The scenario
A batch of supplier documents and identity records arrives as scans, half of them Arabic, some rotated, some with handwritten annotations. Each one is typed into a system by a person today.
Document AI demo
Arabic documents turned into structured, checkable data
Run your hardest files through the pipeline: scanned Arabic, mixed-language forms, Hijri dates, stamps and multi-page tables, extracted, classified and scored against a truth set.
In Saudi Arabia
What turns the Document AI demo into a production system
A working demo and a system a Saudi reviewer will sign off are two different things. These are the questions asked before go-live, and what we hand over against each one.
AuthorityWhat it requiresWhat we buildWhat you keep
SDAIA, PDPL
Lawful basis for every personal-data field the demo touches, before it runs on real records.
Field-level data map produced during scoping, sanitised extract used until that map is signed off.
Data map and lawful-basis note the client's DPO can file before production access is granted.
NDMO
Data classification and quality controls on the sources the demo would depend on in production.
Source-by-source classification, with quality thresholds agreed before the pilot widens.
Classification map and a quality gate report your data office keeps.
NCA ECC
Access control, logging and read/write boundaries clear before wider rollout.
Entitlements scoped per role, every action logged, write paths gated behind approval.
Control matrix and an action log your risk function can replay.
Where it runs
The demo itself can run on a sanitised or representative extract, but where the data cannot leave a system, the build happens inside your environment inside Saudi Arabia, not ours.
Arabic, as it is spoken here
Arabic questions and documents are handled as first-class input, not translated to English before processing, including organisational and Gulf-dialect terms.
Who it is for
Teams re-keying documents by hand: shared services, claims, procurement, records and archives.
Sample workflow
The chain of work behind the answer
1
Prepare
- De-skew, de-noise, page split
- Language and script detection
- Quality gate on unreadable pages
2
Read
- Arabic and English OCR
- Table and layout recovery
- Stamps and signatures located
3
Extract
- Fields typed to a schema
- Hijri and Gregorian dates normalised
- Confidence score per field
4
Validate
- Cross-checks against master data
- Business rules applied
- Low confidence routed to a reviewer
5
Deliver
- Structured records into the target system
- Original linked to every field
- Exception queue with reasons
Nothing is accepted on confidence alone, the reviewer queue is part of the design, and it shrinks as the truth set grows.
What you will see
Two examples, start to finish
A scanned Arabic commercial registration with a stamp over part of the number.
- Behind it
- De-skew → Arabic OCR → registration number field extracted at low confidence → cross-checked against the supplier master → flagged rather than guessed.
- What comes back
- A structured record with the number resolved from master data, the conflict noted, and the original image linked at field level.
A twelve-page contract with a payment schedule table spanning three pages.
- Behind it
- Layout recovery keeps the table as one object → rows typed → totals recomputed and compared with the stated total.
- What comes back
- A payment schedule as rows, plus a flag where the document's own total does not add up.
Deliverables
What you keep after the session
Every item below is an artefact your team can reuse, whether or not the engagement continues.
Extraction run on your batch
Your own documents, including the ones that usually fail.
Accuracy report by field
Scored against a truth set your team confirms, per document type.
Target schema
The field definitions and validation rules used, ready to reuse.
Exception handling design
Thresholds, reviewer workflow and the effort it implies at your volume.
Session agenda
- Bring your worst ten documents, we run them live
- Field-level confidence and where it drops
- Hijri date and Arabic numeral handling
- The reviewer screen for a low-confidence case
What we need from you
- Fifty to two hundred representative documents
- The target fields and where they must land
- A confirmed correct answer for a sample of them
Where data cannot leave a system, the demo is built inside your environment instead of ours.
FAQ
Frequently asked questions
Do the documents leave our environment?
No. The pipeline runs inside your approved environment, which is the point of deploying it this way.
How accurate is Arabic OCR really?
It depends entirely on scan quality and document type, which is why the demo reports a measured number on your files rather than a general claim.
Can it read handwriting?
Partially, and we treat it as a review case by default rather than trusting it silently.
Topic cluster
Explore the topic cluster
Curated links across our Saudi hubs, sector pages and long-form blog research.