Techbanq

AI use case · documents

The last fifteen per cent is the whole project

Clean documents extract cleanly and always did. What decides whether a team can stop double-checking is how confidently the system says "I am not sure about this one" and what happens next when it does.

Serving Saudi Arabia from our Riyadh office and our engineering hubs in Lahore and Karachi

Operational analytics dashboard displayed on a desk monitor
  1. 01

    Market

    Saudi Arabia

  2. 02

    Written for

    COO, CIO

  3. 03

    Decision

    Automate extraction from invoices, contracts or forms

  4. 04

    Languages

    EN · AR

Enterprise delivery credentials

  • ISO

    ISO/IEC 27001

    ISO/IEC 27001 certified

Certifications & registrations

The pipeline

Where a document goes, and where it stops

Five stages, and the interesting one is the third: what the system does when it is not confident. A pipeline without that branch produces silent errors, which are more expensive than the manual process it replaced.

A document pipeline: arrive and classify, extract fields, validate against reference data, post or queue for review, learn from corrections — with human review for low confidence, failed validation or high value.

  1. 01

    Arrive and classify

    From email, portal or scanner. The document type is identified first, because an invoice and a delivery note need different fields and different validation.

  2. 02

    Extract the fields

    Bilingual documents, mixed numerals, stamps over text and photographs taken at an angle are the normal case here rather than the edge case.

  3. 03

    Validate against what you know

    A supplier that exists, a total that adds up, a purchase order that is open. Most errors are caught here by arithmetic and reference data rather than by the model.

  4. 04

    Post, or queue

    Confident and validated documents post straight through. Everything else goes to a review queue with the uncertain field highlighted rather than the whole document returned.

  5. 05

    Learn from the corrections

    Every human correction is recorded against the document type, which is how the exception rate falls over months instead of staying where it started.

Always available

Human review

Anything below the confidence threshold, anything that fails validation, and everything above a value your finance function sets — regardless of how confident the extraction was.

The economics

Automation that still requires checking has saved nothing

If a person opens every document to verify the extraction, the work has moved rather than reduced — and reviewing a machine's output is slower than reading the document, because the reviewer has to reconstruct what the machine saw. The saving only arrives when a defined share of documents is trusted to pass without a human, and that requires a threshold somebody is willing to sign.

That signature is a business decision, not a technical one, and it is usually made by value rather than by confidence alone: documents under a threshold post automatically, documents over it are reviewed however confident the extraction was. Framing it that way is what lets a finance function accept the arrangement, because the exposure is bounded and stated.

What breaks

Four realities of Arabic document estates

  1. 01

    The same document in two languages

    Bilingual invoices and contracts where the two halves disagree slightly. Which one is authoritative is a policy decision, and it has to be made before extraction rather than discovered in a dispute.

  2. 02

    Numerals in two systems

    Arabic-Indic and Latin digits inside the same document, sometimes inside the same total. Handled well this is invisible; handled badly it produces figures that are right in form and wrong in value.

  3. 03

    Photographs, not scans

    A large share of documents now arrive as phone photographs — angled, shadowed, sometimes with a thumb in frame. Any pilot run on clean PDFs will overstate what production will do.

  4. 04

    Stamps and signatures over text

    Common on approvals and delivery notes, and they sit exactly where the fields are. This is where the exception rate concentrates, so it is where the review queue design earns its keep.

Our role

How a document programme runs with us

  1. 01

    Measure before promising

    A week of your real documents, processed and scored, so the automation rate is a measurement rather than a proposal number.

  2. 02

    Build the exception queue first

    The review interface, the highlighting and the correction capture come before the happy path, because they are what the operation actually lives in.

  3. 03

    Write into the system of record

    Into the ERP or records system with the reference data validated, not into an export that somebody re-keys at the end of the month.

The value threshold above which a human always looks is set by your finance function; we build the pipeline so that threshold is enforced rather than assumed.

Common questions

What buyers ask

What accuracy can we expect on Arabic documents?+

It depends on the documents, which is why we measure yours before saying anything. Printed bilingual invoices from known suppliers behave very differently from photographed delivery notes with handwriting on them, and an average across both is a number that describes neither. Two weeks of measurement costs less than one wrong assumption.

Can it read handwriting?+

Printed and structured handwriting in known fields, often well enough. Free handwritten Arabic annotation across a document, rarely well enough to post without review. The practical design is to extract what is reliable and route the annotated documents to the queue rather than to force a single path for both.

Where does the data go?+

Wherever your residency position requires, and that is settled before the build rather than after. Processing can run in-region on managed services or on infrastructure you control; the choice changes cost and latency rather than the design, and it is one your legal and compliance function should make with the map in front of them.

Next step

Process a week of real documents

One document type processed end to end against a real week of volume, with the exception rate measured rather than estimated.