Next-Generation Agricultural Soil Science: From Rhizosphere Ecology and Soil Microbiomes to Climate-Resilient Farming

Artificial intelligence (AI) is increasingly embedded in agronomic sensing, crop diagnosis, yield forecasting, irrigation scheduling, nutrient management, crop protection and field robotics. Yet the literature often treats high predictive accuracy as evidence of decision value, even though agronomic decisions require reliable inference under variable field conditions, explicit objectives and constraints, and a credible link from prediction to intervention. This critical narrative review examines the transition from crop monitoring to prescriptive field management and asks where AI has demonstrated operational agronomic value rather than task-specific technical performance. Literature published principally from 1 January 2010 to 20 July 2026 was examined, with earlier seminal work retained where necessary. Evidence was synthesised across remote and proximal sensing, computer vision, multimodal data fusion, yield prediction, irrigation, nitrogen management, site-specific crop management, selective spraying, robotics and agricultural decision support. The strongest evidence supports AI for perception and state estimation, especially where labelled data are abundant and sensing conditions are controlled. Performance becomes less consistent under domain shift across fields, seasons, cultivars, sensors and management regimes. Predictive models can improve forecasting, but prediction alone does not identify the management action that will improve an outcome. More decision-relevant evidence is emerging from hybrid crop-model-machine-learning systems, in-season nitrogen recommendation, irrigation decision support and perception-to-actuation weed management, although multisite and multi-season intervention validation remains limited. A central conclusion is that agronomic AI should be evaluated as a decision system rather than only as a model. Trustworthy prescriptive field management requires external validation, uncertainty calibration, causal or response-aware reasoning, economically and environmentally meaningful objectives, machinery interoperability, human oversight and feedback from intervention outcomes. Future progress depends on shifting research benchmarks from model accuracy to decision quality, robustness and realised field benefit.