Zero. That’s the number of times you’ll need to open the DoorDash app under the company’s new setup. You text an AI agent inside Apple Messages, say something like “order my usual protein bowl to the office,” and the thing goes off to search local restaurants, scan menus, and build a cart it thinks you want. There’s a US waitlist open now.
I’ve reviewed enough agent demos to know the pattern. The demo sentence is always perfect. “Order my usual protein bowl to the office.” Clean intent, clear item, clear destination. Real humans text like this: “hey can u get me the thing from that place by the gym but no onions and if theyre closed just do thai.” That’s the actual test, and nobody runs the demo on that.
Why the text interface is the smart part
Credit where it’s due. Putting this in Apple Messages instead of a bespoke chat window inside the app is the right call, and it’s the part of this launch I actually respect. Messages is already open on your phone. It already has your keyboard muscle memory. You don’t have to find an icon, wait for a splash screen, dismiss a promo for 30% off wings, or scroll past a carousel of restaurants you’ve never ordered from.
That friction removal matters more than the AI. Most food delivery apps are not designed to get you food quickly. They’re designed to show you things while you’re hungry. Moving the transaction into a plain text thread strips out the merchandising layer, which is a genuinely strange thing for DoorDash to do to itself. Either they’ve decided convenience wins long-term, or the agent will quietly start suggesting upsells once the waitlist crowd settles in. I’d put money on the second.
The trust problem nobody solves with a waitlist
Here’s where I get skeptical. An agent that “builds a suggested cart” is making decisions with your money. Food ordering has a specific failure profile that makes it a rough first outing for autonomous agents:
- Wrong item means wasted money and a wasted meal. Unlike a bad search result, you can’t just hit back. You’ve paid, it’s en route, and you’re eating something you didn’t want.
- Allergies and dietary restrictions are not fuzzy. “My usual” has to encode every modification correctly, every time, across restaurants with inconsistent menu data.
- Substitutions are judgment calls. If the restaurant is closed or the item is sold out, what does the agent do? Guess? Ask? Wait? Each answer annoys a different kind of user.
- Price sensitivity is invisible to the model. Nothing in a text request tells the agent you didn’t want the $19 version of the bowl.
DoorDash says the agent suggests a cart, which implies a confirmation step. Good. That’s the difference between a useful assistant and an expensive liability. But a confirmation step also eats the convenience you came for. If I have to read a cart, verify four line items, check the address, and approve it, I’ve done most of the work of ordering. The agent saved me the browsing, not the deciding.
What this actually signals
Forget the food for a second. The interesting move is DoorDash deciding that a messaging thread is a viable storefront. If that works, the app itself becomes optional infrastructure, which has implications for every company that spent a decade optimizing an app funnel. Your ordering history becomes the product. The interface becomes a sentence.
That’s also the risk. When the interface is a sentence, there’s nowhere to look at competing options. You asked for a protein bowl, the agent picked a restaurant, and you’ll never know which three places it skipped or why. Search ranking becomes invisible, and invisible ranking is where platforms make money. A text thread looks neutral. It isn’t.
My call
Join the waitlist if you order the same thing often enough that “my usual” is a real concept for you. That’s the narrow case where this earns its keep. Repeat orders, simple items, one address, low stakes. For anything involving a group order, a new restaurant, or specific modifications, open the app like a normal person.
The broader question for me is whether this holds up past the honeymoon. Agent products tend to demo beautifully and degrade on edge cases, and food has nothing but edge cases. If DoorDash ships an agent that handles “they’re out of it, pick something close” without producing a plate of regret, that’s a real achievement in applied AI, not just a new front door.
Until I’ve tested it against a messy, badly-typed, mid-afternoon order from a hungry person who can’t commit, I’m treating this as a promising interface with an unproven brain. Text me when it gets my order wrong. I’ll write that one up too.
🕒 Published: