Answers are grounded in this paper and cite their sources.
Background Clinical agentic artificial intelligence (AI) can alter its actions during a clinical task, creating failure pathways that conventional performance metrics may not capture. Whether safeguards interrupt these pathways remains uncertain. Aim To map safeguard effects, residual failures, human recovery, and resource consequences in clinical agentic and related AI workflows. Methods We searched PubMed/MEDLINE, Embase, Scopus, Web of Science Core Collection, and IEEE Xplore, for English-language, peer-reviewed empirical journal reports. This scoping review used independent screening, duplicate extraction, narrative synthesis, and result-specific critical appraisal. Results The 111 reports comprised 56 functionally agentic, 29 related non-agentic, 22 architecture-uncertain, and four human-factors reports; 38 of the 56 agentic reports explicitly documented model-directed subsequent-action selection. Among the 111 selected appraisal results, none both isolated a specified safeguard’s effect and demonstrated reduced measured patient harm. Configurations incorporating safeguards improved selected outcomes, but gains coexisted with omissions, reduced sensitivity, false reassurance, delay, or transferred review burden. Concurrent changes in models, prompts, evidence access, and workflow frequently prevented attribution to individual controls. Some patient-randomised trials reported patient or process benefits, but evaluated bundled interventions rather than the contribution of a safeguard. Human authority was more often documented than successful error detection and recovery; requiring clinician approval did not establish that residual errors were recognised and corrected. Performance varied across models, settings, and information conditions, limiting transfer of favourable findings. Reported efficiency gains depended on measurement boundaries and did not consistently account for review, correction, escalation, and downstream work. Discussion The evidence exposes a gap between local error control and safety across a complete clinical episode. A safeguard can reduce a targeted error while shifting risk to later decisions or human reviewers. We propose a post hoc , non-ordinal evidence profile addressing mechanism fidelity, residual and introduced failures, robustness, joint human–AI performance, and clinical consequences. The profile requires validation; heterogeneous designs and retrospective methodological refinements constrain inference. Conclusion Clinical safety should be evaluated across the complete care pathway, tracing safeguard effects through residual error, human recovery, and patient-relevant outcomes. Selective autonomy requires joint assessment of residual risk and the capacity for timely human review.
No takes yet. Share an insight, caveat, or question.
Arab et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: