Comments live in the PDF, the fixes live in Word
Environmental Assessments and review records go back and forth between the preparer and the reviewers. The comments arrive as annotations on a PDF — often with reply threads between reviewers — while the document the team edits is a Word file. Keeping the two in step, round after round, is where comments get lost.
Step 1 — get every comment out of the PDF
Comment Report reads a marked-up PDF, from SharePoint or from a computer, and extracts every annotation and reply verbatim, in order, with the page it sits on. Reply threads are kept together. Optionally, a language model reads each comment with an image of its page and adds a one-line summary and the passage it refers to — so a comment like “revise per guide” comes with its context.
The result is a web page and a Word report with an Addressed checklist. It is filed on the project in our status board, which shows how many comments are done and closes the task at 100%.
Step 2 — answer in the document, visibly
For each comment we find the exhibit or section it points to, look at how the same item was written in approved documents, and draft the response in the document’s own voice. The rules are strict:
- The proposed text is shown in full and approved before anything goes into the file.
- Every edit lands in a new dated copy; a colleague’s file is never overwritten.
- Inserted text is highlighted, so the team accepts it by clearing the mark.
- Gaps in the sources are flagged for the preparer — never filled with a guess.
Step 3 — prove what changed
When a new draft comes back, a check compares every reviewer comment with the new text and marks each passage unchanged, changed, addressed or with no trace. People make the final call; the check makes sure nobody has to rely on memory.
What it looks like on a real round
On a Community Resilience Center EA, one round held 87 annotations in 47 threads. 27 threads are closed with before-and-after records and images of the revised pages, delivered in three batches of dated copies. On another center, the floodplain text we proposed was lightly edited by the team and signed.
AI’s job here is reading, not deciding. Extraction is exact and deterministic; the summary and the draft response are aids. Whether a comment is addressed is a person’s tick in the checklist.
