Tarnóca Freight
Introduction
Taking the paperwork out of 1,900 shipments a week

The brief
Dispatchers were spending their mornings retyping documents
Tarnóca Freight moves around 1,900 part and full loads a week across Hungary, Austria, Slovakia and northern Italy. Every load produces a trail of CMR consignment notes, delivery confirmations, customs documents and carrier invoices, almost all of it arriving as email attachments. Eleven dispatchers were spending the first two to three hours of each day keying that paperwork into the transport management system, and invoice mismatches were surfacing weeks later at month end.
Before
- Email arrives
- Dispatcher opens attachment
- Retypes fields into TMS
- Files PDF by hand
- Mismatch found at month end
With Zephyr
- Email arrives
- Agent reads and classifies
- Matches to shipment
- Files and updates TMS
- Exceptions go to a dispatcher the same day
What we did
An intake agent with a dispatcher in the loop
We mapped the process with the dispatch team, collected 2,400 real documents and labelled them together to build a test set. The agent we built watches the shared inboxes, classifies each attachment, extracts the fields that matter, and matches the document to a shipment in the TMS. When it is confident it files the document and updates the record. When it is not, or when a weight, price or reference does not match the order, it puts the case in a review queue with the difference highlighted. Dispatchers now clear exceptions instead of typing.
- Process mapping with dispatch and finance
- 2,400-document labelled test set
- Classification and extraction for 14 document types
- Matching against open shipments in the TMS
- Review queue with highlighted mismatches
- Monitoring of accuracy, volume and cost per document
The outcome
Results
71%
of documents filed with no retyping
3 days
earlier detection of invoice mismatches, on average
2.5 FTE
of dispatcher time moved back to customers and carriers
The system went live for the Győr hub in week 14 and was extended to the Sopron depot two months later. The accuracy target agreed at the start was 97 percent on extracted fields; it has run between 97.4 and 98.6 percent since launch. We continue to run it under a monthly support retainer.
“My dispatchers used to start the day with a pile of PDFs. Now they start with a short list of things that actually need a decision.”
What was involved
- Email & TMS integration
- Document extraction
- Review queue
- EU-hosted models
Other projects
All workHave a project in mind?
Get in touch
Tell us what you are trying to fix, in a few lines. We reply within one working day with questions or an outline of how we would approach it.
Give us a call
Head office
Delivery studio
