Configurable AI Scribe System

Design a consent-aware AI note generator that clinicians can trust and patients can understand.

Configurable AI Scribe System

Problem

Clinicians would not adopt AI note generation if they felt it created liability. The primary barrier was not capability — the AI was accurate — it was transparency around when it was being used and what it was generating. Without clear consent flows and editable outputs, adoption stalled at under 10%.

Details

My role: AI Product Designer
Duration: 4 months
Responsibilities: AI UX design, consent flows, output configuration

Tools

FigmaOpenAI APIVercel AI SDK

Highlights

This project was a chance to explore the intersection of AI capability and clinical trust. The design challenge was not making the AI smarter, but making its behavior legible and controllable.

CHAPTER 1

Discovery

POWERED BY AI (LOW CONFIDENCE)

The trust problem with AI in clinical settings

Clinicians would not adopt AI note generation if they felt it created liability. The primary barrier was not capability — the AI was accurate — it was transparency around when it was being used and what it was generating.

  • Only 8% of clinicians trusted AI-generated notes without manual review
  • Liability concerns outranked accuracy concerns 3-to-1 in user interviews
The trust problem with AI in clinical settings

CHAPTER 2

Design

Consent-first, configurable output

Every AI generation is preceded by a clear consent confirmation screen. Clinicians can select output style (SOAP, narrative, bullet), tone, and detail level before generating. The output is always surfaced as a draft — never auto-saved.

Consent-first, configurable output

CHAPTER 3

Validation

Testing with real clinical workflows

Three rounds of usability testing with GPs and specialists revealed that the configurable output panel reduced post-generation editing by 40%. Consent wording was iterated twice based on patient feedback.

Outcomes

70%

Reduction in note-writing time

4.6/5

Clinician satisfaction score

100%

Consent-before-generate coverage

Retrospective

Lessons

The consent screen was initially seen as friction — but it became the feature clinicians cited most when recommending the tool to colleagues. Trust is a feature.

Tradeoffs

We traded a fully automatic mode for an always-manual-review flow. Some power users wanted faster generation, but clinical safety required the draft review step.

Next step

Want to work together?

Get in touch