Skip to Content
HealthcareHealthcare
  • Review
  • Open Access

19 November 2026

29 Pages

Agentic Artificial Intelligence in Healthcare: A Scoping Review of Clinical Value, Safety, and Human Oversight

,
,
and
1
Department of Nursing, Mohammed Al-Mana College for Medical Sciences, Dammam 34222, Saudi Arabia
2
Department of Medical-Surgical Nursing, College of Nursing, University of Ha’il, Ha’il 81451, Saudi Arabia
3
Department of Medical and Surgical Nursing, Faculty of Nursing, University of Tabuk, Tabuk 71491, Saudi Arabia
4
Department of Women’s and Children’s Health, Uppsala University, 751 05 Uppsala, Sweden
This article belongs to the Special Issue AI & ICT in Healthcare

Abstract

Background/Objectives: Healthcare AI systems described as agentic may improve selected tasks, but gains can reflect models, information access, human support or complete intervention packages. We mapped clinical utility, safety and workflow consequences while distinguishing functional agency from evidence isolating its contribution. Methods: We searched PubMed, Embase, CINAHL, Scopus and Web of Science, supplemented by citation searches, through 11 September 2026. Eligible evidence comprised original empirical journal reports of healthcare systems described as agentic or demonstrating potentially agentic functions. Two authors independently screened records, and two independently extracted data, with adjudication. All four authors jointly assigned analytical classifications after a five-report pilot. A structured critical audit and descriptive Clinical Agency Evidence Chain informed synthesis. The review was not registered. Results: The corpus included 123 reports; 90 were classified as core or bounded functionally agentic. Within this subset, technical task performance was the endpoint category in 44/90 reports (48.9%), and clinical decision/process outcomes in 33/90 (36.7%). Three randomised reports in the broader corpus assessed patient health outcomes: one core, one boundary/non-agentic and one indeterminate. All evaluated bundled interventions. Evidence isolating agency and following its consequences within the same comparison was limited. Safety ascertainment, intended-user error correction and whole-pathway workload were incompletely reported. External or temporal validation occurred in 6/90 reports (6.7%), with partial validation in seven others. Predominantly retrospective, simulated or offline evidence limited clinical inference. Conclusions: Intervention-level benefit must be distinguished from the incremental contribution of agency. Technical gains and local or estimated time savings do not establish patient benefit or reduced total human work. Mechanistic comparisons and prospective clinical evaluations address complementary questions. Adoption decisions should consider benefits, harms, implementation conditions and uncertainty.

Article Metrics

Citations

Article Access Statistics

Article metric data becomes available approximately 24 hours after publication online.