Journal / Kirill Naumov
BJC Big C’s Thai-language shopping assistant turns a meal idea into a basket. The retail AI challenge is to keep ingredients, quantities, availability and substitutions honest.
By Kirill Naumov ·
A customer does not always begin with a product name. She begins with dinner: “I want to make tom yum soup for four tonight.” A supermarket site that makes her search separately for prawns, lemongrass, galangal and lime leaves has organised the shop around its catalogue, not around her job. A conversational assistant can make that task easier. It can also make a convincing-looking basket with one crucial item missing.
In May 2026, AWS announced BJC Big C’s Shopping Assistant AI Chat in the retailer’s Thai-language app. The supplier describes separate agents for catalogue search, recipe matching and quantity scaling. It reports a pilot estimate of a 5–10% larger basket associated with customers adding complementary items. That figure comes from a vendor announcement; it is not an independently established causal effect, profit gain or guarantee that another grocer will see the same result.
The strongest idea is not the number of agents. It is the change in unit of search: from one SKU to one meal. That makes product discovery more useful, but it also moves responsibility. When the assistant recommends a dinner rather than displaying a list of items, the customer reasonably expects the basket to be coherent.
Test the ordinary failures. An ingredient has several Thai names or package sizes. A fresh item is shown as available but sold out when the order is picked. Four portions require more than one pack, yet the cart adds just one. A customer avoids shellfish, but the proposed substitution is still unsuitable. Each is a small data problem that can become a bad meal.
The assistant needs a source for recipe quantities, a current view of purchasable SKUs, clear units, stock checks and a visible way to amend the basket. It should distinguish “this is commonly used in the recipe” from “this exact item is available for your store and delivery slot”. Those claims have different owners and different update speeds.
Grocery shoppers often accept a replacement when the purpose stays intact. But a substitution can change taste, cooking time, allergen exposure or total price. Before replacing a key ingredient, the assistant should state the difference, show the price and let the customer choose. For a dietary or safety issue, it should not improvise an assurance. A human support path should remain available.
A good test set would cover common meals, regional names, different household sizes, missing stock, budget limits and dietary constraints. Have a qualified person check the resulting cart. Do not measure only whether the answer sounds natural; measure whether a shopper can actually cook the meal with what arrives.
The vendor-reported uplift is interesting, but bigger baskets can reflect extra helpful items or needless extras. A retailer should also look at completed orders, item removals, substitutions, refunds, wasted perishables, support contacts and repeat use. Compare shoppers and periods carefully; promotions and seasonal menus can move the result. If the assistant adds two ingredients that end up in the bin, a larger basket is a poor victory.
Big C’s public example shows why AI in retail works best when it helps complete a real task. “Plan dinner” is more human than “search for prawns”. The engineering and operating discipline is to make every suggested ingredient, quantity and promise as useful as the conversation sounds.
Explore the English AI for Retail workshop → · Discuss grocery retail AI
Primary source: the May 2026 AWS/BJC Big C announcement. Pilot estimates are vendor-reported and not independently validated; this is not our client case.