Back to all posts

Several Hundred Orders: The First Honest Numbers on Agentic Commerce

2026-11-05
luodaInt
Several Hundred Orders: The First Honest Numbers on Agentic Commerce

Several Hundred Orders: The First Honest Numbers on Agentic Commerce

Most of what you read about agentic commerce is a forecast. Last week produced something rarer: a real number from a company big enough that the number means something. On DoorDash's third quarter 2026 earnings call, CEO Tony Xu said outside AI agents have generated "several hundred orders" across the entire time they have been allowed to order on the platform. DoorDash handled just shy of 1 billion orders in that quarter alone.

That gap between the hype and the order log is not a failure story. It is the most useful data point businesses have gotten yet on how agents actually enter commerce, and it points somewhere specific: the money is showing up in the assistant you embed, not the storefront you open to other people's agents.

What DoorDash actually reported

The numbers come from Ars Technica's report on the earnings call, published October 30, 2026. Speaking about third-party agents, Xu called agent ordering a "very low use case" and said that when it comes to a task as simple as ordering food, customers "tend to see that come directly to DoorDash." Where he does see agent potential is "long consideration purchases," with agents "doing work on your behalf proactively in the background."

Two details keep the number honest:

  • The "several hundred orders" figure covers third-party agents only. It does not include DoorDash's own AI features, which the report treats separately.
  • The comparison base is enormous. A few hundred orders against roughly a billion per quarter is a rounding error. Xu's own description of early usage: "very low use case."

The front door is weeks old

Before drawing big conclusions, look at the timeline. Per the Ars report, DoorDash rolled out its MCP connector less than a month before the report, and access was still limited to a small number of early beta partners, used mainly to streamline corporate lunch orders.

DoorDash's own developer docs, live as of November 5, 2026, list two agentic ordering offerings, both behind waitlists and labeled a limited beta:

  • DoorDash CLI: a command line tool tied to a personal DoorDash account, recommended for headless agent setups that order for yourself, your family, or a small group.
  • DoorDash Corporate Ordering MCP: a DoorDash-hosted MCP server for teams and companies, authorized per user via OAuth, connected to corporate meal budgets, with suggested flows like automated office lunch ordering and Slackbot-coordinated group orders.

So the plumbing for agentic ordering is real, standard (MCP with per-user OAuth), and days or weeks old, not months. A tiny number was always the predictable result of a front door that almost nobody has a key to yet. The interesting question is what happens when the waitlists clear, and whether the corporate flows (recurring team lunches, supply restocks) adopt faster than consumer ones. For the moment, the documented use cases are workplace workflows, which is exactly where the early orders are reported to be concentrated.

First-party AI is where the lift shows up

The sharpest contrast in the same story: DoorDash's own assistant, Ask DoorDash, rolled out over the summer per the report, is producing measurable lift. The company shared that Ask DoorDash baskets get built "five times faster" than without it, that those orders "average nearly 50% higher basket value and about 60% more unique items than the same consumers' traditional orders," and that nearly half of them end up at local spots the consumer has never tried before.

Two different bets, one company

Third-party agents: a connector opened to a handful of beta partners, several hundred total orders. First-party assistant: embedded in the product, no separate authorization step, and the company reports faster basket building and materially higher basket value. Same platform, same quarter, very different traction.

Keep the two claims separate, because the sources do. The several-hundred-orders figure and the Ask DoorDash percentages describe different products. But placed side by side they sketch a pattern worth taking seriously: AI that lives inside the workflow where the purchase already happens shows lift now, while AI that arrives through an external agent's connector is still waiting at the door.

What this means for your agent roadmap

If your business is deciding where agent effort goes, the DoorDash numbers suggest a few practical readings:

  • Embed before you expose. An assistant inside your product, near the task your user is already doing, has a much shorter path to measurable impact than an open connector waiting for outside agents to arrive.
  • Treat agent-facing APIs as early infrastructure, not a channel. Build the MCP server, get the auth model right (per-user OAuth is a good pattern), and expect volumes to be near zero for a while. That is normal for a weeks-old front door.
  • Watch the corporate use cases. The documented agent flows here are workplace workflows: recurring team lunches, group carts, restock lists. Routine, budgeted, repeatable tasks are where agents have an obvious job to do.
  • Measure lift, not orders. The strongest numbers in the story are not counts of agent orders. They are deltas: faster baskets, higher value, more items, new restaurants discovered. Pick metrics that show the assistant working, not just traffic through the connector.

The gravity and the delusion

Xu's own framing, quoted by Ars, is the right posture for anyone building in this space: "you never want to fight gravity. But at the same time, you don't want to be delusional." The report also draws a fair historical analogy: DoorDash's current numbers may resemble Pizza Hut's first tentative online ordering experiments in 1994, years before more than half of Pizza Hut orders were placed online.

Early numbers on a brand-new interface prove very little in either direction. What they do prove is that "agents are about to transact at scale" is still a claim about the future. The present tense looks like this: a big platform ships the standard connector, staffs a waitlist, and counts its first few hundred orders while its embedded assistant quietly does the heavy lifting.

If you are planning where AI agents fit into your business workflows and want help picking the bets that show lift early, get in touch. We help teams automate with clear boundaries instead of forecasts.

Sources