Intelligent Document Processing for Arabic and English Business Documents

Intelligent document processing (IDP) is the use of OCR, layout analysis and AI models to read business documents, extract the fields that matter and pass them to systems such as ERP as structured data. For Saudi organisations, effective IDP must handle Arabic script, bilingual pages, Hijri dates and ZATCA e-invoice formats, with human review for uncertain values.

Most Saudi organisations still move a large share of their work through documents: supplier invoices, purchase orders, contracts, customs paperwork, HR forms and government correspondence. Many arrive as scans or PDFs, and many mix Arabic and English on the same page. Re-typing that information into ERP and case-management systems is slow, costly and error-prone — and it is exactly the kind of work intelligent document processing is built to remove. This guide explains how IDP works, why Arabic documents need special treatment, how ZATCA e-invoicing changes invoice data extraction, and where to start.

How does intelligent document processing work?

Instead of relying on fixed templates, IDP identifies each document type, finds the fields that matter and validates them before anything is posted. A typical pipeline has seven stages:

1. Capture: documents arrive from email inboxes, scanners, supplier portals, shared folders or APIs.
2. Classify: the system recognises what each document is — invoice, purchase order, contract, ID copy, delivery note — and splits multi-document files.
3. Read: OCR converts images to text, while layout models understand tables, key-value pairs, stamps and signatures.
4. Extract: AI models pull the required fields — supplier name, VAT number, invoice number, dates, totals, line items — regardless of where they sit on the page.
5. Validate: business rules check the data against master records, purchase orders and arithmetic, such as whether line items and VAT add up to the total.
6. Review: fields below a confidence threshold, or failing a rule, go to a person in a simple review screen.
7. Integrate and learn: approved data is posted to ERP, finance or document-management systems, and reviewer corrections improve future accuracy.

What is the difference between OCR, IDP and RPA?

The three technologies are often confused, but they solve different parts of the problem and work best together.

Technology	What it does	Handles new layouts?	Typical role
OCR	Converts an image of text into machine-readable characters	Reads any layout, but does not understand meaning	A component inside IDP
Template-based capture	Extracts fields from fixed positions on known forms	No — breaks when the layout changes	Legacy approach for standard forms
Intelligent document processing	Classifies, extracts and validates data using AI	Yes — learns fields, not positions	Turns unstructured documents into structured data
RPA	Automates clicks and keystrokes across applications	Not applicable — works on systems, not documents	Moves extracted data through the surrounding process

In practice, IDP reads the document and RPA carries the result through the workflow — logging in to the ERP, matching to a purchase order, routing for approval. Our article on RPA versus AI automation explains how the two fit together.

Why do Arabic documents need special attention?

Arabic introduces challenges that generic, English-first tools often handle poorly. This is why Arabic OCR accuracy should be tested on your own documents, not taken from a vendor brochure:

• Connected script: Arabic letters change shape depending on their position in a word, and poor scans break the connections OCR relies on.
• Mixed direction: right-to-left Arabic text sits alongside left-to-right numbers, product codes and English terms, often on the same line.
• Two numeral systems: Eastern Arabic numerals (٠١٢٣) and Western digits (0123) can appear on the same invoice and must be normalised before validation.
• Hijri and Gregorian dates: government and older commercial documents often use Hijri dates, which need converting before they reach systems that expect Gregorian.
• Name and address variation: the same supplier may appear in Arabic, in English or in different transliterations, so matching to master data needs fuzzy logic.
• Stamps, signatures and handwriting: official stamps frequently overlap printed text, and handwritten annotations add noise.

A solution built for Saudi operations should be evaluated on a representative sample of real Arabic and bilingual documents, with accuracy measured field by field.

How does ZATCA e-invoicing change invoice data extraction?

The Zakat, Tax and Customs Authority's e-invoicing programme (FATOORA) has changed what a supplier invoice looks like. Phase one, the generation phase, took effect on 4 December 2021. Phase two, the integration phase, began on 1 January 2023 and is being rolled out to taxpayers in waves, connecting their invoicing systems to ZATCA's platform. ZATCA is explicit that a paper invoice converted to electronic form by scanning or copying is not considered an electronic invoice.

