Why OCR for freight documents is harder than for invoices
An invoice from a regular supplier typically has a recognisable layout: logo at the top left, total amount at the bottom right, invoice number in a fixed spot. A CMR freight document works differently. The CMR convention defines the fields, but not the exact layout. Every carrier, freight forwarder, or shipper uses their own form. Field 18 (remarks) appears at the bottom for one carrier and in the margin for another. On top of that, drivers add handwritten notes, customs stamps are applied, receivers sign off, and sometimes a carbon copy has bled through. Classic OCR software reads text based on coordinates: zone X contains the shipment number, zone Y the date. The moment the layout shifts, the software reads the wrong zone or returns no result at all. That is not a flaw in the software; it is the fundamental limitation of a template-based approach.
When is template OCR sufficient for reading freight documents?
Template OCR is not automatically the wrong choice. If your transport runs through one or two fixed carriers and their freight documents use the same form year after year, a template works well. You configure the zones once, the software reads them consistently, and maintenance costs stay low. The same applies if you generate your own freight documents in your TMS: in that case the layout is fixed by definition. Problems arise the moment you deal with changing carriers, international shipments with foreign CMR forms, or documents arriving through multiple channels: scan, photo, email, or upload. One new carrier means one new template to build, test, and maintain. With five carriers you have five templates; with twenty you have a maintenance problem.
What does document AI do differently when processing CMR documents?
Document AI does not read based on coordinates; it understands document context. The system recognises that 'Afzender', 'Expéditeur', and 'Absender' all refer to the same field, regardless of where it appears on the page. It handles varying layouts, multiple languages on the same document, and partially obscured text caused by stamps or stains. For handwritten additions, such as a driver printing a different quantity in block letters, modern systems perform noticeably better than classic OCR, although handwriting recognition remains a point of attention: not every handwriting style is equally legible to a machine, and sometimes not to a human either. The key difference from template OCR is that no per-carrier configuration is needed. New carriers are processed without setting up a new template. Deviations are flagged for human review rather than passed through silently as correct.
How does human review work for automatically processed freight documents?
Both template OCR and document AI require a human check, though the nature of that check differs. With template OCR you verify the output after the fact, because errors occur silently: the software reads the wrong zone without any warning. With document AI, uncertain fields are flagged so a team member only reviews where there is doubt. That is the human-in-the-loop principle: the machine handles the reading and data entry, the person approves. For CMR documents this is especially relevant because deviations on a freight document, damaged goods, shortages, or a load weight that does not match, can have legal consequences. That decision should never be left entirely to an algorithm.
Guidelines for the choice: OCR or document AI for your freight documents?
Choose template OCR if you work with one or two fixed carriers on fixed forms, your volume is low enough to catch occasional errors manually, and your IT team has the capacity to maintain templates. Choose document AI if you work with multiple carriers or varying formats, process international CMR documents in multiple languages, routinely deal with handwritten fields in your document flow, or need scalability without proportional growth in maintenance. A third question to ask: what do you do with the extracted data? OCR that only recognises text but does not link it to a shipment, order, or TMS record does not solve the real problem. What ultimately matters is that the right data ends up in the right place without anyone having to retype it.