THE SHORT VERSION
- DoorDash's chief executive told analysts that third-party AI agents have generated several hundred orders in total, a very low figure against about a billion orders in the third quarter.DoorDash 首席执行官 Tony Xu 称，第三方 AI 智能体累计只产生几百个订单，而公司第三季度订单接近 10 亿。
- The company's connector for outside agents is still limited to early beta partners and mainly used for corporate lunch orders.面向外部智能体的连接器仍只对少量早期测试伙伴开放，主要用于企业午餐订单。
- DoorDash's own AI features, such as Ask DoorDash, show larger effects: faster baskets, higher basket values and more new-restaurant discovery.DoorDash 自家的 Ask DoorDash 效果更大：下单更快、客单价更高，且更多订单流向用户没去过的新店。
Several hundred against a billion几百单与十亿单
On DoorDash's earnings call, co-founder and chief executive Tony Xu answered an analyst question about outside AI agents by putting the number in context: such agents were generating several hundred orders across the time they had been running. Ars Technica reported that he called it a very low use case for a company that handled just shy of one billion orders in the third quarter of 2026 alone.在 DoorDash 财报电话会上，联合创始人兼首席执行官 Tony Xu 回答分析师关于外部 AI 智能体的提问时，先把数字放进背景：这类智能体自上线以来累计只产生几百个订单。Ars Technica 报道称，他把这称为一个非常低的使用场景——而公司在 2026 年第三季度单季就处理了接近 10 亿个订单。
TechCrunch reported the same figure and did the comparison differently: the company said in its 2025 report that it completes an average of three million deliveries a day. Against either baseline, the volume of orders placed by third-party agents is a rounding error rather than a channel.TechCrunch 报道了同一数字，但换了一个比较基准：公司在其 2025 年报告中称平均每天完成 300 万次配送。无论用哪个基准，第三方智能体下的订单量都只是舍入误差，而不是一个渠道。
The door for outside agents is half open给外部智能体开的门只开了一半
Part of the explanation is technical. Ars Technica reported that DoorDash offers a connector that gives third-party AI tools a standardised way to order through its platform, but the connector is still available only to a small number of early beta partners and is mainly used to streamline corporate lunch orders. Without it, an outside agent has to work through a web interface and put in more effort.部分原因在技术层面。Ars Technica 报道称，DoorDash 提供一个连接器，让第三方 AI 工具以标准化方式通过其平台下单，但该连接器仍只对少量早期测试伙伴开放，且主要用于简化企业午餐订单。没有它，外部智能体只能通过网页界面操作，需要付出更多努力。
That matters for how the number should be read. Several hundred orders is not a measure of consumer appetite for agentic ordering; it is a measure of what a small set of beta partners managed to send through a limited integration. A bigger rollout would test the demand, and DoorDash has not said when that will happen.这影响我们如何解读那个数字。几百个订单衡量的不是消费者对智能体下单的兴趣，而是少量测试伙伴通过一个受限集成能送来多少。更大范围的开放才会真正检验需求，而 DoorDash 尚未说明何时开放。
Food is the hard case, not the easy one点餐是难题，而不是简单题
Xu's explanation is more interesting than the number. He said he sees real potential for AI agents in long consideration purchases and major decisions, where an agent does work on your behalf in the background. But when it comes to a task as simple as ordering food, he said, customers tend to come directly to DoorDash.Xu 的解释比数字本身更有意思。他说，他确实看到 AI 智能体在需要长时间考虑的大额购买与重大决策上有潜力——智能体可以在后台替你做功课。但就点餐这样简单的任务而言，他说，顾客往往直接来找 DoorDash。
That reverses the usual assumption about automation, which is that the simplest tasks go first. Ordering dinner is simple enough that adding an intermediary makes it slower, while a large purchase involves research and comparison where an agent can save time. If that holds, agentic commerce will show up in high-consideration categories before it shows up in food delivery, and the several-hundred figure is less a failure than a mismatch between the tool and the task.这颠倒了关于自动化的常见假设，即最简单的任务会先被替代。点晚餐简单到增加一个中间环节反而更慢；而大额购买涉及研究与比较，智能体能省时间。若这一判断成立，智能体商务会先出现在高决策成本的品类，而不是外卖；那几百单与其说是失败，不如说是工具与任务的错配。
DoorDash's own agents do moreDoorDash 自家的智能体效果更大
