverness. It is answering "where is my order" without a human. In an Indian store that question arrives over WhatsApp more than anywhere else, and it arrives constantly.
What makes it work is the boring part: connecting the assistant to your actual order data and your actual shipping partner, so it answers from the record rather than guessing. An assistant with live order status deflects a large share of contacts by itself. An assistant that only has your FAQ page will annoy people and get escalated anyway.
Build the handover before you launch it. A clean path to a human, with the conversation history attached, is what separates automation people tolerate from automation people complain about publicly.
Product data, which is where most catalogues are broken
Almost every catalogue we inherit has the same defects: descriptions copied from a supplier and duplicated across competitors, attributes missing or inconsistent, titles that follow no pattern, images unlabelled. This is high-volume, low-judgement, structured work, which is exactly what current models are good at.
Generating first-draft descriptions from real attributes, normalising sizes and colours, extracting attributes from existing text, and writing alt text across a few thousand images are all jobs where AI does in an afternoon what would otherwise never get done at all. The output needs review, but reviewing a draft is a different job from writing from nothing.
Shopify now ships its own tools for some of this, Shopify Magic and Sidekick, inside the admin. For a store already on Shopify that is the sensible first thing to try before buying anything.
Search and merchandising on a large catalogue
If a shopper searches your store for a common term and gets nothing, you are losing a customer who had already decided to buy. Default store search matches strings, not meaning, so synonyms, misspellings and Hinglish queries fall through. Semantic search fixes a real, measurable problem here, and your own search logs will tell you within a week whether you have that problem.
This is worth the money on a large catalogue. On a store with forty products, it is a solution looking for a problem.
Translation and regional language reach
Machine translation has become good enough that offering your catalogue in another Indian language is now a realistic project rather than an expensive one. The caveat is that product copy carries brand voice and legal claims, so the money you save on volume should partly be spent on a human review of the templates and the top sellers.
Internal work nobody sees
Drafting supplier emails, summarising a month of support tickets into themes, turning a spreadsheet into a first-pass report, writing the ad brief. These save hours a week for a small team and carry almost no risk, because the output never reaches a customer without a person in between. Unglamorous, and the most reliable return in the whole list.
Where it does not pay, or actively costs you
Publishing AI-written content at volume
The idea that you can generate two hundred blog posts and rank is both the most common pitch and the most likely to backfire. Google's own published guidance on generative AI content is explicit that using these tools to generate many pages without adding value for users may violate its spam policy on scaled content abuse.
That does not make AI-assisted writing off limits. It makes volume without substance off limits, which was already true before the tools existed. Using a model to draft an article an expert then rewrites is a workflow. Publishing the draft is a liability.
A chatbot as your only support channel
Deflection is valuable. Deflection with no exit is a customer service problem dressed as a cost saving. Indian buyers, particularly on higher-value or COD orders, want a human when something has gone wrong, and a bot that loops is a faster route to a bad review than no bot at all.
Personalisation on a store with thin traffic
Recommendation and personalisation engines need behavioural volume to produce anything better than "here are some other products". A store doing modest daily sessions will not feed one. The same budget spent on better photography and a faster product page will do more.
Anything pricing-related without margin logic behind it
Automated repricing tools will happily optimise you into selling at a loss, because they optimise the metric you pointed them at. In a market with COD returns and thin margins, price changes need floors, and those floors are business rules rather than model outputs.
AI images of the product you are selling
Generated lifestyle backgrounds and scene composition are useful. Generating the product itself is misrepresentation, and in a market with high return rates driven by expectation gaps it is a self-inflicted wound. The customer finds out when the box opens.
Three questions before you buy any of it
What does this replace, in hours or rupees?
If you cannot name the task and roughly how much of it there is today, you are buying a capability rather than a saving. Capabilities are fine when they are free to try and cost nothing to abandon. They are not fine on an annual contract.
What happens when it is wrong?
Sort every proposed use by the cost of a wrong answer. A wrong alt text is nothing. A wrong description on a regulated product is a legal problem. A wrong order status to an angry customer is a refund and a review. Start where wrong is cheap and reversible, and earn your way towards the rest.
Where does our customer data go?
Any tool touching orders, addresses or conversations is handling your customers' personal data on your behalf. Ask where it is processed, how long it is retained, whether your data is used to train anything, and what happens if you leave. Get the answers in writing before connecting it to your store, not after.
Frequently asked questions
What is the first AI project a small Indian store should run?
Order-status automation on WhatsApp, connected to real order data. It addresses the single highest-volume support question, it is cheap to reverse if it does not work, and a wrong answer is recoverable. If your catalogue is large and messy, product data cleanup is the other honest candidate.
Will AI-assisted product descriptions hurt our SEO?
Not by being AI-assisted. They will hurt if they are thin, duplicated across your own catalogue or copied from a supplier, which is the state most product descriptions are already in. Use the model to escape duplicate supplier copy rather than to produce a new kind of it, and have a human check the top sellers.
Do we need our own model, or a fine-tuned one?
Almost certainly not. For catalogue work, support and content drafting, general-purpose models with your own data supplied as context cover the ground. Training or fine-tuning is a serious project with a serious maintenance cost, and it is very rarely what a store needs to solve the problem in front of it.
How do we stop an assistant inventing an answer to a customer?
Constrain what it is allowed to answer from. An assistant grounded in your order records, policies and catalogue, with explicit instructions to hand over when it does not know, behaves very differently from one asked to be generally helpful. Then read the transcripts weekly. Nobody gets this right on configuration alone.