Four Months Inside a Production AI Agent: What Real Users Changed

Maven Assistant, an AI agent for women’s health, has scaled from a limited rollout to full availability, with conversation volume increasing tenfold over four months. Real-world usage revealed that users prioritize basic health questions—such as pregnancy-related dietary concerns—over the complex administrative tasks initially anticipated. Maintaining reliability requires a robust evaluation harness, including LLM-as-a-judge classifiers and deterministic tool-use checks, to manage clinical scope and empathy. Challenges like conflicting instructions from RAG-retrieved documents and the need for situational awareness within the application highlight the necessity of iterative refinement. Balancing latency with model performance remains critical, as does the strategic decision to prioritize shipping early to gather data rather than over-engineering features before they are proven necessary by actual user behavior.