The mechanical fixes — alternative text, a marked table header row, the document title and its language — are applied for you only after you approve each one. Everything that needs authoring — headings, slide titles, and link text that has to describe where it goes — is flagged for you to make; this build does not correct those automatically.
What was measured: the document's own structure, read directly from the file. Word .docx and PowerPoint .pptx only in this version; PDF files are not read. In a Word file the check finds missing alternative text, heading and title structure, unmarked table header rows, unclear link text, document language, and text contrast where the background is resolvable. In a PowerPoint file it finds missing alternative text on pictures, missing slide titles, the presentation title, document language, and text contrast where the background is resolvable; it does not check table header rows or link text on slides. The body of a Word file and the slides of a deck are read; headers, footers, footnotes and speaker notes are not.
What this does not cover: automated structure checks find a subset of accessibility barriers. Reading order, whether an image the file does not mark as decorative is decorative, and text placed over an image cannot be determined from the file and are listed below as not determined. A document with no findings is not a guarantee that it is accessible — it means these checks found nothing in it.
This is not a determination of legal compliance with the ADA, Section 508, WCAG, or any other standard, and it is not a certification. It describes what this check measured in one document at one time.