Evidence, use cases, and implementation safeguards of large language models in primary care

Recent developments in large language models (LLMs) have created new opportunities to support primary care, where much of clinical work is text-mediated. This narrative review synthesizes evidence on LLM applications relevant to primary care workflows and summarizes implementation safeguards. Across studies, the most consistently supported near-term value is workflow augmentation, particularly documentation and inbox management (e.g., drafting portal replies and summarizing information for clinician review) and communication support, where benefits are reported primarily as process endpoints (time, acceptability, perceived communication quality) rather than hard patient outcomes. Evidence for improvements in clinician diagnostic reasoning, treatment planning, and downstream patient outcomes is more limited and context-dependent, and many evaluations remain simulated or conducted in adjacent settings, limiting generalizability to routine primary care. Accordingly, potential roles in population health and cost reduction should be treated as hypothesis-generating and evaluated prospectively. Challenges related to privacy, security, transparency, and model reliability shape organizational governance requirements and evolving regulatory expectations for the clinical use of generative AI in primary care. We emphasize a pragmatic adoption approach: prioritize high-volume, lower-risk clerical and communication workflows; maintain clinician verification and accountability; and apply governance and equity safeguards (e.g. privacy, security, transparency, auditability, monitoring for drift and error) before scale-up.