Ron Johnson built the Apple Store. This week he said he does not believe people will let an AI agent buy their next laptop.
"Honestly, nobody's going to do that," he told TechCrunch, on the idea of an agent choosing a machine in the $1,000 to $2,000 range without the buyer setting foot in a store.
Google is building the Universal Commerce Protocol. OpenAI is wiring checkout into chat. Billions are moving on the opposite assumption.
One of these positions is wrong, and I do not think it is a clean split.
Two kinds of purchase, one kind of product roadmap
Agentic commerce is being designed as if all buying were the same activity with different price tags on it.
It is not. There is buying where you already know what you want and the process is friction to be removed. Printer ink, dog food, the same running shoe in the same size for the fourth time.
Then there is buying where the process is the product. You are not shopping for a laptop, you are shopping for the feeling of having chosen correctly, and the specs are a proxy for that.
Automation is spectacular at the first kind and structurally poor at the second, because the second is not really an information problem. It is a confidence problem, and a confident answer from a system you cannot interrogate does not resolve it.
I watch this in client accounts constantly. Cart abandonment on a replenishment SKU is a logistics failure. Cart abandonment on a considered purchase is usually someone leaving to get reassurance somewhere else, and then coming back, or not.
Johnson's actual claim is about staff, not software
Read the quote carefully and the argument is not anti-AI. It is about where the reassurance comes from.
"The secret sauce for Apple has always been its people, the people in the store," Johnson said. He designed the stores without sales commission specifically so staff would stop closing and start helping.
That is an engineering decision about incentives, and it is the part most retailers never copied while copying the wood tables and the glass.
His prediction is that agents make people better informed before they arrive, not that agents replace arriving. The research moves to the model, the decision stays where it always was.
If that is right, the agent is not eating the funnel. It is eating the middle of the funnel, which happens to be the part most marketing budgets are spent on.
What gets repriced when research moves
Here is the operational read for anyone selling a considered product.
The comparison content, the buying guides, the spec roundups, the review aggregation pages: that layer is being consumed by models rather than by humans. It still needs to exist, and it will produce far fewer visits.
Which means its value stops being traffic and becomes influence on an answer you never see. That is a hard sell in a quarterly review, and it is where most brands will quietly defund exactly the wrong assets.
Meanwhile the two ends of the journey become more valuable, not less. The top, where the category question first gets asked, and the bottom, where the reassurance happens.
This is the same timing problem I described when looking at how the decision gets made weeks before the transaction. An agent doing the research does not remove that gap, it widens it.
I wrote about your website becoming the source rather than the destination a few editions ago. Johnson is describing the physical version of the same shift.
For most businesses reading this, the reassurance layer is not a store. It is a demo, a sales call, a WhatsApp thread, a real photo from a real customer, a phone number that a human answers within two rings.
Agents default to defaults
There is a second-order effect nobody in the restock column has priced yet.
When a human buys dog food for the fourth time, they are running a habit. When an agent buys it, it is running a selection, and it will make that selection again next month against whatever criteria it holds.
That turns a loyal customer into a recurring competitive event. Your position is no longer protected by the mild inconvenience of switching, because switching now costs the buyer nothing at all.
The brand that becomes the model's default answer in a category gets compounding volume. Everyone else gets to be the alternative that the agent mentions in a clause.
Which is why the restock column is not the safe half of the catalogue. It is the half where the relationship gets renegotiated silently, at machine speed, by a system with no memory of the fact that you sponsored their favourite podcast.
The test that actually settles it
Stop arguing about whether agentic commerce works in general. Run the split on your own catalogue.
Take your top 20 products by revenue and sort them into two columns: purchases a returning customer would let a system make on their behalf, and purchases where they would want to look at the thing, or at least talk to someone about it.
For the first column, your job is machine-readable truth. Clean structured data, accurate stock, unambiguous pricing, no dark patterns in checkout, because an agent will not tolerate what a distracted human tolerates.
For the second column, your job is the opposite. It is proof, specificity and access to a person, and none of that gets better by being compressed into a product feed.
Most catalogues split roughly down the middle, and most marketing teams currently run one strategy across both.
The interesting part of Johnson's position is not that he doubts the technology. It is that the man who built the most profitable retail square footage in the world thinks the bottleneck was never information.
Agents will win the restock. The $2,000 decision still gets made by somebody who wants to be told they are not about to make a mistake.