Sources checked
Layout links text to the correct field
An original invoice workflow should test whether line items remain linked to the correct quantities and prices. A page can contain readable text while still confusing a system about which column or section it belongs to.
Verify important values against the source page
An answer that looks plausible can hide a recognition or layout error. Keep the source page available and review the extracted fields, especially names, dates, and amounts.
What OCR is responsible for
OCR maps pixels to characters. It can be a classical engine or a neural one.
It fails on handwriting, stamps, tiny fonts, and pages photographed at an angle. The output is a string, sometimes with boxes that say where the string sat.
A perfect OCR transcript of a contract is still not an answer to “what is the termination notice period?”
Layout matters. A page has columns, headers, and footnotes.
A naive reading order can merge two columns into nonsense. Tables are worse: a cell’s meaning depends on its row and column, which a flat string throws away.
If you need the table, store the structure, not only the words in the order the engine guessed.
What a later model adds, and what it hides
A document AI step takes the text, and sometimes the image, and fills a schema: invoice total, date, vendor. That step can be wrong even when every character was right, because it attached the number to the wrong field.
It can also be wrong because OCR missed a digit. Log both outputs.
Otherwise you cannot tell which stage to fix.
Do not let the model invent a field that the page does not contain. A missing total should stay missing.
Require the extractor to return the page and the snippet it used. A person reviewing exceptions needs that snippet.
A score with no snippet is not a review.
- Store the OCR text and the boxes, not only the final JSON.
- Test skewed pages, stamps, and the tables you actually receive.
- Refuse to fill a field the page does not support.
- Show the snippet a reviewer would check.
What to record before you rely on it
Save the page image, the OCR text, the boxes, and the extracted JSON as separate objects. When a field is wrong, mark which object failed.
Include skewed photos and tables in the test set, and record the error rate on fields that move money or legal terms. Require a snippet for every filled field.
If the reviewer cannot see the snippet, the field is not ready to post downstream. Keep the model id, the date, and the inputs you used beside that note, so a later change can be compared with the same test.
Common questions
If OCR confidence is high, is the extracted field right?
High confidence means the characters were a good match to a shape. The field can still be the wrong number from the right page.
Check the mapping, not only the character score.
Should I send the page image and skip OCR?
Some multimodal models accept the image directly. You still need a test on your pages, and you still want a quote or a box you can verify.
Skipping the transcript does not skip the error.
Sources & further reading
How this story was made
Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.
Our editorial standards





