A bank rejected someone's loan application. They asked why. The answer: "The model said so."

This is not acceptable. Not legally (under GDPR and the EU AI Act), not ethically, and not practically — because an unexplainable decision can't be appealed or corrected.

Explainability in practice, not theory:

For tabular models: SHAP values are the gold standard. "Your application was rejected primarily because of high debt-to-income ratio (45% weight) and short credit history (30% weight)." Actionable. Understandable. Appeable.

For text models: attention visualization and feature attribution show which words or phrases influenced the prediction. "The review was classified as negative primarily based on the phrases 'would not recommend' and 'disappointing quality.'"

For LLMs in RAG systems: cite your sources. "Based on section 4.2 of the employee handbook, your request is covered under policy X." The user can verify the source.

For image models: Grad-CAM and similar techniques highlight which regions of the image influenced the prediction.

The key insight: explainability isn't one technique. It's matched to the model type, the audience, and the stakes of the decision.

A data scientist wants SHAP values. A loan applicant wants "your income was too low relative to the requested amount." Same model, different explanation, different audience.

Build explanations for your users, not for your team.