The company's internal AI features tell a different story, though they are not counted in the third-party figure. Ars Technica reported that Ask DoorDash can build a cart from a photograph of a recipe or from a vague request such as dinner for a family of four, and that the company says those baskets are built five times faster than ones built without it.公司内部的 AI 功能讲了另一个故事，尽管它们不计入第三方数字。Ars Technica 报道称，Ask DoorDash 可以凭一张食谱照片、或凭“四口之家的晚餐”这类模糊描述生成购物车，公司称这类购物车的生成速度比不使用时快五倍。
The same report said Ask DoorDash orders average nearly 50% higher basket value and about 60% more unique items than the same customers' traditional orders, and that nearly half end up at local spots the consumer has never tried. Those are the kinds of numbers a platform wants from a new interface. They describe an assistant inside the app, not an outside agent acting for the customer, which is the distinction the earnings-call number was about.同一报道称，Ask DoorDash 订单的平均客单价比同一批顾客的传统订单高近 50%，商品种类多约 60%，且近一半订单流向用户从未尝试过的本地店家。这些正是平台希望从新界面获得的数据。它们描述的是 App 内部的助手，而不是替顾客行事的外部智能体——这正是电话会上那个数字所指的区别。
Who controls the last step of ordering谁掌握下单的最后一步
The strategic question is who owns the interface where an order begins. Xu said the company wants to learn more about agentic use cases and develop a healthy relationship with third-party agent makers, and he raised the possibility of advertising inside those agents in future. That would make the agent a distribution channel rather than a competitor to the app.战略问题是谁拥有订单开始的那个界面。Xu 说公司希望更了解智能体使用场景，并与第三方智能体开发者建立健康关系，他还提到未来可能在这些智能体内投放广告。那会让智能体成为分发渠道，而不是 App 的竞争者。
The rest of the industry is split on the same question. TechCrunch noted that Amazon has been blocking personal agents while Walmart has welcomed them. A platform that lets an outside agent complete a purchase gives up some control of the customer relationship in exchange for volume; one that blocks it keeps the relationship and risks being bypassed. This site's coverage of OpenAI's ServiceNow partnership and of GPT-6's move into interactive interfaces describes the same contest over where agent work happens.行业对同一问题的态度是分裂的。TechCrunch 指出，亚马逊一直在封堵个人智能体，而沃尔玛则对它们敞开大门。允许外部智能体完成购买的平台，会用一部分客户关系换取订单量；封堵者保住了关系，却可能被绕过。本站关于OpenAI 与 ServiceNow 合作与GPT-6 进入交互界面的报道，描述的是同一场关于“智能体的工作在哪里发生”的争夺。
What is established, and what is not哪些已经确定，哪些还没有
Established: third-party AI agents generated several hundred DoorDash orders in total; the company handled just under one billion orders in the third quarter; the connector is limited to early beta partners and mainly corporate lunch; Xu expects agents to matter more for complex purchases than simple ones; Ask DoorDash shows faster baskets and higher basket values. Not established: whether agentic ordering grows once the integration widens, how much revenue agents could bring, and whether DoorDash will let agents carry advertising.已经确定的：第三方 AI 智能体累计只带来几百个 DoorDash 订单；公司第三季度处理接近 10 亿订单；连接器只对早期测试伙伴开放、主要面向企业午餐；Xu 认为智能体在复杂购买上比简单任务更重要；Ask DoorDash 显示出更快购物车与更高客单价。尚未确定的：集成扩大后智能体下单是否会增长、智能体能带来多少收入，以及 DoorDash 是否会让智能体承载广告。
For readers, the useful signal is that a large platform has now put a number on agentic commerce and the number is small. That does not settle the direction; it sets a baseline. The next data point will come from a wider connector, and the test is whether the several hundred becomes several hundred thousand without the company changing how it describes the category. This site's report on Claude Sonnet 5.5's everyday use covers the same question from the model side: how much of ordinary work an assistant actually absorbs.对读者而言，有用的信号是：一家大型平台首次给智能体商务报出了数字，而这个数字很小。这并不决定方向，但它设定了基线。下一个数据点会来自更广的连接器，检验标准是：在描述该品类方式不变的情况下，几百单会不会变成几十万单。本站关于Claude Sonnet 5.5 日常使用的报道，从模型一侧提出了同一个问题：助手究竟吸收了多少日常工作。
Sources & reporting limits来源与核查范围
Follow Industry to find related stories on your next visit.关注产业动态，下次访问时阅读相关文章。
