Designing a trustworthy AI workflow for clinical inboxes

TL;dr I designed an AI-assisted workflow to help clinicians triage patient-portal messages according to their own priorities—reducing manual sorting while keeping clinicians in control.

The Problem

The design opportunity: Reduce repetitive triage work while preserving the clinician’s ability to review, override, and control how messages are handled.

Proposed solution

Smartbox MD
A configurable AI-assisted workflow for managing patient-portal messages.
Instead of imposing a single triage system, it adapts to each clinician’s stated preferences.

• Captures preferences for urgency, routing, and delegation.

• Organizes incoming messages according to those preferences

• Recommends next steps while keeping clinicians in control

• Surfaces relevant context before a clinician opens the full thread

Design goal: less repetitive triage, more "this actually works the way I do."

design considerations

I designed for trust through three forms of control:

My process

Social listening on Reddit revealed common patterns

“Way too much access, too many non-urgent messages flagged as urgent… Time suck, time waste, and distracts from taking care of the patient in front of you.”

— Anonymous Reddit user

hypothesis

I hypothesized that AI-assisted prioritization could reduce the cognitive effort required to manage clinical inboxes.

concept evolution

Here’s how my design ideas evolved through iteration while learning more about users, their AI comfort level, and other innovations in Clinical AI SaaS.

Inbox synthesis → separate views → workflow personalization.

Strategic Design Decision 1

✅ Make the inbox easier to scan

I took the limitations of a traditional inbox and designed to pull more content to the surface. The goal was to reduce clicks needed to get an overview of incoming messages and act upon them. Giving a bird’s eye view before opening every message.

Message cards surface key context before opening a thread.

Further Social listening

Key pattern revealed: No two workflows are the same

Strategic Design Decision 2

✅ Let clinicians shape their workflow

I couldn't find a set of categories that could represent every clinician's workflow. Which led me to understand it’s best if we do not try to decide a user’s preference.

Trying to guess in a high-stakes environment is where things can go terribly wrong.

Surfaced design principle: workflows configurable—not assumed

The final system: Smartbox MD

Smartbox MD puts clinicians in control of how AI works in their inbox.

in reflection

What I learned

I came into this project trying to tackle the volume problem.

The more I explored clinical inboxes, the more I saw that cognitive burden also comes from the number of decisions clinicians have to make:

• what is important 

• what can wait 

• who should handle it 

• whether they trust the system making those recommendations

I learned that there isn't one workflow that can solve this for every clinician or organization.

That changed how I think about AI use in healthcare. The opportunity isn't just to automate tasks. It's to build systems that can adapt to the way people work while allowing them to decide how much control they want to give AI.

Next Steps: Validation

Note: The concept has not been validated with practicing clinicians.

What I’d Validate Next

1. Inbox information hierarchy — what clinicians actually need to see at a glance, which filters matter, and what information can be summarized.

2. Workflow variation — how clinicians and organizations actually triage, prioritize, and manage messages.

3. Routing — how routing decisions are currently made, who participates, and whether configuring those rules could become a meaningful time-saving/value proposition.