DoorDash's AI Text-to-Order Agent for Apple Messages

The most consequential read on DoorDash's new agent isn't the convenience pitch, it's defense. Citrini Research argues that AI shopping agents could eventually transact directly with restaurants, routing around delivery apps at a fraction of the cost DoorDash currently charges ^[1]. Alap Shah, the note's co-author, points to DoorDash's roughly 25% take rate on both sides of a transaction as the number an agent-mediated order could compress toward ^[1]. The firm's sharper warning is structural, not just financial: delivery apps have survived on habitual intermediation - the fact that a hungry human reaches for the DoorDash app out of habit - and a machine placing the order has no such habit to exploit, while a 'competent developer' could stand up a rival ordering agent in weeks ^[1].

That's the threat DoorDash CEO Tony Xu has been pushing back on publicly: he positions DoorDash as the orchestrator of the full delivery process even as third-party AI agents might handle the earlier steps of placing an order, arguing the company's logistics and restaurant network stay essential no matter which interface a customer uses to start that order ^[2]. Seen that way, the Messages agent and DoorDash MCP aren't just convenience features, they're DoorDash claiming the agent layer itself before a rival shopping agent claims it instead. Andy Fang's framing leans on the same logic from the user side: 'Every other AI agent is meeting you for the first time. We've known you for years,' he said, pointing to DoorDash's years of order history and stored payment data as a moat a generic agent can't match on day one ^[3].

DoorDash's own rollout copy concedes the agent isn't fully reliable: the company tells users to 'check their carts before paying because the AI can make mistakes' ^[4]. Bloomberg's early testing bore that out, turning up prices in the chat window that didn't match the DoorDash app and recommended photos that didn't match what was actually being ordered ^[5]. Handing a system the company itself flags as error-prone access to stored payment methods is a different risk profile than a human misreading a menu - a wrong price in a chat window is easier to wave through than one sitting in a line-item cart ^[5].

That gap between the pitch and the testing is exactly what soured the feature's reception in its own community thread: commenters questioned why a probabilistic AI agent should replace the deterministic 'reorder' button that already exists in the app, and one specifically flagged that it 'won't be long until someone complains the AI is ordering a crap ton of food on their account that they did not authorize' - a direct read on what happens when an error-prone agent sits on top of a saved card. Several more joked about the agent confidently trying to order from a restaurant that's been closed for years, an organic callback to the same pricing and image mismatches DoorDash has already acknowledged.

The text-to-order agent is live only in the U.S., even though DoorDash owns Deliveroo - acquired for £2.9 billion - and Wolt, which together operate in more than 40 countries ^[5]. The blocker is regulatory, not technical: the EU AI Act has required since August 2 that people be told when they're dealing with an AI system, and Europe's Unfair Commercial Practices Directive creates liability for misleading pricing, a real exposure given that testers were already seeing price discrepancies between the agent and the app ^[5]. That combination means the same testing flaws that are merely embarrassing in the U.S. pilot would be a compliance problem in Europe, which is likely why DoorDash is keeping the agent confined to a 20,000-user U.S. iOS pilot and a fresh waitlist rather than pushing it into Deliveroo or Wolt markets yet ^[3].

DoorDash's own numbers complicate the growth story. Its in-app 'Ask DoorDash' chatbot has driven grocery baskets nearly 50% larger than traditional ordering and sent 46% of restaurant orders between June and August 2026 to merchants the customer hadn't ordered from before, which is why DoorDash is betting conversational ordering drives more spending, not less ^[3]. But early traffic through the new text agent tells a different story: baskets placed through it have come in about two-thirds smaller than a typical DoorDash order ^[3]. The gap suggests the two conversational surfaces are being used differently - Ask DoorDash nudges people toward bigger, exploratory grocery runs inside the app, while a quick text message is being used for small, habitual top-ups rather than full baskets. DoorDash's corporate MCP connector, by contrast, targets a very different basket profile entirely: bulk office orders and supply restocking placed automatically through tools like Slack bots at companies including SpaceX, Cognition, and Mercor ^[6], a B2B channel where order size is set by headcount rather than impulse.