Tony Xu was asked, on his company’s third-quarter earnings call, how much of DoorDash’s business now arrives through artificial-intelligence agents built by other companies. His answer was “maybe several hundred orders in total.”
DoorDash handled 993 million orders in the same quarter. The distance between those two figures is the most concrete public measurement anyone has offered of agentic commerce, a category that has absorbed billions of dollars of venture money and a large share of the technology industry’s confidence about how people will shop.
Xu, the co-founder and chief executive, framed the number with care. “You never want to fight gravity,” he said of the technology. “But at the same time, you don’t want to be delusional.” He named the outside agents that have connected to the platform, among them Muse, Instinct and Fo, and described the area as “very dynamic.” Orders still arrive overwhelmingly through DoorDash’s own app.
The company has not been passive about the shift. It shipped Ask DoorDash in June, a chatbot that takes orders from prompts and photographs; a command-line ordering tool in July, aimed at developers; a text-to-order agent in September; and a connector built on the Model Context Protocol so that other chatbots, including corporate assistants, can place orders on a user’s behalf. The internal work is real. What it has not produced is volume.
That gap deserves attention because the investment case for agentic commerce rests on a specific claim: that the chat window becomes the front door to retail, and that the apps behind it become plumbing. Amazon has responded by blocking outside agents from its storefront. Walmart has done the opposite, opening its catalog and explicitly courting them. DoorDash sits in the middle, dependent on merchants it does not control and customers it reaches through an app it does control.
Food delivery is a hard place to prove the thesis. A restaurant order is a dense packet of preferences: item modifiers, substitutions, allergies, tip amounts, delivery instructions, apartment codes, and a refund policy when the fries arrive cold. Each of those is a place where an agent’s small error turns into a customer-service cost that the platform absorbs. A chatbot that books a hotel room can survive a wrong date. One that sends shellfish to a customer with an allergy cannot.
The economics are equally awkward. DoorDash collected about $4.6 billion of revenue against $33.9 billion of marketplace gross order value in the quarter, a take rate near 14%, or roughly $4.60 on the average order of about $34. An outside agent that originates the order will want a share of that, and the merchant will resist paying more, which leaves the platform’s margin as the funding source. Xu said the company is exploring advertising placements inside third-party agents, a model that would let it earn from the agent channel rather than subsidize it.
The rest of the quarter was strong enough to make the agent question look like a rounding error. Revenue rose 34% to $4.6 billion, ahead of the $4.56 billion analysts expected. Total orders climbed 28%, helped by the acquisition of Deliveroo and by growth in monthly active users. Marketplace gross order value rose 36%. Net income attributable to shareholders reached $466 million, or $1.06 a share, up 91% from $244 million a year earlier. Adjusted EBITDA rose 41%.
Growth by acquisition is now part of the story, and so is the regulatory risk that comes with it. DoorDash is still seeking approval for its purchase of Instacart, a deal that would extend its reach from restaurant meals into grocery, where basket sizes are larger, margins are thinner, and the operational complexity is considerably higher. Antitrust review of a combination that touches both restaurant delivery and household grocery could take most of a year.
Inside the company, the bet on agents is less about near-term orders than about defaults. If a consumer’s assistant learns to order lunch without opening an app, the platform that holds the payment credential and the address book becomes the supplier, and the supplier’s brand fades. That is the outcome DoorDash is trying to insure against by building its own agents first and negotiating with everyone else’s. The company is also exposed to a second question it cannot answer alone: whether agent-initiated orders carry the same fraud liability as human ones, and who pays when an agent orders to the wrong address at 2 a.m.
Xu’s phrasing suggests he has thought about the asymmetry. A platform that dismisses agents risks being disintermediated by them; a platform that reorganizes around them before they have users risks burning capital on a channel that stays a curiosity.
Nine hundred and ninety-three million orders is a large enough base that a few hundred agent transactions change nothing about the quarter. It does change the argument. Until this week, the debate over agentic commerce was conducted almost entirely in projections. DoorDash published a unit count, and the count was small.
The next data point will come from the other side of the transaction, when the agent makers report how many of their users have handed over a saved card and asked for dinner. Until then, the most useful thing in the quarter is the honesty of the number itself, offered by a chief executive who had every incentive to round up.
