AI in Retail & E-commerceSearch · Recommendations · Assistant · Forecasting

Which retail AI features ship in a Miracuves base, and which are custom?

One is confirmed as shipping today. The Amazon, Flipkart and Etsy bases, one product at $2,799, include an AI module that drafts product descriptions, SEO titles and meta descriptions, tag suggestions and image alt text through OpenAI. Vendors see it as AI-assisted descriptions in the seller panel; the operator sets token quotas per feature and sees tokens and cost logged for every call. It runs on your own OpenAI key, so it is integration-required rather than included.

The Alibaba page also lists AI product recommendations, a chatbot, an AI negotiation bot and trend analytics; the written scope confirms what runs in the version you receive. Everything else on this page is custom work built on top: intent and visual search, recommendations, a shopping assistant, forecasting, return and review insights and support automation. Each is scoped in writing and takes 2-8 weeks. The base search is MySQL full-text; a dedicated search engine is a quoted add-on.

Search, recommendations or a shopping assistant: where should a store start?

Start where shoppers already fail. If search logs show many queries returning nothing, or queries that describe a need rather than a product name, intent search pays back first, because every visitor who searches touches it. If shoppers find products but leave with one item, recommendations come next. An assistant fits catalogs where buyers ask before they buy: sizing, compatibility, materials, delivery.

AI is the wrong first step when the catalog itself is the problem. Missing attributes, duplicate listings and inconsistent categories make every model worse, so catalog enrichment comes before search. With little order history, recommendations start from product similarity and simple rules, and learn from behavior once orders build up.

What does retail AI cost to run each month?

The build is quoted once; running costs are separate and belong to you. They come from three places. Model usage is billed per call by the provider you choose, so it grows with traffic and with how much text each call carries. A search index and vector store need hosting. Forecasting and ranking jobs need scheduled compute.

Three habits keep the bill predictable: cache answers to repeated questions, send the model only the fields it needs, and set token quotas per feature. The base AI module already enforces quotas and logs the cost of every call, and the custom layer keeps the same log. A self-hosted open model trades per-call fees for server cost, which suits high-volume stores. Running costs are estimated from your traffic before you approve the scope.

What data should a retailer have ready before AI work starts?

The 6 working days of a ready-made base are Miracuves build time; AI work runs 2-8 weeks and depends on your data, so start gathering it early. Miracuves is based in Mumbai and works remotely, so read-only access is all that is needed:

• A product export with titles, attributes, categories and image links
• Search logs, including the queries that returned no results
• Order and return history, with return reasons where you record them
• Product reviews and support tickets, if the assistant or review insights are in scope
• Stock by location and supplier lead times, for forecasting
• An account with your chosen model provider, such as OpenAI

How do you stop an AI shopping assistant from making things up?

By never letting the model be the source of a fact. The assistant retrieves products, prices, stock and policies from your database at the moment it answers, and the model only phrases the reply. If nothing in the catalog matches, it says the store does not carry that item instead of inventing one. Prices and delivery dates are inserted from live data, never generated.

Before launch, the assistant is scored on an evaluation set of real shopper questions, including trick questions about products you do not stock. Actions such as changing an order or starting a return run through your existing APIs with the shopper signed in, and anything unusual hands off to a person. The same retrieval approach runs through our RAG development work.

How does demand and inventory forecasting work for an online store?

A forecasting model learns each product's pattern from sales history, then adjusts for what it can see coming: seasonality, scheduled promotions, price changes and, where useful, holidays or weather. The output is a forecast per product and location with a range, not a single number, because the range tells a buyer how much safety stock to hold.

In the operator console that becomes reorder suggestions with supplier lead times applied, waiting for a person to approve. It needs at least one full seasonal cycle of clean sales data to be worth trusting; new products borrow patterns from similar items. On the Alibaba base the same approach drives reorder prompts for trade buyers. It is custom work, built in 2-8 weeks.

How should you evaluate a retail AI vendor?

Ask questions that separate a demo from a system. Will the model keys and usage sit in your accounts? Do you receive the prompts, the evaluation set and the code, or only access to a hosted tool? How will quality be measured, against which baseline, and who agrees the metric before build? What happens when the model provider retires a model or has an outage?

Then ask for the plain evidence: a cost log per AI call, a written list of what the product ships today versus what will be built, and a test on your own catalog rather than a curated dataset. Miracuves answers each of these in the written scope, and AI consulting can run the assessment if you are still choosing between approaches.