“About 3 minutes”
Patients wanted to know what they were signing up for. So the intro leads with a time estimate, before the first question.
Previsit.ai · 2024 · AI health-tech · Solo designer
Redesigning an AI medical assistant to keep patients talking
Problem
Previsit.ai runs the intake chat before a doctor’s appointment. Nearly half of patients quit it partway, so doctors started visits without the basics.
My role
Solo product designer. Conversation analysis, patient interviews, flow redesign, UI and copy. Also designed the dashboard doctors use to review the answers.
Outcome
Completion went from 57% to 77% in the first 47 live conversations after launch. Doctors rated answer quality 4 of 5, up from 3.
The chat only works if people finish it. Nearly half of them didn’t.
57% → 77%Conversation completion rate
3 → 4Answer quality out of 5, rated by 8 doctors
Before: one month of live conversations before the redesign. After: the first 47 conversations, two weeks post-launch. Each doctor rated at least 4 answers.
Before & after
before: chat state · 57% finished
One message. A generic hello, then every question at once. No sense of how long it would take or how far you’d got. Patients saw a wall and left.
after: chat state · 77% finished
An intro that names your doctor. A time estimate. One question at a time. A bar.
The insight
I assumed the chat asked too much. 30 conversations and 6 patient interviews said otherwise.
People stayed when they could see the end coming. Drop-off clustered at question 6 whether the questions were easy or hard. The finish line mattered more than the length.
Constraint
So the brief became: 6 questions from the doctor, plus AI follow-ups whenever an answer is too vague to use. That breaks the obvious progress indicator. You can’t show “3 of 8” when the total keeps changing.
the cliff at question 6 · the 30 conversations I read
Completion per question · research sample
30 conversations, read one by one
30 conversations, chosen to include drop-offs. The 57% baseline comes from product analytics, not this sample.
Three fixes that only work together
The bar never moves backward
“3 of 8” breaks the moment the AI adds a follow-up. Now it’s 3 of 9 and the finish line just moved, and trust goes with it. The bar only moves forward, however many questions get asked.
Follow-ups stay inside the bar
“I smoke” is useless without the frequency. The AI asks until the answer is usable, and the bar keeps moving. The patient never sees the goalposts shift.
How the AI asks
Then there was the AI itself. I worked out how it behaves with our prompt engineer, by going through the conversations that had already happened.
Follow-ups fire on habits. Daily medication, smoking, alcohol: answers where a doctor needs the how much and the how often. There’s no hard cap. If an answer still isn’t usable, the AI asks again.
“I don’t know” is a fine answer. The AI suggests finding out before the visit and moves on, with no pressure.
No medical language. The AI talks the way these doctors’ patients talk. A lot of that call was intuition, built from reading how those patients actually wrote.
Decisions
Four calls, what each replaced, and what it cost.
Keeping the AI follow-ups
instead of: Cutting them to shorten the chat
cost: Some chats got longer
A bar that only moves forward
instead of: A “3 of 8” count
cost: An exact question count
One question per message
instead of: Every question in one message
cost: More taps for the same answers
A time estimate set against the long case
instead of: A shorter, optimistic estimate
cost: Short chats look longer than they are
57% → 77%Conversation completion rate
Measured across 47 patient conversations in the two weeks after launch. The people who used to quit at the wall were reaching the end.
Mostly a story about restraint: taking things off the screen until the patient felt sure about what was left. If you’re building patient-facing AI and fighting drop-off, say hi.
natalia.wlwsk@gmail.com →