The person with the strongest track record in modern retail isn’t reading your AI shopping think pieces, and that should bother the people writing them.
Ron Johnson, 66, joined Apple in 2000 to build its retail business. He has heard the physical-retail-is-finished prediction before, more than once, and he’s still here talking about stores. Right now his attention is on his retail insights and a new book, not on the current news cycle coming out of Silicon Valley. That gap between what he’s paying attention to and what the industry insists is urgent is the most interesting thing in this story.
I build bots for a living. I write the retrieval layers, the tool-calling loops, the fallback logic for when a model confidently returns a product SKU that doesn’t exist. So when I say the AI shopping bet looks thin, it’s not skepticism about the technology. It’s familiarity with it.
Demos are not behavior
Every agentic shopping demo follows the same script. A user types a vague request in natural language. The agent searches, compares, picks, and checks out. Thirty seconds, no friction, applause.
Now try building it. The parts that break aren’t the model calls:
- Product catalogs that disagree with themselves across feeds, APIs, and the actual warehouse
- Pricing and availability that change faster than your cache
- Payment and identity flows that legitimately do not want an automated client touching them
- Returns, which are a human negotiation wearing an API costume
- Any purchase where the buyer wants to hold the thing first
None of that is a model problem. It’s a systems problem, and shipping an agent into it means absorbing every inconsistency in the chain. The demo hides the chain. Production is the chain.
The trust budget nobody models
The thing I’ve learned building bots that spend money or take consequential actions: users have a trust budget, and it’s smaller than you’d like. One wrong order, one duplicate charge, one item shipped to an old address, and the user goes back to tapping through the app themselves. Not because the app is better. Because the app is legible. They can see what’s happening and stop it.
A physical store is the most legible interface ever built for commerce. You walk in, you see everything, you touch the product, a person answers your question, you leave with the thing. Zero latency, zero hallucination, no confirmation dialog. What Johnson built at Apple was essentially a high-trust interface with a very short path from question to answer. That’s a hard baseline to beat with a text box.
What the Apple pattern actually teaches
Apple’s retail bet in 2000 was contrarian in the same direction Johnson still seems to lean: bet on the experience people actually have, not the one the industry says is coming. And Apple’s leadership choices suggest that instinct hasn’t been abandoned. Tim Cook, never positioned as the AI guy, announced in April 2026 that he would step down. His successor is John Ternus, a 25-year Apple veteran who runs hardware engineering. That’s not the pick you make if you believe the future is a chat interface that replaces everything.
For those of us building agents, the lesson isn’t “don’t build agents.” It’s about where the agent belongs in the stack.
Where shopping bots are genuinely good
The wins I’ve seen hold up in production are unglamorous and narrow:
- Research, not checkout. Agents are strong at collapsing forty tabs into a short comparison. Let the human press buy.
- Reorders and known items. Low ambiguity, low regret, easy rollback. Ideal agent territory.
- Support triage. Reading order history, matching a policy, drafting a reply for a human to send.
- Internal tooling. Inventory questions, merchandising analysis, pulling the numbers a store manager needs before opening.
Notice that none of those replace a store. They reduce the busywork surrounding one. That’s a smaller story than the one being pitched, and it’s the one that survives contact with real users.
Build for the boring middle
My practical take, as someone who has to make these things work on a Tuesday: design your shopping agent as a well-instrumented assistant with hard limits. Keep a human in the loop for anything irreversible. Log every tool call. Make the agent’s state visible to the user in plain language. Treat catalog data as the actual product and the model as a thin layer on top, because that’s the ratio that determines whether your bot is useful or embarrassing.
Johnson’s disinterest in the daily hype cycle reads less like stubbornness and more like a filter. He’s spent a career on the part of retail that doesn’t change every eighteen months. If your agent architecture can’t survive that same test, it isn’t a product yet. It’s a demo with good lighting.
🕒 Published: