Healthcare AI is rapidly moving beyond chatbots toward retrieval-augmented and agentic systems that search external sources, call tools, rank evidence, and generate recommendations. But a response can appear well grounded while still relying on poor evidence, incorrect citations, or inappropriate tool behavior.
This session presents a practical framework for stress-testing agentic healthcare AI across evidence provenance, retrieval quality, citation fidelity, and tool use. Using a working healthcare discovery prototype, we demonstrate how controlled conflicting and low-quality evidence can expose weaknesses in retrieval and reasoning pipelines, and how safeguards such as authority-aware ranking, provenance tracking, citation verification, and evidence conflict detection can improve trustworthiness. The session focuses on what organizations should evaluate before moving healthcare AI systems from experimentation into real-world use.