A supplier sends a PDF packet with a quote, product drawings, and installation notes. You ask AI for a summary before a project meeting. The answer lists the products and delivery terms clearly. Before using it, check whether the drawing notes and tables were actually available to the model.

Datum recommends adding a short document intake record before an AI summary: which files and pages were supplied, what could be read, and which questions remain blocked by missing or unclear material. Keep that record beside the summary so the next person can see its limits.

Confirm how your tool handles the PDF

Anthropic’s current PDF documentation describes processing each page as an image alongside extracted text. It also warns that dense documents can exceed the model’s capacity before reaching the page limit, and that PDF analysis shares the limitations of vision tasks. These are documented capabilities and constraints, not proof that a particular project packet was understood.

Ask whoever maintains your AI tool whether it passes page images, extracted text, or both. Anthropic documents a specific Amazon Bedrock Converse configuration that uses text extraction without visual analysis when citations are disabled. That detail does not apply to every AI product, but it shows why the upload button alone cannot tell you how a file is processed.

Make an inventory outside the summary

Have the person preparing the packet list the expected files and revisions. Compare that list with the actual upload or processing record. Where available, use the PDF reader’s page count and the tool’s processing logs to check what arrived. Asking the model whether it read everything is not an independent check.

  • File identity: the original filename, document date or revision, and where the approved copy is stored.
  • Coverage: the expected files and pages, the pages supplied, and any upload or processing failures.
  • Readability: pages with unclear text, small drawing notes, cropped details, or tables that need checking.
  • Follow-up: the question affected by each gap, the person who will resolve it, and the replacement source needed.

Page count is only a coverage check. A packet can contain every page and still have an illegible note. Ask AI to flag uncertainty, then have a reviewer compare important extracted details with the original page. Preserve both the PDF page number and the printed sheet identifier when they differ.

Tie each gap to the decision it affects

Consider a hypothetical cabinet project. The quote clearly names an accessory, but a small note on its installation drawing is unreadable. AI may be able to summarize what was quoted. It cannot establish the requirement hidden in that note. The intake record should identify the drawing, the unreadable area, and the question that must go back to the supplier before anyone relies on the summary for ordering.

Request a clearer original or a legible detail from the source. If you enlarge or crop a page for another attempt, retain its filename, revision, page reference, and the surrounding drawing. A detached detail can lose the labels needed to interpret it. Have the responsible reviewer confirm the relevant requirement before the summary informs an order or installation decision.

Use precise status language. “Not found in the pages reviewed” describes a search result. “Page unavailable” describes missing input. “Text unreadable” describes a visibility problem. None of those statements establishes that a requirement does not exist. Keep unresolved items visible when the summary moves into a meeting agenda or handoff.

Try the intake check on a known packet

Choose a completed packet that an experienced person can review in your company’s approved AI environment. Before running the tool, have that person record the pages and details needed to answer a specific question. Include a working copy with a missing page or an unclear scan, clearly labeled for the trial so it cannot be mistaken for the project record.

Compare the intake record and summary with the reviewer’s answer. Record omitted pages, unsupported conclusions, incorrect page references, and time spent checking or requesting better material. These are proposed trial measures, not reported performance results. If the tool repeatedly misses the same kind of drawing detail, keep that part of the review with a person until a revised process passes the same check.

For the next live packet, name the person who clears each open question. Keep the source inventory, summary, and unresolved items together in the project workspace. A colleague should be able to open the cited page and understand what still needs confirmation without repeating the entire review.

Sources Read

  • PDF supportAnthropic · current platform documentation; accessed September 22, 2026

Next step, if this note maps to a problem on your desk: book a conversation with Adam. Thirty minutes, weekday evenings, Eastern time.

Related Field Notes