DoorDash Says AI Agents Placed Only Several Hundred Orders

Agentic commerce has been sold as the next front door to the internet. DoorDash's third-quarter earnings call put a number on it: several hundred orders. Not millions. Several hundred.

What DoorDash actually said

On the October 29 call, CEO Tony Xu was asked how much of the agentic experience is real for food ordering. His answer was blunt. "You never want to fight gravity," he said, "but at the same time, you don't want to be delusional. If you look at where AI agents are today, we're seeing maybe several hundred orders in total from some of these third-party agents."

Those agents include Muse, Instinct and Fo, per TechCrunch's account of the call. The comparison is lopsided: DoorDash said in its 2025 report that it averages about 3 million deliveries a day, and it processed roughly 970 million orders in the second quarter alone, according to The Verge. Third-party agents are, for now, a rounding error.

Xu also noted that most agent traffic it does see runs through the company's own surfaces. DoorDash has been shipping them steadily: Ask DoorDash in June, a CLI agent in July, Text DoorDash in September, plus a Model Context Protocol server for corporate ordering so other chatbots can place group orders.

The plumbing problem behind the number

The small figure is not a story about model quality. It is a story about distribution and permissions. Consumers are not yet in the habit of handing a meal order to a third-party assistant, and the platforms that own the customer relationship have little incentive to make it frictionless. Amazon blocks outside shopping agents. Walmart has gone the other way and courted them.

Xu left the door open, though, in a way worth noting: he said that if agents do become a real channel, there is an opportunity to advertise inside them. That is the tell. Platforms will integrate agents when they can monetize the surface, not simply because the capability exists.

Why it matters for builders

The lesson for anyone wiring agents into commerce is sequencing. Building the connector is the easy half: MCP servers, OAuth flows and tool schemas are now commodity work. The hard half is demand - whether a human will actually delegate the transaction, and whether the counterparty will allow it at scale.

Two practical takeaways. First, instrument agent traffic separately from human traffic from day one; DoorDash could answer that analyst question precisely because it tracks the channel. Second, do not build a business model that assumes agent-mediated volume in 2026. Build the plumbing now, expect the payoff later, and watch which platforms open their checkout and which close it.

The agents are coming. They are just ordering very few sandwiches so far.