Most AI retail decks describe what the technology will do. A FashionUnited essay supplies the missing slide: three retail problems no model touches. The list is culture, judgment, and accountability, and each one is what gets the rollout into trouble after six months. The deck’s silence on these is what makes the omission expensive.

Start with culture. “AI works best when the foundations are already strong.” Read backwards, the sentence tells the retailer with weak foundations to wait, which is precisely the retailer the vendor is most eager to sell to. A bad process, automated, is the same bad process moving faster, with fewer witnesses left in the room. Strategic thinking cannot be purchased; pretending otherwise is what a confident vendor calls transformation.

Now judgment. The model has no capacity for genuine empathy, very little awareness of when it might be wrong and it rarely admits uncertainty. That describes the moment the brand cares about most: customer recovery. A complaint at midnight requires a decision about which apology to offer, which concession to extend, and which loyal customer to keep. The model cannot make that decision, and the brand’s reputation hangs on it. The retailer who bought judgment as a feature discovers, by month six, that it was never in the box.

A HubSpot survey cited in the essay found that 91% of unhappy customers leave without complaining, which explains why the judgment cost is hidden. It does not show up in the ticket queue, which is the only place the dashboard looks. The retailer cannot automate a complaint it never receives. The departure is the entire signal, and it leaves no trace in the chatbot’s logs.

Then accountability. “The technology can support decision making, but it cannot take responsibility for it.” Vendors do not put this on the slide — their indemnity clauses, characteristically, are more forthcoming on the point than their sales teams. Accountability stays with the executive who signed the procurement order, not with the vendor on the other end. The buyer’s general counsel reads this six months later, in the incident report, by which time the vendor’s case manager has already explained that the model behaved as designed.

When something goes wrong, the question is never who built the model; it is who signs the apology.

The guest contributor who wrote this is not an AI critic, and that is the point. The objection is coming from the people who run the floors. Their list is unsentimental: three problems, none of them solvable by a model, all of them now expensed to the budget that was supposed to cover them.