Explainable AI describes methods and controls that make an AI system's outputs understandable enough for people to evaluate and use responsibly. The required explanation depends on the decision, audience, model and risk.
What can an explanation show?
- Inputs that materially influenced the output
- Rules, evidence or retrieved sources used
- Confidence, uncertainty and limitations
- Reason codes for classifications or recommendations
- Changes that could alter the result
- Model and data version in effect
Explanation vs. transparency
Transparency describes what the system is, how it is governed and where it is used. Explainability helps a person understand a particular output or behavior. A transparent system can still produce outputs that are difficult to explain.
Why does it matter in commerce?
AI may help classify products, extract documents, score suppliers or recommend prices. Users need enough context to detect errors, challenge an outcome and know when manual review is required.
What makes an explanation useful?
It should be faithful to the actual system, specific to the decision and understandable to the intended user. A plausible narrative generated after the fact is not evidence of the model's real reasoning.
What should be retained?
Record the input, output, supporting evidence, confidence, explanation method, model version, reviewer action and final decision. Sensitive or proprietary details should be protected without making accountability impossible.
