Visual analysis plays a critical role in the grain receival and outloading process by ensuring the safety of food products and maintaining quality across diverse applications. Automation of this analysis promises improvements in repeatability over human sampling, enhanced transparency for growers, and increased throughput for site operators. However, a major challenge is the extensive labelling required during data collection. Each grain type includes dozens of varieties, numerous damage and unsound classifications, and hundreds of contaminants and foreign materials that must be identified, with the additional complexity of acquiring hundreds or thousands of examples per category to effectively train modern computer vision models. This is especially challenging given the relative rarity of damaged and unsound grains compared to healthy grains. This paper presents a novel approach that employs Multiple Instance Learning (MIL) to significantly reduce the labelling burden. By integrating MIL with conventional sampling techniques, labelled data can be collected with minimal additional overhead, yielding a model that can be either deployed directly or used to enhance the data acquisition workflow.