Abstract Background Systematic reviews (SRs) are essential for the development of clinical practice guidelines but require time and human resources, raising concerns about sustainability and timeliness. Recent advances in machine learning (ML) and large language models (LLMs) offer promising opportunities to automate SR tasks, including citation screening. However, optimal strategies for integrating these technologies into guideline development workflows remain unclear. Main body Based on experiences from the SR automation team for the Japanese Clinical Practice Guidelines for Management of Sepsis and Septic Shock 2024, this perspective describes the practical application of ML‐ and LLM‐assisted approaches in guideline development. Semi‐automated citation screening using ML‐based tools reduced screening time, although performance varied across clinical questions. More recently, LLM‐assisted citation screening demonstrated further efficiency gains and improved flexibility without requiring task‐specific training data. Prompt engineering played a critical role in optimizing model performance, enabling sensitivity to be increased while preserving specificity. Comparative evaluations across multiple LLMs revealed inherent trade‐offs between sensitivity and specificity, highlighting the importance of model selection based on task priorities. Beyond citation screening, emerging evidence supports the potential role of ML and LLMs in search strategy development, data extraction, and quality assessment, although full automation remains limited by interpretability, domain variability, and the need for expert oversight. Conclusions ML‐ and LLM‐assisted automation is expected to reduce workload and enhance the efficiency of SRs in clinical guideline development. A hybrid human–AI approach, combining automated processes with expert judgment, represents a practical and safe pathway toward sustainable, timely, and reproducible guideline development.
Oami et al. (2026) studied this question.