Invoices and forms read into fields your system trusts, with a person on the doubtful ones

Three steps sit behind every extraction. First the image is cleaned: rotated upright, deskewed, shadows and noise reduced. Second the text is read, which is the OCR step, and the layout is understood so the system knows which text is a header, which is a table cell, which is a stamp. Third the values are matched to fields: invoice number, issue date, seller and buyer tax numbers, subtotal, VAT, total and line items. Cloud services from Google, Microsoft and AWS provide invoice-aware extraction, and we choose between them, an open-source stack and a model trained on your documents, based on your volumes, document variety and where your data must stay.

Arabic raises the difficulty. Letters change shape by position, diacritics are optional, numbers may be written in Arabic-Indic or Western digits, and lines mix right-to-left words with left-to-right codes and amounts. Vendor claims about Arabic support differ a lot between providers and document types, so we test candidate engines on a sample of your own documents before we recommend one. We do not quote a general accuracy figure, because the honest answer depends on your suppliers' paper, and we report results per field type instead.

The design principle is simple: read confidently what is clear, and send what is doubtful to a person. Every extracted field carries a confidence score, and you set the threshold per field. A total that does not equal the sum of lines, a tax number with the wrong length, a date in the future or a supplier not in your master data goes to the review screen. The reviewer fixes it in seconds, and the correction is stored so repeated problems with one supplier become visible. Stamps over amounts, handwritten edits and low-light phone photos are the hard cases, and they are routed to review by default, not hidden.

ZATCA shapes the context. Under the e-invoicing rules, many invoices are now created electronically, with structured fields and a QR code, and the authority publishes technical specifications for them. Where you receive electronic invoices, we can read the structured data and the QR directly instead of doing OCR at all. Paper and PDF invoices from smaller suppliers, delivery notes, customs documents and bank advices still arrive as images, and that is where OCR earns its place. Sending your own invoices to ZATCA is a separate job, covered by our ZATCA integration work, and this service does not generate or submit them.

Invoices and IDs contain company financials and personal data. Under the Personal Data Protection Law, you need a lawful basis, a retention period for source images, and a deletion path. We set retention on stored scans, restrict who can open the review screen, and log every correction for audit. Cloud extraction services send images to the provider's servers, so for sensitive documents we can use a self-hosted engine on infrastructure in the Kingdom, with the trade-offs in cost and tuning effort explained plainly.

A first pipeline for one document type, such as supplier invoices, usually takes four to eight weeks, including a sample test, validation rules and the review screen. Adding document types, ERP posting for more entities and a self-hosted engine takes three to five months. We start with the document type that costs you the most clerk hours, prove it on a month of real paper, and then widen the scope.