Executive Summary

Key Figures

Four numbers from this earnings call carry the rest. The first two are what DoorDash said or reported, the third is the ratio between them, and the fourth is a separate announcement from the same day.

Sources: TechCrunch and DoorDash investor announcements (October 29–30, 2026)

On the October 29 third-quarter call, an analyst asked how many orders were arriving through AI assistants. Tony Xu's answer was short: across some of these third-party agents, the orders came to maybe several hundred in total. A follow-up question drew the same answer again. Everything those assistants had accumulated across the time they have been up and running added up to several hundred.

Most of the coverage that picked up the figure set 3 million a day beside it. It looks like a usable comparison until you follow the source. The 3 million comes from the Dash Report that DoorDash itself publishes, it covers calendar year 2024, and its scope is deliveries completed by Dashers inside the United States. The 993 million, meanwhile, covers July through September 2026 across the whole platform, Deliveroo included. Two values with a different period and a different scope cannot be measured by the same yardstick.

Match the denominator to the same quarter's reported figure and the daily average becomes roughly 10.8 million. That is more than triple the 3 million the coverage used. Correcting the denominator does not narrow the gap; it widens it. The conclusion holds either way, but what remains is how easily a comparison travels when nobody states the scope of its numbers.

The ratio is simple. Put the upper end of several hundred generously at 900 and it still falls short of one millionth of the quarter's 993 million orders. And 900 is not one quarter's worth but everything accumulated across the whole stretch those assistants have been running, so holding it against a quarterly figure is already arithmetic that favors the agents.

The three Tony Xu named are all personal-assistant services that arrived this year. Muse, which Meta released on September 8, picks out goods by clicking through shopping sites in its own browser the way a person would. Spear Street Technology's Instinct has no app screen at all: call it by text or phone and it works a handset and a computer the way a person does, and it is still invitation-only. Fo, which Wajo opened up on September 28, handles work by phone and email, hands off to a trained human assistant wherever the AI cannot finish, and pays with a virtual card minted fresh for each purchase.

The three are built differently, but from the store's side they all arrive looking the same. They are customers using the screens and the phone numbers meant for people, with no dedicated connection of their own. Whether the menu changed, whether that store is open right now, whether anything is left in stock — none of it is handed to them as an official feed. They have to read it off the screen.

None of the three has disclosed an agreement with DoorDash to exchange data. So what this number measures is less the skill of a model than how much a guest standing at someone else's front entrance, with no official credential, can actually get done.

DoorDash has never said it shut these assistants out. Tony Xu said his standing view on all things technology is that you never want to fight gravity, and attached the qualifier that you don't want to be delusional either. He also explained that the company is learning which features on the third-party assistants' side actually bring in more orders, and that this is how a healthy partnership gets built. Asked about putting advertising inside the assistants, he answered that the path to consumer value has to come first, and that the value is not there yet.

Its own tools, by contrast, it added briskly. In June it launched the Ask DoorDash chatbot, which takes orders from a prompt or a photo; in July it added an agent callable from the command line, and in September an agent that orders by text message. It also opened a Model Context Protocol server so corporate orders can be placed from other chatbots. This is a company that has already accepted the picture of an agent placing an order.

A different account of why the number is small surfaced the same day. Amazon chief executive Andy Jassy said on his own company's October 29 third-quarter call that few customers are yet buying through outside shopping assistants. His reason was not access terms but the quality of the experience: today's assistant experience is still not good, and it is no more convenient than going to the retailer directly. In the same session Jassy also said the main assistant used for shopping on Amazon will be its own Alexa, so this diagnosis cannot be taken as a neutral observation. Still, the same number was being explained elsewhere in the industry by how finished the assistants are.

Facing the same question, two of the largest US retailers moved in opposite directions. From the night of September 20, Amazon raised a notice window for Muse users telling them that continued access by an unauthorized AI assistant violates its terms of use, and cut the access off. Amazon's account was that Meta had never asked permission and that the assistant does not identify itself on each request. We covered that episode on this blog as the moment a contract became the place where access rights get decided.

Walmart opened the opposite door on October 21. It stood up a server called Agent Gateway so that approved assistants connect over the Model Context Protocol, and it has customers register an assistant inside their own account as a separate profile. A customer can require approval on every purchase or set a ceiling the assistant may spend on its own, and the profile can be deleted at any time. A condition came with it: assistants that break the terms get blocked. The explanation that followed likened the whole thing to adding one more user to a credit card.

These companies are not the only ones with a position already settled. A survey the consultancy BCG released on October 21, the same day as Walmart's announcement, asked 50 of the world's largest retailers and 100 of its largest brands how they intend to receive outside assistants. In it, 42% said they have arranged their data so that any assistant can find their products, and 32% said they feed data directly only to assistants they have chosen. Blocking outright was the rarest of the three at 6%, and 20% had not decided. Add the three together and eight in ten already have a position. Shutting the door the way Amazon did is the minority choice; building a counter like Walmart, or simply leaving things open, is the majority.

Having a position and having a passage that actually works are two different things. A merchant survey PYMNTS Intelligence published with Visa in June of this year found that 15% had organized product data an assistant could read. The share that could tell an incoming visit came from an assistant and tie it through to a purchase was 23%, and 37% said they cannot distinguish one at all. That is the distance between answering that the door is open and having an open passage.

