From Answers to Agents: What Must Change Before Generative and Agentic AI Become Clinical Infrastructure in Hematology.

Abstract

Generative artificial intelligence in hematology is entering a new phase. The dominant question, whether large language models are accurate enough for clinical decision support, is being overtaken by a harder one, as systems shift from answering questions to acting: extracting structured cases, routing them, classifying variants, and grounding recommendations in guidelines and case memory. In 2026, a hematology agent achieved roughly 83% concordance with tumor-board decisions in a prospective silent trial, with hallucinations in 0.3%, suggesting that for well-structured tasks the binding constraint is shifting from accuracy to governability. Moving from experimental support to reliable infrastructure requires four changes: validating trajectories rather than answers; allocating autonomy inversely to case complexity; treating deployed models as regulated instruments under continuous surveillance; and engineering against automation bias, with data-privacy and consent safeguards throughout. For hematologists, this reframes AI adoption as a governance and human-factors program rather than a search for a better model.