Early AI results are promising, but some outputs could harm user trust. Would you launch?

Fictional support-drafting product: outputs go to a human agent before being sent. In 1,000 independently reviewed test drafts, 920 were usable without edits, 70 had minor errors and 10 contained another customer's information. Categories are mutually exclusive. The approved release policy blocks any external pilot with a confirmed cross-customer disclosure; average quality cannot waive that rule. The security lead owns incident assessment; product and security jointly authorize release. No root-cause fix or clean retest exists yet.