Founders email me about AI features the way they'd email me about magic. They want to know which model, which provider, which framework. By the second email I'm gently steering us toward the actual problem, which almost never lives in the model.
Most AI features I build take roughly a week of model work and three weeks of interface design. That ratio shocks people. It shouldn't. The model is a component. The product is the work that goes around it.
Here is what "the work that goes around it" actually looks like. The empty state before the user has typed anything. The loading state while the model is thinking. The state where the model returned nonsense and we have to admit it. The state where the model returned something good but the user wants to push back on it. The undo. The history. The handoff to a human.
None of those states are in the model. All of them are in the product. And all of them are where users decide whether they trust the feature or not. Trust is the deliverable. The model is just the engine.
The other thing the model rarely solves: the input. Most AI features fail because the user couldn't describe what they wanted in the way the system wanted to hear it. The interface design problem is the gap between the user's intention and the prompt the model needs. Closing that gap is most of the engagement.
I've learned to start every AI build by asking what the feature looks like when the model gets it wrong. If the answer is "it just doesn't work," we're not ready. If the answer is "the user can edit, retry, or fall back to the manual flow," we're closer. Designing for the failure mode is designing for the feature.
The model itself? GPT, Claude, Gemini, the open-source ones — they're all close enough that the choice rarely changes the product. Pick the one that matches your latency, cost, and policy constraints, then move on. The interesting work is upstream and downstream of the call.
If you're scoping an AI feature and the conversation is mostly about the model, the conversation is in the wrong place. Move it to the user, the failure mode, and the moment of trust. That's where the engagement actually lives.