Too much information was competing for attention.
The CDI workflow contained a large amount of useful clinical and coding information, but usefulness did not automatically translate into usability. The design challenge was to help CDI specialists reach the information they need for the task at hand without making them process everything on the screen at once.
Reduce cognitive load
Prioritize the most relevant clinical and coding context instead of giving all information equal visual weight.
Reduce context switching
Keep evidence, patient context and documentation close to the task rather than forcing users to move between disconnected views.
Make actions obvious
Clarify what users can review, add, dismiss, verify or continue manually.
Designing for appropriate trust — not simply making AI look trustworthy.
The AI UX research focused on AI-generated recommendations, summaries, insights, drafts and actions inside existing healthcare workflows. The core idea was that users should have enough transparency, evidence and control to decide whether an AI result deserves their trust.
Five questions every consequential AI result should answer
Research translated into product requirements
AI supports judgment rather than silently replacing it. Review and confirmation should increase with consequence.
Consequential recommendations should expose patient-specific evidence and make source verification easy.
AI-generated content should remain distinguishable from verified source-of-truth data.
Missing, conflicting, incomplete or stale information should be visible rather than hidden behind false precision.
Users need meaningful ways to modify, reject, retry, cancel, undo or continue manually.
AI should appear where it supports the task — not become another disconnected alert stream.
Lead with the useful result, expose supporting evidence and limitations, let the user decide, then perform downstream actions with appropriate review or confirmation.
Keeping the workflow visible without making the interface feel overloaded.
I explored multiple ways to structure the experience and compared the trade-offs between seeing the whole workflow, separating dedicated work areas and focusing on one step at a time. The handoff direction uses clear top-level navigation and task-specific views so the user retains context while the screen remains focused.
Turning research findings into testable UI behavior.
I selected two research findings for deeper exploration: explicit AI failure and recovery states, and prevention of alert fatigue. These were intentionally framed as UX proposals for discussion and validation — not as final engineering requirements.
Different problems need different explanations.
A generic “Something went wrong” state does not tell a user whether AI lacked evidence, found conflicting information, is stale or is technically unavailable.
Not every AI insight should become an alert.
If every finding receives the same prominence, important recommendations compete with lower-value noise. The proposal introduces contextual hierarchy and groups related findings.