CMR OCR · transport & logistics
CMR waybill OCR: your consignment notes extracted automatically, all the way into your TMS.
A signed CMR comes in as a driver’s photo, a scan or an email. The AI reads every box — shipper, consignee, goods, weight, handwritten reservations — matches it against the transport order and writes the result into your TMS. Flagged cases go to a person. The engine is deployed on your server, and no page is ever billed.
Definition
The CMR, in plain terms.
The CMR consignment note (international road waybill) is the document that travels with any goods carried by road between two countries party to the CMR Convention. It embodies the contract of carriage and serves as evidence at every stage: taking over, delivery, any reservations. Issued in several copies (shipper, consignee, carrier), it follows the truck and comes back signed.
Its layout is standardized in numbered boxes: sender/shipper, consignee, place of delivery, place and date of taking over, documents attached, marks and numbers, number of packages, method of packing, nature of goods, gross weight, volume, sender’s instructions, carriage charges, carrier and successive carriers, carrier’s reservations and observations, then the signatures of the shipper, the carrier and the consignee with the date of receipt. It is this fixed structure, filled in very inconsistently, that the OCR has to read.
The problem
Why the CMR is hard to read.
On paper, the CMR is a form. In practice, it is one of the most hostile documents a conventional OCR can meet. Carbon-copy sets produce faint or smudged copies; the reservations box is filled in by hand, in pen, on the truck’s tailgate; the consignee’s stamps sit over the signatures and sometimes over the weights; and the copy that reaches the office is often a phone photo taken in a hurry on the loading dock, at an angle, under a strip light.
Then there is language: box labels are printed in two or three languages depending on the issuer, and handwritten remarks may be in Polish, Spanish or Italian. A generic OCR returns raw text without knowing which box it is reading or what a reservation is. The result: the document is rekeyed by hand, or worse, filed without being read, and the reservation only surfaces when the claim lands.
CMR data extraction
What the AI extracts from a CMR.
Out of the box, the engine recognizes a CMR among 36 transport document types and extracts its number, date and issuer. Full extraction of the boxes below is obtained by fine-tuning the model on your own CMRs during the pilot. It does not return text: it returns fields, structured and normalized, ready to be compared with your transport order. Every field carries a confidence score; below a threshold agreed together, it goes to human-in-the-loop validation.
Shipper and consignee
Company name, address and country from boxes 1 and 2, normalized to match your customer records and delivery sites.
Carrier and subcontractors
The main carrier and, where present, the successive carriers named on the document.
Taking over and delivery
Place and date of taking over, the planned place of delivery, and the actual date entered at reception.
Goods
Description, nature of goods, marks and numbers, number of packages, method of packing, gross weight and volume, line by line.
References
CMR number, transport order or purchase order number, truck and trailer registration when they appear on the document.
Reservations and signatures
Handwritten remarks at reception (missing packages, damaged packing), presence of the shipper, carrier and consignee signatures and stamps.
The workflow
From the driver’s photo to the TMS.
Automating a consignment note does not stop at reading it: its value lies in what it triggers next. Five steps, only one of which involves a person, and only on the cases that deserve it.
Step 1
Capture
The driver photographs the signed CMR from a phone, or the document arrives scanned by email, SFTP drop or through your existing mobile app.
Step 2
Extraction
The specialized model reads the document — printed, handwritten, stamped — and turns each box into structured, usable data.
Step 3
Checks and matching
The extracted data is compared with the transport order in your TMS: right customer, right site, right package count, right weight.
Step 4
Review of flagged cases
A hard-to-read box, a package count mismatch, a handwritten reservation: the document is shown to a person with the field in question highlighted. The human decides.
Step 5
Write to the TMS and archive
Validated data feeds the TMS or ERP, the CMR is archived with its metadata, and the proof of delivery is available for invoicing.
The AI does the extraction and the first check; your dispatchers decide on whatever falls outside the rules. The processing is a deliberate best effort: on field documents, the right measure is not “fully automatic”, it is “nothing wrong gets through without a person having seen it”.
Proof of delivery
From the signed CMR to the invoice.
