A Scanned-PDF AI Research Workflow That Keeps the Original in View

Open the document and try to select a sentence. If the selection follows characters cleanly, the file likely has a usable text layer. If the cursor selects the whole page as an image, or copied text is empty or scrambled, treat it as a scan. Some PDFs are mixed: a clean cover and contents page may sit beside scanned body pages, or an old OCR layer may be present but unreliable.

Adobe's current Acrobat documentation explains the core distinction: a paper scan begins as image data, and OCR adds selectable, searchable text. That text is an interpretation of the image, not a replacement for it. Save an untouched copy before changing the file so later reviewers can compare every extracted passage with the source as received.

Do not evaluate OCR only on the cleanest paragraph. Pick five to ten pages that represent the document's failure modes: small type, two columns, a table, footnotes, a skewed page, handwriting, faded text, and a page with proper names or numbers. If the document has hundreds of pages, this small set tells you where broad extraction will be risky.

Write a control sheet with page number, visible feature, expected phrase, extracted phrase, and error type. You are not trying to calculate a universal accuracy score from a tiny sample. You are identifying patterns that affect the planned research question. A report about financial totals demands special attention to digits; an oral-history scan demands attention to names and speaker labels.

If you run OCR, select the correct page range and language, then retain the page image beneath or beside the recognized text. Adobe recommends reviewing the result for accuracy and completeness after recognition. That manual review is especially important when the document contains tables, mathematical notation, unusual fonts, marginalia, or low-contrast pages.

Avoid silently correcting the archival source. Keep corrections in a separate note: original image reading, OCR output, corrected reading, reviewer, and date. If a character cannot be resolved, mark it as uncertain instead of guessing. The same discipline should carry into an AI prompt: tell the system which pages have suspect OCR and instruct it to label unreadable passages rather than completing them from context.

Create a short map of page ranges: front matter, definitions, methods, findings, appendices, tables, and references. Use printed page numbers when available, but also record the PDF viewer page because front matter can shift the count. A locator such as “printed page 17, viewer page 23” is much easier to reproduce than a floating quotation.

Ask the AI for document structure and candidate passages, not a definitive summary on the first turn. Then inspect representative pages from every section it relies on. The public NIST AI RMF is a useful practice file because its numbered sections, tables, and defined functions make locator checks concrete; use the official PDF rather than an unattributed mirror.

Each note should contain one claim, the page locator, a short paraphrase, any necessary quotation, and an OCR status. Use statuses such as clean text, image checked, corrected OCR, uncertain, and non-text element. A note marked uncertain should not become a confident sentence in the final draft.

For tables, describe the row and column that support the claim and inspect the image yourself. For footnotes, preserve the marker and the note text together. For a name or number, compare at least the relevant line with the page image. These checks take longer than a one-click summary, but they target the exact details most likely to produce consequential errors.

Pause the workflow when a material passage cannot be read confidently, when tables are systematically scrambled, or when the file lacks stable page identity. Look for a born-digital edition, a higher-quality scan, an accessible transcript, or another official copy. The right outcome may be “not verifiable from this file.”