Model Explainability (XAI) Procedure

How to use this document

• What it is for. This is a base technical procedure to decide which explainability techniques to use by model type, how to validate and document them, and which technical outputs (main factors, reason codes, counterfactuals) to hand to those who explain decisions to people, and to reviewers, auditors and authorities. It does not govern how requests for explanation from people are handled: that belongs to the "AI Transparency and User Explanation Policy."
• Who should own it. Usually the data science technical team, on behalf of system owners and coordinated by the AI lead. Enter the function in [RESPONSIBLE FUNCTION] and the person in [OWNER].
• Decisions to take first. Before adopting it, decide how you classify your systems by explainability level (Section 4), which technical tools the team uses, who reviews the reason code texts (Section 9) and what information you treat as a trade secret (Section 12).
• Fields to complete. Replace the bracketed fields, such as [ORGANIZATION NAME], [AI COMMITTEE], [REVIEW FREQUENCY] and [COMPLETE], and prepare an explainability sheet (aid E) for each level E2 or E3 system.
• Country variants. This document has no country variants. The minimum content and deadlines for explanations given to people differ by law; they are handled in the "AI Transparency and User Explanation Policy" and the "Automated Decision Review Procedure," which contains the country variants.
• Review. Ask legal and the data protection officer to review the reason code texts and the handling of trade secrets, and the AI lead to validate the level classification.
• Approval and publication. Have it approved by [AI COMMITTEE] or data leadership, record the approval in "Version history" and share it with data science teams and system owners.

Model Explainability (XAI) Procedure

Field	Value
Organization	[LEGAL ENTITY NAME]
Owning function	[RESPONSIBLE FUNCTION]
Procedure owner	[OWNER]
Approved by	[APPROVED BY]
Effective date	[DATE]
Review frequency	[REVIEW FREQUENCY]

1. Purpose

1.1 This procedure sets out how [ORGANIZATION NAME] selects, applies, validates and documents explainability mechanisms for its AI systems, and how it turns their results into useful explanations for each audience.

1.2 Its aim is that the organization can explain, where appropriate, what each system does, why it produced a given result, how it is controlled and what the affected person can do, without claiming more than the technique can support.

2. Scope

2.1 It applies to AI systems developed, fine-tuned or integrated by the organization whose outputs:

a) are used to make or support decisions about people;
b) must be justified to clients, auditors, supervisors or authorities;
c) feed regulated processes, such as creditworthiness assessment, insurance, hiring or workforce management.

2.2 For third-party systems, the organization obtains from the vendor the information needed to apply this procedure, or uses techniques that do not require access to the inside of the model.