02 / 1.TALK AI AGENT · 0 → 1
Designing clinic AI for adoption: 432 clinics in eight months.
I defined AI behavior boundaries, human handoff, and repeatable QA so clinics could control how AI joined their workflow. Eight months after launch, the product had 432 clinics onboarded and 54 paid subscriptions.
01 THE CONTEXT
An answer is only useful if a clinic can trust it.
Clinics wanted help with repetitive questions, but could not risk incorrect fees, false promises, or missed complaints. The design challenge was making AI useful while keeping staff in control.
Trust starts with knowing the limits.
02 FAILURE & TURNING POINT
A false promise changed how I defined AI boundaries.
The initial rules limited what AI could discuss, but did not fully limit the actions it could promise. That gap surfaced shortly after launch.
Where it broke down
AI told a patient it would pass their request to the front desk, but the product had no notification mechanism. The patient believed the clinic knew; staff received nothing.
What revealed the gap
Comparing the reply with product capabilities exposed the gap: an empathetic answer had promised follow-through the system could not deliver.
What I changed
I added a hard rule against promising unsupported actions, with clear signals on conversations that needed staff attention so the front desk could recognize handoff needs.
Review whether a promise can be fulfilled. Updated rules still need expected replies, failure records, and repeat testing to check their behavior.
THE TRADEOFFHelpful-sounding replies were less important than reliable expectations. Higher-risk conversations needed a visible route back to a person.
Explore the supporting design
03 THE SOLUTION
Give clinics control they already understand.
I separated control into a global switch, a per-conversation switch, and a human handoff signal. I also translated prompts into familiar settings: assistant identity, reply style, and clinic-specific instructions.
Global control
Let each clinic decide when AI participates in its workflow.
Conversation control
Handle exceptions individually with a per-conversation switch.
Human handoff
Make the need for staff attention explicit in sensitive conversations.
THE TRADEOFFClinics operate differently. A single automation setting could not cover every schedule, exception, or sensitive conversation.
04 MAKING IT WORK
Turn one failure into a repeatable test.
Because AI replies vary, one successful test is not a stable acceptance standard. I built an eight-category test management system with QA, recording expected and actual replies, failure reasons, and retest status.
A repeatable review process takes ongoing effort. It makes behavior changes observable instead of relying on a one-time sign-off.
- 01
Define expectations
Group cases into eight categories and specify the expected behavior.
- 02
Find the gap
Compare actual replies and record failure reasons and risks.
- 03
Adjust and retest
Revise the rules and track whether the correction holds.
Explore the process and collaboration
05 IMPACT & OWNERSHIP
What changed through this work?
Within eight months of launch, the product reached 432 clinics across Taiwan and Japan, with 54 paid subscriptions. The behavior specification and testing process also supported onboarding the first 16 Japanese clinics.
Product-level results, reported through August 2026. My ownership: behavior boundaries, risk layering, control architecture, and the testing loop.
A CLOSER LOOK
A little more context.
MVP scope & research
The MVP answered published clinic information and identified available sessions, but did not complete bookings. Three internal testing rounds and a two-month closed beta with nearly 30 clinics informed the risk model: answer low-risk questions, constrain medium-risk wording, and hand high-risk conversations to staff.
Team & measurement context
One PM, one QA, one frontend engineer, two backend engineers, and one designer (me). Development began in May 2025; the product launched in January 2026. Adoption and paid subscriptions are product-level results reported through August 2026.
HIRING A SENIOR OR LEAD PRODUCT DESIGNER?
If this is the kind of work your team needs, let’s talk.
Based in Taipei (GMT+8), open to relocating for the right role. The form reaches me directly.