AI UX- Domain Expert AI Design Prototype — KENNETH HUNG | Design Leader

school Learning Mode

Domain Expert AI — TL;DR — Ken Hung

Prototype 02 Medical AI · High-Stakes Domain Designing for Different AI Archetypes

Domain Expert AI

High-Stakes Decision Support

Showing confidence badly makes outcomes worse, not better.

User Problem

Experts are being misled by the interfaces meant to help them.

Domain experts — clinicians, lawyers, financial analysts — use AI to augment decisions that carry real consequences. The instinct is to show a confidence score and call it transparency. Research says otherwise: displaying confidence badly causes expert accuracy to drop, not rise. A radiology study found accuracy fell from 82% to 46% when AI confidence was shown incorrectly.

The problem isn't the AI model — it's the interface. Experts who trust a high-confidence wrong answer more than a low-confidence right one aren't making bad decisions. They're responding rationally to a badly designed signal. In specialized AI, UX is not decoration — it is a safety layer.

Design Principles

Five principles of Domain Expert AI UX

Domain Guardrails

Specialized models have boundaries, and those boundaries must be visible. When the AI declines to evaluate something outside its training, users should understand why — not receive a hallucinated answer that exceeds its scope. An explicit AI Scope indicator prevents both over-reliance and false confidence in the model's range.

Transparent Reasoning Chains

Experts don't just need the conclusion — they need the logic. Expandable reasoning panels reveal which data points were evaluated, what thresholds were applied, and where uncertainty entered the chain. Experts can agree with the conclusion while disagreeing with a step — and act accordingly.

Contextual Confidence Scores

A single percentage means very little without context. Confidence display must answer three questions: how certain is the model, what is that certainty based on, and what should the user do differently at this confidence level? The score should change the interface state — not just decorate it.

Verifiable Citations

Every output should be traceable to its source. Model version, training data, data lineage, and audit ID aren't compliance overhead — they're the foundation of institutional trust. In legal, medical, and financial AI, an unverifiable output is an unusable one.

Human Override Options

The expert must always have full authority over the AI's recommendation. Approve, modify, and reject actions should be equally prominent — not visually weighted toward acceptance. Every override should be documented with reasoning, creating an audit trail that improves the system and protects the human who made the call.

Design Solution

Confidence shapes the entire interface — not just one element

The prototype uses a medical diagnosis context to show how the same interface must respond differently at three confidence levels. The UX adapts: evidence framing, alert severity, action emphasis, and the reasoning chain all shift with certainty.

Impaired Glucose Tolerance vs Type 2 Diabetes

UX response: Borderline evidence with caution flags. Physician required to review before proceeding.

Type 2 Diabetes Mellitus, Uncontrolled

UX response: All evidence confirmed. Clear path to treatment with streamlined approval flow.

Prediabetes — Further Testing Required

UX response: Urgent alert surfaced. Missing data gaps explicit. System gates strongly toward review first.

Beyond the three scenarios, the prototype includes a sticky patient panel that keeps context always visible, an expandable reasoning chain that shows each diagnostic step with evidence weights, a data lineage section with model version and audit ID, and a Review & Modify panel that requires documented reasoning before an override is accepted — creating an audit trail without slowing the expert down.

Key Insight

"In specialized AI, UX is not decoration — it is a safety layer."