Ben Thompson on why Amazon will win the agentic commerce war — and why personal agents misunderstand what consumers actually want

Amazon wins the agent commerce war — and consumers don't want what personal agents are selling

Ben Thompson's central argument is that agentic commerce will consolidate around Amazon, not because Amazon is particularly consumer-friendly, but because logistics trust is the actual moat that no one else can replicate cheaply.

Why Amazon wins by default

Thompson's revealed preference test is simple: even when ChatGPT or Claude surfaces product links, Amazon never appears because it doesn't permit them. He still goes to Amazon and types it in. The friction of not having a direct link is real but not fatal, because Amazon owns something more valuable than the click — certainty. Customers know the package will arrive, know where to track it, and know where to process the return. Every alternative introduces what he calls "massive amounts of real world friction just because you're too lazy to find a click over to the other browser."

Shopify is structurally disqualified from closing that gap. It tried logistics, the stock chart shows where that went, and it retreated to its actual model: asset-light, merchant-first, present across every platform. Shop Pay works precisely because it eliminates checkout friction without touching fulfillment. That's a different business than rivaling Amazon's last-mile guarantee.

“We're not helping you be more productive. We're helping you get rid of annoying crap in your life. It needs to be framed more as a pain relief... At the end of the day, my personal revealed preference is I would almost always want it from Amazon because I trust it's going to get there. Shopify tried to do the logistics thing — go look at the stock chart.”

The Walmart exception

Walmart is the only credible rival on logistics, and Thompson says most people in tech don't realize how good its e-commerce and in-store app already are. When OpenAI launched its checkout integration, Walmart jumped in and Amazon refused. Thompson read that as a real competitive move. Six to nine months later, Walmart pulled back — abandoned carts spiked, basket sizes shrank, and conversion suffered when the experience moved off Walmart's own site. The lesson: even Walmart needs to own the session to convert it, which means the agent-as-checkout-layer thesis ran into the same wall for them that it runs into everywhere else.

Consumers don't want agents to shop for them

The sharper argument Thompson makes is that the personal-agent category is solving a problem that mostly doesn't exist. People enjoy shopping. Fast fashion, endless product research, the ritual of building a cart — these are entertainment, not inefficiencies to be eliminated. The package arriving is part of the experience, not an afterthought.

He draws the same line on flight booking: he has had a personal assistant for a decade and that assistant has never come close to booking his flights. He knows which routes delay, which seats he wants, which trade-offs to make. Delegating that to an agent is not a productivity gain; it's a loss of something he values.

The framing he thinks actually works for personal agents is pain relief, not productivity. Not "I'll make you more efficient" but "you're leaving money on the table with unprocessed returns, and I can fix that while you do whatever you actually want to do." The subscription-cancellation use case is the closest thing to a genuine consumer pain point, though he notes it creates downstream support headaches for subscription businesses — including his own — when agents cancel things users didn't intend to cancel, or mass-unsubscribe from transactional emails that are legally distinct from marketing lists.

The physical world gap

Running through the whole conversation is a problem that neither agents nor e-commerce platforms have solved: the embodied task. Thompson's assistant doesn't just manage inbound communications — he processes packages, opens boxes, catalogs where things are stored, handles returns. The home inventory app Thompson built works because a human physically labeled everything and logged it; the bot is the interface, not the labor.

His own internal tool, Geckobot, captures a version of this. He leaves notes and requests asynchronously over weekends; his assistant arrives Monday to a structured briefing instead of a weekend's worth of intrusive texts. The assistant's reaction, described as the greatest week of work of his life, illustrates how narrow but real those software-shaped holes are — and how far agentic software still sits from filling them without a human in the loop.

The actual end state, in Thompson's read, is a humanoid robot that maintains a fleet of single-purpose physical robots: the pool cleaner, the window washer, the mail slicer. Until that exists, the shortcut is paying a person. The agent layer is useful at the margins but not a replacement for someone who can go park the car, run into a store, or put things away in a place you'll be able to find them later.