Governance, Ethics & Legal Practical insights
Before customer chats become AI training material
Review the purpose, permission, minimization and retention of customer conversations before using them for AI marketing analysis, retrieval or training.
Customer conversations can reveal recurring questions and confusing product language. They can also contain addresses, health details and information about people who never contacted your organization. Before sending transcripts to an AI tool, define the improvement you need and the least revealing material that can support it.
Distinguish the three uses
Summarizing a batch for an internal report, making conversations searchable and training a model are different operations. List which people and suppliers receive the text in each case. A system that retrieves individual tickets may expose identifiable passages even if no model training occurs. ‘We do not train on your data’ answers only one part of the review.
Revisit the collection purpose
Ask the privacy owner to assess the proposed use against the original purpose, notices and lawful basis. GDPR purpose limitation and transparency requirements remain relevant to AI processing. [2] Do not infer permission from the fact that the organization already holds the conversations. Record the specific decision for the selected workflow and provider.
Reduce detail before transfer
For a project about confusing delivery wording, a reviewed list of recurring questions may be enough. Remove irrelevant personal information at the source. Replacing names with customer numbers is not necessarily anonymization: the remaining context may identify someone. The EDPB emphasizes case-specific assessment of anonymity in AI models. [1] Apply the same caution to broad claims about a supposedly anonymous dataset.
Make the improvement measurable
Use a permissioned sample to identify a concrete wording problem, revise the relevant help content and track whether the same question recurs. Restrict access to original transcripts and define deletion dates for working copies. This keeps the project focused on better customer communication rather than building an indefinite transcript archive simply because an AI tool can ingest it.
Sources and evidence
Sources checked on 4 October 2026. Proposed workflows and hypothetical examples are editorial analysis.
From insight to practice