DoorDash’s Tony Xu has put a modest number beside a substantial promise: some third-party AI agents had generated “maybe several hundred orders in total”.
The company’s co-founder, chairman and chief executive made the observation on its October 29 earnings call, while discussing how AI assistants might change food and grocery shopping. He said customers predominantly used DoorDash’s app for straightforward food orders.
The remark makes a useful starting point for India, where Swiggy has already opened its services to external AI assistants.
An agent can search menus, assemble a cart and initiate a purchase. Persuading someone to make that their regular way of ordering dinner is a separate commercial task.
WHAT THE SMALL NUMBER TELLS US #
Xu’s observation concerned some third-party agents. It was not a count of all orders involving AI, including DoorDash’s own assistants.
Look Who's Talking Regime Change #
23 Oct 2026 - Vol 05 | Issue 43
Read Now Look Who's Talking Regime Change The accessible transcript supplies no common measurement period or eligible audience for calculating an adoption rate. Several hundred orders from a tightly restricted trial would mean something different from the same volume after widespread distribution.
The number establishes that the third-party channel was small in Xu’s account. It cannot settle whether customers reject AI ordering, or how large the opportunity might become.
It does, however, shift attention towards an observable question: after being offered the option, how many customers complete an order and choose it again?
SWIGGY HAS BUILT THE CONNECTION #
Swiggy announced external AI ordering on January 27, 2026, covering food delivery, Instamart and Dineout. It described tools for searching, comparing options, building carts, applying coupons, placing orders and tracking deliveries.
The connection uses Model Context Protocol, or MCP, a standard through which an AI assistant can call another service’s tools. Swiggy’s current documentation lists restaurant discovery, menus, ordering and tracking among its capabilities.
Its authentication documentation also describes connecting a user’s account through a sign-in and consent process.
Those are meaningful building blocks. They establish that external ordering is a supported proposition, rather than simply a chatbot suggesting restaurants.
The materials reviewed for this analysis do not provide completed-order volumes, repeat rates or conversion figures for Swiggy’s external AI ordering. DoorDash’s experience supplies a question for the Indian business, rather than an estimate of its adoption.
THE RIVAL IS AN EXISTING HABIT #
Consider a customer who knows the restaurant, dish and delivery address. A familiar app offers a relatively simple task for an assistant to improve upon.
Typing a request, reading an interpretation and correcting the result could add work if the ordinary route is already quick.
The proposition becomes stronger when the purchase contains several decisions. A group dinner with different preferences requires more coordination than repeating a familiar lunch. Turning a recipe into a grocery basket involves quantities, available products and substitutions.
These are plausible advantages to test. They are not evidence that customers have adopted them.
The relevant measure is the entire journey from intention to confirmed order. A fluent first response provides little convenience if the customer then spends several minutes correcting the basket.
DOORDASH’S OWN AI TELLS A DIFFERENT STORY #
DoorDash has published more encouraging indicators for its own assistant.
In its September 30 product announcement, it said grocery baskets were built five times faster with Ask DoorDash, based on marketplace data from June. It also said Ask grocery orders had nearly 50% higher basket value and around 60% more unique items than the same consumers’ traditional orders, using orders from August through early September.
These are company-reported comparisons. They do not establish that AI caused the larger baskets. Customers may choose an assistant for more complicated or substantial shopping trips.
Nevertheless, they complicate any sweeping conclusion that people do not want AI involved in shopping.
They also suggest a different competitive possibility: AI could improve a delivery platform’s existing experience before a separate general-purpose assistant becomes a preferred shopping destination.
For the customer, success may look like an easier grocery trip. It need not involve abandoning the platform they already use.
THE OFFICE LUNCH MAY BE A BETTER START #
DoorDash’s corporate-ordering connector offers a concrete example of a more involved task.
Announced on September 30, it lets companies connect compatible internal AI tools to DoorDash for activities such as team lunches and office restocking. The company described a limited beta, with employees connecting their own accounts before an agent acts on their behalf.
Here, an assistant could reduce coordination across people and systems. A workplace conversation can contain the requests; a connected service can assemble and execute the order.
That is a more specific problem than making every individual takeaway purchase conversational.
For Indian developers building on Swiggy’s tools, the opportunity may similarly lie in a narrowly defined workflow. Its value would depend on the work removed, the reliability of execution and the frequency with which customers need it.
WHO GETS SOMETHING EXTRA? #
An order routed through an assistant can have several meanings.
It could be an existing customer’s usual purchase, made through another interface. It could be an additional purchase that would otherwise not have happened. It might also shift business between competing delivery platforms.
Each can matter commercially, but they justify different claims.
Additional demand helps support spending on a new channel. An existing order moved into a more expensive process needs another benefit, such as better retention or lower servicing costs.
For an independent agent developer, the economics also require a way to cover model calls, integrations, support and failed attempts. The public materials reviewed here do not establish a sustainable cost per successful order. A convenient demonstration can leave that question unanswered.
THE MEASUREMENTS THAT WOULD CHANGE THE STORY #
For Swiggy and external agent developers, completed orders should be the beginning of the assessment. A useful comparison would follow customers offered the feature, record how many try it, and measure completion and repeat use. It should count corrections, abandoned checkouts and time spent reaching a confirmed order.
The comparison also needs to reflect the task. A large grocery shop should be assessed against a similar shop through the ordinary interface. Otherwise, differences in basket size or time can be mistaken for the effect of AI.
Finally, the purchase must work beyond checkout.
DoorDash’s results statement emphasises its physical fulfilment network alongside its AI products. An assistant’s successful order still depends on available stock, restaurant execution and delivery. ir.doordash.com
The evidence so far supports a selective opportunity, with substantial gaps in public adoption data. Xu’s small third-party order count is a reason to examine the customer task more closely.
The business becomes compelling when the assistant earns a second order because the first one was easier.