What did DoorDash say about AI agent orders?
DoorDash reported third-quarter results on October 29 and then took questions from analysts. CEO Tony Xu used one answer to put a number on agentic commerce. Third-party AI agents had produced "several hundred orders in total," he said.
He repeated the figure when another analyst asked about consumer agents. Agents had generated "several hundred orders across the time that they've been up and running," he said. Neither answer included a start date, a market list or a dollar value.
"Third-party agents" means external assistants that act for a user, not DoorDash's own tools. Buyers can reach them through a chat app, a browser extension or a protocol connection. The category is broad, and DoorDash's number covers all of it.
The scale is tiny against the rest of the business. DoorDash reported 993 million total orders in the quarter, up 28% from a year earlier. Marketplace GOV, which counts order value including taxes, tips and fees, reached $33.9 billion. Revenue rose 34% to $4.6 billion.
Even 1,000 agent orders would work out to roughly one order in a million at that volume. TechCrunch, which reported the remark on October 30, named Meta's Muse among the agents involved. Muse is Meta's consumer agent, launched in early September with shopping and task features.
The company also reported $466 million of net income and $1.06 billion of adjusted EBITDA. It described the quarter as a record for Marketplace GOV, revenue and net income. The agent figure did not appear in the results release or the shareholder letter.
Xu framed the number as early rather than final. "You never want to fight gravity," he said. "But at the same time, you don't want to be delusional." That mix of an open door and a clear-eyed view of demand runs through the rest of DoorDash's AI strategy.
Why are AI agents placing so few food delivery orders?
Xu's explanation was that customers use agents for the harder parts of a task, not the easy ones. Ordering lunch is a few taps in an app that many people already use. "The predominant use case of that is actually through our app," he said.
DoorDash also argues that the ordering step is not the difficult part of delivery. "The hard part is making sure that they actually show up," Xu said, meaning the food arrives as expected and in good condition. Substitutions, missing items, refunds and cold meals all need a system that can act quickly.
The company's own fulfilment data makes the point. Xu said deliveries from DashMarts are "near perfect." DoorDash manages the inventory in those facilities. Deliveries from third-party stores run in the low 90s for accuracy. An agent can place an order, but it cannot fix a store that is out of stock.
Access method matters for security. A browser agent can click through a site like a person, which is what Amazon objected to in its Perplexity suit. A protocol connection such as MCP gives the platform a way to set terms, log activity and revoke access.
Ordering is also a small share of the work. Couriers handle handoffs, gate codes and substitutions, and support teams resolve refunds. An agent that cannot recover from a missing item creates work rather than removing it.
Payment trust matters too. An agent needs permission to spend, an address, a payment method and a way to handle disputes. Walmart's answer is a separate profile with spending limits. DoorDash's answer is partly its own MCP connection, which it says can serve corporate ordering.
Speed matters as well. A person can answer a substitution prompt in seconds. An agent needs rules in advance, or it will pause the order and wait.
What is DoorDash's "atoms vs bits" strategy?
The shareholder letter published with the results includes a section titled "Atoms vs Bits." It says that when ChatGPT launched, DoorDash saw two possible battles. One was to build the leading digital personal assistant. The other was to build the leading physical network that fulfils tasks in the real world. DoorDash chose the second.
The letter says the company remains "focused on winning the battle for atoms" as personal agents become more mainstream. That does not mean refusing to build AI products. DoorDash launched Ask DoorDash in June, a chatbot that turns photos and prompts into grocery and restaurant orders. It later added Text DoorDash, a messaging route that works outside the app.
The company also shipped a Model Context Protocol server that lets external agents place orders, including for corporate lunches. The message is that DoorDash wants to be the fulfilment layer for agents, not the assistant that sits in front of them.
The financial case is straightforward. DoorDash reported $4.6 billion of revenue in the quarter, up 34%, and $466 million of net income. For the fourth quarter, it guided to Marketplace GOV of $35.4 billion to $36.4 billion. Adjusted EBITDA guidance was $900 million to $1.1 billion.
The biggest strategic bet is Instacart. DoorDash agreed to buy the company for $65.00 per share, or about $17 billion in equity value. That is a 26% premium to Instacart's October 22 close. Management sees a multi-year path to $500 million or more in annual synergies. It plans to reinvest more than 100% of them in the first year.
DoorDash also said its US restaurant category grew faster quarter over quarter than in any third quarter since 2020. That core business funds the agent experiments, even if agents add little revenue today.
The letter also describes heavy spending on warehousing and autonomous delivery. Neither project depends on agents becoming popular. Both make the physical network cheaper and more reliable, which is where DoorDash expects the durable advantage to sit.
How do Walmart and Amazon treat AI shopping agents?
Walmart has taken the opposite approach to Amazon. On October 21, it announced an Agent Gateway for Walmart and Sam's Club. The gateway is an MCP server that lets approved agents search inventory and buy on a customer's behalf. CEO John Furner spoke about it at the Masters of Scale Summit in San Francisco.
Walmart also said it would roll out Agent Profiles next month. Those profiles give each agent a separate identity inside a Walmart account. Customers can require approval for every purchase or set a spending cap. Walmart says its returns policy applies even when an agent misbehaves, and that agents which break its terms will be blocked.
