Every ecommerce leader we talk to is sitting in some version of the same meeting. A slide goes up. There are three columns. Build. Buy. And now, since the middle of 2025, a third: vibe code.

The third column is new and loud. It says: we don't need to build a platform or buy one — we can prototype something in a weekend with a small team and a handful of agents that write the code for us. And increasingly, teams are shipping those prototypes.

So the question on every slide has sharpened. It isn't "which stack do we adopt?" anymore. It's: when you can ship a working AI feature in 72 hours, what exactly are you choosing between?

What the three options actually mean

Build means you hire the team, own the roadmap, and accept that the first 18 months are infrastructure work before you see any revenue impact. You own everything, including the parts you'd rather not.

Buy means you pay a vendor to have solved the infrastructure problem already. You inherit their shape, their opinions, and their velocity.

Vibe code means you (or a small internal team) use AI-native tooling to prototype and ship a version yourself, quickly, without the full build cost. The surface area is small. The speed is real. So is the fragility.

Where vibe-coded AI breaks in production

Vibe coding could be a genuinely important unlock — the ability to go from idea to working prototype in a week is a structural change, not a fad. But prototypes and production systems are different, and the gap is where most of the interesting engineering lives.

A few questions worth asking before a prototype touches real customer revenue: What happens on the fortieth turn of a conversation, when the happy-path demos are long over and a shopper is asking something no one scripted for? What happens when your model confidently hallucinates a claim that your brand legally cannot make? Who catches a price that was invented rather than retrieved? Six months from now, when the model has seen tens of thousands of real shoppers, has it actually gotten better — or has it plateaued because nothing closed the loop?

None of these are unsolvable. All of them are survivable with the right guardrails, observability, and feedback mechanisms. The question is whether that infrastructure lives inside the prototype you shipped in 72 hours.