Nordstrom, AI and the right to change your mind

A customer who bought black shoes last month may want black shoes again. Or they may be on a one-person campaign to avoid black shoes forever. Both are normal. That is why fashion recommendations are not merely a matching problem. They are a conversation about an occasion, a mood, a budget and the occasionally brilliant decision to buy something unlike anything in the purchase history.

Nordstrom announced investments in BevyUp and MessageYes in 2018. Its release described BevyUp as a clienteling tool for extending a sales associate’s customer relationship across channels, and described MessageYes as using AI and customer feedback in mobile messaging. The company’s 2024 annual report also describes its connected physical and digital model and services including styling, pickup, returns and alterations. These are company descriptions of its platform and investments. They are not proof that a specific recommendation will be welcome, accurate or commercially successful.

The valuable design question is therefore not ‘what else will this person buy?’ It is ‘what choice would be useful to consider now, and why?’ An answer that exposes its reasoning gives a stylist something to work with. An answer that merely says ‘people like you bought this’ turns a customer into a statistical shadow of their past self.

A recommendation needs an editor

Give the adviser a compact recommendation card: the occasion it assumes, the evidence it used, the constraints it did not know and three genuinely different options. One may be a safe repeat. One may be a sensible stretch. One should be explicitly unlike the customer’s history but appropriate to the stated need. The adviser can remove an item, replace it or ask a clarifying question before anything reaches the customer.

That editability matters for both trust and learning. If a stylist rejects a recommendation because it ignores climate, fit, dress code or a customer’s stated preference, the system should record the reason as feedback on the suggestion—not as a reason to make the stylist comply next time. Human taste is not dirty data to be cleaned out of the funnel.

Marketing teams should set boundaries before measuring click-through. Do not infer sensitive attributes. Do not pretend a recommendation was hand-picked if it was automated. Do not let an old return or a gift purchase become a permanent identity. Retain only the data needed for the service, give people meaningful controls and make it easy to start a conversation without a profile at all.

Test for useful surprise, not automated pressure

A small pilot could use one category and volunteer advisers. Compare an ordinary product list with an assisted list that includes reasons, uncertainty and an edit step. Review whether advisers accept, change or reject suggestions; whether customers ask for alternatives; and whether a recommendation creates follow-up work through poor fit or unavailable stock. Do not call every accepted recommendation a success. A customer may accept it because the queue is long.

The qualitative review is often the useful part. Read the cases in which a stylist says, ‘This is technically relevant, but it is not right for this person today.’ Those are not edge cases. They describe the exact judgement that makes clienteling feel personal instead of invasive.

The best retail recommendation engine does not try to win every choice. It helps the human in the conversation make a better one, gives the shopper room to disagree and remembers that a preference can be a sentence, not a life sentence.