Every value, traced to the pagepage 1 it came from.
Documentu reads Arabic and English documents, traces every value to the exact place on the page, and puts it to work in the systems you already run — in our cloud, or sealed inside yours.
02Trust
Extraction is easy to demo and hard to trust.
Any model will pull a total off an invoice. The question that decides whether you can post it automatically is a different one: how do you know it is right, and what happens when it is not?
Every value comes with its trail: the page it was read from, the box drawn around it, the words it was found in, and the engine that read it.
Arabic-first: script identification, bidirectional reading order, and tables that span pages.
03Doubt
47 invoices. Line items do not sum to the subtotal.
The documents that need a person arrive grouped by what went wrong, not as a queue to open one at a time. Every group names a cause someone can act on: clear the group in one move, or fix the cause once and watch it stop coming back.
Every correction teaches the next run. The work that reaches a person shrinks month after month — which is the whole point of the product.
04Benchmark
Reads documents better than the flagships. At a fraction of their size.
On public document-parsing benchmarks, the engine behind Documentu outscores GPT-4o, Gemini 2.5 Pro and Mistral OCR — and it is small enough to run on a single GPU, which is what lets it run inside your building instead of someone else’s cloud.
Whole-document parsing
OmniDocBench · overall · English- Documentu87.5
- Gemini 2.5 Pro85.2
- GPT-4o76.7
- Mistral OCR73.2
Tables
OmniDocBench · table TEDS- Documentu88.6
- Gemini 2.5 Pro85.8
- Mistral OCR75.8
- GPT-4o72.0
126 languages, Arabic included
Multilingual document parsing · overall- Documentu82.3
- Gemini 2.5 Pro74.9
Real-world PDFs
olmOCR-Bench · overall- Documentu79.1
- Mistral OCR72.0
- GPT-4o69.9
- Gemini 2.0 Flash63.8
Published results on OmniDocBench, olmOCR-Bench and a 126-language document benchmark; competitor figures as reported in the same evaluations. Higher is better. OmniDocBench scores are 100 × (1 − edit distance). Bars start at 50.
05Route
Three engines. Each page goes to the cheapest one that can read it.
The routing decision is recorded per page, so the cost of a document is a number you can look at rather than a line on an invoice.
Point at a property to see where it lives.
06The Black Box
Nothing ever leaves your building.
The Black Box is Documentu sealed inside your own network — every model included. No call goes out: not to us, not to an AI provider, not to anyone. What runs, what it stores and what it could ever send is yours to inspect.
It is possible because the engine is small enough to run on one GPU in your rack. The flagship models behind most document AI are cloud APIs; they cannot come inside.
Built for the sectors that have to ask: Banks · Hospitals · Government · Defense
Our cloud
Managed by us. The fastest way to start, and the same system as the other two.
Your servers
Your hardware or a dedicated box. Your Postgres, your storage, your keys. Reachable for updates, nothing else.
Air-gapped
No route to the internet, ever. Updates arrive as signed images, handed over offline.
07How it works
Intake. Read.
Understand.
Check. Act.
- 01
Intake
Through the API, an email address we give you, the mailbox you already have, or the shared drives and ERP inboxes where documents already land.
- 02
Read
Every page goes to the engine that reads it best — layout, tables, handwriting, stamps and reading order, in Arabic and English.
- 03
Understand
It knows what kind of document it is holding, and pulls the fields your business needs into your schema — normalised, typed, ready to use.
- 04
Check
Totals that must add up, suppliers that must exist, amounts that must match the purchase order. What does not pass goes to a person, with the reason.
- 05
Act
Straight into your ERP, CRM and approval flows — or out as signed webhooks and JSON — every value carrying its page and coordinates with it.
08Automation
Then it does the work.
Reading a document is the beginning. Documentu’s automation layer turns what it read into finished work: records created, matched, approved and filed across your ERP, CRM and inboxes — each one linked back to the page it came from, so anyone can check where a number came from, months later.
- Create the vendor bill in your ERP, with the invoice attached and every line traced to its page.
- Match each document against your purchase orders, suppliers and customers before anything posts.
- Route the exceptions to the right person, in the tool they already live in.
- Update the customer record in your CRM the moment their contract or ID arrives.
ERP & accounting
- SAP
- Oracle NetSuite
- Microsoft Dynamics 365
- Odoo
- Sage
- QuickBooks
- Xero
- Zoho Books
CRM
- Salesforce
- HubSpot
- Zoho CRM
- Dynamics 365 Sales
Where documents arrive
- Outlook
- Gmail
- SharePoint
- Google Drive
- Dropbox
- WhatsApp Business
Where people work
- Microsoft Teams
- Slack
- Approvals
- Webhooks & API
Deep, document-aware integrations — not a pipe. Your system not listed? It is on the roadmap the day a customer needs it.
09Developers
An API, and the parts around it you would have had to build.
Signed webhooks with retry, backoff, and a delivery log you can replay. A sandbox mode that runs a real document against your workflow’s own unpublished draft without it counting against your usage.
- Sent
- No answer
- Trying again
- Delivered
10Questions
Short answers.
Yes — that is the Black Box. The whole system, every model included, runs on your servers or fully air-gapped, with no call out to us or to any AI provider. It is possible because the engine is small enough to run on a single GPU.
Early access
Start with one document type.
We are opening Documentu to a small group of early customers across the region. Tell us which documents you process and where they need to end up — we will come back with a plan built on your real documents.