Purpose To review the current evidence on artificial intelligence (AI)-based diagnostic systems for ophthalmic emergencies and to evaluate their performance, clinical applicability, and implementation challenges in emergency department settings. Methods A narrative review was conducted with a systematic literature search of PubMed/MEDLINE and the Cochrane Library (January 2015–June 2025). Studies were included if they reported AI models for central retinal artery occlusion (CRAO), anterior ischemic optic neuropathy (AION), rhegmatogenous retinal detachment (RRD)/retinal breaks, acute angle closure (AAC), or infectious keratitis using readily available emergency-department imaging modalities (anterior-segment photographs, colour fundus photography, or retinal optical coherence tomography). AI architecture, imaging modality and specific devices, dataset size and source, level of validation, and reported performance metrics were extracted and critically appraised. Results Twenty-four studies met inclusion criteria. Deep learning models achieved high diagnostic performance in CRAO (area under curve [AUC] 0.96–0.99 from 2 studies), AION (AUC 0.97 from 1 study), RRD/ retinal break (AUC 0.888–1.00 from 7 studies), AAC (AUC 1.00 from 1 study), and infectious keratitis (AUC 0.65–0.997 from 13 studies). Promising results were also observed for smartphone-based fundus photography and portable OCT devices. However, most studies relied on retrospective, single-centre datasets with limited external or prospective validation. Conclusions AI demonstrates considerable potential to support rapid, accurate triage of sight-threatening ophthalmic emergencies, especially in resource-limited settings. Nevertheless, generalisability, data privacy, and integration into clinical pathways remain major barriers to routine adoption. Prospective multicentre trials with adequate external validation and privacy-by-design solutions are needed before widespread clinical implementation.