Layered on top of the split between the two companies is the absence of a standard. Three protocols for how an assistant reaches a store have appeared between last autumn and this year, and the three take different layers while their participant lists overlap. Poppy, the latest, is still a 0.1 draft, and even the two that started running earlier have not settled which of them takes the reference seat.

That the field has not converged on one standard means an assistant wanting to reach a given store has to strike a separate arrangement with that store. The same thing happened first on the payment side, and we laid that process out in a post on the agent payment protocol Google put forward.

A food delivery platform's quarterly results are not something Pebblous follows. The reason for looking into this remark is where the several-hundred figure sits. It looks like a place where the intelligence of a model gets measured, but what actually gets decided there is on what credential an assistant reaches the catalog, the inventory and the order history. Amazon withdrew that credential through its terms; Walmart started issuing it in the form of a profile and a limit. DoorDash has done neither.

Another announcement from the same day supports the picture. Alongside its third-quarter results, DoorDash said it had agreed to acquire Instacart at $65 a share, an equity value of about $17 billion. Instacart carries close to 100,000 stores today, and DoorDash says the combination would let Instacart customers shop at more than 50,000 additional stores. On one side, several hundred orders come in while the credential to hand an outside assistant remains unsettled; on the other, a company with an entire store list on its books gets bought largely for cash. How the market prices the work of reaching a catalog shows only when the two numbers sit side by side. Reading the contrast as a pair is this article's interpretation, and DoorDash has never said it acquired for that reason.

For anyone who works with data, the situation is familiar. If the route by which outsiders reach your data is undefined, then both blocking and opening get done in a hurry after something has already happened. The decisions banks made together this autumn had the same shape. If you run a service, three things are worth writing down first.

• Is the line between the data you will hand over and the data you will refuse written in one document? If it is not written down, the judgment made on the day an assistant shows up becomes the policy
• Can you tell whether the visitor is a person or an assistant? The first condition the Amazon and Walmart terms both demand is that an assistant declare what it is
• Who grants the permission and when is it withdrawn? Walmart's profile and limit sit in the customer's hands and Amazon's terms sit in the company's. The two designs put accountability in different places when something goes wrong

Where to draw the first line already has a proposal on the table. The BCG survey from the previous section recommends leaving product, inventory, price, promotion and delivery information for assistants to read, while the store keeps customer identity, preferences, loyalty history and behavioral records. In the same survey, 87% said they have never handed customer data to an assistant, and 86% said they have handed over product data or plan to. The line is already being drawn, and whoever decides late ends up working from a line someone else drew.

Several hundred is, for now, a number small enough to ignore. That is exactly why drawing the line while the number is still small is the better move. Walmart could open a counter because it had defined approval, limits and blocking first, and Amazon could cut access off overnight because a sentence was already written into its terms.

Thank you for reading this far. The several-hundred remark and Tony Xu's other answers were checked against DoorDash's October 29 third-quarter earnings call and the reporting that carried it; the 993 million orders, the 776 million from the same quarter a year earlier and the terms of the Instacart deal come from the materials DoorDash issued that same day. The source and scope of the 3 million a day come from DoorDash's Dash Report, and Amazon's September 20 block and Walmart's October 21 announcement were checked against each company's public statements and reporting in several outlets. Andy Jassy's remark that few customers yet buy through outside assistants comes from Amazon's third-quarter earnings call on that same October 29; the 42, 32 and 6 percent distribution of retailer positions and the recommendation separating product data from customer data come from the survey BCG published on October 21; and the 15% with organized product data and the 23% able to identify an assistant's visit come from the merchant survey PYMNTS Intelligence and Visa published in June of this year. Section 1's swap of the denominator for the same quarter's reported figure, and section 5's reading of the gap as a question of data access credentials, are this article's interpretation. If your organization has settled what it will hand over and what it will refuse when an outside AI assistant arrives, we would be glad to hear the standard you set.

Pebblous Data Communication Team
October 31, 2026

Primary Announcements & Earnings

• 1.DoorDash. (2026). "DoorDash to Acquire Instacart to Help Local Stores Win." DoorDash Investor Relations, Oct. 29, 2026.
• 2.The Motley Fool. (2026). "DoorDash (DASH) Q3 2026 Earnings Call Transcript." Fool.com, Oct. 29, 2026.
• 3.DoorDash. (2025). "Dash Report 2025: Delivery in America."
• 4.Taylor, B. & Bavor, C. (2026). "Sharing a draft of Personal Agent Protocol." Sierra, Oct. 9, 2026.

News Coverage

• 5.Mehta, I. (2026). "DoorDash said it's only received 'several hundred' orders from AI agents." TechCrunch, Oct. 30, 2026.
• 6.Herrera, S. (2026). "Amazon CEO says shopping agents like Meta's Muse 'not great'." Fortune, Oct. 29, 2026.
• 7.Perez, S. (2026). "While Amazon blocks AI agents, Walmart rolls out the welcome mat." TechCrunch, Oct. 21, 2026.

Industry Research

• 8.Boston Consulting Group. (2026). "How Retailers Can Win With AI Shopping Agents in 2026."
• 9.PYMNTS Intelligence & Visa Acceptance Solutions. (2026). "2026 Global Digital Shopping Index: Merchant Edition." June 2026.