Client documents, treated like client documents.

Azure OpenAI receives relevant full-page images and extracted text for structure and layout analysis in the EU Data Zone. Requests use storage disabled (store=false), so normal Responses API conversation state is not retained. Customer material is not used to train generative models. Microsoft performs automated abuse monitoring. Material flagged by the abuse-monitoring system, or identified as part of a potentially abusive pattern of use, may be stored in a Microsoft-controlled abuse-monitoring store and reviewed by authorised Microsoft staff. Documents and OCR remain in UK South.

Uploaded scans are scheduled for deletion from active storage two hours after upload. Azure Storage lifecycle rules make overdue source and standard extracted-text cache blobs eligible for deletion once they are more than one day old. Microsoft controls the exact lifecycle processing time. Deleted blobs may remain recoverable to authorised administrators for up to seven days under Azure soft delete.

Two conversion paths use AI differently. On Reconstruct, the standard path, Azure Document Intelligence OCR supplies every word and AI decides structure and layout only. On Restyle, an experimental path a workspace chooses deliberately, an AI model transcribes the words directly from full-page images, including handwriting, and a second focused AI call may propose a correction to an individual disputed span. A correction is kept only when it exactly matches the independent OCR reading. Every Restyle draft is held for human review rather than released as accepted.

During house-style setup, AI may suggest labels for a firm's own Word styles using style names and formatting only, never document content. A suggestion takes effect only after a person confirms it.

The full policy is in our privacy policy.