A QA workflow for teams shipping AI-assisted documents

Where generated documents break

Specifications, proposals and policies generated with an assistant tend to fail in the seams rather than the sentences. The prose is fluent; the structure is not. Headings skip a level, two sections share a name, a table is introduced and never referenced, and a placeholder survives to the final page.

The second failure is factual drift: the document asserts something the source material never said. Because the writing is confident and grammatical, nobody checks. This is the expensive one, and it is why a document review is a reading exercise rather than a formatting one.

The fifteen-minute pass

Do these in order. The structural items are quick and they make the reading pass more reliable, because a document with a clean outline is easier to check for factual continuity.

• Read only the headings first. If they do not form a coherent argument on their own, the outline is wrong.
• Check the heading levels: one H1, no skipped ranks, no duplicated names.
• Search for placeholders: double braces, [brackets], TBD, XXX, and the word 'example'.
• Check every link resolves, including anchors to headings that were renamed.
• Verify date formats are consistent across the whole document, especially in tables.
• For data files: confirm every row has the same number of columns as the header, and that no column is unnamed.
• Read every number against its source. Numbers are where a confident document does the most damage.

Handoff rules

A reviewed document should carry three things: who reviewed it, what they checked, and what they did not. The third is the one teams skip and the one that matters in a dispute.

Keep the automated report next to the final file. It costs nothing, it takes a second to produce, and it turns 'we think this is fine' into 'these specific checks ran on this specific version'.

What to record

Record the rule set version, the date, the reviewer's name and any finding you consciously accepted. An accepted finding is a decision, not an oversight — write it down as one and the next reviewer will not re-litigate it.

• Rule set version and score at the time of review.
• Reviewer name and discipline.
• Explicit out-of-scope list.
• Each accepted finding with the reason it was accepted.
• The hash or version identifier of the document reviewed.