What the research found
Researchers reviewed 111 reports on clinical AI workflows, 56 of them on functionally agentic systems that choose their own actions, looking at safeguards such as knowledge grounding, rule gates, verification steps and human review. None of the appraised reports both isolated a specified safeguard's effect and showed reduced measured patient harm. Some patient-randomised trials reported patient or process benefits, but they evaluated bundled interventions rather than the contribution of any one safeguard.
As AI tools start taking actions rather than just making suggestions, providers should ask vendors for evidence that their guardrails protect patients, not just that they exist.