Under phase two, invoices are generated as structured XML, and they can be shared with buyers as XML or as PDF/A-3 files with the XML embedded. Standard (business-to-business) tax invoices are cleared by ZATCA before they are shared, while simplified (business-to-consumer) invoices are reported within 24 hours. For accounts payable teams, this has three practical consequences:

• Read the XML first. When a supplier sends a compliant PDF/A-3 or XML file, IDP should extract the embedded structured data directly rather than OCR the visible page — it is faster and exact.
• Keep OCR for everything else. Suppliers not yet in scope, foreign vendors, older invoices, delivery notes and contracts still arrive as scans and PDFs.
• Validate tax data. Check the supplier's 15-digit VAT registration number against master data and confirm that VAT at the standard 15% rate reconciles with the totals.

If you issue invoices as well as receive them, our ZATCA e-invoicing integration service connects your ERP or POS to the FATOORA platform, and our e-invoicing guide for Saudi Arabia covers the requirements in detail.

Which documents should you automate first?

Document type	Typical fields	Complexity	Good first candidate?
Supplier invoices	Supplier, VAT number, invoice number, date, line items, VAT, total	Medium — varied layouts, consistent content	Yes — high volume, clear ROI
Purchase orders and delivery notes	PO number, items, quantities, delivery date	Medium	Yes, alongside invoices for three-way matching
Customs and shipping documents	HS codes, values, origin, consignee	Medium to high	Yes, for import-heavy businesses
HR forms and ID copies	Names, ID numbers, dates, nationality	Low to medium, but personal data	Only with strong PDPL controls
Contracts	Parties, dates, values, renewal and penalty clauses	High — long, unstructured text	Later, using language models for clause extraction

How do you keep document processing compliant with PDPL?

Many business documents contain personal data — employee files, national ID and Iqama copies, customer details. The Personal Data Protection Law, fully enforceable since 14 September 2024 and supervised by SDAIA through the National Data Governance Platform, applies to that processing. Sending such documents to an external cloud OCR or AI service can also raise questions about transfers outside the Kingdom. Practical controls include processing sensitive documents on infrastructure inside the Kingdom or on-premise, masking fields that downstream systems do not need, applying role-based access to the review queue, setting retention periods for source images, and keeping audit trails showing who approved each extracted value.

What does an IDP implementation look like?

A focused implementation usually runs in four phases. Discovery samples a few hundred real documents, maps the fields each downstream system needs and records the current cost and cycle time. Build configures classification, extraction and validation rules, tunes Arabic OCR on your samples and creates the review screen. Integration connects the pipeline to your ERP, finance or document-management platform through APIs where they exist, or through RPA where they do not — our ERP and custom software service covers systems that need new integration points. Hypercare runs the solution alongside the existing process, compares results and adjusts confidence thresholds before the manual step is retired. Choosing where the pipeline runs — in-Kingdom cloud or on-premise — is part of discovery, driven by the sensitivity of the documents.

How do you measure the return on IDP?

Measure the current process before you automate, then track the same metrics afterwards:

• Straight-through processing rate: the share of documents posted with no human touch.
• Field-level accuracy: measured on a held-out sample, per field and per document type.
• Cycle time: from document receipt to posting or payment approval.
• Cost per document: including review effort and exception handling.
• Exception rate and causes: so you can fix the rules or supplier behaviour behind them.

Automate one document type end to end, prove the numbers, then expand. In our process automation project for a Saudi industrial group, RPA with AI decision layers integrated across operational ERP systems delivered a 72% reduction in processing time and a near-zero error rate, with staff redeployed to higher-value work. Amwaj Solutions designs bilingual document-processing pipelines integrated with your ERP and finance systems — explore our intelligent automation and RPA service to see how we approach it.