Walmart uses large language models to create or improve catalog data

The problem

A retail catalog with a fast-growing marketplace needs accurate attributes and descriptions for every item. Search, inventory placement and order delivery all depend on that data. Doing it by hand across a catalog of this size does not scale.

What they built

Walmart used multiple large language models to create or improve catalog data, including attributes and characteristics of items. On the call, CEO Doug McMillon said Walmart uses data and models from others and builds its own.

How they did it

The sources do not describe the pipeline, the models or the checks. They say only that the models were used on the catalog and that the better data helped match items to what customers intend to buy, per John Furner, CEO of Walmart U.S.

Results

McMillon said the models were used to create or improve over 850 million pieces of catalog data. He said that without generative AI the work would have needed nearly 100 times the current head count in the same time. In the Q&A, Furner said using generative AI is about 100 times more productive than having people work through each product display page. Both comparisons are management estimates.