Bilingual document automation: invoices, contracts and forms in Arabic and English

Organisations in Qatar handle a constant stream of documents in two languages: supplier invoices, purchase orders, contracts, ID copies, certificates, official letters and application forms. Many arrive as scans or phone photos. Re-typing them into systems is slow, and the errors surface later as payment disputes or compliance problems.

Why Arabic documents are harder

• Connected script: Arabic letters change shape depending on their position in a word, and scans blur the joins between them.
• Mixed direction: Arabic runs right to left, while numbers, codes and English words inside it run left to right. Simple text extraction often scrambles the order.
• Two digit systems: one document can use Western digits (0–9) and Arabic-Indic digits (٠–٩), sometimes in the same table.
• Two calendars: dates may be Gregorian or Hijri, and a system must know which before it can compare them.
• Names and spelling: one name can be transliterated into English in many ways, which complicates matching with existing records.
• Stamps, signatures and handwriting overlapping the printed text.

How a modern pipeline works

Traditional OCR turned images into characters and left the rest to templates, which broke whenever a supplier changed its layout. Current vision-language models read a page more like a person does: they understand layout, tables and labels in both languages, and can extract fields into a defined structure without a template per supplier. A robust pipeline has five steps:

• Intake: documents arrive by email, upload or scanner; the system identifies the type (invoice, contract, form) and the language.
• Extraction: the model fills a defined structure, such as supplier, invoice number, date, line items and totals, and links each value to where it appears on the page.
• Normalisation: digits converted to one system, Hijri dates converted with the original kept, names matched against your supplier or customer records.
• Validation: business rules check the result. Line items add up to the total, the purchase order exists, the supplier is approved, the date is plausible.
• Routing: documents that pass go straight to the target system; anything uncertain goes to a person, with the doubtful fields highlighted.

People stay in control

Full automation is rarely the goal on day one. Set a confidence threshold per field: confident values pass automatically, and the rest are confirmed by a person on a review screen that shows each extracted value next to the original. Every correction is logged, which gives you an audit trail and the examples needed to improve the system. As accuracy is proven, the share of documents processed without review grows.

Contracts are different from invoices

Invoices are about fields; contracts are about meaning. For contracts, AI is most useful for summarising key terms, flagging clauses that differ from your standard, comparing the Arabic and English versions of a bilingual agreement, and answering questions with a reference to the exact clause. Legal judgement stays with your lawyers, and the system shows its sources so they can check quickly.

Measure before you go live

Take 100 to 200 real documents from your own suppliers and forms, including poor scans, and record the correct values. Measure accuracy per field, not per document: a system that reads totals perfectly but misreads dates needs a different fix from one that struggles with handwriting. Run the same set again whenever the model or the rules change.

Where to start

Pick the document type with the highest volume and the clearest rules, often supplier invoices, and connect the result to the system where the data is needed. Our AI automation work covers this from intake to integration, or talk to us with a sample of your documents.