Trust-building AI workflow for clinical inboxes

TL;dr I designed an AI-assisted workflow to help clinicians manage patient portal message fatigue, while staying in control of their preferences.

Skills demonstrated:

• Social listening via Reddit for pain points, sentiment analysis

• MVP of white label feature which would sit within an existing portal inbox.

• Cursor prototype hosted by Vercel

• Capstone demo presentation to class of 30+

The Problem

Traditional inbox

A high-volume stream of messages requiring clinicians to manually determine:

What needs critical attention?
Which messages need to be handled by whom?
Does this need an appointment?

Smartbox MD

A configurable AI layer that:

Prioritizes messages according to clinician-defined rules
Routes work according to organizational policies
Summarizes context before the clinician opens the full thread
Drafts responses when appropriate
Keeps clinicians in control of AI autonomy

Building from Scratch

There were no clear requirements, so I made my own.

I fed the roadmap and user research into a Claude project to define the onboarding strategy from the ground up.

Key Feature Challenges

• Support 50+ Offerings with intent signals

• Reduce manual setup and cognitive overload

• Create scalable onboarding for enterprise admins

• Design for complex multi-state recommendation flows
  

The feature concept

Strategic Shift
The old Intent setup experience relied on manual topic selection.

I reframed Intent from a configuration tool into a guided AI-powered onboarding system.
This shifted admins from building setup manually to reviewing optimized recommendations.

AI accelerated early exploration

Google Stitch helped me visualize the optimization partner direction V1.
This became the foundation that influenced the final design!

Designing a full system

Once AI recommendations became central to onboarding, the problem shifted from screen design to systems design.

Intent became a living system with complex states, edge cases, and evolving logic.

19 scenarios
across 5 trigger categories

Onboarding, updates, edge cases, partial and total failures, and multi-admin workflows.

To reduce design debt and simplify implementation,
I aligned with Engineering around a centralized scenario map.
(instead of dozens of disconnected Figma flows)

STRATEGIC DESIGN DECISION 1:

Two Types of Topics

The PM initially wanted one table for both offering-based and general intent topics.

As I mapped the workflows, it became clear there were two different mental models.

🥵 Trying to force both topic types into a single table created cognitive overload and made large-scale setup harder for admins.

✅ I proposed separate optimized workflows over one bloated setup.

Although it increased development scope to code two table arrangements, the simplification reduces scanning effort and makes Intent setup onboarding easier to manage at enterprise scale.

Offerings configured
During onboarding, Admin sees all their offerings have intent topics set up.

Without Offerings
During onboarding, Admin sees they can add offerings to tailor their intent topics.
Topics for their business have been automatically set up.

STRATEGIC DESIGN DECISION 2:

Creating a clear path through hierarchy

As the team explored ways to organize recommendations, education, and advanced tools, tabs emerged as a common solution.

I chose progressive disclosure instead, organizing the experience around a predictable content hierarchy:

• Context and onboarding status

• Recommendations to review

• Supporting education

• Advanced tools and configuration

This kept recommendations visible, reduced navigation overhead, and made the experience easier to learn and return to over time.

Usability Testing

Balancing simplicity with flexibility

Strategic design decision 3

Designing for cascading system states

What started as a removal UX uncovered a larger systems problem.

Since topics could belong to multiple Offerings, approval and removal actions became interconnected. Working closely with the lead developer, I mapped these relationships and uncovered gaps in how shared topics were shown in the UI.

As a result I included chip-level approval indicators. This clearly communicated topic states visible across Offerings.

Claude Code helped us rapidly test edge cases and validate system behavior throughout the process.

In reflection

This project reminded me that the most common solution isn't always the right one.

Throughout the project, there were moments where the obvious answer was another tab, or another table. Pushing back on those patterns led to simpler workflows and experiences that were easier to understand and trust.

It reinforced the importance of trusting my instincts when something feels overcomplicated.

This project improved my ability to think in systems rather than screens. It made me focus on designing relationships between workflows, states, and recommendations.

Claude and Claude code helped with some of the tedious work; rapidly mapping scenarios, uncovering edge cases, and validating system behavior.