In the context of AI applications, it is important to design an AI system so that its decisions are understandable from the beginning, rather than adding an explanation afterward.
This is called interpretable-by-design or explainability-by-design. Here, explanation is treated as a core design requirement, not an afterthought.
For example: Therefore, AI system should be designed so that its decisions can be understood, traced, and challenged as part of the system itself. So we should see the AI decisions as a system-design problem.
Each stage of system is deliberately designed to preserve enough information to answer “Why did the system do this?”
An explainable AI should ask:
“Can we understand why it produced that output?”
And, humans come in the loop as -
“Can someone verify, challenge, and correct that decision?”.
Let's talk about AI system that evaluates loan applications.
A poorly designed system might produce decision: Reject.
An system which follows explanable-by-design could preserve a decision path such as:
Application data → eligibility checks → risk model → risk score → decision policy → rejection
The user-facing explanation could then be:
“Your application was not approved because the calculated risk exceeded our Bank's threshold. The primary factors were X and Y.”
Meanwhile, an auditor could inspect the inputs, model version, policy version, and decision records etc.
So there are actually different explanations for different audiences who are asking different questions -
User: Why was I rejected?
Developer: Why did the model behave this way?
Auditor: Can we reconstruct and validate the decision?
Designer: Is the overall decision process fair and appropriate?
The questions become actions, evidence, and feedback loops.
That is where explainability meets system design. What happens after those questions is the crucial part of an explainable system. The questions become actions, evidence, and feedback loops. It is designing a system where explanations lead somewhere.