M*Modal · Enterprise healthcare AI · 2 years3M's AI had the answers.Clinicians had a workflow.I designed where the two meet.
After I redesigned how clinicians review its suggestions, 3M's AI saw 25% more of its flagged diagnoses accepted, recovering millions of dollars more in missed billing.
The problem
AI output without information hierarchy is just more work, particularly in a complex space like medical billing
3M's AI-based HCC Management software (dating back to 2016) scans medical documentation and prior billing to identify evidence of missed diagnoses. The evidence is reviewed by specialists before being forwarded to doctors.
Prior to the redesign I worked on, the AI just showed each finding as an identical row in a table, which was inefficient for medical specialists to work from.
I worked with subject matter experts (learning the ropes how HCC medical coding works) and conducted contextual inquiries, shadowing medical documentation specialists to identify problems and opportunities.
Wasted review
No Hierarchical Condition Categories (HCCs) group related diagnoses for reimbursement. Within a hierarchy, only the most severe category counts, so the lower ones are outranked., so outranked categories still demanded attention.
Hidden status
Prior-year, current-year, and billing state sat one click deep.
Tool switching
Specialists were spending most of their time away from our app and in other tools.
Convincing stakeholders
The data said we weren't the main character. I made the case for us to avoid the trap of doing everything.
Only about 20% of specialists worked from our worklist. The rest lived in Epic, and in Excel or Access lists their team leads assigned. That number, backed by what I saw at every site visit, convinced leadership to stop competing for the center of the workflow.
So instead of replacing the lists specialists already trusted, the redesign would plug into them and stay out of the way.
The strategy argument behind the redesign: leadership wanted us to be the main tool in the clinician's ecosystem. My research showed the faster route to their own adoption goal was to be a great supporting tool.
The redesign
Meeting clinicians where they work
After testing prototypes with engineering stakeholders and vocal power users, I gave the flat diagnosis list the hierarchy it was missing: HCC grouping and status at a glance. Then, rather than competing with team leads' lists, I pushed leadership toward smaller changes that plug into them, earning power users major time savings at a fraction of the engineering cost.
A hierarchy built around HCCs
Specialists think in HCCs, the categories that decide reimbursement. Grouping by HCC and flagging outranked categories ended review work that could never matter.
Status that can be read at a glance
Prior-year, current-year, and billing state used to hide behind a click into every diagnosis. I tested five treatments; tinted icon chips shipped.
Quick views from existing tools
Upload the CSV or Access export a team lead already maintains, and every specialist gets a one-click, pre-filtered worklist.
Stay in flow across a patient list
A sticky pager (“Reviewing 1 of 5 patients”) moves straight to the next chart. No round-trips to the worklist between patients.
Same AI, two contexts
Designing for different attention contexts
The worklist you just played with is a full-day professional tool, information-dense on purpose. But the same AI also surfaces nudges to doctors mid-visit, while a patient sits across from them.
I led both ends: a compact desktop widget that earns a glance without competing with the patient, and the dense review workspace it hands off to. Same engine, opposite information densities, because the attention economics of the two users are opposite.
Specialist · a desk and all day
~40 data points on one screen
Comparing beats calm.
Doctor · a patient in the room
3 data points, one decision
Calm beats completeness.
The hardest call
How loudly should the AI announce itself?
Meeting clinicians where they are includes their trust posture. Doctors asked us to label what the AI found versus what humans found. Too loud, and they dismissed the AI wholesale. Too quiet, and they couldn't calibrate trust. I tested three levels:
A · Separate AI section
RejectedDoctors disregarded AI content entirely. Hard separation turned into bias against the machine.
B · Icon on AI items
RejectedFewer dismissals, but the icon needed training to understand and collided with existing icons.
C · Italics for AI items
ShippedSubtle provenance. Power users could read the source; nobody was dissuaded from acting.
The middle ground existed. Finding it took behavioral evidence, not opinion.
The new 3M redesign is a lot better to look at compared to the old design. It has been much easier for the team to prioritize what to work on.
This work contributed to a 25% lift in chronic-condition diagnosis capture, recovering millions of dollars in missed diagnoses and contributing to 3M retaining major contracts from large world-class healthcare organizations.
Widespread adoption at
Learnings
Three things 3M taught me
Clicks are a buyer's metric, not a user's
Stakeholders judged designs by click count, because clicks are countable. I learned to re-anchor debates on what specialists actually lose: errors, wasted reviews, broken flow.
Not knowing the domain is a research tool
I had no medical training in a field full of experts. Asking the basic questions surfaced assumptions everyone else had stopped seeing, and made the tool legible to newcomers too.
Spotting a problem isn't the same as fixing it
As the only designer among ~30, I owned whatever I flagged: research recruiting, analytics, advocacy. That instinct is why this rebuild exists at all.
