Casey Drummond
All work
Case study 01 · Conversational AI · Healthcare

Optimizing care delivery with AI

I led three designers and research on a brand-new, AI-enabled intake for orthopedic patients and their care teams. My job was to make getting started easy for first-time, often senior users, and to make sure care teams heard about risk the moment it appeared.

Role
Associate UX Director
Led 3 designers + research
Company
Zimmer Biomet
Patient & care-team products
Users
Patients, care teams,
surgeons, hospital admins
Tools · Timeline
Figma, Miro, Claude Code
Concept → validated in 20 days
Confidentiality notice: The screens shown are from Project Cora, an independent concept prototype I created for a Maven AI design course. They contain no Zimmer Biomet confidential information, designs, data or code. My work at Zimmer Biomet is described at a high level, consistent with my confidentiality obligations.
Project Cora concept prototype: home hub, AI conversation answering a swelling question with a clarifying safety check, and the safety escalation screen with Call 911 and Call Sarah
Try Cora yourselfAsk a question, or type “I fell” to see the safety screen. The profile menu switches research variants.
Open the live prototype
−18%Fewer patient calls to care teams, projected from A/B tests of the chat
+20%Larger potential market, once the product passed hospital security review
3 stepsPlain-language steps to get started, built for first-time, older patients
20Patients tested it, and every one finished each core task on the first try
01 · The challenge

A 50-page handbook was the only guide

Older patients self-managed orthopedic recovery from a 50-page handbook. They called the wrong numbers and reached the wrong care teams, while clinicians had no early-warning signal.

Make it effortlessReplace the handbook with one plain-language AI conversation any patient can use on the first try.
Design for the most vulnerableI owned flows, logic and usability testing so seniors, not just power users, finish every core task.
How I think

In healthcare AI, I design the behavior before the screens. What the assistant refuses to say matters as much as what it says.

02 · How I worked

My process, step by step

What I did, why I chose it, and what it changed.

Step 01

I listened before I designed

What I did

Quick surveys, then contextual inquiry with patients and care teams.

Why this step

Surveys were fast but shallow, so I sat in on clinic visits with patients and their care team.

What it changed

Three insights set the direction: overloaded staff, access barriers for seniors, and too many contact points.

Handwritten field notes from a clinic visit with a male patient over 65, including: no questions, but difficult to get in touch; the phone number was the main appointment center.
My clinic notes, patient #2: “No questions… difficult to get in touch.” The phone number was the main appointment center.
Step 02

I agreed on success before spending on design

What I did

Built a capability map with journeys and KPIs from both organizations’ qual and quant data.

Why this step

Two organizations had to fund one direction. Sharing the evidence before meetings cut decision time in half.

What it changed

Cross-functional leaders signed on to one plan.

Step 03

I designed the AI’s behavior before its screens

What I did

Wrote the system map, prompt and component spec first, then built the screens in Project Cora.

Why this step

Screens built before the rules end up decorating the wrong behavior.

What it changed

Three screens with clear jobs: a calm home hub, a plain-language chat, and a safety screen the AI can’t override.

System prompt · behavior rulesProject Cora
RoleInform, reassure and route. Never practice medicine or change a treatment plan.
Voice6th-grade reading level. Short sentences. One question at a time.
GroundingAnswer only from the patient’s care plan and approved patient-education material. If it isn’t there, say so and route to the clinic.
NeverDiagnose, interpret symptoms, change medication, or predict outcomes.
Escalation rules · tested phrasingProject Cora
Escalate nowBleeding · any fall, including “I’ve fallen” and “I can’t get up” · a bare “help”. Input locks; the safety screen opens with Call 911 and Call Sarah.
Ask firstPain and swelling are expected after surgery. Ask one clarifying question (fever, redness, drainage, calf tenderness, chest pain), then decide.
Answer normally“I can’t stand the pain” · “When can I get up from a chair?” · “Help me understand my exercises.”
Every channelThe same checks run on the care-team message box, because a reply there can take a business day.
Step 04

I wrote the guardrails, then tried to break them

What I did

Listed emergencies, ambiguous symptoms and near-miss phrases, then tested real patient wording against each rule.

Why this step

Patients must never be trapped in automation, and constant false alarms teach people to ignore them.

What it changed

Escalate instantly on emergencies, ask one question on expected symptoms, answer everything else normally.

A dead end I learned from

Treating “pain” as an emergency

My first prompt escalated on the word “pain.” But pain is expected after joint replacement, so nearly every chat would have ended in “call your clinic,” eroding trust. Now Cora asks one clarifying question from real post-op guidance and escalates only on warning signs.

Testing also showed “fallen” missed a rule written for “fall,” and “I can’t get up” had no fall word at all. I rewrote the triggers around what patients actually say.

Cora chat: the patient writes “My knee pain is worse today.” Cora replies that some pain is expected, asks whether it is getting worse, there at rest, or suddenly sharp, offers two answer buttons, and cites its source.
Step 05

I let patients and data settle the details

What I did

A/B tested the chat and persona voice with Product and Data Science; usability tested with 20 patients.

Why this step

Tone is measurable. I wanted evidence, not taste, to decide how Cora sounds to a 70-year-old with low digital literacy.

What it changed

Non-technical patients finished every core task on the first try; A/B results backed the projected 18% drop in calls.

Variant A · drawn iconsVariant B · real photos · chosen
Two versions of Cora’s task list side by side. Variant A uses small two-tone drawn icons; Variant B uses real photos of a knee stretch, an ice pack, a walk and a pain log.
Why Variant B won
  • Real photos patients recognized at a glance
  • Slower, reading-pace typing
  • Short bullets in a softer tone, not walls of text
Switch between both in the live prototype’s profile menu.
03 · The direction

“I redesigned the point of entry, not the patient.”

With staffing constrained, I ran the usability test myself and set the design direction.

Automated intake & verificationRebuilt clinical capacity by removing manual steps.
Accessibility-first patternsThe AI helps patients find and understand information specific to their condition and history.
One streamlined intake pathCare-team info sits front and center and is one tap from the AI chat.
04 · Key decision

From a 50-page handbook to one AI conversation

My decision: replace the handbook maze with an assistant that gives tailored, plain-language answers, shows its source, and always offers a path to a human.

05 · Outcome

Validated results, delivered in full

Every result above was validated with patients, and engineering received a production-ready handoff: complete Figma specs and token documentation.

A corporate restructuring affected the timeline after design was complete.