The example is Sam at Kestrel Design Studio, a made-up firm that gets about 60 vendor invoices a month. Some arrive as PDFs from billing software, some as phone photos or scans.
How do you tell which kind of PDF you have?
Try to select a word.
| Kind | Example in the lesson | What it means |
|---|---|---|
| Digital PDF | Printcraft invoice 3318.pdf, made by billing software | The words are stored as text. Any AI tool can read them straight away |
| Scanned PDF | Courier receipt (scan).pdf, a photo of paper | The file is a picture of a page. Nothing can read it until OCR turns it into text |
What does OCR actually do?
Three steps:
- Find: spot where text sits on the page (lines, boxes, table cells) and in what order to read it. Tables and two-column layouts are where this goes wrong most.
- Recognize: turn each shape into a letter or number, with a confidence score. Blurry photos, stamps and handwriting lower the score.
- Output: plain text, often laid invisibly over the original image, so the PDF looks the same but you can search and copy it.
Try it free, step by step
How AI Reads a PDF: OCR, Document Parsing and Field Extraction. Follow two vendor invoices, one digital and one scanned, through OCR, parsing rules vs AI extraction, and the four checks that stop a misread total from reaching your books. You do every step yourself in a practice copy of Stepthrough, it checks your work as you go, and “Do it for me” finishes any step you get stuck on. Free, and nothing touches your real accounts.
Start the free lessonFrequently asked questions
- Is OCR always accurate?
Usually right, but numbers on scans are where small mistakes hide. In the lesson, OCR on the courier receipt read a receipt number with the letter O where a zero should be, and swapped the comma and period in the total. That is why later steps check totals.
- Is OCR text the same as data?
No. Text is not data yet. Sam needs five fields from every invoice for the bills sheet, and parsing is the step that finds them in the text.
- What are the two ways to parse a document?
Rules: find the text after a label, using a template or parsing rules. Fast, cheap and exact when every invoice from a vendor looks the same. Docparser works this way, and a new layout breaks the rule. AI extraction: ask a model for the fields by name. It finds them wherever they sit and in whatever words. Flexible, but it must be told not to guess. The data extraction entry covers both.
- When should I use rules, and when AI?
Count how many vendors send the same layout every time. Those suit parsing rules. The rest suit AI extraction, such as an n8n Information Extractor. Either way, add the checks.
- What should happen when a field isn't on the invoice?
It stays blank. In the lesson, the Printcraft invoice has no due date, so the field is left empty, not filled with "30 days" or a made-up date. A blank you can see beats a guess you can't.
- Which checks should run before a row reaches the books?
Four, in the lesson:
- Do the lines add up to the total? This would have flagged the courier receipt's swapped total as a format problem.
- Have we seen this invoice number before? A duplicate check on vendor plus invoice number stops the same bill being paid twice.
- Any required field blank? A blank due date goes to a short review list, and a person fills it in from the vendor's terms.
- Over the approval limit? Anything over your limit waits for the owner's yes.
- Does document AI decide what gets paid?
No. In the lesson's words, document AI saves the typing, not the judgment. Software reads the invoice; a person still decides to pay it. See what to check before AI approves invoices.
- What are the other document jobs AI can do?
The lesson's cheat sheet lists four: read and summarize a long PDF with an AI PDF tool (Smallpdf, KDAN), extract fields from many documents into a spreadsheet, compare two versions with a compare tool (Foxit), and fill and sign a flat PDF without printing.
- Why can't my document bot read my PDF?
If it's a scan, there's no text in it until OCR runs. The retrieval FAQ explains how a bot answers from your files and why scanned PDFs come back empty.
- Where should I start with my own invoices?
Open three recent invoices and try to select a word in each. Write down the five or six fields you retype most, in the order your sheet wants them. Count the vendors with a fixed layout. Decide your checks before you automate.
Key takeaways
- If you can't select text in a PDF, it needs OCR first.
- OCR gives text. Parsing turns text into the fields you need.
- Rules suit fixed layouts. AI extraction suits varied layouts but must not guess.
- Leave missing fields blank and flag them.
- Check totals, duplicates, blanks and an approval limit before anything counts.
See what usually goes wrong in document AI mistakes, or turn this into a real workflow with extract invoice data with AI.