Selected work
A design investigation · TraceFoxUX/UI Design · Case study

AI can generate insights in seconds.

The harder problem is knowing whether they’re worth acting on.

In product discussions, teams rarely questioned what the AI concluded. They questioned how it got there. Without evidence, even accurate insights were hard to defend, and important decisions still meant reopening interviews and documents by hand. The challenge was never generating better summaries. It was designing an interface that made AI reasoning inspectable.

TraceFox is a product-discovery platform. I joined as the UX/UI designer, shaping its interface, flows and interaction logic. This is the investigation behind it — the questions, the rejected directions, and the decisions that stuck.

How did the AI get here?
AI-generated claim
Users abandon onboarding at the workspace-setup step.
Confidence 0.78show the evidence →
Source · Interview 04 · 11:32

“I got as far as the part where you set up the workspace and I just closed the tab, it wanted a bunch of decisions I wasn’t ready to make yet. I never went back.”

Verbatim · the words the claim came from
The design question
How can users verify every insight in seconds?
What I owned
The evidence model, the propose–review–commit loop, traceability and the brief workbench.
Constraint
The AI pipeline was evolving fast, so interaction patterns had to stay stable even as outputs changed.
Status
In active development — Munich, 2026.
01The challenge

Finding information stopped being the bottleneck.

Modern product teams have more research than ever — interviews, support conversations, analytics, product feedback, internal docs. The bottleneck moved. It’s no longer finding the information. It’s turning it into decisions everyone trusts.

Customer interviewsSupport conversationsAnalyticsProduct feedbackInternal documentation

Existing tools failed for opposite reasons.

AI tools
✕ answers without context

Polished summaries, but little confidence. Impressive to read — impossible to defend in a review.

Research repositories
tag · onboardingtag · pricingtag · churn
✕ context without answers

Evidence preserved, but only through constant manual upkeep — so the knowledge quietly dies.

The opportunity was to combine both: answers and the context behind them.

Why it mattered commercially: discovery only pays off when teams can act on research. An insight that can’t be defended in a roadmap review rarely changes the roadmap. So traceability wasn’t a UX nicety — it decided whether research influenced product direction at all.

02Reframing the problem

The model wasn't the problem. The interaction model was.

My first instinct was to improve the AI’s outputs. But after mapping how teams actually worked, something became clear: users couldn’t inspect how conclusions were formed. Without that visibility, every AI-generated insight became just another opinion. That shifted the design question.

We stopped asking
How can AI produce better insights?
We started asking
How can users verify every insight in seconds?

That single shift influenced every design decision that followed. Trust wasn’t going to come from a higher score. It would come from showing the work.

03Design principles

Four principles, not a feature list.

Rather than designing individual features, we set the rules the whole product had to obey.

01
Every claim should be explainable.

Users should always be able to trace an insight back to its original source.

02
AI should propose, not decide.

Generated knowledge starts as a draft until someone reviews the supporting evidence.

03
Uncertainty should stay visible.

Confidence is communicated instead of hidden. Conflicting evidence is surfaced instead of ignored.

04
Knowledge should emerge naturally.

Documentation shouldn't require extra work. The repository grows as teams do their everyday jobs.

04Exploring the solution

How much evidence belongs on screen?

I started from the opposite assumption: that transparency meant showing as much evidence as possible. My first prototypes attached citations to every generated sentence. In review sessions, people trusted the output more — but read it noticeably slower. That finding pushed me toward progressive disclosure. Step through the two directions I rejected and the one that shipped.

Why I shipped it

Evidence stayed one interaction away while preserving reading flow. Users stayed focused until they needed to verify — this became the foundation of the product.

Concept C · expandable evidence — shipped

Users abandon onboarding at the workspace-setup step.

trace to source →
Interview 04 · 11:32 · expanded on demand

“…closed the tab, too many decisions I wasn’t ready to make.”

✓ evidence one interaction away — reading flow intact
05The core interaction

One pattern the whole product repeats.

Instead of automatically turning AI output into documentation, every claim passes through human review. The same four steps happen everywhere, so the mental model never changes.

From AI proposal to shared knowledgeStep 1 / 4
01AI generates a claimDrafted from the sources, shown with its supporting excerpt.
02User reviews evidenceThe claim is opened alongside the words behind it.
03Knowledge is approvedA person edits, accepts or rejects — AI proposes, never decides.
04Repository updatesOnly approved knowledge joins the shared, traceable graph.

