Case study · Conversational AI
Grace: giving a health system a voice
Grace is an AI voice assistant that helps patients manage their appointments: verify, schedule, reschedule, and cancel. This is how I built the persona, the standards, and the flows that let it schedule care with dignity — without pretending to be human.
ROLE
Lead conversation designer
SCOPE
Persona · Standards · Flows · Pattern library
MODALITIES
Voice · Chat · SMS
STATUS
Live in five markets since June 2026
Note: employer details anonymized; dialogue samples use fictional providers and facilities.
01 · THE BRIEF
Patients wait on hold. An AI agent shouldn't make things worse.
One of the largest nonprofit health systems in the U.S. wanted an AI agent that could schedule, reschedule, cancel, and verify appointments over the phone. Every conversation happens with someone who may be sick, scared, or grieving. Conversation design didn't exist as a discipline inside the organization. So before designing the flows, I had to design the rules.
What I delivered
Conversation design standards in the Helix design systemSTANDARDS
Grace persona: voice, tone ladder, behavioral rulesPERSONA
Voice flows: schedule, reschedule, verify, cancelFLOWS
SMS dialogue spec across all four flowsSPEC
A living conversation pattern libraryLIBRARY
02 · THE STANDARDS
Before Grace could speak, the system needed a grammar.
I authored the conversation design standards housed in Helix, the organization's design system: persona, voice and tone, twelve conversation archetypes, and modality rules for voice, chat, and SMS. Written for designers, but used by research, product, engineering, and legal.
The persona
Caregiver meets Sage
Grace is built on two Jungian archetypes. The Caregiver shows up as warmth and an instinct to protect. The Sage shows up as clarity and honesty about uncertainty. Grace acknowledges what users share without claiming feelings it doesn't have.
Do“Got it. Let's find that information right away.”
Don't“I'm so sorry, that sounds horrible.”
The tone ladder
RUNG +2Relief or celebrationmatch the energy, keep it brief
RUNG +1Encouragementacknowledge lightly, no over-praise
RUNG 0Neutralwhere most interactions live
RUNG −1Mild frustrationslow down, validate, don't over-apologize
RUNG −2High emotionshorter sentences, one step at a time
Five registers, so any writer on any team can find the right one, and defend it.
THE BOLDEST RULE
Grace never says “I.”
First-person pronouns make people attribute consciousness and feeling to software: misplaced trust with real risk in healthcare. Grounded in anthropomorphism research and where AI transparency regulation is heading, I made pronoun removal a standard, not a preference. Copy gets harder to write. Trust gets easier to keep.
I found two appointments for you.
There are two upcoming appointments on file.
There are two upcoming appointments on file.
I didn't catch that.
That wasn't clear. Can you say it again?
That wasn't clear. Can you say it again?
Voice
Design for the ear. Real-time and ephemeral: front-load key information, limit options per turn, confirm verbally. Mishearing is a design case, not an edge case.
Chat
Design for the eye. Persistent and visual: users control the pace, so lean on buttons, cards, and progressive disclosure. Never emoji.
SMS
Design for brevity. 160 characters, one action per message, links as escape hatches. The user might answer three hours later. Design for that.
Principles are modality-agnostic. Artifacts must be modality-specific.
THE STANDARDS IN 3:46
What’s in the design standards? Here’s a summary. Yes, the presenter is AI too.
Created with HeyGen, because an explainer about conversational AI standards should be comfortable letting AI do some of the talking.
03 · THE PATTERN LIBRARY
Standards nobody can browse are standards nobody uses.
So I built the Grace Conversation Pattern Library: an interactive companion to the Helix standards where teams can explore archetypes, tone, and recovery patterns as working examples instead of documentation.
Explore the live library →
04 · THE FLOWS
Then the standards went to work: four flows, three modalities.
I designed the schedule, reschedule, verify, and cancel flows: every turn, decision, and failure path. Below: samples from schedule, reschedule, and the SMS spec. Verify and cancel follow the same closing, survey, and escalation logic.
The mechanics under the dialogue
Three ways in, one flow. Callers arrive knowing a reason, a provider, or a specialty; each entry pattern gets its own path to the same outcome.
