Founders tend to email me about AI features the way you'd email someone about magic. Which model, which provider, which framework. By the second email I'm usually trying to steer us gently toward the actual problem, which almost never turns out to live in the model.
The AI features I build take roughly a week of model work and about three weeks on everything else. That ratio surprises people. I don't think it should.
Here's what everything else looks like in practice. The empty state before anyone has typed anything. The loading state while the model is working. The state where it returned something wrong and you have to admit that. The state where it returned something decent but the user wants to argue with it. Undo. History. Handing off to a person when it's clearly out of its depth.
None of that is in the model, and all of it is where someone decides whether they trust the feature enough to use it again tomorrow.
The other thing the model doesn't solve for you is the input. Most AI features I've seen fail did so because the user couldn't describe what they wanted in the way the system needed to hear it. Closing the gap between what someone means and what the model needs is most of the design work in these projects.
So I've got into the habit of starting every AI build by asking what the feature looks like when the model gets it wrong. If the answer is that it just doesn't work, we're not ready to build it. If the answer is that the user can edit it, retry, or drop back to doing it manually, we're much closer.
The model itself matters less than people expect. GPT, Claude, Gemini, the open-source ones, they're all close enough now that the choice rarely changes the product much. Pick the one that fits your latency, your costs and whatever policy constraints you have, then get on with it. The interesting work is either side of the call, not in it.
If you're scoping an AI feature and the conversation is mostly about which model, the conversation is in the wrong place. Move it to the user, to what happens when it goes wrong, and to the moment where someone decides whether to trust it.