AI automation for freight document processing

The problem

The administrative team processed several hundred documents a week by hand: carrier invoices, consignment notes, delivery confirmations. Data was re-entered into the billing system, and errors only surfaced at reconciliation.

What we built

• A Python service that receives documents from email and a shared folder.

• Field extraction with a language model, validated against data already in the system.

• A confidence score per document; below the threshold, it goes to human review.

• A review interface showing the document and the proposed values side by side.

• A full log: what the model extracted, what the human corrected, when and why.

Measurement

We ran the system in parallel with the manual process for six weeks, comparing results on every document before letting it write to the billing system. Automation was extended only to document types where measured accuracy passed the threshold agreed with the client.

Privacy

The documents contain commercial and personal data. Processing happens in European Union regions, with no model training on client content, under a data processing agreement signed in line with GDPR.