For all the talk of AI agents that will shop for you, the real numbers so far are tiny. DoorDash's chief executive Tony Xu told analysts that third-party AI agents have generated only several hundred orders across the whole time they have been able to, against the nearly 1 billion orders the app handled in the third quarter of 2026 alone. That is a useful reality check on agentic shopping: for a quick, habitual purchase, people still open the app themselves.
Xu's own framing is the most telling part. He called agent orders a very low use case, and said he sees genuine potential for AI agents in long consideration purchases, the major decisions where an agent could work in the background on your behalf. Ordering dinner is the opposite kind of task: simple, frequent and fast, and he said that for those, customers tend to come straight to DoorDash. The gap between a weekly food order and a months-long purchase decision is where the value of an agent actually sits.
In fairness to the agent makers, the plumbing is barely in place. DoorDash only launched its Model Context Protocol connector less than a month before the call, a standard way for outside AI tools to order through its platform, and it is still limited to a small number of early beta partners, mostly used to streamline corporate lunch orders. Without that link, an agent has to drive the website by hand, which is slower and more fragile. So the few hundred orders say more about how new and gated the connector is than about any ceiling on demand.
DoorDash is not standing still on AI itself. Its own Ask DoorDash assistant, launched in the summer, can do things like build a shopping cart from a photo of a recipe or suggest an order from a vague description, and the company says that kind of tool helps customers discover restaurants they have not tried. The point is that DoorDash sees more value in AI inside its own app than in handing the customer to an outside agent.
The honest reading is that the number is a snapshot, not a verdict. A connector weeks old and open to a handful of partners cannot show what demand would be at scale, and food delivery may simply be the wrong test case, since Xu himself points to bigger, slower purchases as the better fit. Still, the burden of proof sits with the agent makers. The comparison worth remembering is Pizza Hut's first tentative online ordering in 1994, which looked trivial then and is now most of its orders, so today's few hundred may read very differently in a decade, or may not.