Anu Narayanan, president of women’s and home at Anthropologie, said at Shoptalk Spring that her company “marries data and intuition” in merchandising. The phrase is common enough to pass unnoticed. It shouldn’t.

“Data doesn’t tell you everything,” Narayanan told Modern Retail. “It won’t tell you what’s next, and it won’t tell you where there are strategic opportunities.”

That is not a caveat about data’s limits. It is a description of who decides.

Most companies that invoke “data and intuition” have organized them hierarchically: data moves first, intuition is available for appeal. The human layer fires when confidence is low or a governance requirement demands review. The architecture, underneath, still belongs to the model.

Narayanan described a different structure. At Anthropologie, data arrives at a merchant who reads it alongside social signals, customer anecdotes, and store observations — and then makes a judgment. Not an approval. A judgment.

The examples do the work. When the brand spotted TikTok interest in ’90s silhouettes — “little mini oval sunglasses and loafers and khaki pants” — and cross-referenced it against sales data and cultural signals ahead of the FX series Love Story’s premiere, they launched a dedicated section within a week. That timeline belongs to a merchant who has already decided. It is not a committee-approval cadence.

The footwear story is more instructive. High return rates had kept footwear largely off store shelves, with only eight locations carrying it. Data showed those eight locations attracting more new customers, buying across more categories, at higher lifetime value. Anthropologie identified Charlotte as a test market through a data partner, ran the expansion, and footwear is now in roughly 200 physical locations. Anthropologie Group net sales grew 8.7% in fiscal 2025, per URBN’s earnings release.

“Data showed” is the wrong reading of that sequence. Someone looked at a return-rate problem and decided the new-customer signal outweighed the friction signal. The data created the conditions; the merchant assigned the weights. Those are not the same operation.

There is also a third example, quieter than the others. Narayanan noted that Anthropologie risks over-serving its existing customer if it relies too heavily on existing purchase data. “If we are trying to open the aperture as we are a multi-generational retailer,” she said, “how are we ensuring that we have something for every generation?” That question cannot be answered by an optimisation model. It requires someone to decide that the current customer is not the only customer worth having — and then act on that before the data catches up.

Most AI transformation roadmaps are built the other way. The model optimises; the human approves or overrides. That architecture produces decisions. What it cannot produce is a decision that reframes the question — one that says the return problem is a distribution opportunity, or that the core customer is a constraint, not a mandate.

An approval architecture optimises within the space the model already understands. A judgment-first architecture can decide the space itself needs to change.

Narayanan framed none of this as a critique of AI. She described how Anthropologie operates. The industry will read it as a story about balance. That is not what she said.

She said the merchant is at the node. Most AI roadmaps have removed the node, redistributed its functions across dashboards and confidence intervals, and called the resulting vacancy efficiency.