Why does AI misread my tables and reports?
A PDF stores layout, not meaning.
You upload last quarter's report as a PDF and ask: "What was the revenue for the North region in Q2?"
AI gives you a number confidently. It's the South region's Q3 number.
The model may be fine. The problem probably happened before the model saw anything.
A Word document knows "this is a heading", "this is a table", "this cell is row 3, column 2".
A PDF mostly doesn't. It stores instructions like "draw the letters 'North' at this spot on the page" and "draw '4.2' over here". It's built to look right when printed, not to be read in order.
As one write-up on the problem puts it: "PDFs most often do not encode the semantic structure of their content." In plain English, the file doesn't record what each piece of text is.
Before the model can read a PDF, a tool has to turn it back into plain text, working out the order from where things sit on the page.
The model only ever reads what the extraction step gives it.
Two columns get read straight across, so a line from the left column runs into a line from the right.
Tables lose their grid. Numbers come out in a row, and which number belongs to which heading is left to guesswork.
Headers and footers ("Confidential – Page 7 of 40") end up in the middle of sentences.
Charts are pictures, so the numbers in them may not come through at all.
What you see:
| Region | Q1 | Q2 |
|---|---|---|
| North | 3.1 | 4.2 |
| South | 2.8 | 3.9 |
What can come out of the extraction:
Region Q1 Q2 North South 3.1 2.8 4.2 3.9
From that, "North Q2" could reasonably be read as 2.8 or as 4.2.
Made-up example.
A scanned document, or a photo of a page, has no text in it at all. It's just an image.
To read it, the tool has to recognise the letters first, and that step makes mistakes: 8 becomes B, 0 becomes O, decimal points disappear.
OCR (optical character recognition): software that turns a picture of text into actual text. It's the same thing your phone does when you copy text out of a photo.
Give it the original data, not the PDF. If the table started life in Excel or a database, share that, or paste it in as plain text.
Ask it to show you what it read first. "Before answering, reproduce the table on page 4 exactly as you see it." If the table comes back scrambled, you've found the problem.
Spot-check a few numbers against the page yourself before relying on its analysis.