In most retail businesses, the vast majority of questions that customers ask are questions that were asked in the past by previous customers.
What I propose is an AI, with voice analysis software, that eavesdrops on customer / employee interactions, and builds a database of questions and their typical
answers.
After a suitable learning period, an interface would be set up which would allow customers to ask their questions to the AI, which would then display a list of the most common employee responses.
A typical database entry would be something like, "What aisle are lightbulbs in?" -> { "Lightbulbs are in aisle ten." , "Aisle ten on the left" , "One aisle to your left, on the left hand side" }.
In it's simplest form, the AI wouldn't need to actually understand the questions or the answers, merely store what it's heard in the past, and regurgitate responses.
A more advanced form would have a map of the aisles of the store, with a pre-programmed knowledge of what front, back, left, right, and distances (like what "halfway down the aisle"). While learning, it would recognize directions to products in the store, and instead of learning the phrases used to give those directions, it would learn the products' locations. Once product locations are known, creating answer phrases that give customers directions to products should be relatively easy.
Another advanced option would be integration with inventory tracking, and a database of what products are scheduled to arrive when. Thus, if a customer asks for directions to a product which is out of stock, the AI could tell him or her, "Sorry, we're out of stock of Windex; we should have more in by next tuesday. The usual location for Windex is in aisle eight on the left."
Mics for eavesdropping would be placed unobtrusively throughout the store, customer interaction with the AI might be through something like an Asimo.