Artificial intelligence (AI) is increasingly applied to clinical microbiology workflows, but strong technical performance does not necessarily establish clinical usefulness. This narrative review, based on literature searched through 30 September 2026, critically examines AI-assisted culture-plate analysis, microscopy, MALDI-TOF mass spectrometry, genomic and rapid phenotypic antimicrobial susceptibility prediction, molecular-result interpretation, and laboratory-linked surveillance. Evidence was evaluated according to the independence of the test data, validation design, measurement unit, clinical comparator, and implementation outcomes. Culture-image screening has been assessed in a three-laboratory study involving 104,730 specimens, whereas some promising spectral models rely on very few independent test isolates. Machine-learning-assisted nanomotion AST has demonstrated 89.5–98.9% accuracy in an independent test series, but performance varies across antibiotic–organism combinations. A published PCR-interpretation study reported agreement across 22,200 interpretations and reduced manual analysis time. Separately, total laboratory automation shortened urine-culture reporting time in a before–after study, although this improvement cannot be attributed to AI alone. Evidence for genomic resistance prediction, prescribing support, and outbreak prediction remains heterogeneous, predominantly retrospective, or insufficiently externally validated. Important barriers include dataset shift, in adequate reference standards, sample-level data leakage, limited prospective clinical-impact studies, workflow costs, equity, and regulatory oversight. The central priority is therefore not higher accuracy alone but transparent reporting, independent multicentre validation, and evaluation of patient and laboratory outcomes. AI should be introduced as a monitored addition to microbiologist-led practice, with implementation decisions matched to the strength of evidence for each specific task.