DoorDash said it’s only received ‘several hundred’ orders from AI agents

We have been hearing for a year that AI agents are going to replace our fingers. The promise was simple: you tell your phone you are hungry, and a sandwich shows up thirty minutes later without you ever opening an app. It sounds like the logical conclusion of the convenience economy. But the reality, as revealed by DoorDash recently, is a lot quieter.

DoorDash leadership confirmed that they have only seen a few hundred orders processed through third-party AI agents. In a company that moves millions of orders a day, a few hundred is effectively zero. It is a rounding error. For those of us building in the AI space, this is a necessary reality check. The gap between a cool tech demo and a reliable consumer habit is currently a canyon.

The Friction Problem Nobody Talks About

When you use an app like DoorDash, you aren't just clicking a button. You are navigating a complex web of real-time variables. Is the restaurant open? Is the specific item in stock? Does the price on the screen match the price at checkout? Are there dynamic delivery fees based on a rainstorm happening three miles away?

Humans are surprisingly good at filtering this noise. We see a "sold out" banner and pivot to our second choice in two seconds. Current AI agents, mostly built on top of Large Language Models, struggle with this level of dynamic uncertainty. Most of these "hundreds" of orders likely came from developers testing their own scripts or power users willing to babysit a buggy automation. For the average person, it is still faster to just tap the screen themselves.

Infrastructure Over Hype

DoorDash isn't being hostile to the tech. In fact, they have stated they are open to working with third-party agent builders. This is a smart move. They don't care if a human or a bot places the order, as long as the transaction clears and the food gets delivered. The bottleneck isn't gatekeeping; it's the lack of standardized APIs for agents to talk to platforms.

Right now, most agents try to "see" the web like a human does, using computer vision to click buttons. It’s brittle. If DoorDash moves a button three pixels to the left, the agent breaks. For builders, the takeaway is clear: stop trying to hack the front end. The future of the agentic economy depends on back-end integration. We need a way for agents to query data directly without the overhead of a visual interface.

The Trust Gap in Autonomous Commerce

There is also a significant psychological barrier. Giving an AI agent access to your credit card to make purchases on your behalf requires a level of trust we haven't reached yet. If an agent hallucinates and orders forty pizzas instead of one, who is responsible? DoorDash isn't going to eat that cost, and the agent developer probably won't either.

Until there are clear guardrails and "proof of intent" protocols, people will be hesitant to let agents loose with their bank accounts. We are seeing the same hesitation in the crypto space with automated trading bots. The tech works until it doesn't, and when it doesn't, the loss is permanent.

What This Means for Builders

If you are building in AI or Web3, don't be discouraged by these low numbers. Be encouraged by the lack of competition. Everyone is busy making wrappers that summarize PDFs. Very few people are solving the hard problem of reliable, cross-platform execution.

• Focus on reliability: A tool that works 90% of the time is useless for commerce. It has to be 99.9%.
• Standardization is key: We need open protocols that allow agents to verify prices and availability without scraping.
• Solve for the edge cases: Building an agent that orders a burger is easy. Building one that knows what to do when the burger is out of stock is the real business.

The Long Game

We are in the "dial-up" phase of AI agents. Just because the numbers are low today doesn't mean the category is dead. It means the current approach isn't working. DoorDash admitting they are open to collaborators is a massive green flag for founders. They are effectively asking for someone to build a better bridge.

The first company that creates a truly invisible, reliable ordering layer will capture a massive chunk of the market. But that won't happen through better prompts. It will happen through better engineering and deeper integration with the platforms that actually hold the inventory.

The infrastructure for autonomous commerce is still being laid. Don't mistake a slow start for a lack of destination.

For now, the manual tap remains king. But keep an eye on those "several hundred" orders. When that number hits several hundred thousand, we will know the bridge has finally been built.

Read the original at TechCrunch AI →