Basically no one is using AI agents for their DoorDash orders

On Thursday's third-quarter earnings call, DoorDash co-founder and CEO Tony Xu told an analyst that outside AI agents have produced "several hundred orders" on the platform "across the time that they've been up and running." Set against a quarter in which the company handled just shy of 1 billion total orders, Xu characterized agent-driven ordering as a "very low use case." The comparison is stark: several hundred orders is a rounding error next to a billion, a gap of roughly six orders of magnitude.
Xu drew a distinction between task types when explaining why. He said agents have real potential for "long consideration purchases" and major decisions, scenarios where an agent is "doing work on your behalf proactively in the background." But for something as simple as ordering food, he said, "we tend to see that come directly to DoorDash." That framing suggests the company views agent traffic as better suited to research-heavy or multi-step buying decisions than to habitual, low-friction reordering.
Two scope caveats sit alongside the number. First, DoorDash only launched its Model Context Protocol connector less than a month before the call. MCP is an open standard, originally developed by Anthropic, that gives third-party AI tools a common way to talk to external services; a connector built on it lets an assistant query menus and place orders without scraping a web interface. DoorDash's connector remains limited to a small set of "early beta partners" and is mainly used to streamline corporate lunch orders. Before that link existed, agents had to work through a web interface, which is more fragile and less reliable.
Second, Xu's figure covers third-party agents only, not DoorDash's own AI features. The company rolled out "Ask DoorDash" this summer; it can assemble a cart from a photograph of a recipe or suggest an order from a vague prompt such as "dinner for a family of four." DoorDash hasn't published detailed usage statistics for that feature, but last month it said Ask DoorDash baskets are built "five times faster" than ordinary ones. It also said those orders average nearly 50 percent higher basket value and about 60 percent more unique items than the same customers' traditional orders, and that nearly half land at local restaurants the customer has never tried.
The contrast between the two channels is the most instructive part of the disclosure. Native assistance embedded in the app is producing measurable changes in basket size and merchant discovery, while external agents barely register. This suggests the friction that matters isn't the intelligence of the ordering assistant but the number of steps between intent and checkout. A feature inside the app can inherit saved addresses, payment methods and order history; an outside agent has to negotiate all of that through an API surface that is still in beta. A likely trade-off for DoorDash is that opening its catalog to third-party agents means ceding some of the customer relationship and the merchandising real estate that drives upsells, even as it captures orders it might otherwise lose.
There is a historical parallel worth keeping in mind, though it is context rather than a prediction. Pizza Hut began experimenting with online ordering in 1994, when the web was exotic and few customers had a reason to use it; today more than half of its orders arrive online. Early adoption curves for new ordering channels have been misleading before. Xu's own posture reflects that tension. On AI agents broadly, he said, "My take on all things technology is you never want to fight gravity. But at the same time, you don't want to be delusional."
In practice, the numbers put a hard bound on near-term expectations for agentic commerce in high-frequency, low-consideration categories. Grocery, rides and restaurant delivery are exactly where builders have most aggressively pitched agent integrations, and the DoorDash data suggests consumers aren't yet routing those habits through an assistant. Developers building agent-to-merchant integrations should expect adoption to track the maturity of the connector ecosystem, which here means a beta limited to corporate accounts. A reasonable inference is that enterprise and team ordering, where a coordinator is already delegating and where bulk lunch runs are routine, will be the beachhead rather than consumer dinner.
Why it matters: For product teams weighing investment in agent-facing commerce APIs, DoorDash's disclosure is a useful reality check — a full quarter of agent access produced hundreds of orders against a billion. Teams building MCP connectors or agent checkout flows should treat corporate and delegated purchasing as the likeliest early demand, and measure success against native in-app assistance, which is where DoorDash is already seeing higher basket values and new-restaurant discovery. Consumer adoption of third-party ordering agents remains unproven at this scale, though a beta that is weeks old is a thin basis for a verdict either way.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
But it's still early days for the relatively new e-commerce use case.