Google Research has introduced Agile Deliberation, a framework that treats building an AI image classifier for a subjective concept as an iterative conversation rather than a one-off labelling task. The researchers argue that standard training assumes people already have a clear, stable definition of a category before they start labelling. That assumption holds for objects such as stop signs, but breaks down for concepts like healthy food, inspiring artwork or content moderation, where boundaries shift with culture, personal views and changing policy. The workflow has two stages. In concept scoping, the user breaks a broad idea into positive sub-categories and explicit exclusions, such as splitting healthy food into salads and lean proteins while excluding processed foods. In concept iteration, vision-language models surface borderline images, such as a fruit smoothie, that test those definitions. The user's decisions on these edge cases feed an automatic prompt-optimisation loop that adjusts the classifier. The team tested the approach in 18 live sessions of about 1.5 hours with human participants rather than simulated benchmarks, and reported less user fatigue and more consistent models that outperformed standard baselines. The underlying code and algorithms have been released on GitHub. For Australian businesses and professionals, this is a research prototype and changes nothing immediately. Teams building image classifiers for moderation or curation may find the method useful for writing down rules and edge cases before labelling, and the open-source code makes it cheap to try. If the images contain identifiable people or personal information, obligations under the Privacy Act 1988 apply to how that data is collected and used, whatever method is employed.