Everyone is treating DoorDash’s new texting agent as the moment AI agents finally got practical. I think it’s close to the opposite. The model work here is the least interesting part of the announcement. What DoorDash actually shipped is a distribution move dressed up as an agent, and that distinction matters a lot if you’re deciding whether to care.
Here’s what we know. DoorDash has launched an AI agent that takes food orders through Apple Messages. You text something like “order my usual protein bowl to the office,” and the agent searches local restaurants, scans menus, and builds a suggested cart from your request. A US waitlist is open now. The whole thing runs on AI behind the scenes.
Read that again and notice what the agent is doing. Search. Parse menus. Assemble a cart. Those are three capabilities DoorDash has had in production for years, exposed through its app’s search bar and recommendation engine. The new layer is a natural language front door and a messaging channel. That’s a real product decision with real consequences, but it’s not a new capability.
Why the channel is the actual story
The reason I find this worth writing about is that DoorDash put its ordering flow inside the app people already have open. Not a new app. Not a browser tab. The same thread where you coordinate lunch with a coworker. That’s a meaningful reduction in friction, and friction is the entire competitive game in food delivery.
If you’ve reviewed enough of these tools, you start to notice a pattern. The agents that stick are almost never the ones with the smartest reasoning. They’re the ones that removed a step you didn’t realize you were tolerating. Opening an app, waiting for the home feed to load, scrolling past promoted restaurants, searching, filtering, tapping through modifiers. The texting agent collapses that into one sentence. Whether the AI underneath is impressive is almost beside the point.
The flip side is that messaging is a terrible interface for correction. Which brings me to my actual concern.
What I’d want to test before trusting it
A suggested cart is a prediction, and predictions in food ordering fail in expensive, annoying ways. DoorDash says the agent builds a cart based on your request. It does not say the agent is right. Those are different claims, and the gap between them is where most agent tools fall apart in daily use.
The things I’d stress-test on day one:
- Modifier handling. “My usual protein bowl” implies stored preferences. What happens when the restaurant changes its menu, retires an item, or renames a size? Does the agent substitute silently or stop and ask?
- Confirmation before money moves. Text is a low-ceremony medium. A cart that becomes an order without an unambiguous confirmation step is a refund request waiting to happen.
- Ambiguity behavior. If three nearby places serve something matching your description, does the agent pick, or does it present options? Picking is faster. Picking is also how you end up with lunch from a restaurant you actively dislike.
- Correction cost. Fixing a wrong cart in an app takes two taps. Fixing it over text takes a sentence, a reinterpretation, and a hope. If the repair loop is slower than the original flow, the convenience evaporates.
- Who the suggestions serve. DoorDash is an ad business as much as a logistics business. An agent that chooses for you is an agent with enormous power over what gets chosen. I’d watch for whether promoted merchants show up in suggested carts, and whether that’s disclosed.
The waitlist tells you something too
A US-only waitlist is a sensible way to ship this, and it’s also a signal. You gate a launch when you’re not confident about edge cases at scale, or when the cost per interaction is high enough that you want to meter it. Either reason is fine. Both mean the version early users get will not be the version that matters. I’d hold judgment on quality until the gate opens.
Reports also point to DoorDash building conversational ordering elsewhere in its product, which suggests the texting agent is one surface in a broader push rather than a standalone experiment. That’s the more useful frame. Companies with existing search, catalog, and payment infrastructure are now wrapping conversational layers around them and shipping into whatever channel their users already live in. Food delivery is just an early, legible example because the task is simple and the stakes per transaction are low.
My call
Worth joining the waitlist if you order from DoorDash regularly and your orders are repetitive. Repetition is where this design wins, because the prediction problem is nearly solved for you already.
Skip it if you browse before ordering, change your mind often, or care about which specific restaurant fulfills your request. For you, a text thread is a narrower interface than the app, not a wider one. The agent framing makes it sound like you’re gaining a capability. In practice you’re trading visibility for speed, and that trade only pays off when you already know exactly what you want.
đź•’ Published: