Notes on calm AI interfaces

Most AI workflow UI tries to hide the model. The output appears, polished and confident, and the user has to take it on faith. That works until it doesn't — and when it breaks, there's nowhere to go.

I think the better posture is calm exposure. Show every extraction. Show every confidence score. Show the model call as a step on a timeline. The user doesn't need to understand the internals; they just need to know that there are internals, and that they can dig.

Three patterns I keep reaching for

Citations everywhere. Every claim the model makes should be attached to the source it came from — a span in the document, a row in the database, a tweet. The citation is a hyperlink to the evidence. This single move closes most of the trust gap.

Confidence as a visible attribute. Not a single global "confidence: 87%" number — that's noise. Instead, badge each extraction with its own score and let the reviewer set a threshold for what they want to manually check.

Process steps as a timeline. When the AI does a multi-step thing (read doc, extract entities, classify, summarise), show those steps. Each one is hover-able to reveal its inputs and outputs.

The fintech idiom

I styled the Underwriter prototype in a fintech idiom on purpose. That audience is allergic to surprise. The visual restraint forces the AI behaviour to be the interesting thing, not the chrome.

There's a deeper reason: fintech UIs have spent decades earning trust by being explicit. Pending. Cleared. Settled. Every state has a name and a colour and a place in the timeline. Calm AI interfaces want to learn that vocabulary.

The rule I'm landing on: never make the model speak in the first person. Every "I" is a chance for the AI to overclaim. Better to caption: Extraction · 94% · Source: pp. 12–14. Same information, none of the personality.