Are the rules for AI shopping agents arriving before the shoppers?

How many orders does it take to justify a rulebook? Uber says it handled around 320 million trips and orders in Halloween week. According to its CFO, four rides and about 500 Uber Eats orders came from third-party AI agents.

That contrast runs through this edition's news. The early evidence suggests that the terms of access and trust for shopping agents are being written ahead of demand for them, not in response to it. Platforms and payment companies are specifying how an agent may buy, while the reported number of agent purchases remains tiny.

Start with the rules. On 21 October, Walmart announced (opens in a new tab) per-agent profiles tied to a customer account. Walmart says customers can require approval for every purchase or let an agent spend within a limit they set, with $50, $100 or $200 given as examples, and that agents that do not follow its terms will be blocked from its site. Digital Commerce 360 (opens in a new tab) also reports the preset spending limits. The Agent Gateway is a limited beta, and Walmart's Agent Profiles had yet to be confirmed as live at the time of writing.

American Express went down a similar path on 13 October, with an Intent Intelligence framework that records what a customer asked an agent to buy and uses that record to obtain a scoped payment credential in place of the card number. Amex confirmed to Digital Commerce 360 (opens in a new tab) that there are no live integrations yet, and that its Agent Purchase Protection is not available because it depends on those integrations. Both designs treat explicit, recorded, delegated authority as the basis of trust. Neither has shown adoption.

Now the demand. Uber's CEO called traction "very, very early and very limited" (opens in a new tab), and the CFO's figures, as transcribed (opens in a new tab), put agent volume far below a thousandth of one percent of the week's total. DoorDash's CEO said the company is seeing maybe several hundred orders in total (opens in a new tab) from third-party agents, and that users still mostly order in DoorDash's own app.

The counter-evidence deserves weight. Uber's CEO still expects agents to be a net positive for demand, and small early volumes are normal for any new channel. Rules written early could be sensible preparation, not premature effort. Friction may also be part of the story: Walmart's beta is limited, and its approval-by-default design could hold orders down. Low usage is not clear proof that shoppers don't want agents. Where a native app is already quick, a third-party agent may simply add little. An NMI survey of 1,000 US adults in August, cited by Digital Commerce 360, found 58% worried an agent could buy the wrong item, size, color or brand, which suggests trust is a live concern.

The numbers also carry caveats. Both are company-reported, and the definition of an "agent order" is not given. DoorDash's figure is cumulative over an unspecified period, while Uber's covers one week, so the two cannot be compared. Uber's count also covers only third-party agents.

One reading is that this is a bet on option value. Opening a channel to agents and setting terms early looks cheap, and the platform that writes the terms holds the leverage over access. Whether that is foresight or over-engineering depends on volume that has not yet appeared.

The next few weeks should show more. Watch whether Walmart's profiles go live and which agents sign up, and whether any of the three companies publishes fresh agent order counts covering a longer stretch or a bigger base once holiday trading is in the figures. Google's expansion of UCP-powered checkout into Canada is another signal. A tenfold jump would suggest the rules are following demand after all.