Agentic AI earns more physician trust than standard models in clinical settings - but that trust is not always warranted.
Researchers had three physicians evaluate 315 multimodal clinical cases, comparing an agentic AI system against non-agentic baselines. The agentic model won on both cognitive trust - how much doctors believed the reasoning - and behavioral reliance - how often they actually followed it. On treatment planning specifically, physicians preferred the agentic output in 89.57% of cases, a gap the study deems statistically significant (P < 0.001). The key mechanic: agentic AI calls external tools mid-reasoning and shows its work, making the process legible in a way that black-box outputs do not.
The deeper finding is the one buried in the caveats. Transparency in reasoning made doctors more confident, but confidence and correctness are not the same thing. Measurable over-reliance on incorrect agentic outputs still occurred, meaning showing your work can make a wrong answer more persuasive, not less. That is a meaningful limitation for any setting where errors carry clinical consequences.
Three physicians across 315 cases is a narrow sample, and the study comes from arXiv without peer review yet. Still, it puts a number on something the AI-in-medicine debate has talked around for years: a more explainable model is not automatically a safer one.