Purpose: To evaluate artificial intelligence (AI) models for surgical phase recognition in general surgery and assess readiness for real-time intraoperative use. Methodology: PRISMA 2020 systematic review. PubMed, Embase, MEDLINE and Cochrane were searched (1 Jan 2016–11 Dec 2024). Included studies used intraoperative video from general surgical procedures for automatic phase recognition and reported ≥1 quantitative metric. Two reviewers independently screened studies and extracted procedure type, dataset characteristics, model architecture, validation setting and inference speed (frames/s) where reported. Results: Of 3,201 records, 32 studies met inclusion criteria, comprising ~3,000 videos from 17 institutions. Procedural groups included laparoscopic cholecystectomy (n=16), colorectal/hernia (n=5), upper gastrointestinal (n=7) and hepatopancreatobiliary surgery (n=4). Accuracy ranged 0.50–0.96; complex procedures showed lower and more variable performance (0.50–0.70). Architectures evolved from Convolution Neural Network (CNN)–recurrent hybrids to temporal convolutional networks, transformer-based models and 3D CNNs, with newer models reporting improved temporal stability, reduced phase-switching noise and higher accuracy. Only 6 studies (19%) reported inference speeds compatible with real-time use, and only 1 described prospective live operating room deployment. Conclusions: AI phase recognition is accurate for standardised general surgical procedures in controlled settings, but limited external validation, inconsistent real-time evaluation and minimal prospective deployment remain major barriers to routine adoption. Multicentre datasets, standardised phase definitions, robustness to real-world video artefacts and theatre-ready benchmarking are priorities for safe clinical translation.