A document AI pipeline in production: what actually keeps it running

Hundreds of scanned paper timesheets every month. Manual keying, transcription errors, no systematic check. That was the starting point for a project delivered for an international industrial group, and the document extraction pipeline I designed and put into production addresses a very concrete problem: turning a scanned document into reliable data inside a business system, without removing human control where it actually matters.

The pipeline, in four steps

Automatic ingestion. Scanned PDFs dropped onto a network folder enter the processing queue on their own. No user action is required to trigger processing.

Two-pass extraction. A first OCR pass in the cloud captures text and tables. A second pass, using a vision-capable generative AI model, makes this extraction more reliable: field normalization, a confidence score for each extracted value, automatic matching against the employee directory.

Systematic human review. A validation interface shows the source PDF and the extracted data side by side. There is never a fully automatic import: every batch goes through an operator who confirms or corrects it.

Transactional import. The database write happens atomically, with a full rollback on the smallest error, and complete traceability for every processed document.

What actually matters is not the model

What keeps this system running in production is not the choice of AI model. It comes down to three architectural decisions:

• Human validation is mandatory, a non-negotiable auditability requirement in this industrial context.
• Every correction an operator makes is stored and fed back into the next extractions: the system improves with use, without heavy retraining.
• The extraction building blocks are interchangeable. If a better OCR engine or a better model ships tomorrow, it plugs in without touching the rest of the pipeline.

A system, not a proof of concept

This is not a prototype running in a demo. It is a system wired into the existing IT estate (network folders, employee directory, business database) and used every single day. This is exactly the pattern I find in most AI integration projects that hold up over time: an AI layer that proposes, a deterministic layer that validates, a human who decides on ambiguous cases. I cover this architecture pattern in a dedicated article.

My conviction, confirmed project after project: AI that creates real value for a business is AI wired into what already exists, with a human in the loop at the right point, not AI bolted on next to the system in the hope that it replaces oversight.