Governance, Ethics & Legal Practical insights
Explain AI customer decisions without inventing reasons
Give customers grounded explanations of AI-assisted decisions, protect other people’s information and provide a practical route to correction and review.
A fluent explanation is not necessarily an accurate explanation. If a customer asks why they received a particular offer, an AI assistant should not invent a plausible reason from general marketing knowledge. It needs the actual decision record and a defined boundary around what can be explained.
Identify the kind of decision
Distinguish a product recommendation from a decision with legal or similarly significant effects. GDPR Article 22 concerns decisions based solely on automated processing that meet that threshold, with specified exceptions and safeguards. Articles 13 to 15 contain relevant information provisions. [1] Have specialists assess the actual workflow rather than treating every recommendation as identical.
Explain from recorded facts
Capture the rule or model version, relevant inputs, available offer conditions and any human intervention. Use approved reason codes when they faithfully describe the outcome. For example, explain that an offer was limited to existing subscribers only if that condition actually drove the decision. A model’s generated narrative should not override contradictory system records.
Avoid exposing other people
Provide enough information to make the response useful without disclosing other customers’ records or confidential third-party material. If the available evidence cannot support a reliable explanation, say what can be verified and send the case to a person with authority to investigate. Do not fill a missing audit trail with confident language.
Make correction possible
Tell the customer how to challenge incorrect data or request the relevant review. Give the reviewer access to the underlying record and the power to change an outcome when appropriate. Track recurring explanation failures as a product defect: perhaps a rule was not logged or a field has an unclear source. Better explanations begin with a system that preserves meaningful reasons, not with a more eloquent response prompt.
Sources and evidence
Sources checked on 4 October 2026. Proposed workflows and hypothetical examples are editorial analysis.
From insight to practice