When I talk to teams building AI products, I notice a common blind spot: they think about the model and the UI, and they skip everything in between. But the layers between "the model can do this" and "users love this product" are where most of the product work lives.
Here's how I think about the AI product stack.
Layer 1: Model capability
This is the foundation. What can the model actually do reliably? Not what it can do in a cherry-picked demo—what can it do at the p50 and p90 of real-world inputs?
Most teams overestimate this layer. They see the best outputs and assume that's the typical experience. The honest assessment of model capability is the first and most important product decision.
Layer 2: Task framing
Given what the model can do, what user task should it be applied to? This is where most teams go wrong. They frame the task too broadly ("summarize anything") instead of narrowly ("summarize this specific type of document for this specific user need").
Narrow framing is a feature, not a limitation. The tighter you scope the task, the higher the quality, and the easier it is for users to build trust.
Layer 3: Interaction design
How does the user interact with the AI? This isn't just UI—it's the entire interaction model. Does the user trigger the AI or does it run automatically? Can the user edit the output? How does the user provide feedback?
The interaction design determines whether the AI feels like a tool the user controls or a black box they're subjected to. Users adopt tools. They resist black boxes.
Layer 4: Trust calibration
How do you help the user understand when to trust the output and when to verify it? This includes confidence indicators, source attribution, and honest uncertainty.
The goal isn't to make users trust the AI blindly. It's to give them enough signal to make fast decisions about when the output is good enough and when it needs review.
Layer 5: Integration context
Where does this feature live in the user's existing workflow? The best AI features don't ask users to go somewhere new. They show up exactly where the user already is, at the moment they need help.
Using the stack
When I evaluate an AI product or feature, I work through these layers in order. If a team can't clearly articulate their answers at each layer, the feature isn't ready to ship.
The model is the easy part. The product is the hard part.