Would you let an AI do all your shopping for you? More precisely: would you let an AI do the eleven or so shopping-ish things you do each month that it could plausibly help with? Those turn out to be different questions, and the gap between them is the difference between a industry-reordering revolution and a nice little feature.
Here is the bull case, as sketched by Peter Kafka at Business Insider. A new class of AI agents — Meta’s Muse, and Instinct, the startup that didn’t exist a few weeks ago and is now valued at $10 billion after a $1 billion Series C reported by TechCrunch — promises to go find and buy the things you want so you never have to think about shopping again. Investors like this because they imagine it hollowing out e-commerce, and specifically Amazon: ruthless agents comb the internet for the best deals with zero loyalty to any retailer, and the toll booth moves from the storefront to the agent. Mark Zuckerberg has said Meta would make its money in that world by taking a small fee from transactions. (Amazon, for its part, has already thrown up barriers against Muse.)
Now the mechanics of that fee. A toll collector needs traffic. Analysts at MoffettNathanson, in a new report arguing that “we do not expect personal agents to be the disruptive forces to ecommerce” that some investors expect, did the wonderfully deflating exercise of counting the traffic. The average American consumer makes about 47 payments a month, per the Atlanta Fed. Start crossing off the ones an agent can’t help with: automated bill payments, gas stations, anything bought in an actual physical store. What remains, by their reckoning, is 11 payments a month that a Muse-like agent might plausibly handle for you.
Eleven is not nothing. But now run the business model through it. Imagine, purely for illustration, that you run one of these agents and you somehow get a 1% fee on every agent-assisted purchase a user makes, and that these users spend maybe $40 a pop on those 11 monthly transactions. That is $4.40 a month per fully converted, enthusiastically delegating user — and real fees would likely have to be smaller, because consumers are the ones being asked to agree to them. To build a $10 billion company on that math, you don’t need some adoption; you need essentially everyone, forever, doing it every month. Which is why the MoffettNathanson view is less “agents fail” than “agents don’t dent Amazon.”
There is a certain comedy in the personal experiment Kafka ran alongside the analysis. He went from “I’ll never give a bot my credit card” to discovering that Stripe’s Link service already had his credit card and that he was fine with Stripe, so he was fine letting Instinct transact through Link. The agent then planned a San Francisco itinerary, suggested Bounce — a luggage-storage service he’d never heard of — and paid for it, $17, without his Visa leaving his pocket. It worked. The payment rail turned out to be the easy part, because it turns out he had already handed a payment company his card years ago. The trust barrier, the one everyone assumes is the hard problem, dissolved in about two weeks.
So if trust isn’t the bottleneck and the transaction plumbing works, the bottleneck is the one nobody can engineer around: there just aren’t that many purchases. People don’t shop that often, and a large share of what they do buy happens at gas pumps and grocery aisles where no agent is invited. The pitch deck says “the agent handles your commercial life”; the Atlanta Fed says your commercial life, the delegable part, is 11 line items a month.
Which is why the MoffettNathanson analysts suspect the endgame is more familiar. Muse, they think, ultimately becomes another surface on which Meta sells ads — a business Meta is quite good at, whether or not it has expressed any interest in that yet. The alternative theory is that the agents get so useful you pay a subscription for them, thereby converting the fantasy of disintermediating Amazon into the reality of charging you $10 a month to save you a few clicks eleven times. Kafka’s own verdict on the products as they exist: while they’re free, he’ll keep playing with them. After that, no idea.
One way to read the whole boom is that investors are paying for the premise that shopping is a burden people are desperate to offload. The evidence so far is that people will happily offload it when it’s free, when Stripe already has their card, and when the stakes are a $17 luggage deposit. The $10 billion question is whether “free and convenient” was ever the business model, or just the demo.