Ariadesigning an AI copilot people can actually trust
A self-initiated exploration of agentic AI for product teams. What would it take for an AI to do real work on a team's behalf, while every person stays in control of what goes out the door?
AI features in SaaS tools look great in demos, then get quietly ignored by week two. Usually it's not because they're wrong. It's because people can't tell what the AI did or why.
I made up a product called Lumen and designed Aria inside it: a copilot that drafts, flags and connects research, but never sends or publishes anything without a person signing off.
A seven-screen system built on one idea: show the reasoning, ask before acting, and make every action reversible.
Most AI features feel like a black box
This project started with a question I kept running into in my own work: why do AI features in SaaS tools feel so unsatisfying to use? They impress in a demo, then people stop opening them. I wanted to design an answer from scratch, with no constraints except being honest about the problem.
To do that, I created a fictional B2B product platform called Lumen. The scenario is a product team in the middle of a launch, spending a big part of their week on coordination work: synthesizing research, writing PRDs and tracking sprint health. The AI tools they have feel too generic to trust with anything important. Lumen doesn't exist. The problem it stands for does.
…let an AI take real work off a product team's plate, without anyone losing track of what it did, why it did it, or how to undo it?
Grounding a concept in real patterns
Since this was a concept, I didn't have a user base to study. So I pulled from three places and turned what I found into design scenarios, realistic situations that every screen had to hold up against.
Four patterns kept coming up, and each one shaped a part of the design.
"I don't want AI to tell me what to build. I want it to do the stuff I have to do before I can think."
A feeling that came up again and again in conversations with product practitionersDesigning for the cautious adopter
I combined the research into one primary persona to anchor every decision. Sofia is a composite, not a real person, but everything about her comes from patterns I heard and read.
Mid-size B2B SaaS · 80-person product org
Cautiously optimistic. Trust has to be earned a little at a time.
Seven years in UX research. About 40% of her time goes to synthesis, write-ups and stakeholder updates instead of research. She's tried three AI tools in the past year and kept none of them past week two.
- Get research into product decisions faster
- Spend more time talking to users, less time writing up
- Stay credible with her PM and eng leads
- Control how her research gets represented
- AI that oversimplifies nuanced findings
- No way to check what an AI did or why
- Pressure to use AI she doesn't trust with final work
- Insights scattered across Dovetail, Notion and Jira
Four rules I set before designing a single screen
Every design decision after this traces back to one of these.
The higher the stakes, the more a person is involved
The real question in agentic AI isn't what the AI can do. It's what it should do without asking. I sorted every Aria behavior into four levels, so the line between helping and overstepping was a design decision, not an accident.
Following one finding through the whole product
The seven screens aren't separate features. They tell one story: a team discovering that new users are overwhelmed by notifications, and Aria carrying that finding from raw interviews all the way to a PRD, with a record at every step.
Seven screens, one system
The hardest balance was making Aria visible enough to be useful and quiet enough to stay out of the way. Here's how each screen handles it. These are the live prototypes, so the panels scroll.
Dashboard
Aria is always there, but never in the way.
- 01Familiar on purposeThe main layout stays familiar. Aria lives in a panel on the right and never covers the actual work.
- 02Reasoning on every cardFour insight types (risk, synthesis ready, draft ready and pattern found) each end with a one-line reasoning note.
- 03Teal chips show provenance"Aria" chips on tasks show at a glance what the AI touched and what a person did.
Research synthesis
A first draft the researcher signs off on.
- 01The researcher owns the conclusionA banner says exactly what Aria did and when. The main action is "Approve synthesis."
- 02Themes tie back to peopleEvery theme links to real quotes from named participants in the side panel.
- 03Honest about the dataOpen questions call out what the data can't answer, like the fact that only new users were interviewed.
Sprint board
Aria shows up where the work already is.
- 01Edges, not a new tabCards Aria worked on get a teal edge and a short note. Risks, like a review with no meeting booked, turn coral.
- 02It asks before actingWhen four tasks stall for five days, Aria asks before flagging them and drafting a status update.
PRD editor
An AI that's willing to say "I don't know."
- 01Plain on purposeThe editor feels like a normal doc. A banner shows who drafted it and what it was based on.
- 02Open questions, not guessesInstead of filling gaps with confident guesses, Aria lists open questions and routes each one to a named owner.
- 03Every requirement is traceableThe sources panel traces requirements back to the research they came from.
Insights repo
Research that doesn't die in a slide deck.
- 01Evidence and follow-throughEach insight shows how many sessions and quotes back it up, and whether it's linked to a PRD yet.
- 02Aria connects the dotsIt spotted that notification fatigue and onboarding drop-off overlap, and suggested fixing both in the same onboarding step.
Aria setup
Teams turn Aria on one capability at a time.
- 01Plain-language togglesEach toggle says what Aria will actually do, not just "enable AI features."
- 02A preview you can pictureA preview column shows the cards you'll start seeing, because people trust what they can picture.
- 03A promise up frontAria stays in the workspace, and everything is logged and reversible.
Activity log
Proof that Aria has nothing to hide.
- 01Status and undo on every actionEvery action shows whether it's waiting on you, done or undone, and can be reversed right there.
- 02Misses are shown, not hiddenOne entry shows a teammate undoing a summary because she wanted to write it herself.
- 03A full reasoning traceIt ends with Aria keeping the PRD in draft and notifying no one until a person reviews it.
Three calls that defined the product
What the design delivers, and how I'd measure it
Concept project. These are design outcomes, not results from a live product.
Design goals and where they live
| Goal | How it's handled | Where |
|---|---|---|
| AI actions are clear | Reasoning note on every insight and draft | Every screen |
| People stay in control | Nothing goes out without approval | System-wide |
| Mistakes are recoverable | Full log with undo on any action | Screen 7 |
| Research reaches decisions | Interviews → synthesis → repo → PRD | Screens 2–5 |
| Helpful, not pushy | Aria suggests and flags. People decide. | Screens 1, 3, 5 |
If it shipped, I'd track
| Metric | What it tells me |
|---|---|
| Use after week 4 | Whether Aria survives the point where most AI tools get dropped |
| Draft approval rate | Whether drafts are a real head start or get rewritten from scratch |
| Undo rate by capability | Which capabilities people don't trust yet |
| Interview to approved synthesis | The actual time saved on the most tedious step |
What I'd do differently
AI design is a systems problem, not a feature problem
The hardest questions here weren't visual. They were about trust, transparency and control, and where helpful automation turns into unwanted interference.
It also shows how I take something as open-ended as "AI for product teams" and give it structure: a clear scenario, grounded research, a set of principles, and a story that holds together across every screen.