BackBuilding a Healthcare AI Chatbot: Meet “Connie”, the Contraceptive Care Chatbot
How do you build a chatbot that is empathetic, personalized, and private, which are three things that AI chatbots notoriously struggle with? When a healthcare client approached us to build an AI chatbot specialized in contraceptive care, we had to face this question head on. Healthcare conversations require towing the line between providing patient-centered educational resources without crossing into clinical decision-making territory. Here's what I learned, as a Product Manager on the team that built “Connie,” an educational chatbot for a digital contraceptive healthcare product.
Understanding the client’s needs
Like any good software project, the first step in building the AI chat was understanding what the client wanted to achieve, and how it would fit into the rest of the digital product. This part felt familiar, even if the technology behind it was new to me. What would success look like? How would this chatbot fit into the larger experience?
Personalization and contextualization
One of the client's most important requirements challenged the generic chatbot paradigm entirely. They didn't want users to always simply be greeted with "Hi, what can I do for you?" Instead, they wanted Connie to be contextually aware of the current page that the user was viewing. Imagine you just clicked into the page with information about the IUD. When you open the chatbot and ask, “tell me about this,” Connie responds: "I see you're looking at the Hormonal IUD. Would you like to ask a question about side effects or insertion?" That's a fundamentally different (and better) user experience. We track which pages the user is on, and which pages the user has visited. Including that in the chatbot prompt allows Connie to say things like, "I noticed you were curious about IUD pain. Would you like to discuss pain management options?" This approach transforms a reactive help tool into a proactive guide that meets users where they are in their research journey.
Patient-enabled decision making tone
Trust is everything in healthcare. Providers are trained to maintain a patient-centered, empathetic, and calm tone of voice while giving care. Because we were working within the context of contraceptive care, which has its own host of stigma, it was even more important to us to think critically about gaining the trust of a potential user interacting with an AI chatbot. Rather than writing tone guidelines from scratch, we used actual physician training materials as part of the chatbot's context. We also provided examples of "good" versus "bad" contraception support language, with specific phrases to use and avoid. A core design principle was making the chatbot feel like a knowledgeable friend or sister rather than a clinical authority. By allowing casual language, affirming phrases, and even the occasional emoji, the experience feels approachable and human. Connie meets users where they are instead of where a doctor's office expects them to be. The anonymity of a chatbot adds another layer of comfort, removing the self-consciousness that can come with discussing sensitive topics face-to-face. For women who find conversations about reproductive health intimidating, that combination of warmth and privacy can make all the difference.
Security and Privacy Are Non-Negotiable
Healthcare data requires extreme caution, even in a public educational tool. We had to address several security concerns: Session Management: We store conversations server-side rather than in the browser to prevent users from editing their message history and manipulating the chatbot into violating its safety guidelines. HIPAA Compliance: We treated HIPAA compliance as a foundational requirement throughout the project, which meant thinking carefully about potential data exposure at every step of the process. Because the application does not register users, users are not identified as they navigate the site. Additionally, each chatbot session gets a unique token, so there is no persistent tracking across sessions. Beyond the user-facing design, we evaluated our entire infrastructure for compliance risks. This included ensuring our LLM provider has a BAA in place, and configuring our analytics and error-tracking tools to avoid logging identifying details. FreeBC is fully HIPAA compliant, so chat histories are stored in a HIPAA-compliant database.
Testing is key
The client wants their team, their clinical team, and our internal team to actively try to "break" Connie before MVP launch. The most helpful feedback format we established was: Query: What the user asked Response: What the chatbot said Expected Response: What it should have said This structured feedback loop helped us iterate on the prompt engineering and refine the chatbot's expected behavior. One helpful decision we made was to separate the system prompt such that the chatbot on the production site used its own file, completely isolated from staging. This allowed the engineering team to iterate freely on the system prompt themselves in staging leading up to the MVP launch, without any risk of affecting the live product. What this testing looked like in practice was hands-on and iterative. It involved testing system knowledge, adjusting the chatbot’s tone, and refining edge cases. It was deliberately a team effort of trial and error: each session surfaced gaps or unexpected behaviors, which were addressed by refining the system prompt.
The Bottom Line
The process of building healthcare AI chatbots successfully comes from respecting three things: your users' context, your domain experts' knowledge, and your users' privacy. Most importantly, remember that in healthcare a chatbot isn't just answering questions, it's supporting people through potentially life-changing decisions. That responsibility should guide every design choice you make.
