A Comprehensive Perspective on Explainable AI across the Machine Learning Workflow

Explainability requires attention throughout the entire AI workflow, not just at the final prediction stage. The paper identifies six critical components—data, analysis set-up, learning process, model output, model quality, and communication—that all need explanation. Focusing only on explaining individual predictions misses upstream decisions about data collection and downstream quality checks that determine whether AI insights can actually be trusted.

Different stakeholders need different explanations tailored to their roles. The framework distinguishes between domain experts, data analysts, and data scientists, each with distinct information needs. A one-size-fits-all explanation approach fails to serve these different users effectively, suggesting organizations should customize how they communicate AI decisions based on who needs to understand them.

Current AI explanation tools have significant gaps in coverage. The paper's survey of contemporary tools against a 112-item question bank reveals that existing toolchains don't adequately address the full range of stakeholder needs across the workflow. This suggests that organizations relying on standard explanation tools may have blind spots in their ability to verify and trust AI systems.

Effective explanations must be clear, actionable, and cognitively manageable. Drawing on human-computer interaction research and user studies, the paper identifies specific characteristics that make explanations useful in practice—not just technically accurate, but actually understandable and usable by the people who need to act on them.

AI systems can help bridge the gap between technical complexity and stakeholder understanding. The paper demonstrates that large language models can orchestrate multiple explanation techniques to translate technical details into stakeholder-specific narratives, potentially making AI systems more transparent and trustworthy to non-technical users.