Retail has no shortage of impressive AI demos. Ask a chatbot to recommend a jacket. Show a camera spotting an empty shelf. Watch an agent generate a replenishment task. Everybody applauds, somebody says “game changer,” and then Monday morning arrives.

Now try making that experience work across 800 stores, with different layouts, inconsistent inventory data, stretched store teams, and customers who have absolutely no interest in participating in your technology experiment.

That’s where the conversation gets interesting.

On October 28, Rethink Retail’s AiR Virtual Summit: The Agentic Store brings together retail and technology leaders to examine what AI can actually do inside physical stores. I’m listed as a participant in a breakout discussion titled “What Would It Take to Scale AI Across Your Stores?” alongside Matt Redwood of Diebold Nixdorf.

That question is more consequential than the latest AI product announcement. Because the hard part isn’t imagining the store of the future. It’s operating one.

A clever pilot isn’t a retail operating model

In a pilot, you can give an AI project extra attention, handpick a location, clean up the data, and put your most enthusiastic associates on the floor. That’s useful for learning. It is not proof that the same system will work in every location.

Scale introduces the things a demo can conveniently ignore: uneven connectivity, legacy systems, regional assortment differences, training needs, and the wonderfully unpredictable behavior of actual humans.

My rule of thumb? Before asking whether an AI tool can perform a task, ask who owns the result when it gets the task wrong. If nobody can answer that, you haven’t designed a scalable operation. You’ve designed an exciting meeting.

The associate is not an implementation detail

One of the summit’s central themes is the AI-powered frontline. That’s the right emphasis. Retail associates don’t need another dashboard that makes them spend more time looking at a screen while a customer waits for help.

They need tools that eliminate tedious work, surface reliable information, and make the next decision easier. A useful associate copilot might help locate an item, answer a product question, or flag an inventory discrepancy. But the value isn’t the AI response itself. It’s the better customer experience or the time returned to the employee.

That also means measuring adoption honestly. A tool installed in 500 stores isn’t necessarily a tool being used effectively in 500 stores.

Personalization needs a trust budget

Another summit discussion, “The Store That Knows You,” explores the line between helpful and intrusive. It’s a line retailers should take seriously.

Just because a store can infer something about a shopper doesn’t mean it should immediately act on that inference. Personalization can be valuable when it reduces friction. It becomes creepy when it surprises customers with information they didn’t expect the retailer to have.

Good agentic retail will need restraint, clear customer benefits, and an understandable way to correct mistakes. Trust is much harder to rebuild than a recommendation engine.

Measure the work, not the magic

I’d rather see a retailer show a modest reduction in out-of-stocks, fewer minutes spent on repetitive tasks, or more accurate product availability than announce a sweeping AI transformation with no operational evidence.

The agentic store won’t be won by whoever can produce the flashiest demo. It’ll be won by whoever can make the technology reliable, useful, and boring enough to work on a busy Saturday.

That’s a less glamorous headline. It’s also where the money is.

Explore the October 28 AiR Virtual Summit agenda and speakers.