Walmart: the customer came for groceries, not a treasure hunt

A customer arrives for five things. Four are easy. The app confidently sends them towards the fifth, which turns out to be somewhere entirely different. At this point, the conversation about personalised customer experience becomes rather academic. These small expeditions deserve a place at the beginning of any serious discussion of AI in retail.

Walmart already provides item finding and store maps. That is the documented feature; what follows is my independent proposal, not a Walmart engagement or a claimed implementation. The commercial question is simple: can we reduce the number of customers who need to ask twice for the same help?

A database balance is not a shelf

The map only helps if its product location is connected to the quality of the availability signal. Stock on the books, a recent shelf check and stock awaiting replenishment are different things. Sometimes the honest answer is that an associate needs to check. That answer may be commercially more useful than a beautifully presented guess.

An unsuccessful search should leave behind a specific task for an appropriate team. Send the item, location and problem, rather than announcing a general emergency. An associate closes the task with a reason: sold out, moved, incorrect label. Their work then improves the next customer's answer instead of disappearing into a conversation.

Then comes a question easily missed on a process diagram: whose existing work are we interrupting? Pulling someone away from checkout may improve shelf assistance while making ten other people wait. Task suggestions therefore need staffing and workload context. A shift leader should see that trade-off and retain the decision. People are not spare processing capacity waiting for an algorithm to discover them.

Test the shopping outcome

One department with recurring availability questions is enough to find out whether this is worth doing. Establish a baseline, then compare similar shifts with and without the assistance. Account for promotions and deliveries. An extra truckload of stock can produce a more impressive improvement than a model, and it would be useful to know which one did the work.

To judge whether the assistance earns its place, measure time to actually finding the item, repeat requests for the same search and purchases lost because the item was unavailable. Checkout waiting time becomes a guardrail. A department does not win if the rest of the store pays for its improvement. Map clicks are not a substitute for completed shopping.

This trial does not need personal movement histories. Anonymous events and an optional request for help are enough; staff access should be limited to their task. A human route remains visible when the map is wrong. The best outcome is wonderfully uneventful: the customer finishes the list, unaware that finding detergent involved any intelligence at all.