The example business is Kestrel Design Studio, a made-up firm whose office person, Sam, handles about 60 vendor invoices a month.
What are the most common document AI mistakes?
| # | Mistake | Fix |
|---|---|---|
| 1 | Sending a scanned PDF straight to an AI tool | Check for selectable text; run OCR first |
| 2 | Trusting numbers OCR read from a scan | Check totals against the line items |
| 3 | Treating OCR text as finished data | Parse it into the fields you need |
| 4 | Parsing rules on vendors whose layout changes | Use AI extraction for varied layouts |
| 5 | Letting AI extraction guess missing fields | Leave them blank and flag them |
| 6 | No duplicate check | Check vendor plus invoice number |
| 7 | No approval step | Software reads; a person decides to pay |
What goes wrong before the text is even read?
1. Skipping OCR on scans. A scanned PDF is a picture of a page. Try to select a word and you get nothing, so no AI tool can read it until OCR turns it into text. Phone photos of receipts are the usual culprit.
2. Trusting scanned numbers. In the lesson, OCR on a courier receipt read the letter O instead of a zero in the receipt number, and swapped the comma and period in the total. Each shape gets a confidence score, and blurry photos, stamps and handwriting lower it. Tables and two-column layouts are where reading order goes wrong most.
Which parsing mistakes cost the most time?
3. Treating text as data. OCR gives you text. Sam needs five fields from every invoice for the bills sheet. Document parsing is the step that finds them.
4. Rules on changing layouts. Parsing rules ("the invoice number is whatever follows Invoice #") are fast, cheap and exact when every invoice from a vendor looks the same. Docparser works this way. A new layout breaks the rule. Count the vendors whose layout never changes: those suit rules, and the rest suit AI extraction.
5. Letting AI guess. AI extraction finds fields wherever they sit and in whatever words, but it must be told not to guess. The Printcraft invoice in the lesson has no due date. The right result is a blank, not "30 days" or a made-up date. A blank you can see beats a guess you can't.
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 lessonWhich checks do teams skip?
6. No duplicate check. Vendors resend invoices. A check on vendor plus invoice number stops the same bill being paid twice.
7. No approval step. In the lesson, anything over the owner's approval limit waits for a yes. Software reads the invoice; a person still decides to pay it. The same check list also catches the totals that don't add up (which would have flagged the courier receipt) and the required fields left blank, which go to a short review list.
The lesson sums it up: document AI saves the typing, not the judgment. The checks are what make it safe to trust 60 invoices a month you didn't type yourself. For a longer list, see what to check before AI approves invoices.
Key takeaways
- Check for selectable text first. Scans need OCR.
- Numbers on scans are where OCR mistakes hide, so check totals.
- Use rules for fixed layouts and AI extraction for varied ones.
- A missing field stays blank and gets flagged.
- Run four checks: totals, duplicates, blanks and an approval limit.
More answers are in the OCR and document parsing FAQ.