Key Takeaways

  • Large language models (LLMs) offer superior recommendation capabilities for online commerce.
  • AI-powered ordering agents are a real opportunity to create a new "top-of-funnel" for services.
  • However, frictionless AI ordering alone is insufficient without comprehensive end-to-end fulfillment.
  • The enduring challenge for AI in commerce is mastering the messy, real-world logistics after an order is placed.

The Disagreement

John Collison, Stripe's co-founder, champions the front-end power of AI in commerce. He sees LLMs as natural recommenders, capable of generating far better product suggestions than current interfaces. Collison predicts a shift towards natural language queries, envisioning an "agentic commerce" where users simply ask an AI to handle their shopping.

“Shouldn't we be somehow using the fact that LLMs are pretty good recommenders within products?” Collison asks. He believes this capability is "underutilized" across almost every product.

Tony Xu, CEO of DoorDash, largely agrees with Collison's vision for AI as a new ordering interface. He acknowledges an "opportunity" for products like DoorDash to have their "own ordering agents." Xu sees AI as a powerful new way for customers to interact and initiate orders, acting as a different kind of discovery funnel.

But Xu quickly pivots to the practical reality. He cautions that this exciting front-end development masks a recurring, difficult problem: fulfillment. Past services have offered slick ordering without comprehensive logistics, and they have failed. The challenge isn't just making ordering easier; it's ensuring the order arrives on time, accurately, and with full communication if something goes wrong.

“There's a lot of things that happen after the checkout button,” Xu states. Contacting drivers, managing out-of-stock items, handling delays—these are the real hurdles that AI ordering agents, however advanced, do not automatically solve.

Who's Right (and When They're Wrong)

Both Collison and Xu are right, but Xu's point is the critical reality check that ambitious founders often miss. Collison identifies a powerful new capability. The ease and intelligence of AI ordering agents will indeed reshape how people initiate purchases. That much seems inevitable.

However, Xu reminds us that the fundamental problems of commerce—moving physical goods or delivering services reliably—remain. A beautiful AI interface for ordering a meal means nothing if the food arrives cold, late, or wrong. The promise of "agentic commerce" is thrilling, but its success hinges entirely on the mundane, difficult work of logistics and operations. Founders building an AI product that touches the physical world need to remember this.

Collison's vision is a North Star, but Xu's caution highlights the difficult path to get there. Prioritize the backend before the front-end magic scales.

What to Do With This

If you're building an AI-powered commerce or service, don't just build the LLM interface. Before you even write the first line of AI code, map out every single point of potential failure after the user clicks "order" or "confirm." Pick one specific "agentic commerce" feature for your product. Then, spend two full days with your team role-playing every single post-order scenario, from the simplest success to the worst-case failure. Document precisely what your system (human or automated) would need to do in each case to prevent a customer from having a bad experience. This non-AI work is where your product will actually live or die.