How to work with scanned Urdu documents using AI

What AI is good for here

Government offices hold large amounts of Urdu material on paper: old files, registers, notifications, applications and handwritten notes. Officers often receive them as scans or phone photos. Optical character recognition (OCR) is the technology that turns an image of text into text you can edit and search. Many AI tools now include it.

Used carefully, it can save hours of retyping. It can also help you search old files, pull key details into a table or get a rough translation. But Urdu is harder for these tools than English. The mistakes are easy to miss.

Why Nastaliq is difficult

Most printed and handwritten Urdu uses the Nastaliq style. Letters join in flowing, sloping shapes, sit at different heights and overlap. Many tools were built mainly for simpler scripts. As a result, OCR on Urdu may:

• confuse letters that differ only by dots, such as ب, ت and ث
• drop or add dots, changing one word into another
• break or join words in the wrong places
• jumble the order of text in tables, columns or margins
• misread numbers, especially where Urdu and English digits are mixed
• struggle badly with handwriting, stamps, faded ink or poor photos

Older documents typed in other styles, such as Naskh, may give better results. Always assume errors until you have checked.

Protect the data first

Scanned files often contain personal data: names, CNIC numbers, addresses, property details, signatures and photos. Before you upload anything:

1. Use only tools your department has approved for official material. Do not upload official documents to a public chatbot or a free online converter. See data leakage.
2. Leave out classified or sensitive files, such as personnel, legal, police or medical records, unless an approved tool is available.
3. Share only the pages you need. Crop or cover parts that are not relevant. See data minimisation.
4. Keep the original. The paper or scanned original remains the official record, not the extracted text.

Step by step

1. Get a good image. Lay the page flat in good light. Hold the phone straight above it. Avoid shadows, glare and cut-off edges. One page per image works best.
2. Run the text extraction in your approved tool. Ask it to keep the original line order and mark anything it cannot read as "[unclear]" rather than guessing.
3. Compare line by line. Put the image and the extracted text side by side. Read slowly. Pay special attention to dots, numbers and proper names.
4. Ask a fluent Urdu reader to check anything important. Errors in Urdu text can be invisible to someone who reads it quickly.
5. Correct the text yourself and note on file that it was produced with OCR and checked.

What to check

• Names, CNIC numbers, dates and amounts. Check every one against the image. One wrong digit can change a case.
• Legal or official wording. If exact wording matters, rely on the original document, not the extracted text.
• Tables. Make sure rows and columns have not shifted.
• Gaps. Look for lines or words the tool skipped without saying so.
• Invented text. Some AI tools fill unreadable parts with plausible words. See hallucination.

Translation and summaries

Once you have checked the Urdu text, you can ask for an English summary or translation. Remember that every error in the extracted text will carry through. Check important passages with a fluent reader. See our guides on translation and extracting data.

What not to use it for

• Treating extracted text as an official copy of a document
• Relying on OCR alone for legal, financial or land records
• Uploading sensitive files to tools your department has not approved

For pulling checked details into a table, see the extract to table prompt.