The CMR signed by the consignee is the proof of delivery (POD). Until it has come back, been read and matched, the invoice waits, and so does your cash. Reading the CMR on the day of delivery, instead of the week the driver hands in the paperwork, shortens the gap between delivery and invoicing by the same amount.
It is also the best moment to deal with claims. A handwritten reservation detected at extraction — two packages missing, a damaged pallet — is raised immediately to your claims team, the file is flagged before invoicing, and you hold the documentary evidence without digging it out of an archive box weeks later. A CMR with no reservation that matches the transport order goes straight to invoicing.
The engine
A model specialized in transport documents.
GTEK does not wrap a general-purpose OCR in rules. The engine is a specialized model combining character recognition with language-model reasoning, measured in production on 195,514 real delivery notes and 2,365 layouts: 99.3 % of documents validated (97.9 % with no human touch), and, against human-verified ground truth, the issuer correct on 99.4 % and the date on 98.9 % of the automatically validated documents.
The same engine recognizes CMRs and extracts their number, date and issuer out of the box; it is then fine-tuned on your consignment notes to read every box, the handwritten remarks, the stamps and the field photos. Accuracy on your own CMRs is not a brochure promise: it is measured during the pilot, on a sample of your documents, with the same method used on delivery notes.
Deployment & license
On-premise, connected to your TMS.
The engine is deployed on your server, on your premises or in your hosting. Your CMRs, your customers’ details and your freight rates never pass through a third-party service. You keep your TMS or ERP: the engine connects to it by API, webhook, SFTP drop or mailbox, wherever your documents already flow.
The pricing model follows the same logic: a flat annual license, unlimited volume within your server’s capacity. No per-page pricing, no credits, so no meter running away in peak season, and no reason to choose which documents deserve to be read.
Comparison
Reading CMRs: three approaches.
| GTEK transport engine | Manual data entry | Generic OCR with per-page pricing | |
|---|---|---|---|
| Handwriting and stamps | Read by a specialized model fine-tuned on field documents | Deciphered by the clerk, legibility permitting | Often skipped or misread |
| Reservations detected | Spotted and raised as cases to handle | If the clerk notices them | Not interpreted |
| Transport order matching | Automatic, mismatches flagged | Manual, file by file | Rarely included |
| Cost at high volume | Flat annual license, unlimited volume within your server’s capacity | Grows with every document | Per-page or credit-based pricing |
| Where the documents go | They stay on your server | In-house | To a cloud vendor |
Frequently asked questions
CMR OCR: your questions.
Does the OCR read handwriting on a CMR?
Yes. Reservations, dates and names written by hand at reception are read by the specialized model, fine-tuned on field documents. When a remark stays ambiguous, the field is flagged for review and shown to a person rather than guessed.
Is a photo taken by the driver good enough?
Yes, it is the most common case: phone photo, rough framing, loading-dock lighting. The engine is built for this kind of capture. A photo that is truly unreadable is flagged for a retake, not silently interpreted.
Are CMRs in several languages supported?
Yes. The CMR is an international document by nature: box labels are printed in several languages and handwritten remarks may be in the driver’s or the consignee’s language. The engine identifies fields by position and content, not only by their label.
How does the data get into my TMS?
Through whichever channel suits you: API, webhook, SFTP drop or mailbox. You keep your TMS or ERP; the engine plugs into it, it does not replace it.
How are reservations handled?
A reservation at reception is detected, extracted and raised as a case to handle: the file is flagged before invoicing, and your claims team is told the same day instead of when the invoice is disputed.
Where do my CMRs and my customers’ data go?
Nowhere. The engine is deployed on your own server, on your premises or in your hosting. Documents, extracted data and models stay with you; nothing is sent to a third-party service.
How much does GTEK’s CMR OCR cost?
A flat annual license, unlimited volume within your server’s capacity, with no per-page or credit-based fees. The amount is quoted after a free 30-minute call, based on your document flows and your integration.
Your CMRs read on the day of delivery.
30 minutes to look at your consignment note flows, your TMS, and what a pilot on your own documents would let us measure.