Erbal
A conversational AI that replaces wellness product browsing with a guided consultation, matching shoppers to Erbal products and partner services, built on a trust model that decides what the AI explains, admits, and refuses at each stage.
ROLE
Solo UX Designer on the project
1 Design Director (review), 1 Engineer
2024.10-2025.2
Website
Chat on the left, interactive panels on the right, both updating in real time. Users who would rather point than type can use the symptom checker instead, and any point in the report or any recommendation can be questioned back in the conversation.
Leadership wanted AI to make wellness shopping personalized and proactive. Personalization needs private answers, and wellness is privacy sensitive, so an AI that recommends what it sells has to earn credibility before it asks. Trust lands first. Nothing gets personalized if nobody answers the questions.
I mapped the happy path onto a doctor and patient consultation, then traced the emotional curve beside it. Curiosity about a new AI starts high, drops at every request for something private, and drops again when the answer looks like a sales pitch. Trust is not one feeling. It fails at three moments, so each one got a principle and a mechanism.
Assistant AI → fast, no visible context, users suspect the intent
Collaborator AI → slower, room to explain, correct, and clarify
The assistants were quicker and trusted less. A collaborator has to introduce itself, so the first message names its role, its scope of 40+ products, and its limits: not a doctor, no diagnosis, no treatment.
Why the data is needed → every question states its purpose before it asks
How it is used → a memory notice says what is stored, how it is used, and that it is never shared
Control over what is shared → any question can be declined, every field stays editable and removable
Helper, not collector → the AI returns its analysis after each input, so the data goes back to the user immediately
Interviews returned two requirements: reasoning that informs rather than sells, and a visible line from the user's own data to the suggestion.
The wellness report quotes back the symptoms and lifestyle factors the user reported
Each herb is explained by what it does, not what it sells for
Every product card carries a short "why we chose this for you" block, and an Ask AI button to push back
Leadership took the work as Erbal's position in the wellness e-commerce market.
The three trust moments became the unit the team plans in.
Users tied their willingness to buy to the reasoning behind each recommendation.
Designing for AI meant learning by doing. There was no pattern to copy for a wellness AI that has to earn trust before it asks anything. I built, tested, and refined the insights as new findings came in, instead of settling the model up front. From 0 to 1 the ambiguity never fully cleared, so quick user testing at each trust moment was what turned it into decisions I could design against.
