Background: Clinical documentation burden is a major contributor to burnout in surgery. Artificial intelligence (AI) tools, such as automatic speech recognition (ASR) and large language models (LLMs), may streamline documentation without sac­rificing quality.

Objective: We systematically reviewed the performance of ASR- and LLM-based docu­mentation tools in surgical settings.

Methods: Following the Preferred Reporting Items for Systematic Reviews and Me­ta-analyses, MEDLINE, Embase, CENTRAL, and Scopus (January 2015–October 2025) were searched for studies evaluating AI-enabled documentation (e.g., ambient scribes, advanced ASR, LLM-assisted drafting) in surgical care. Dual reviewers screened, ex­tracted, and assessed risk of bias using the Risk of Bias in Non-randomized Studies of Exposures tool. Heterogeneity of included studies precluded meta-analysis, and results are presented narratively.

Results: Seven studies published between 2023 and 2025 across otolaryngology, neuro­surgery, plastic surgery, and urology were included. Tools such as LLM-assisted operative reports, ambient clinic scribes, and ASR dictation were employed. Findings revealed that AI scribes improved documentation efficiency (5.16 min vs. 10.58 min) and reduced documentation time (5–50 s vs. 7.1–7.4 min), with hybrid clinician-in-the-loop work­flows achieving the best balance of speed and quality. AI scribe notes were non-inferi­or to clinician notes on the Physician Documentation Quality Instrument-9 (33.6/45). Operative note quality was highest with hybrid attending-reviewed generative pre-trained transformer drafts (79% as-is approval) and lowest with generative pre-trained transformer-only notes (23%). Whisper ASR was non-inferior to Dragon Medical One for word error rate and superior when linguistic errors were excluded.

Conclusion: Early evidence suggests clinician-supervised AI documentation may ac­celerate note generation while maintaining comparable quality, with hybrid use out­performing AI-only approaches. However, the evidence base is early, heterogeneous, and largely non-randomized, and downstream outcomes—including burnout—remain unmeasured. Real-world trials incorporating patient, workflow, safety, and governance outcomes are needed to guide supervised implementation.