AI proposes. A person reviews the evidence and approves. Only then does the shared repository update. There’s never a moment the system quietly decides something on its own.

06Designing traceability

Every insight is connected to its origin.

A claim is assembled in four layers, each grounded in the one below. Selecting any claim reveals the participant quote, interview metadata, supporting observations, confidence, and related evidence. Step it up one layer at a time.

A real sentence from a real interview. Nothing is inferred yet — this is the ground truth every layer above must point back to.

Interview quote · the raw words
Interview 04 · 11:32

“I got as far as the workspace setup and I just closed the tab, too many decisions I wasn’t ready to make. I never went back.”

Rather than asking users to trust the AI, the interface gives them enough context to judge it themselves. Any reader can open a claim and land on the exact sentence a real person said.

Anatomy of a claim

Every element on a claim earns its place. Why each one exists:

Confidence 0.78

Users abandon onboarding at the workspace-setup step.

trace to source → Edit claim
⚠ recalculates when a source conflicts ④
  1. Confidence sits up front, not buried — it invites discussion instead of blind acceptance.
  2. Evidence stays one interaction away — transparency without slowing the read.
  3. The claim stays editable — AI proposes, the person decides what the brief says.
  4. The score recalculates automatically on conflict — the brief never goes stale in silence.
07Designing for contradictions

Research rarely produces one clear answer.

Participants disagree. Markets evolve. Evidence changes. Instead of hiding these inconsistencies, the product treats them as valuable signals. When new evidence conflicts with an existing claim, confidence decreases automatically and the contradictory sources are surfaced for review.

Claim · H1

Onboarding drop-off is concentrated at workspace setup.

⚠ Interview 07 conflicts with H1confidence 0.780.61

The claim remains, but its score falls and the conflicting source is named — nothing gets quietly overwritten.

The conversation shifts
“Is the AI wrong?”
“Do we have enough evidence?”
08A typical workflow

A PM investigating a drop in onboarding completion.

  1. 01

    Instead of rereading dozens of interviews, she asks TraceFox a question. The system proposes several possible explanations.

  2. 02

    One hypothesis points to confusion during workspace setup. Before accepting it, she opens the supporting evidence.

  3. 03

    Only after reviewing excerpts, timestamps and confidence does the insight become shared knowledge.

  4. 04 · weeks later

    New interviews introduce conflicting evidence. The claim remains, but confidence changes automatically. Knowledge evolves without manual maintenance.

Evidence drawer · opened
Confusion during workspace setup
Interview 0411:32
Interview 0904:07
Observations3 linked
Confidence0.78

“…too many decisions I wasn’t ready to make.”

09Key design decisions

A few decisions shaped everything else.

DecisionWhy it mattered
Human approval before publishingPrevented AI outputs from becoming unquestioned facts.
Expandable evidenceBalanced readability with transparency.
Visible confidenceEncouraged discussion instead of blind acceptance.
Automatic contradiction detectionHelped research stay current over time.
Progressive disclosureKept complex workflows approachable.
Outcome

AI positioned as a reasoning partner, not an answer engine.

The clearest signal came from how discussions changed. Product conversations stopped stalling on where an AI answer came from and moved to whether the evidence behind it was strong enough. That shift — from validating summaries to weighing evidence — was the outcome I was designing for.

What changed for the work
  • A claim is defended by opening its source, not re-reading the transcript.
  • Contradictions surface on their own instead of being missed.
  • The repository builds itself instead of needing upkeep.
  • Shared briefs stay checkable — every reader can trace the reasoning.

AI products don’t become trustworthy by making fewer mistakes. They become trustworthy when people can see where a conclusion came from — and designing that moment became the real product.

Reflection

The biggest lesson wasn't about AI. It was about information architecture.

Before this project I thought trust was mostly a function of model quality. Designing TraceFox changed that assumption. Users rarely know how accurate a model actually is — they judge whether they can independently verify its conclusions. So designing that verification layer mattered more than designing the summaries themselves. Explainability wasn’t a machine-learning problem; it was an information-architecture one. That insight still shapes how I approach AI products today.

TraceFox is in active development, heading toward launch from Munich in 2026. This case describes the product as designed and the thinking behind it, rather than adoption to date.

Building something complex? Let’s make it clear.

See more work