The BARMER AI chatbot answers questions about health and long-term care insurance, health benefits, membership, and many other topics related to the health insurance provider—directly on barmer.de, around the clock. Its knowledge base is drawn from all of BARMER’s website content, which is organized in a structured manner and updated regularly. Unlike a traditional FAQ system, the assistant understands the context of the conversation: follow-up questions are not evaluated in isolation but are answered within the context of the conversation’s flow.

It is equally clear what it does not answer: Information on individual cases and medical diagnoses are deliberately outside its scope—such inquiries are identified and redirected to relevant information pages or expert advisors. In emergency situations, the chatbot directly refers users to appropriate sources of assistance. The user interface guides users precisely to the right point of contact.

The original plan was an intent-based chatbot: predefined paths, manually maintained response structures, and clear routing logic. It quickly turned our, however, that this model is not viable for the scope of knowledge required by a large health insurance provider: too many topics, too many variations, and too much maintenance effort. The system would have become a second CMS—with the same structural problem: growing complexity, decreasing manageability.

Thus, the decision was logical: intent recognition and routing are no longer handled through rule-based systems, but rather through an agent-based approach. A Large Language Model (LLM) classifies user inquiries and forwards them—within the context of the conversation—to the agent responsible for the task, always adhering to the implemented guardrails, which are absolutely essential in a regulated environment such as BARMER’s. The result: a system that becomes more precise—not more cumbersome—as its knowledge base grows.

The goal of the project was not merely to launch the chatbot—but to ensure that BARMER would be able to continue developing the system independently even after the project was completed. What is usually addressed in workshops and seminars was an integral part of the project early on.

This meant involving BARMER’s internal team from the very beginning and empowering them—not just informing them and handing over the final product. The technical setup and architectural design were developed collaboratively. Through pair programming, knowledge-transfer workshops, and joint architectural decisions, sustainable internal AI expertise was built not afterward, but directly within the project, leading to independent operation within the organization.

AI development follows a different logic than traditional software projects. Defined code produces predictable, consistent behavior—an LLM does not. Its behavior can change with every query, depending on the context, wording, and available data. The system must therefore be continuously developed in a test-driven manner at all levels simultaneously: from experience design through the model and infrastructure to the knowledge base.

We turned this reality into a methodology to deliver the first testable prototype as quickly as possible and move into the evaluation phase. Whereas traditional projects typically consist of 80% development and 20% testing, the development of the chatbot followed the reverse approach. Internal test groups provided reliable feedback early on. The phased rollout initially targeted 950 employees at the call center; subsequently, all BARMER employees were onboarded. Optimization thus becomes a permanent process.