The retailer had already partnered with Meta's Muse and offers integrations with ChatGPT and Gemini. Its terms say agents cannot access payment information or act beyond the shopper's intent. Walmart's terms combine broad access with tight controls.
Muse, launched in early September, connects to a user's apps and services, including email and calendars. Meta is also bringing it to smart glasses and Macs. Both would widen the number of places where a food order can begin.
Amazon has gone the other way. It blocks dozens of shopping agents, including ChatGPT, and it sued Perplexity in November 2025 over its Comet browser. In March 2026, a federal judge granted a temporary injunction. The judge found Amazon had shown "strong evidence" that Comet accessed its site without authorization. The order included a weeklong stay so Perplexity could appeal.
Amazon's complaint also raised an advertising problem. Agent traffic can look like a shopper to an ad system, which complicates billing advertisers for real human impressions. That argument matters for DoorDash, whose ad business depends on being the place where customers browse.
Amazon is not standing still on its own products. It has invested in Rufus, a shopping assistant built into its website and app. The dispute is less about whether AI should shop and more about who controls the account and the data.
- •Walmart: open access through an MCP gateway, with per-agent profiles, spending limits and its returns policy.
- •Amazon: blocked agents, a court injunction against Perplexity, and concerns about advertising billing.
- •Meta: a free consumer agent for a high token allowance, with an expected fee on transactions.
What does this mean for restaurants, grocers and advertisers?
For restaurants, the risk is losing the customer relationship. If an agent recommends a burrito and handles checkout, the restaurant becomes a supplier behind someone else's interface. DoorDash's app keeps the brand, the upsell and the reorder button in its own hands.
For grocers, agents are harder because baskets are complex. A recipe needs substitutions and ripeness judgments. DoorDash's answer is inventory control through DashMarts and, soon, Instacart's storefront software. Xu said the goal is to manage inventory for retailers of every size.
Grocery and retail are also growing faster than restaurants for DoorDash. The company said segment GOV grew at more than double the pace of its US restaurant category. That makes the grocery question more urgent than the restaurant one.
Advertising is the most valuable question. It is already a meaningful contributor to DoorDash's EBITDA, as one analyst noted on the call. Xu said he could imagine advertising inside third-party agents, but added that "there isn't yet that value" at current volumes.
Agent makers have their own economics. Meta says Muse will be free for a large number of tokens. The company expects to profit by taking "a small fee from transactions." If agents become the front door, they will want a share of the order value that DoorDash currently keeps.
Scale also matters for advertising. DoorDash told investors that Instacart's grocery audience would give packaged goods advertisers more reach. That pitch depends on DoorDash keeping a direct line to shoppers, whether they arrive through its app or an agent.
Couriers are less exposed in the near term. Every order, however it is placed, still needs a Dasher. The immediate effect is on order flow and advertising rather than on the delivery work itself.
What are the limits of DoorDash's disclosure?
The number is vague. "Several hundred" could mean 300 orders or 900, and DoorDash gave no period, market list or gross order value. The figure came in an answer to an analyst question, not as an audited operating metric. It did not appear in the results release or the shareholder letter.
It also covers only third-party agents. Orders through Ask DoorDash, Text DoorDash or the MCP server are not part of the comparison. The figure therefore measures external distribution, not AI use across the company.
Reported growth also includes Deliveroo, which DoorDash bought in 2025. Total orders rose 17% excluding Deliveroo, rather than 28%, and the company will stop publishing the ex-Deliveroo comparison from the fourth quarter.
Finally, DoorDash has reasons to describe agent demand cautiously. A low number supports its argument that fulfilment, not the chat interface, is the durable advantage. That does not make the figure wrong, but readers should treat it as one company's view of an early market.
The context also cuts both ways. Several hundred orders is a small base, so a jump to a few thousand would sound dramatic without being material. The useful measure will be agent orders as a share of total orders over several quarters.
What should you watch next?
The fourth-quarter report will be the first test of whether agent orders grow during the holiday season. DoorDash has guided to $35.4 billion to $36.4 billion of Marketplace GOV. It also guided to $900 million to $1.1 billion of adjusted EBITDA. Analysts will ask whether advertising revenue is holding up.
Walmart's Agent Profiles are due next month. If they work as described, they will give other retailers a template for letting agents in without handing over payment credentials. Watch whether DoorDash builds a similar control layer for its own marketplace.
Meta's Muse is the other variable. It connects to user accounts and is coming to smart glasses and Macs. A transaction fee would give Meta a direct stake in where its users order food. That is the scenario DoorDash is preparing for.
The legal fight between Amazon and Perplexity will set boundaries for agent access. An appeal is pending, and other platforms are watching. A ruling that agents need explicit permission would favour companies that can negotiate terms, such as Walmart and DoorDash.
Inside DoorDash, the Instacart deal and the single technology stack are the projects to track. The stack is due to be finished in the first half of 2027. If agents do become a real channel, DoorDash wants the plumbing ready before the demand arrives.
The company has also said it will stop reporting results excluding Deliveroo from the fourth quarter. That will make year-over-year comparisons cleaner. It will also make it harder to see how much of the growth came from the acquisition.
One smaller signal is corporate ordering. The MCP server is aimed at office lunches and similar orders. That use case fits the technology better than a single hungry customer, and it may show up in the numbers sooner.