Document intelligence
Invoices, contracts, delivery notes, tenders and civil IDs read, checked and filed — in Arabic and English — with a person approving anything that involves money.
- KD 1,450
- build, once
- KD 195
- a month to run
- Arabic
- and English documents
The work this removes
Someone opens a supplier invoice, reads the number, types it into the accounting system, files the PDF, and repeats that ninety times a month. It is slow, it is boring, and boredom is exactly when transposition errors happen.
Document processing takes the reading and typing. It does not take the approving. Every extracted figure that affects a payment goes to a person before it becomes real, and the system’s job is to make that review take seconds instead of minutes.
Why Arabic documents are genuinely harder
This is the part most proposals skip, so here is what actually breaks.
Mixed numerals. Some invoices write amounts in Arabic-Indic figures, others in Western, and a fair number use both in the same document — one set in the header, another in the totals. Conversion looks reliable until a single digit is silently transposed, and the failure is quiet.
Dates. 03/04/2026 is either the third of April or the fourth of March, and no model can resolve that from the document alone, because the information is not in the document. It needs a rule tied to the supplier, or a convention you enforce, or a person.
Names that never match. The same supplier appears as a trade name, a legal name and an abbreviation, in two scripts, with inconsistent spacing. Matching that to one record in your system is a data problem, not a reading problem, and it is usually the largest part of the work.
Scans. Photographed at an angle, stamped over the total, faxed at some point in their history. Accuracy on a clean PDF tells you very little about accuracy on what actually arrives.
How we build it
Against your real documents, including the difficult ones, from the first week. We measure per-field accuracy — supplier, date, total, tax, line items — because a single overall percentage hides the fields that matter.
Every field carries a confidence level, and anything below the threshold is routed to a person rather than passed through. Every extraction keeps a link back to the page and position it came from, so a reviewer can check it in one glance instead of reopening the original.
Where the honest limit is
Handwriting in Arabic remains unreliable and we will not tell you otherwise. Documents whose meaning depends on an attachment nobody scanned cannot be resolved by any model. And a process where nobody can define what “correct” looks like is not ready to be automated.
What you get
Extraction running against your live document flow, a review screen your team actually uses, per-field accuracy figures measured on your own documents, and the integration into whatever system the data has to land in.
What it costs
Two parts, kept separate so you can see exactly what you are paying for.
Build — KD 1,450, once. One document type read end to end: extraction running on your real documents, a review screen your team will actually use, per-field accuracy figures measured on your own files, and the integration into whatever system the data has to land in. Paid in advance, before work starts.
Running — KD 195 a month. Processing for ordinary monthly volume, monitoring, and the corrections that new supplier formats force — and they always do. A twelve-month agreement billed quarterly at KD 585 every three months, each quarter paid in advance, renewing automatically.
The running cost is higher here than on our other services for one reason: documents keep arriving, and every new supplier is a new layout to handle.
A second document type is quoted separately. An invoice and a delivery note share almost nothing but the paper.
This is the service most often scoped as a single month of real invoices first, because a month of real documents tells you more than any demo. Send us a description of the document you handle most and roughly how many arrive a month.