Useful AI features need context, but context is not a license to forward entire customer records. Minimization starts with naming the fields that actually change model output quality.
Prefer derived signals over raw documents when the task allows it. Summaries generated inside your boundary, redacted excerpts, and role-based views reduce exposure.
Write retention rules for prompts and completions the same way you would for application logs. If nobody will review the data, do not keep it by default.
Make vendor data-handling terms part of the technical design review, not a late legal checkbox. Engineering choices about batching, caching, and logging often decide what is practical to promise.