An explanation is useful when it helps someone decide whether to trust and act on a prediction. Technical detail is rarely what they need.
Principles
- Explain in business terms: 'order frequency dropped and two support complaints are open'.
- Show the few strongest drivers, not dozens.
- Say how confident the model is.
- Make it easy to see the underlying records.
- Let users flag explanations that don't make sense.
Test explanations with the actual users. If they cannot say what they would do differently, rewrite them.