DoorDash AI Agent Orders Still in the Hundreds, Says Company

What happened

DoorDash has said that orders placed through outside AI agents have so far reached only the hundreds, a small figure for a food-delivery platform of its scale. Reporting by Ars Technica, which carried the story under the headline "Basically no one is using AI agents for their DoorDash orders", frames the finding as evidence that agent-led ordering is still at an early stage.

The development matters because it tests a widely discussed idea: that AI agents acting on a customer's behalf will become a significant new route into e-commerce. Ordering a meal is a useful test case. It is frequent, low-value and time-sensitive, so it should be one of the easier tasks to hand to an agent. The early numbers suggest that the capability to place an order and actual customer use are still some distance apart.

Why it matters

For leaders working on AI and digital transformation, the gap between what an agent can technically do and what customers choose to do with it is the central issue. An agent can only complete an order reliably if the platform exposes accurate menus, prices, availability, delivery options and payment flows in a form the agent can use. Each of those touchpoints has to be designed, governed and maintained. Where they are not, the agent becomes another intermediary that adds friction rather than removing it.

For experience and service design teams, the more important question is trust. A customer who hands over an order to an agent is delegating a decision and a payment. Unless they can see what was ordered, confirm it easily, and correct it without difficulty, the convenience of automation is quickly outweighed by the effort of checking. Low early uptake may therefore reflect the cost of building that confidence as much as any shortfall in the technology.

The Renascence take

Early adoption figures for agentic commerce tend to be read as a verdict on the technology, when they are more usefully read as a verdict on the experience around it. The question for operators is not whether agents will eventually order food, but whether a customer who tries one once will want to do it again. Designing for that second order is where the real work sits.

Most readers will focus on the word "hundreds" and conclude that agents have failed. The more useful reading is that a new channel has barely been designed for, because it has not yet been given the trust signals, confirmation steps and recovery paths that habitual ordering relies on. Behavioural principles such as default settings, clear feedback and easy reversal will decide whether an agent feels like a helper or a risk. A customer-obsessed operator should make its menus, pricing and order states legible to machines and reassuring to humans, and should measure repeat use rather than first use.