Summary
DoorDash said it has received only “several hundred” orders from third-party AI agents, while most customers still order through its app. Zest’s new text-based restaurant concierge shows how agents are moving from recommendations toward reservations and transactions, but the market is still working through trust, access, authorization and accountability issues.
From answering questions to executing transactions
The conversation around AI agents taking over commerce has grown quickly, but the latest operating data suggests that actual adoption is still limited. During its quarterly earnings call, DoorDash said it had seen only “several hundred” orders from third-party agents, including Muse, Instinct, Fo and others. That is a small early signal compared with the delivery platform’s previously reported average of roughly 3 million deliveries per day.
The disclosure does not mean agentic commerce lacks potential. It does show that there is a meaningful gap between an agent being technically capable of placing an order and consumers routinely delegating transactions to one. For now, DoorDash said users are still predominantly ordering through its application, where they can browse options, review the cart, confirm delivery details and complete payment directly.
An agent may reduce the number of steps in that process, but it has not yet displaced the dominant consumer interface. The current market is therefore better understood as an experimentation phase: platforms and startups are building the connective tissue for agent-driven transactions, while usage patterns remain largely app-based.
Zest moves restaurant discovery into text messages
Zest is approaching the opportunity from the restaurant-discovery side. The company has introduced a personal AI-powered restaurant concierge that lives in text messages. Users can ask for restaurant suggestions in a conversational format and then move directly to the next step of booking a table.
The product reflects a broader effort to make conversational agents a replacement for some stand-alone mobile-app interactions. Rather than opening a restaurant application, searching through lists and then switching to a booking flow, a user can interact with the service through a familiar messaging channel. Discovery, recommendation and reservation are presented as parts of one continuous exchange.
Zest’s personalization model is based on users’ real-world dining activity. The app imports credit-card transaction data through Plaid and says it focuses on transactions in the food and beverage categories. That information helps it identify where users have eaten and build a picture of their preferences. It can then offer recommendations in a user’s home city or while traveling.
This approach highlights one of the defining characteristics of agentic commerce: an agent is not limited to searching public information. With authorization, it can combine a user’s history, preferences and context to determine what action should come next. Zest founder and CEO Mario Gomez-Hall described the company’s direction by saying that “the agent is the product now.” In practical terms, the product is no longer just a restaurant list; it is the conversational layer that connects personal data to a recommendation and then to a booking.
Small order volumes offer an important reality check
DoorDash’s disclosure provides a useful counterweight to the more expansive claims surrounding agent-led commerce. Agents can interpret requests, identify products or restaurants and initiate transactions, but real-world use remains constrained by several factors.
Consumers may not yet be comfortable delegating decisions to an unfamiliar system. They may also want to inspect prices, delivery times, substitutions and fees themselves before submitting an order. In more complicated cases, an agent needs to resolve ambiguity: which location should it use, what counts as an acceptable substitute, and when should it ask for confirmation rather than proceed?
DoorDash CEO Tony Xu characterized agent ordering as a dynamic area and argued that companies should neither resist technological change nor become detached from current reality. That position reflects the platform’s dual incentive. DoorDash needs to understand new agent-driven use cases and maintain a workable relationship with third-party developers, but it also has to protect the reliability and economics of its core marketplace.
Agents could eventually change how platforms compete for consumer attention. In a traditional app, companies influence decisions through search rankings, recommendations, page design and advertising placements. If an agent becomes the primary interface, some of that relationship may move into an intermediary layer. DoorDash has suggested that advertising inside agents could become an opportunity in the future, although the commercial model would have to address a basic trust problem: users may not want paid placement to be confused with an agent’s supposedly personalized recommendation.
DoorDash is building both first-party and external pathways
DoorDash is not waiting for third-party agents to define the category. The company has launched its own text-to-order agent, a command-line agent and Ask DoorDash, a chatbot that allows users to place orders using prompts and photos. It has also released an MCP to support ordering from other chatbots for corporate ordering.
Together, these efforts suggest a two-track strategy. A first-party agent gives DoorDash more direct control over the customer relationship, order logic, permissions and commercial rules. External access, meanwhile, allows the company to participate if consumers begin using other assistants as their main interface. Standardized access can expand reach, but it also creates operational questions around authentication, order errors, customer support, data handling and liability.
The outcome may depend on whether platforms can make agent access predictable and auditable. A human navigating an app can see what is happening at each step. An agent may perform several actions behind the scenes, making it more important for the service to disclose what was requested, what was authorized and what was actually completed.
Payment and data authorization remain central challenges
The examples from Zest and DoorDash show that agentic commerce is not only a natural-language problem. A reliable consumer agent must understand intent, access the right information, obtain appropriate authorization and clearly report the result. When payment is involved, the system also needs to handle identity verification, spending limits, confirmation requirements, cancellations, refunds and disputes.
Zest’s use of card transaction data illustrates the privacy trade-off. A richer history can produce more relevant recommendations, but it also raises questions about consent, data minimization, retention and the possibility of incorrect inferences. Users need to understand what data is being imported, which categories are being used and how authorization can be changed or withdrawn.
The more an agent relies on personal history, the more important those controls become. A recommendation can feel helpful when it reflects genuine preferences, but the same personalization can become opaque if users cannot tell why a venue was suggested or whether commercial incentives influenced the result.
Access rules are also unsettled. Websites and platforms are still deciding whether to allow personal agents to access their content and services. Some companies have restricted such access, while others see agents as a way to capture additional business. The long-term direction will depend not only on model capability, but also on whether platforms create clear machine-access policies and transaction interfaces that can support permissions, confirmation and audit trails.
Early infrastructure, not mass adoption
The evidence available so far is best read as a sign that the infrastructure and product formats for agentic commerce are taking shape, rather than proof that consumers have already shifted away from apps. DoorDash’s several hundred third-party-agent orders indicate that the transition from direct app interaction to delegated ordering has not happened at scale. Zest’s text-based concierge shows, however, how startups are trying to combine discovery, personalization and reservation into one conversational workflow.
The next indicators will extend beyond raw order counts. The market will need to show whether agent orders are completed reliably, whether users return to the experience, how often errors or cancellations occur, and how platforms, merchants and agents divide responsibility when something goes wrong.
If agents can make authorization transparent, outcomes controllable and service performance dependable, they may gradually become a new layer of consumer infrastructure. For now, the clearer conclusion is more modest: AI agents have begun to execute consumer orders, but the habits, safeguards and business relationships required for mainstream adoption are still being built.
Source: link