Check a scanned PDF before relying on an AI answer
The scan says £18.50. The text the AI reads may say something else, and it will calculate from that without blinking.
What you'll make
an AI answer from a scanned PDF in which every figure that matters has been checked against the image, with anything unreadable left unresolved.
What you'll need
Adobe Acrobat with the Scan & OCR tools (Acrobat Pro or equivalent) and, for questions, AI Assistant; a scanned PDF you’re allowed to use; and about 15 minutes. Work on a copy and keep the original. To practise, download the image-only practice scan (PDF).
Example result.
Page 1, row 1: display block, £18.50 each (checked against the image), quantity 3. Total: 3 × £18.50 = £55.50. Delivery and tax not shown on this page. Text layer originally read £16.50; corrected with Correct recognized text.
A scanned PDF can be just a picture of text. OCR (optical character recognition) adds a searchable text layer, and that layer is what AI tools read. It can misread characters, decimal points, minus signs and table order. A number you can select isn’t proof that it matches what’s printed.
1. Test whether the page has text
Open the PDF, press Ctrl+F (Windows) or Cmd+F (Mac) and search for a word you can see on the page. If nothing is found, the page is an image and needs OCR. In a mixed PDF, repeat on each important page: one searchable page doesn’t mean every page is.
Then find the figure you care about, with its label, unit and row, and decide which characters matter. A decimal point, a minus sign or a unit can change the answer.
Why this works: it takes twenty seconds and tells you whether the AI has anything real to read.
What you should see: “found” or “not found” for each important page, and the figures you’ll check.
2. Create or correct the recognised text
On your copy, go to All tools › Scan & OCR › In this file, choose the page range and language, then Recognize text.
To review the result, choose Correct recognized text in Scan & OCR and tick Review recognized text. Acrobat boxes the words it’s unsure of; select one, type the right value in Recognized as and select Accept. Check the figures that matter even if they aren’t boxed: errors don’t always get flagged.
If the image itself is unreadable, neither OCR nor a more confident prompt can establish the value. Ask for a better copy, or mark it unreadable. Don’t remove document protections to force access.
Why this works: fixing the text layer once fixes every answer drawn from it, including other people’s later.
What you should see: the selectable text matching the image for each important figure.
3. Ask a bounded question
Open AI Assistant on the corrected copy and ask:
From [document and page], report the item, currency, unit price and quantity, with their locations. Then show any calculation separately. Keep decimal points and units exactly as written. Mark unreadable or missing values; don't infer them. Use only this document and say if any page couldn't be read.[document and page] is the file name and the page you checked.
Why this works: separating the figures from the arithmetic shows you which went wrong if the answer is off.
What you should see: each figure with its page and row, then the calculation on its own line.
Check the result
- Each figure in the answer matches the image, not just the text layer.
- Units and “each” or “per” are kept, so the arithmetic uses the right basis.
- Figures come from the right row and column; tables can extract in the wrong order.
- Anything unreadable is marked, not guessed.
The usual failure: a single misread character, such as an 8 read as a 6, which produces a total that looks entirely plausible. Spot it by comparing every figure you rely on with the image at a readable zoom.
At [page and row], the scan shows [checked value and unit], but the answer used [wrong value]. Use the checked value, recalculate the affected result and cite the location. Leave anything unreadable unresolved; don't guess a missing digit.Worked example: the 8 that became a 6
Fictional example. The practice scan shows one line on page 1:
Display block £18.50 each Qty 3The Ctrl+F test finds nothing, so the page is image-only. In this example, after Recognize text, the answer to the step 3 prompt was: “Display block, £16.50 each, quantity 3. Total £49.50.”
£49.50 is a perfectly believable total. At 200% zoom, the image clearly reads £18.50: the OCR read the 8 as a 6. Correct recognized text didn’t box it, so it was fixed by hand in Recognized as. (Your own OCR run on the practice scan may read it correctly; the check is the same.) Rerunning the question, or using the repair prompt, gave 3 × £18.50 = £55.50.
- The total looked fine. Only comparing the unit price with the image caught the £6 difference.
- The fix went into the text layer, not just the answer, so the next question about this file starts from the right figure.
- Delivery and tax stayed unstated. The page doesn’t show them, so the answer doesn’t invent them.
Task kit for this job: Download (ZIP) · Reusable instructions
Go deeper
Check reading order and keep a record
A table can extract the right numbers in the wrong order. Confirm which row and column each figure belongs to, especially on multi-column pages. When a figure feeds a payment or a report, keep a short record:
Source, page and row:
Visible value, unit and label:
Text-layer value before correction:
Correction made, or manual transcription and its basis:
Checked calculation:
Unreadable or missing points:For checks across several documents, see compare several documents in an Acrobat PDF Space. For invoices, see extract invoice data and reconcile it.