A small number from a grocery slotting project caught me this week and I want to write it down before I forget it.
A grocery distributor, one of the big ones, ran a generative AI model across twelve months of order lines and its own warehouse slotting. The model came back with an odd suggestion: move the ice cream to be right next to the frozen pizza.
Why? Because it had noticed that 78 percent of pizza orders also contain ice cream.
Seventy. Eight. Percent.
The distributor actually did it. Pick time dropped by 31 percent. Everyone went home. The 78 percent of us who apparently cannot bake a pizza without reaching for a tub of something cold went home too. Which, I want to emphasise, is a completely normal thing to do.
This isn't a one-off
Another food distributor used AI dynamic slotting to fit 2,500 additional SKUs into the same facility, avoiding a 2 million dollar expansion project. Amazon's predictive inventory system dropped stockouts by 35 percent and cut handling costs by 10 to 15 percent. Walmart uses AI to forecast labor needs and saved 15 percent on labor cost.
The bigger one: generative AI is producing warehouse layouts in four hours that used to take a human team three weeks. The best AI-designed layouts benchmark as 40 percent more efficient than the best human-designed ones. Walmart is piloting this in its micro-fulfillment network.
The category is real now
Gartner's April 2026 report puts agentic-AI-enabled SCM software at under 2 billion dollars in 2025, and forecasts 53 billion by 2030. Adoption moves from 5 percent of SCM software users today to 60 percent by 2030. You don't argue with that shape of curve.
How I feel about it, honestly
I've spent a lot of years telling people that warehouse slotting is craft, that it takes experience, that you need to know the operation, that the human in the room is the best pattern-matcher. Most of that is probably still true.
But the models are reading the receipts now. They are making better calls than our slotting engineers, and the gap is growing monthly. The sensible move is to put the AI in the loop, let it do the routine analysis, and save the human experts for the edge cases the model doesn't know about yet. Things like "the freezer door thermal profile" that the model isn't thinking about but the shift supervisor is.
And, you know, the pizza thing. The model knows about the pizza thing. I still don't know how to feel about that.