A friend of mine is building an AI-powered learning app. During a conversation about his product roadmap, he mentioned something from a podcast that stuck: design your AI product for six months from now.
The logic is straightforward. The model underneath his app will be meaningfully different in six months. GPT-4 to GPT-4o took about a year. Claude 3 to Claude 3.5 took less. Each jump changes what’s possible: cost per query drops, multimodal capabilities improve, reasoning gets more reliable. His users will be different too. Six months of living with AI tools changes how people prompt, what they expect, and what feels like baseline versus what feels like magic.
Most product teams don’t build this way. They design for the current model’s capabilities and the current user’s behavior. That makes sense for stable technology. It’s a trap for AI products.
Consider the onboarding problem. Right now, many AI apps invest heavily in teaching users how to prompt well: guided templates, example queries, curated starting points. That’s a real friction point today. But as AI literacy spreads and models get better at interpreting vague inputs, that entire onboarding layer becomes dead weight. You’ve built scaffolding for a wall that no longer exists.
My friend’s learning app faces the same question. Right now, users don’t know what to ask. So his team built structured “cocktail learning” paths that package topics into digestible modules. That’s a valuable product today. But in six months, if models can generate adaptive curricula on the fly from a single prompt, the structured packaging becomes less important. What remains valuable is the pedagogical framework underneath: when to challenge, when to support, how to sequence difficulty. That’s the stable layer.
This is the practical filter: for every product decision, ask whether it still matters if the model gets twice as good. If the answer is yes, you’re building something durable. Core insights about your user, the real problem you’re solving, the workflow that matters regardless of what the model can do. Those are worth overinvesting in. If the answer is no, hold the design loosely. Build it to be replaced.
The companies that get this right will look obvious in hindsight. They’ll be the ones who seemed slightly ahead of their users instead of perfectly matched to them.