Conversations

Sarah Miguel Cournane is a data scientist in Hello Retail's R&D department, splitting her time equally between understanding what has already happened in ecommerce data and deciding what needs to happen next. This is a special episode — recorded the week Hello Retail launched Product Agents — where the conversation pulls back the curtain on the data science that makes the product work. That means going into product intelligence, the vector space behind it, and how predictive and generative AI are being combined in a way Sarah describes as the perfect mix.

The technical centrepiece is product intelligence — Hello Retail's approach to personalisation that does not require knowing who a user is. The foundation is a vector space with more than 700 dimensions, where every product occupies a point defined entirely by its behaviour: what it is bought with, what it is not bought with, how it relates to everything else in the catalogue. When a new customer arrives, a single click is enough to draw on all the collective purchasing intelligence around similar products across thousands of stores. Two numbers anchor this: 30% of products drive roughly 70-80% of revenue, and 50% of products active in any catalogue today will not be there in six months. Product intelligence is what stops that constant churn from erasing everything the system has already learned.

Product Agents combines both sides of AI: predictive to choose the right product and the right moment, generative to write the message. The examples Sarah walks through are concrete — price drop emails for products a shopper viewed but never bought, replenishment timing based on package size (a 1kg dog food order signals a different reorder window than a 2kg one), and alternative recommendations when the original product is no longer in stock. Every trigger is decided by product intelligence. Every email is written by generative AI, in the brand's own tone of voice, in the customer's language.

That last point turned out to be more interesting than expected. Friendly in Spanish is not the same as friendly in German. Formal in Swedish carries different connotations than formal in Danish. The solution was building a matrix of tone-by-language combinations and verifying them at scale through automated quality checks — because reading thousands of emails across every language by hand is not possible. The end result, as Sarah frames it: the shop coming to the shopper, rather than the shopper having to go back to the shop.