Large Language Models in Clinical Workflows: A Critical Review of Evidence from Documentation to Decision Support

Despite explosive growth in the clinical large language model (LLM) literature, the distribution of evidence remains heavily skewed toward simulations and retrospective benchmarks rather than prospective workflow integration studies. This narrative synthesis evaluates the reviewed literature through the lens of evidence tiers (from benchmarks to prospective controlled evaluations) and safety failure modes, focusing on three recurring domains: documentation/summarization, patient-facing communication, and clinical decision support. Across documentation tasks, head-to-head reader studies show that LLM-generated drafts can be comparable to clinician-authored text on rubric-based measures, yet clinically relevant omissions and hallucinations remain setting-dependent — particularly in emergency department (ED) contexts where uncertainty is high. In decision support, a large ED evaluation using real clinical notes found LLM recommendations underperformed a resident physician comparator with an overly cautious profile, whereas constrained specialty applications show potential for augmenting physician diagnosis. Patient-facing studies demonstrate improved readability but reveal divergence between patient-perceived helpfulness and clinician-rated fidelity. Safety research has advanced from descriptive concern to measurable failure modes, including adversarial prompt contamination and social-pressure sensitivity that increase harmful outputs despite mitigation prompts. Current evidence supports deploying LLMs for supervised, low-to-moderate risk documentation drafting with clinician review and explicit safety check-points. However, evidence does not support autonomous documentation in acute settings or high-stakes decision support without substantial safeguards. Implementation remains limited by the scarcity of prospective studies, standardized error taxonomies, and operational monitoring frameworks. High-risk decision support should remain research-only in most settings pending further prospective validation.