In testingAs Associate Product Manager · Built with AI
An assistant that knows when to send you to a doctor.
- My part
- What it must and must not do
- Code
- AI coding tools
- Status
- In testing
The problem
The chat bot on our website, a different one from my WhatsApp bot, answers only inside fixed boxes. Customers do not ask in fixed boxes.
Three questions
- Hair questionProducts. No doctor needed.
- CholesterolProducts, plus a link to our online consultation.
- Chest painAn urgent-care message. No products.
What I decided
- It answers in the customer’s language, from our 248 products and 22 approved answers.
- A vague concern like “bloating” gets one clarifying question first.
- It remembers what a customer shares about themselves during the conversation.
- Code checks every reply: price, product name, link and language. Anything that fails a check is dropped before the customer sees it.
How I tested it
About 5,000 test examples, plus about 100 scripted turns in nine languages, including urgent symptoms and attempts to trick it.
Before customers use it
A clinician review, a native-speaker review of each language, a legal review of product claims and a security check.
How it was built
I defined what it must and must not do. AI coding tools built it.