A Human Corrected the AI. Then the Data Changed (www.zingg.ai)

In the development of Zingg 1.0 Enterprise, a significant challenge arose when integrating human approval with AI decision-making in data management. Traditionally, humans would evaluate exceptions and apply rules to data, but with AI agents now processing records concurrently, questions emerged about decision ownership. Specifically, when both a human and an AI act on the same data, determining whose judgment prevails becomes complex, especially as the underlying data changes. This interaction between human judgment and AI adaptation can disrupt established practices, necessitating new governance frameworks for entity resolution and master data management (MDM). Zingg's open-source platform utilizes machine learning for entity resolution, enabling it to learn from labeled examples and handle complex data scenarios such as typos and variations in naming conventions. As businesses increasingly demand cooperative systems where both AI and humans contribute, Zingg encountered challenges in designing features that affirm the integrity of human decisions amidst dynamic data changes. The platform’s evolution led to key principles governing how human approvals affect subsequent AI actions. For instance, while a machine may add new records, it cannot override a human decision, thus ensuring accountability without sacrificing flexibility. This novel approach to human-plus-agent collaboration offers valuable insights into developing more reliable and intuitive AI systems within the AI/ML community.