Legal has ruled that no customer record may appear in a prompt at inference time. The ML team proposes fine-tuning on those records instead. Adjudicate.

While legal concerns about data in prompts are valid due to request logs and vendor retention, fine-tuning is not a safer alternative. Facts embedded in model weights cannot be easily deleted or audited. The recommended approach is to keep records in a secure, deletable datastore and implement technical controls like redacted prompt logs, zero-retention agreements, and inference within a private boundary.