Document extraction for CMR, T1 and eCMR: where AI saves time without putting the file at risk

Document handling is the work process nobody owns

Route optimisation has a planner, inventory management has a buyer, but the document flow often belongs to everyone and to no one. Part of it happens in dispatch, part in transport administration, part in the customs department or at the freight forwarder, and part with the driver at the loading bay. That is precisely why this work rarely features in an improvement project, even though it recurs with every shipment and is costly when mistakes happen: a wrong number of packages, a missing reference number or a declaration that is not discharged costs more time than the data entry itself. It is no coincidence that our overview of the processes where AI pays off first in wholesale and logistics names order processing and administration, alongside better-known applications such as route optimisation.

What is more, the amount of data per shipment is increasing, not decreasing. On 29 October 2024, the Douane (Dutch Customs) replaced the old NCTS in one go with the Douane Vervoer Applicatie (DVA, the Dutch customs transit application, also known as NCTS phase 5), without a transition period. The declaration structure went from two to four levels, and the six-digit commodity code is now mandatory at item level (evofenedex, the Dutch association for trade and logistics businesses). If nothing changes in the way of working, more data elements per shipment simply mean more data entry.

What document extraction is and is not

It is useful to keep two steps apart. Text recognition (OCR) reads what is on a document; that has existed for decades and works well on clean, fixed forms. What AI adds is the second step: understanding which field means what, even when the layout varies by sender, by freight forwarder and by country. Which number is the MRN (Master Reference Number), which address is the consignee's and which line contains the number of packages, on a scanned consignment note with a stamp half across it?

One thing extraction is not: validation. A model that neatly reads 24 pallets on a consignment note has only established what is on the paper. Whether there really are 24 pallets on the trailer, and in what condition, is a matter of observation at the loading bay. That distinction sounds obvious, but it is exactly the point at which document automation in logistics goes wrong: the system is trusted on something it can say nothing about.

Where the time savings actually are

The gain rarely lies in a spectacular new process; it lies in removing the repetition and searching that are now scattered through the day. Four places stand out:

• Retyping: data from the consignment note, declaration and invoice arrives pre-filled in your TMS or ERP, so that the administration team checks instead of typing.
• Consistency checks between documents: discrepancies between packing slip, consignment note, commercial invoice and declaration come to light before departure, instead of during an inspection or at discharge.
• Building the file: documents are automatically linked to the right shipment and the right reference number, so that a file is complete when a customer, an insurer or the Douane asks for it.
• Making exceptions visible: the system flags what it is not sure about or what differs from earlier shipments from the same trading partner, so that attention is directed there.

Accuracy is a percentage, your file is not

Vendors like to present extraction with a single number: an accuracy percentage. Always press for more, because a number like that says little without context. Accurate on which document types, on which fields, measured on whose documents, and does a field that stays empty count as an error or not? A model that performs excellently on neatly formatted in-house forms can stumble on a customer's handwritten note or on a scan of a scan.

More important than the average is the distribution of the errors. In document handling, not all fields are equal: a typo in a street name corrects itself along the way, a wrong number of packages or a wrong commodity code does not. So measure two things side by side: how many documents went through correctly without intervention, and how much time the documents that did need intervention took. If a system is unsure about ten percent of the documents but cannot say which ten percent, in practice you check one hundred percent and it gains you nothing.

So ask for a trial on your own documents, from your own senders and customs agents, instead of a demo on sample files. And record beforehand how many documents you currently process per week and how much time that takes; without that baseline measurement, a promise of time savings cannot be verified afterwards.

CMR: why a signature without reservations carries so much weight

International road haulage is governed by the CMR Convention, which turns the consignment note into evidence. Article 8 requires the carrier, on taking over the goods, to check the accuracy of the statements as to the number of packages and their marks and numbers, and the apparent condition of the goods and their packaging. Article 9(1) provides that the consignment note is prima facie evidence, that is, full evidence unless the contrary is proved, of the conditions of the contract of carriage and of the receipt of the goods by the carrier.

Article 9(2) makes clear why that weighs so heavily in practice: in the absence of specific reservations in the consignment note, it is presumed that the goods and their packaging appeared to be in good condition when they were taken over and that the number of packages corresponded with the statements. Anyone who signs without reservations therefore gives away their own evidential position. An extraction model can prepare that consignment note down to the last line, but it is not at the loading bay and does not see a damaged pallet. The check and any reservation therefore remain the work of the driver or the dispatch operative, not of the system that filled in the form.

