← All work Case Study 01 · AI · Trust & Human Control

Clinical Triage AI — Who Decides First?

Before this AI ever offers a recommendation, the clinician has already reviewed the case, corrected the data, and recorded their own judgment — on purpose.

Role
Concept, research, design — solo
Type
Self-initiated critique & redesign
Year
2026
Focus
Human-AI decision sequencing

The Product

What it is
A conceptual clinical triage application that helps emergency department clinicians prioritize incoming patients and review clinical observations before AI decision support appears.
Its goal
To speed up triage prioritization under time pressure without letting the AI become the first — or only — voice in a clinical judgment.
Who it serves
Emergency physicians and triage nurses managing a moving queue of patients, illustrated here through Dr. Maya Chen’s day shift and one patient, ER-061.
Where the stakes live
Getting triage priority wrong costs patients time they don’t have. Letting AI shape that judgment too early costs clinicians their independent reasoning — and the accountability that comes with it.
Annotated Shift Triage Overview screen explaining the active tab, sorting control, patient cards, review button, and selected patient ER-061
The Shift Triage Overview screen, annotated — the interface Dr. Chen works from all shift.

Why This Case Study

What I wanted to learn
Whether the sequencing of AI output — appearing before or after the clinician’s own judgment — actually changes how much of the decision the human still owns.
Tools I set out to test
AI-assisted UI generation for a high-fidelity clinical prototype, plus AI image composition to turn raw product screens into portfolio-ready case study visuals.
Workflow I wanted to adopt
Writing the product argument and scope discipline first, then generating only the screens needed to prove that argument — rather than designing a full hospital platform and picking highlights afterward.
Why this product, specifically
High-stakes clinical software makes design judgment visible in a way a convenience app doesn’t. Every sequencing decision here has a real consequence — which is exactly what makes the reasoning worth showing.
Diagram of the emergency triage problem space across awaiting triage, in assessment, and awaiting treatment queues
The problem space: emergency triage is a moving queue of incomplete decisions, scattered across three shifting states.
The Argument
Most AI-assisted workflows put the recommendation first and the human’s agreement second — which quietly trains people to defer rather than judge. Clinical Triage AI does the opposite: it requires the clinician to review the case, correct the data, and record a provisional category and reason before any AI support appears. The AI still gets to contribute — but it speaks second, into a decision the human already made and can now compare against, not accept blindly.
01

Reviewing before the machine does

Before Dr. Maya Chen can see any AI input on patient ER-061 (41-year-old male, abdominal pain), she has to review five clinical observations one by one — confirming or correcting each. Nothing moves forward until this step is complete.

Six-step flow of Dr. Maya Chen reviewing patient ER-061 before AI support appears
The six-step pre-AI review pathway for ER-061 — from the shift queue to a recorded provisional category, with no AI input yet visible.

The Decision

The system requires explicit confirm/change on each observation rather than defaulting to “auto-accept unless flagged” — a faster pattern that would have skipped forcing genuine clinical attention onto each value.

What I Learned

A small interaction detail — an in-app keyboard that doesn’t let the tablet’s system keyboard cover the data panel — turned out to matter as much as the AI-sequencing decision itself. Clinical software lives or dies on bedside mechanics like this.

Looking Forward

Whether this five-observation review actually feels necessary to a working clinician, or like friction, is the first thing worth testing with real users.

02

Making the clinician’s reasoning explicit

Once observations are reviewed, the clinician doesn’t just pick a triage urgency category — Immediate through Non-urgent — she has to write the clinical reasoning behind it, before the AI is shown.

Clinician holding a tablet recording a provisional triage category before AI support is shown
Recording a provisional triage category and the reasoning behind it, on the tablet, before AI support appears.

The Decision

I rejected a plain category-selector — fast, but leaves no record of why — in favor of a required reason field, turning a click into a documented judgment that becomes part of the audit trail before AI ever appears.

What I Learned

The line “Your pre-AI category is recorded in the audit trail before AI support appears” does more design work than any visual element on the screen — it’s the moment the product’s whole argument becomes legible to the person using it.

Looking Forward

Whether writing a reason under time pressure feels like protection or paperwork can only be answered by watching a real clinician do it, not by reasoning about it in Figma.

03

Deciding where the AI doesn’t go yet

The idea could have grown into a full hospital platform — ICU mode, routine ward, mobile handoff, analytics. I scoped the first prototype down to one moment: Emergency Mode, ending right before AI support is introduced.

Product scope diagram showing Emergency Mode steps in scope and Routine Ward, ICU Mode, Mobile Companion, Analytics as future extensions
Product scope: Emergency Mode’s five in-scope steps, with Routine Ward, ICU Mode, Mobile Companion, and Analytics deliberately left for later.

The Decision

I chose to stop the prototype at “AI support later” — deliberately not designing the AI recommendation screen itself — over building the complete loop, because the argument here is about the sequence leading up to AI, not the AI’s output.

What I Learned

Scope discipline is a design decision worth defending in a portfolio, not just a project-management footnote. Knowing where to stop is as much judgment as knowing what to build.

Looking Forward

The next step isn’t more features — it’s validation: testing whether the pre-AI sequence actually earns clinician trust, or just adds friction, before ever designing what AI support looks like.

Three design principle cards: pre-AI human review, visible reasoning, audit-ready decision

The Principle

AI supports judgment. It does not replace responsibility.