Cover · source-attribution chat (16:9)
Every claim carries the passage it came from, one click away.

AI · Trust & Verification · 2025

Designing for AI — trust, verification, and getting out of the PM's way

Two problems that turned out to be the same problem: people won’t use an answer they can’t verify, and they won’t opt into a feature they can’t scope.

Problem

Two new problems with no pattern to copy

How do you design an AI chat experience people will actually trust, rather than treat as a black box they have to fact-check somewhere else? In a regulated industry, a confident-sounding answer with no provenance costs more to verify than it saves.

The second problem looked unrelated at first. Our AI features all sat behind a single global switch a PM had to flip manually — feature by feature, week after week — just to manage beta access. Neither problem had an existing UI pattern to borrow.

Process

Attribution as an interface, not a footnote

For AI chat, the design problem wasn’t the conversation UI — that part is becoming a solved pattern industry-wide. It was trust: making it obvious that an answer was actually grounded in the user’s own documents, not hallucinated. I designed source attribution directly into the chat experience, so verification is part of reading the answer rather than a separate step someone has to remember to take.

For the feature switch, I didn’t start with implementation. I prototyped the concept first, specifically to get PM, design and engineering aligned on the right approach before any real engineering time got spent — replacing the single global toggle with granular, per-feature access.

Before — a single global beta toggle
Before
After — per-feature access, scoped per team
After
Trust first. Then get out of the way.
People don’t want less AI. They want to know where the answer came from, and to decide how much of it they’re signing up for.

Technical depth

What the work actually involved

  • Designed and built AI chat with inline source attribution, tying generated answers back to their originating documents.
  • Replaced a global feature flag with granular per-feature toggles, giving users self-serve access instead of manual PM intervention.
  • Built our first real AI-powered search experience in the product — another zero-to-one UI problem with no internal precedent.

Outcome

The chat experience gives users a built-in way to verify AI output instead of trusting it blindly, and the feature switch redesign eliminated a recurring manual workflow entirely. Both shipped as genuinely new patterns, not adaptations of something that already existed.

Motion clip — hovering a claim to reveal its source