In practice, that comes down to one working agreement: pre-filled data is a draft until someone has seen it. The better the extraction works, the more important that agreement becomes, because a document that looks complete and tidy is an invitation to sign it off unread.

eCMR and eFTI: going digital does not lighten the requirements

The electronic consignment note has its own legal basis: the Additional Protocol to the CMR concerning the Electronic Consignment Note (Geneva, 20 February 2008), which has been in force for the Netherlands since 5 June 2011. Article 3 requires authentication by means of a reliable electronic signature: uniquely linked to the signatory, capable of identifying the signatory, created using means that are under the signatory's sole control, and linked to the data in such a way that any subsequent change can be detected. Article 4 requires the procedure to ensure the integrity of the data from the time when the consignment note was first generated in its final form.

For the way you set up AI extraction, that is a specific design requirement, not a legal footnote. Extraction belongs before the moment at which the consignment note is finalised and signed. Any field a model still updates after that must be visible and traceable, otherwise you lose exactly the integrity the Protocol asks for. Also bear in mind that the eCMR only works between parties in countries that have ratified the Protocol; according to Stichting Vervoeradres (a Dutch foundation for transport documents and conditions), there are now 43 of them, while Belgium and Norway have signed but not yet ratified, and a separate arrangement applies to the Benelux.

Meanwhile, the direction is clear. Under the eFTI Regulation (Regulation (EU) 2020/1056), the implementing act for authorities was published on 20 December 2024, and from 9 July 2027 competent authorities will be obliged to accept legally required freight information electronically during inspections (TLN, the Dutch transport and logistics trade association). For businesses, that is a right and not an obligation, but it does mean that anyone who puts their document data in structured order now will have less work later than anyone who waits.

T1 and customs transit: the declarant remains responsible

In customs transit, the division of responsibility is even more explicit. Article 15 of the Union Customs Code (Regulation (EU) No 952/2013) provides that whoever lodges a declaration is responsible for the accuracy and completeness of the information given in it and for the authenticity, accuracy and validity of the documents supporting it. That responsibility does not shift because a model has pre-filled the fields.

The risk in a T1 declaration rarely lies in an entry that is obviously wrong, because that stands out. It lies in an entry that looks plausible: a commodity code close to the correct one, or a value copied from an earlier shipment from the same supplier. That is exactly where reusing earlier data both saves time and makes checking lazier. So agree which fields are always confirmed by a human, regardless of how certain the system claims to be.

On top of that, the evidence is increasingly digital. Since the introduction of DVA, the transit accompanying document has no longer been mandatory during transport, and showing the MRN is sufficient (Douane). Your digital file is then the file: if the link between document, shipment and reference number in your system is not right, you will notice it not at the counter but only when something goes wrong.

The human decides and signs

At AIBA, our starting point is that the human decides and signs. For document handling, that means in practice: the system pre-fills, a human confirms, and who confirmed can be traced. When setting it up, specify which fields, below which confidence threshold, always go to a human, and make sure it remains visible what the model proposed and what the employee changed. That trail is not only useful in a claim or an inspection; it is also the only way to see whether the system gets better or worse over time.

Article 4 of the EU AI Act (Regulation (EU) 2024/1689) has also applied since 2 February 2025. Since the amendment by Regulation (EU) 2026/1744, in force from 27 July 2026, it has asked organisations to take measures that support the AI literacy of employees who work with AI systems. They do not have to guarantee a specific level. For transport administration, what matters in practice is that the people who work with extraction understand where the model is reliable, where it typically stumbles, and when they should not accept a proposal.

Finally, think about the data itself. A consignment note contains personal data: the driver's name and signature, often a vehicle registration number and contact details. Having documents read by an external AI service is processing for which you need a data processing agreement, and for which questions about storage location, retention period and the use of your documents for model training should be answered in advance. The same consideration applies to AI tools in the workplace in general.

Where to start in practice

Do not start with all document flows at once, but with one document type that recurs often and whose layout is reasonably stable. For one week, count how many of them you process, how much time that takes and how often a correction was needed afterwards. That is your baseline measurement and also the yardstick against which you measure every vendor's proposal.

Then, before you choose a tool, decide which fields may never enter the system without human confirmation. In logistics, there are usually fewer of them than employees fear and more than vendors suggest. Our free Practice Scan helps you choose that first document type and sharpen the baseline measurement, so that an investment in extraction rests on your own figures and not on an accuracy percentage from a sales pitch.