Clinicians often joke that they went into medicine to help people, then met the EHR and started helping dropdown menus. This systematic review and meta-analysis asks a very practical question: when AI tools help generate clinical notes and related documents, do they actually reduce the documentation burden and the time clinicians spend writing, and do they do so without making documentation quality worse (Zhao et al., 2025).
The authors synthesized evidence across 23 studies that evaluated AI-assisted clinical documentation in real or simulated clinical workflows, focusing on front-line health professionals and outcomes that matter for day-to-day work: perceived documentation burden and workload, burnout-related measures, and time spent documenting (Zhao et al., 2025). Most studies were conducted in the United States, with a smaller number from Germany, the UK, the Netherlands, and a joint Sweden–Switzerland study; the settings and specialties varied widely, but ambulatory care dominated (Zhao et al., 2025). The tools were not all the same: some studies used general-purpose large language models such as GPT-4 with structured prompts, while many others evaluated purpose-built “ambient clinical intelligence” systems that listen to the encounter and draft notes, often integrated into EHR workflows, with Nuance DAX and Abridge appearing repeatedly (Zhao et al., 2025). A key practical detail is that in most studies clinicians reviewed and edited the AI draft rather than accepting it as-is, which aligns with how these tools are intended to be used in real care delivery (Zhao et al., 2025).
The headline finding is that AI tools show a moderate overall reduction in documentation burden and closely related outcomes, including burnout proxies, compared with usual practice or baseline periods. Pooling results from 14 studies that reported documentation-burden-related outcomes yielded a standardized mean difference (SMD) of -0.71, with a 95% confidence interval from -0.93 to -0.49 (Zhao et al., 2025). In plain terms, across heterogeneous settings and measures, AI assistance tended to shift workload and burden downward by an amount typically interpreted as “moderate” on Cohen’s conventional thresholds (Zhao et al., 2025). Importantly, heterogeneity was substantial (I² about 75%), meaning effect sizes varied widely across studies, so the average benefit should be read as “often helpful, sometimes less so,” rather than a guaranteed improvement in every context (Zhao et al., 2025).
A second core finding is that AI support generally reduces documentation time, even when clinicians must review and edit the draft. When the authors restricted the time meta-analysis to studies where clinicians edited AI-generated drafts, they still found a moderate overall reduction in documentation time: SMD -0.72 (95% CI -0.99 to -0.45) across 17 studies, again with very high heterogeneity (I² over 90%) (Zhao et al., 2025). The practical implication is that time savings were commonly reported, but the size of the savings likely depends on the tool, the clinical setting, local workflow design, and what exactly is counted as “documentation time” (time on notes, after-hours “pajama time,” or related EHR metrics) (Zhao et al., 2025). The review also notes an important exception pattern: while most consultation-note workflows showed reductions, at least one study focusing on AI-drafted replies to patient inbox messages did not show a significant time change, suggesting that “documentation” is not a single uniform task and AI’s value may differ by task type (Zhao et al., 2025).
The subgroup analyses sharpen what might be driving differences. For documentation burden outcomes, effects were directionally larger in consultation-note tasks than in other tasks (such as discharge summaries or inbox messages), and purpose-built tools tended to outperform general GPT-style prompting, although some subgroup comparisons did not reach conventional statistical significance (Zhao et al., 2025). For time outcomes, the difference between purpose-built tools and general GPT approaches was clearer: specially designed clinical documentation tools showed larger time-saving effects than general GPT technologies (test for subgroup differences p=0.04) (Zhao et al., 2025). This is an actionable takeaway: tools engineered for clinical note production, integrated into workflow and optimized for medical context, appear more reliably time-saving than “prompting a general chatbot” and pasting outputs into documentation (Zhao et al., 2025). The review also observes that studies conducted in real practice tended to show greater benefits than simulated-case studies for documentation burden, although these differences were not always statistically decisive, partly because the number of simulated studies was small (Zhao et al., 2025).
A third major finding concerns documentation quality and safety: across 10 studies assessing note quality, AI-generated clinical documents were at least comparable to clinician-written notes on the quality measures used, including structured instruments such as PDQI-9 variants and other assessment approaches (Zhao et al., 2025). However, “comparable quality on average” does not mean “error-free.” The review repeatedly underscores that AI drafts can contain inaccuracies, omissions, or unsuitable content, and therefore clinician review and editing remains necessary for safe deployment (Zhao et al., 2025). This matters because the value proposition of AI documentation is not merely speed; it is offloading cognitive work. If clinicians must spend substantial time policing errors or rewriting, gains may evaporate. The review’s stance is that AI can reduce cognitive load and burden by producing structured drafts, but quality control is not optional and should be treated as a core implementation requirement rather than a “later” feature (Zhao et al., 2025).