Triage without whiplash. If clinical triage becomes required mid-reschedule, a softening line brings the caller into it. No blindsiding.
Failure leaves no scars. Reschedules execute as one atomic call. If it fails, the original appointment stays intact and a human takes over with full context.
Silent intelligence. Telehealth eligibility and slot windows are filtered behind the scenes; the caller is never quizzed about what the system already knows.
Escalation from any turn. Trigger words route immediately to a human. The exit is never more than one sentence away.
SMS confirmation · schedule160 CHARS MAX
Appointment confirmed with Steve Paulsen, MD, Wed Dec 31 at 10:20am. (1/2)
Meadowbrook Medical Group, 1700 N Bridge St, Elgin. Map: msys.io/m/elgin · Provider: msys.io/d/paulsen (2/2)
Two purposeful messages, numbered. Declarative, confirmational, never promotional. No emoji: mixed-age recipients, screen readers, healthcare context.
A spec, not a suggestion
[date] → Wed Dec 31
[time] → 10:20am
[provider] → no “Dr.” prefix
[facility] → spelled out in full
Every variable has a format rule; every send has a failure plan (retry once, then the voice agent tells the caller aloud). When message 2 runs long, the street address drops; the map link already knows the way. TCPA compliance is designed in, not bolted on.
05 · LIVE, AND LEARNING
Grace is in production. The findings are already reshaping the work.
Following an initial Grace 2.0 pilot, Grace went live across five markets in June 2026 — two weeks of hypercare, nothing egregious. Launch is where conversation design stops being theoretical: real callers, real transcripts, and a few findings that matter more than any single metric.
SCALE
Five markets
live since June 30, primary care first
CONTAINMENT
14–15%
of calls resolved without a human — early, and honest
SATISFACTION
56%
rate Grace 4 or 5, vs. 85% for human navigators — and escalated callers never see the survey, so the unhappiest voices aren't in that number
HEAR IT FOR YOURSELF
A real call with Grace, de-identified.
An audiogram of an actual production call. Patient details removed; the conversation is otherwise untouched.
The hardest problem isn't in the flows. It's the first five seconds.
The biggest design finding so far: callers resist engaging with Grace at all. Some hang up at the intro; others ask for a human immediately. The intro is now the highest-stakes copy in the system.
The boldest rule is being tested in production.
Leadership overrode the no-first-person rule at launch, so Grace says “I” in production. Rather than relitigate it by opinion, it became a research question — enforcement is paused while the study runs.
Listening got targeted.
QA moved from random call sampling to targeted listening driven by friction, after-hours, and escalation tags — and graders now review the entire call path a patient experiences, not just Grace's part of it.
06 · THE RESEARCH TRACK
Ship first. Then find out what's true.
I'm partnering with the research team to validate what the live data suggests — and to answer the questions it can't. The research agenda is grounded in five weeks of real transcripts, not hypotheticals, and it ties directly to what we see happening in real time.
Questions on the table
First person vs. personified third person. Semantic differential scales separate three things we'd been treating as one target — conversational, warm, and transparently a machine — with participants rating synthetic call samples on each dimension.
How Grace is described before anyone engages. What people are told about Grace shapes whether they'll engage with it at all.
Intro optimization, tested live. Run as A/B tests in production, because we can now segment callers by the exact moment they drop out.
And abandonment finally has an instrument. Escalation-trigger logging segments hang-ups by where in the call they happen — so intro resistance and mid-task frustration stop looking like the same problem.
07 · WHAT THIS WORK TAUGHT ME
Over-confirmation is a failure mode.
The test I wrote into the standards: does this action create or change something outside the conversation? If yes, confirm explicitly. If asking again costs nothing, don't make the user verify every step.
Trust is a design material.
The more fluent a system sounds, the faster people trust it, which raises the cost of implying feelings it doesn't have. Grace acknowledges emotion; it never performs it.
Standards are a leadership tool.
Persona, archetype, and modality are usually defined by different teams at different times, and the gaps show up in the experience. Housing the rules in the design system closed those gaps.
Dignity scales.
Treat users as capable adults, acknowledge hard moments without drama, give real information without euphemism. Rules like these survive every flow, every modality, every release.
© 2026 Marilyn Frank