DoorDash Reveals Only Hundreds of Orders from AI Agents, Open Collaboration Seen as Key

DoorDash recently disclosed that its platform has received only "hundreds" of orders generated through AI agents. This figure stands in stark contrast to expectations that AI shopping assistants are about to disrupt the food delivery industry, and it reveals that agentic commerce is still in an extremely early stage.

What Does "Hundreds of Orders" Mean?

"Hundreds of orders" is a highly informative qualitative signal. For a food delivery platform that handles millions of orders per day, this number is almost negligible. It shows that users who actually complete end-to-end ordering through AI agents are still at the experimental and small-scale trial stage.

This does not mean AI agents have no value. Rather, it shows that between "AI recommends a restaurant" and "AI completes payment and places the order on behalf of the user," there remain a series of practical barriers: trust, payment authorization, address confirmation, and after-sales responsibility. DoorDash's decision to disclose this data appears to be an effort to cool industry expectations: the inflection point for scaling agentic commerce has not yet arrived.

AI Agent here refers to a software system capable of autonomously perceiving its environment, planning steps, and invoking tools (such as browsers, APIs, and payment interfaces) to accomplish user goals, as distinct from chatbots that only provide suggestions. A typical agent ordering workflow includes: understanding user preferences, retrieving merchants and menus, assembling an order, confirming address and payment method, submitting the order, and tracking status. Its core capability is "task orchestration," that is, chaining multiple atomic operations into an end-to-end workflow. Current mainstream implementations are mostly based on the function calling capabilities of large language models (LLMs), combined with memory and planning modules. This is why the article emphasizes the gap between "recommendation and ordering"—recommendation only requires generating text, while ordering requires agents to have reliable tool calling, exception handling, and permission management capabilities.

Platform Attitude: Open Collaboration Rather Than Closed Self-Building

More noteworthy than the order numbers is DoorDash's statement that it is willing to collaborate with third-party agent developers. This means DoorDash does not intend to keep the AI ordering entry point firmly in its own hands, but instead tends to view agents as a new traffic channel.

This strategy has its rationale. The core assets of a food delivery platform are merchant supply, delivery network, and user account system, while agent developers excel at conversational interaction and task orchestration. Collaboration between the two can let agents handle "helping users decide what to eat and completing the operation," while the platform handles fulfillment. For DoorDash, an additional entry point means more orders, without bearing the R&D costs of the agents themselves.

But openness also brings new problems: when third-party agents place orders on behalf of users, how should order attribution, exception handling, and refund responsibility be divided? How should the commercial revenue-sharing model between platforms and agent developers be designed? There are currently no public answers to these questions, yet they are precisely what determine whether agent orders can move from "hundreds" to scale.

Agentic commerce can be understood as a new transaction paradigm: users no longer directly operate apps, but delegate shopping intent to AI agents, which interact with e-commerce platforms (or through APIs) to complete browsing, decision-making, and payment. Its difference from traditional e-commerce lies in the fact that the "interacting subject" changes from a person to a proxy program. This brings a series of new infrastructure requirements, such as identity authentication for agents (how to prove the agent has obtained user authorization), machine-readable product and inventory interfaces, and audit logs of agent behavior. Currently there is no unified industry standard; some platforms adopt an open API strategy, while others build closed agents in-house. DoorDash's stance belongs to the former, treating agents as an external traffic entry point similar to search engines or social platforms, which also explains its open attitude toward third-party developer collaboration.

The Real Bottleneck of Agentic Commerce

From DoorDash's data, one can infer the current state of the entire agentic commerce track. AI agents are already quite mature in information retrieval, price comparison, recommendation, and other areas, but once actual transactions are involved, users often fall back to manual operation. The reason is that the transaction stage has far higher requirements for certainty and accountability than the information stage.

Another underestimated bottleneck is user habits. Letting AI order food on one's behalf requires users to first build trust in the quality of the agent's decisions, while food ordering itself is a high-frequency, low-decision-cost scenario—users may order food themselves faster than describing their needs to an AI. This partly explains why agent orders are growing slowly.

Payment authorization and accountability are key obstacles to the implementation of agent transactions. In traditional online payments, the user clicking "confirm payment" constitutes authorization; when an agent acts on the user's behalf, a mechanism is needed to prove that the operation obtained the user's prior or in-process consent, such as tokenized payment credentials, spending-limit authorization, or biometric secondary confirmation. At the same time, once an order goes wrong (such as wrong delivery address or duplicate order), responsibility must be clearly divided among the user, the agent developer, and the platform; otherwise the platform will bear all customer complaint and refund risks. In addition, the sharing of personal data such as addresses and contact information also involves privacy compliance. These mechanisms are currently mostly exploratory and lack common industry standards. It is these, rather than model capabilities, that determine whether agent orders can scale.

Three Implications for the Industry

First, agentic commerce is still in the infrastructure-building phase. Open platform interfaces, clear responsibility boundaries, and settlement mechanisms—these "dirty work" tasks—determine the ceiling of agent orders more than model capabilities.

Second, early data should not be overinterpreted. Hundreds of orders is both the current state and a starting point. As agents become integrated with platform account systems, order volume may grow nonlinearly, but only if trust mechanisms are established first.

Third, the platform's open and collaborative attitude deserves attention. DoorDash's choice shows that food delivery platforms are more willing to treat agents as a channel rather than a threat. If this positioning is adopted by more platforms, agent developers will gain a clearer commercialization path.

Between "hundreds of orders" and "millions of orders" for agent ordering, what is missing is not model capability, but transaction trust and ecosystem collaboration. DoorDash's disclosure has precisely put this gap on the table.

One-line Summary

DoorDash disclosed only hundreds of AI agent orders, revealing that agentic commerce is still extremely early and bottlenecked by transaction trust and ecosystem collaboration.

Paragraph Summary

DoorDash recently disclosed that its platform has received only "hundreds" of orders generated through AI agents, in stark contrast to expectations that agent shopping assistants will disrupt the food delivery industry. The article argues that this figure shows agentic commerce remains at the experimental and trial stage, and that between AI recommendation and AI completing payment and ordering on behalf of users lie practical barriers such as payment authorization, address confirmation, after-sales responsibility, and user trust. More noteworthy than the order numbers is DoorDash's willingness to collaborate with third-party agent developers, viewing agents as a new traffic channel rather than a threat. The article points out that the real bottleneck for agent ordering is not model capability, but certainty and accountability in the transaction stage, as well as user habits. Infrastructure building—open platform interfaces, clear responsibility boundaries, and settlement mechanisms—determines the ceiling of agent orders more than model capabilities.