The evidence base, while promising, has real limitations that shape how strongly one should interpret these results. Methodological quality was generally low: many studies were non-randomized or before-and-after designs, sample sizes were often small, comparability between intervention and control participants was frequently unclear, follow-up was often incomplete, and none of the studies had multiple repeated measurements both before and after implementation, which limits causal inference about changes over time (Zhao et al., 2025). The setting distribution also limits generalizability: most evidence comes from the United States and other high-income countries, with no direct evidence from low- and middle-income contexts, where EHR maturity, staffing models, language differences, and workflow constraints may alter both feasibility and impact (Zhao et al., 2025). In addition, because AI tools evolve rapidly, some evaluated systems may become outdated quickly, and results can drift as models, interfaces, and integration quality improve or change (Zhao et al., 2025).
The review also probes publication bias. For documentation burden outcomes, funnel plot asymmetry was not statistically significant, which is mildly reassuring (Zhao et al., 2025). For time outcomes, however, funnel plot asymmetry was statistically significant (Egger’s test p=0.02), suggesting that smaller studies with weaker or negative time effects may be underrepresented, or that other small-study effects are present (Zhao et al., 2025). This does not negate the observed time savings, but it does argue for caution: the average time benefit might be overstated relative to what a fully representative set of studies would show, especially in early adoption phases where enthusiastic pilot sites are more likely to publish (Zhao et al., 2025).
One of the most practically insightful interpretations offered is that AI’s impact is deeply dependent on the surrounding digital infrastructure. The authors argue that AI should not be treated as a standalone fix for documentation burden. The maturity and interoperability of the EHR ecosystem, the absence of data silos, and thoughtful workflow integration influence whether AI reduces burden or simply adds another layer of technical complexity (Zhao et al., 2025). This point reframes AI documentation from a “tool purchase” to an organizational intervention: it requires training, governance, monitoring, and continuous evaluation in real clinical conditions, not only accuracy testing in controlled demonstrations (Zhao et al., 2025). The authors explicitly call for rigorous quality control and ongoing evaluation to optimize effectiveness while safeguarding patient care outcomes (Zhao et al., 2025).
Taken together, the most important message is balanced but clear. Across 23 studies, AI-assisted clinical documentation is associated with moderate reductions in documentation burden and time, with documentation quality generally comparable to human-written notes when clinicians review and edit drafts (Zhao et al., 2025). At the same time, heterogeneity is high, study quality is often limited, and errors in AI-generated text remain a real risk, making clinician oversight and robust implementation governance non-negotiable (Zhao et al., 2025). If you are thinking about what this means in practice, the evidence favors piloting purpose-built, workflow-integrated documentation systems, measuring outcomes that matter (time, after-hours work, perceived burden, burnout proxies, and quality), and treating “review and edit” as the default safety architecture rather than an optional workflow step (Zhao et al., 2025).
References: Zhao, J., Liu, H., Chen, Y., & Song, F. (2025). Application of artificial intelligence tools and clinical documentation burden: A systematic review and meta-analysis. BMC Medical Informatics and Decision Making. https://doi.org/10.1186/s12911-025-03324-w
Mini dicitonary:
Clinical documentation: the comprehensive recording of a patient’s medical history, diagnoses, treatment plans, test results, and related information used to support continuity of care and communication among providers (Zhao et al., 2025).
Documentation burden: the workload and cognitive load created by documentation tasks, including their contribution to stress, emotional exhaustion, and burnout-related outcomes (Zhao et al., 2025).
Electronic health record (EHR): the digital system in which clinical documentation is created and stored; inefficiencies in EHR interfaces and regulatory coding requirements are described as key contributors to documentation burden (Zhao et al., 2025).
Pajama time: documentation performed outside normal working hours, commonly because EHR and documentation tasks spill over beyond the clinical day (Zhao et al., 2025).
Ambient Clinical Intelligence (ACI): AI-enabled tools that can draft structured clinical notes by listening to patient–clinician conversations, moving beyond basic speech-to-text dictation (Zhao et al., 2025).
AI-assisted note creation: the use of AI technologies to produce drafts of clinical notes or related documents (for example discharge summaries and letters to patients), typically followed by clinician review and editing (Zhao et al., 2025).
Systematic review: a structured method of identifying, selecting, and synthesizing all relevant studies addressing a question, using predefined eligibility criteria and a transparent process (Zhao et al., 2025).
Meta-analysis: a statistical approach that quantitatively pools results from multiple studies to estimate an overall effect, here applied to outcomes such as documentation burden and documentation time (Zhao et al., 2025).
Standardized mean difference (SMD): an effect size that expresses the difference between intervention and control groups relative to variability across studies, used to combine outcomes measured on different scales (Zhao et al., 2025).
Random-effects model: a pooling approach that assumes true effects vary across studies and weights studies using within-study and between-study variance (Zhao et al., 2025).
Heterogeneity (I²): a statistic quantifying how much variability in effect estimates is due to real differences between studies rather than chance; higher values indicate more inconsistency across studies (Zhao et al., 2025).
Funnel plot asymmetry and Egger’s test: methods used to assess whether small-study effects or publication bias might be present in a meta-analysis when enough studies are available (Zhao et al., 2025).
References: Zhao, J., Liu, H., Chen, Y., & Song, F. (2025). Application of artificial intelligence tools and clinical documentation burden: A systematic review and meta-analysis. BMC Medical Informatics and Decision Making. https://doi.org/10.1186/s12911-025-03324-w