Fazley RabbiWhen AI "reads" a document, two things usually happen. First, optical character recognition (OCR) turns the picture of a page into machine-readable text. Then an extraction step finds the specific details you care about, such as the invoice number, date, supplier and total, and puts them in the right fields. Modern AI models often do both at once, but the weak points remain the same: blurry photos, handwriting, stamps, tables and Bangla script all cause errors. The only way to trust the output is to test it against documents you have checked by hand, and to keep a person reviewing anything uncertain.
Step one: OCR, from picture to text
A scanned page or a phone photo is just an image: coloured dots, no words. OCR looks at the shapes and works out which letters they are. Older OCR tools did this letter by letter and struggled with anything unusual. Newer AI-based tools look at whole words and lines in context, which helps with smudges and odd fonts.
Bangla is harder than English for OCR. Conjunct letters, vowel signs that sit above, below or beside a letter, and the headline that joins letters together all give the software more ways to go wrong. A missing or misplaced vowel sign can turn one word into another. Mixed Bangla and English on the same page, which is common on invoices and forms in Bangladesh, adds another challenge.
Step two: extraction, from text to useful data
Raw text is not yet useful. You need "Supplier: …, Invoice no: …, Total: …" in your accounts or spreadsheet. Extraction is the step that finds those details.
- Template-based extraction says, "The total is always in the bottom-right box." It works well when every document looks the same, and fails when a supplier changes their layout.
- AI-based extraction understands that the number next to "Grand total", or its Bangla equivalent, is probably the total, wherever it sits. It copes better with variety, but it can also guess when a field is unclear, and its guesses look as confident as its correct answers.
What affects accuracy
| Factor | Easier | Harder |
|---|---|---|
| Image | Flat scan, good light | Angled phone photo, shadow, blur |
| Text | Printed, clear font | Handwriting, faded thermal receipts |
| Language | One language | Bangla and English mixed |
| Layout | Simple and consistent | Tables, stamps or signatures over text |
| File | Digital PDF with real text | Photocopy of a photocopy |
| Old Bangla files | Unicode text | Older non-Unicode fonts |
A tip that costs nothing: a digital PDF produced by a supplier's system already contains real text, so it skips OCR entirely and is usually far more reliable than a photo of the printout. Ask suppliers to email PDFs where they can.
How to check accuracy before relying on it
Do not judge a document AI tool by a demo with a perfect sample. Test it on your own paperwork.
- Build a test set. Collect thirty to fifty real documents of the kind you want to process, including bad ones: creased, handwritten, mixed-language, photographed on a phone.
- Create the correct answers by hand. For each document, type the true values of the fields you need.
- Run the tool and compare field by field. Do not stop at "it looks right". Check each number, date and name against your hand-made answers.
- Note error types. Is it misreading digits, swapping day and month, dropping Bangla vowel signs, or confusing the subtotal with the total?
- Look at confidence signals. Many tools can flag fields they are unsure about. Check whether those flags actually line up with the errors.
Building checks into everyday use
Even after a good test, keep safety nets in the workflow:
- Arithmetic checks. Do the line items add up to the total? Does the VAT line match the rate printed on the document?
- Cross-checks with your records. Is the supplier already in your list? Has this invoice number been entered before?
- Human review where it matters. Send uncertain fields, large amounts and new suppliers to a person before anything is paid or posted.
- Keep the original. Store the image alongside the extracted data so anyone can check it later.
These checks follow a broader principle explained in human review in AI workflows.
A practical start
Pick the one document type that eats the most typing time, often supplier invoices or delivery notes, and run the test above on a small batch. If the results are good with checks in place, the invoice data entry automation approach shows how to connect it to your accounts. If your documents are mostly in Bangla, read how AI handles Bangla first. I build document workflows like these as part of my AI and workflow automation service.