PDF document preparation: how the workflow works

Which page needs OCR

Sending every page down the same path adds operations and makes it harder to find where a read failed. The gateway checks the text layer of each page. A page with readable text goes on without OCR, a scanned page goes to OCR, and a corrupted text layer stops the document for a person's review. The file's history shows the source, the reading method chosen, the result of the check and the decision.

On 10 October 2026 we ran the OCR function on 12 generated documents with fictional data: a case number, a date and an amount. Six images were sharp, two slightly rotated, two blurred, and in two the identifier was covered. In the 10 documents with visible fields, all three values matched the reference. In the two with the number covered, the parser returned no identifier, while the date and the amount were correct.

What it costs and when it pays back

The count covers the work around a case: searching, reading, writing and handing over. What remains is the time to check the result, 20% of today's work in the scenario. Before the quote, we measure both times on 20 to 100 cases the team has already handled, in a free test that takes 5 working days.

In the demonstrator, a separate read of one document cost about €0.0008. The gateway's most expensive error is a page passed as readable when its text is corrupted, so at acceptance we count it apart from unnecessary OCR calls.

When OCR of the whole file is enough

If every document is a scan, or every document has a text layer, a page-by-page check adds little. A gateway makes sense with a mixed stream: contracts with scanned annexes, invoices with attachments, case files with photocopies. Open tools for converting documents, such as Docling, read PDFs with their page layout, so the gateway uses them as one component, and the part we build is the routing rules and the review queue. Acceptance keeps the choice of path apart from the accuracy of the fields and their approval.

Where to start

Collect representative files and judge the text layer of each page by hand. We measure missed OCR and unnecessary calls separately. Before a pilot you need a set of the client's documents, a retention policy and a manual queue for uncertain reads. We build this kind of rollout as part of AI automations. For the free analysis, a description of one case from recent weeks and of the tools it passed through is enough. You do not need to send confidential documents, and you get a written plan within 2 working days.