Larry Contratti ← Writing
AI-Native Design

What AI-Native Design Actually Means

It isn't an AI feature bolted onto an old interface. It's designing the product around what the model can do — and it usually means the interface does less, not more.

"AI-native" gets used to mean almost anything now. Here's the definition I design by: an AI-native product is one whose core experience only makes sense because a model is doing the work — where the interface is shaped around the model's capability instead of the model being stapled onto an interface that already existed. In practice, that shift shows up as the interface asking for less, explaining more, and behaving like a colleague rather than a form.

Most "AI features" I see are the opposite. They take a screen that already worked — a form, a search box, a dashboard — and add a button that says "✨ do it with AI." The workflow underneath is unchanged. The user still has to know what to ask, still has to translate their intent into the app's schema, still does the structural work. The model is a garnish. That's AI-enabled. It's not AI-native, and users can feel the difference immediately.

AI-native means the interface does less

The clearest signal of AI-native design is subtraction. When the model can read, extract, and structure information, the fields a person used to fill in should disappear.

On VOCAL — an AI-native intake tool I led for Amazon's government business — the old process was a roughly 30-minute Salesforce form that almost nobody completed. The AI-native version flipped it: the user hands off a note (typed, or even a handwritten scan) and an assistant reads it, extracts the fields, and fills the record. The user never fills a field. That single design decision moved intake from 26 records to 450 in an hour and adoption from 4 users to 57. The win wasn't a smarter form. It was deleting the form.

Explainability isn't a feature — it's the interface

When a model makes a consequential call, the design problem is no longer "show the answer." It's "let a person trust, question, and override the answer." If they can't, the model is a liability no matter how accurate it is.

SCRAM, a supply-chain risk platform I designed, replaced a ~$650K-a-year tool whose scoring was a black box — and that opacity was itself the audit risk. The AI-native move was to make every score interrogable: a person can ask why a supplier scored the way it did and get a real answer, dimension by dimension. Trust came from transparency, not from a confidence percentage. Risk-to-mitigation went from 4 days to 2.5 hours — because people could act on results they actually believed.

If AI drives behavior, you have to version it like code

Here's a problem traditional design never had to solve: when a feature's behavior comes from a model, that behavior changes, and teams need to ship, compare, and roll it back with the same confidence they have with code. That's a design problem as much as an engineering one — it needs an authoring experience and a mental model, not just a config file.

On AI Feature Versioning at Capital One, the original spec imagined one kind of user. The real system had two: producers who ship versions and consumers who have to trust them. Designing for both — a way to publish, and a way to evaluate before adopting — took platform satisfaction from 3.1 to 8.9 and quadrupled sharing. AI-native products are living things; the design has to account for their version history.

An AI-native product gives advice, not results

Search is what the market knows how to build. But once a model can reason over a domain, users stop wanting ten blue links — they want a point of view. Fedlane, a product I designed, built, and shipped solo, surfaces federal opportunities months before they become public RFPs and tells you what to do about them. The interface is a strategic advisor, not a search box — and the whole design bet was that 97% answer accuracy plus a clear rationale beats a list every time.

What this means for how you design

If you're moving from AI-enabled to AI-native, the questions change. Instead of "where do we add AI to this screen," ask:

None of this is about the model itself. It's about designing the product so the model's capability becomes the user's experience — which is, and has always been, the actual job. AI just raised the ceiling on how much of the work the interface can quietly take off the user's plate.

LC
Larry Contratti
Design lead working on AI-native products — turning complex internal systems at companies like Amazon and Capital One into tools that unlock real business value. Get in touch.