Kern
The AI copilot that works from draft to deliverable
Paula La Rosa, Product & Service Designer and Design Engineer
- Sector
- AI Product · Mobile
- Role
- Product Design · UX/UI · Design System
- Year
- 2026
Context
Kern is a mobile AI copilot for product and design teams. It talks to your sources, generates editable artifacts like PRDs, flows and syntheses, and runs tasks with agents. The visual language is polished glass over silver, with streaming answers, source citations, voice mode and human approval, in a prototype you can navigate end to end.
Problem
Today's AI tools are chat boxes. You ask, copy the answer and paste it somewhere else. The middle is missing: turning the conversation into a deliverable, knowing where the answer came from so you can trust it, and delegating the repetitive work without letting go of control. And almost all of it lives on desktop, far from the people deciding in meetings.
Solution
I designed Kern around three patterns that AI in 2026 demands and almost nobody solves well on mobile. Streaming with source citations, for trust. An editable artifact one tap from the conversation, to become a deliverable. Agents with a human in the loop, to delegate without letting go of the wheel. All on a light design system of glass and silver, with tokens, soft elevation and consistent icons.
What the research showed
How it was done. Contextual interviews with product and design people, following how each one uses AI in a normal week, plus a diary of chat use on the phone during meetings. The findings became an opportunity map, and each solution was tested as a prototype before going into the app.
Findings
- The answer is not the end, it is where the work starts: People who use AI at work copy the answer and rewrite it somewhere else. The value was in leaving the conversation with a document ready to edit.
- Without a source, nobody takes the answer to a meeting: People trust AI for a draft, not for a decision. To carry an answer forward, they need to show where it came from.
- Delegating is scary when you cannot see the middle: Long agent tasks felt like a black box. The wish to delegate is there, but only with checkpoints for approval.
Personas
Lia, 34, Product manager on a team of six. Spends the day in meetings and writes PRDs and syntheses in between, almost always on her phone. Needs: Turn the meeting conversation into a document before she forgets. Gets in the way: Copying and pasting between chat and editor, without knowing where the AI got each sentence.
Rafael, 41, Head of design. Reviews the team's work, sets priorities and uses AI for quick research. Needs: Delegate repetitive research without losing control over what reaches the team. Gets in the way: Agents that act on their own and results with nothing behind them.
Personas composed from the research. Names and ages are illustrative.
What was left out
I discarded the assistant that acts without asking: in testing, people undid more than they approved. Chat as the main screen was also left out. It became the way into the artifact, not the destination.
How to know it worked
- How many conversations become an edited, shared artifact.
- How many answers have their source opened before being used.
- How many agent tasks finish with approval instead of being cancelled midway.
Design decisions
Trust as a requirement
Every answer carries its source citation and an honest "generating" state. The person sees where the information came from before trusting it. The opposite of the black box that writes beautifully with nothing behind it.
From chat to artifact
The AI does not just answer. It generates an editable, versioned document with a severity tag, which opens from the conversation and gets refined by talking. It closes the loop a plain chat leaves open.
Agents with a human in the middle
Tasks run as visible steps on a timeline and pause for approval before acting. The person sees each step and decides what goes in. Automation without losing control.
Polished glass, no AI look
A cool silver world, with glass only on what floats: the tab bar and the agent notice. Content sits in solid lists with hairlines. Ink and a single steel accent, metadata in mono and agent state in functional color with checked contrast. Tokens with light and dark modes and components ready for code.
Results
- From question to deliverable: The editable artifact closes the loop the chat leaves open. The answer is born as a PRD, flow or synthesis ready to use.
- Trust you can see: Honest streaming and a source citation in every answer. People trust it because they see the origin, not on faith.
- Delegate without letting go: Multi-step agents with human approval. The machine does the work, the person approves what matters.