A method of enabling a remotely hosted AI to enhance a document while withholding sensitive information from the AI includes identifying sensitive terminology associated with the sensitive information and replacing by an exclusion filter of the sensitive terminology with redaction markers that cannot be interpreted as misspelled words or coined terms, thereby creating a redacted draft that is submitted to the AI. Upon receiving an enhanced, redacted draft from the AI, the original sensitive terminology is restored in place of the redaction markers, and the resulting enhanced document is delivered to a user. Sensitive terms can be local or global. Globally sensitive terms can be stored in databases directed to categories of sensitive information. Sensitive terms can include indicators directed to sensitive numerical quantities and/or other targets. In embodiments, the exclusion filter automatically corrects any grammatical errors arising from replacement of the redaction markers by the original sensitive terminology.