Document processing in logistics: reading CMRs and packing slips automatically

Document processing in logistics is reading consignment notes, packing slips, proof of delivery and supplier documents, so that what is on them gets into your WMS, TMS or ERP without anyone retyping it. With AI a system can take over most of that reading. A person then only looks at what does not match. Today it often still goes like this. A driver stands at dock 4 with a packing slip and a CMR with a stamp on it. Someone from goods in signs and puts the paper on a pile. At the end of the day a colleague in the office types the quantities in.

Which logistics documents can you process automatically?

Any document whose data someone now retypes, or holds next to a screen to check, can be read automatically. In a warehouse that is usually a handful, and you want something different from each.

• The supplier’s packing slip at goods in: item numbers, quantities, batch and best before date, to set against the advance shipping notice or the purchase order.
• The CMR or consignment note: sender, consignee, number of packages, weight and the remarks written on it at loading or unloading.
• The proof of delivery: who signed, when, and whether there is a reservation on it.
• Delivery notes and documents from foreign suppliers, often in English, German or Polish.
• Certificates, analysis reports and other attachments that belong to a shipment and now disappear into a folder.

If orders arrive as mail or PDF, that is the same technique at the front of the process, in order processing. That is covered in orders from mail into your WMS.

What is the difference between OCR and AI document recognition?

Classic OCR turns pixels into characters and reads a field from a fixed position, while AI document recognition reads the whole document and finds a field by what it means. That difference decides how much upkeep you have.

OCR with a template

With template OCR you draw a box per supplier: the order number is top right, the lines start two thirds down the page. That works as long as the supplier does not change its form. Move a logo, add a column or bring in a new supplier with its own layout, and the system reads the wrong box or nothing. With eighty suppliers you have eighty templates to maintain.

OCR with a language model

Intelligent document processing combines OCR with a language model. The OCR supplies the text and its position on the page, and the model works out which value is a quantity, which is an item number and which line belongs to which column. A new layout then usually means no new template. The risk changes too. A template reads nothing when the box is empty, while a language model can fill in a value that is not there. So you have the model point out, for each field, where on the page it read that value, and you check the result against what your WMS already knows.

How does reading a CMR or a packing slip work, from scan to WMS?

The document arrives, the system reads it, sets it against the order and either passes it on or holds it.

1. Arrival. A scan from the printer at goods in, a photo from the driver’s app, or a PDF in the mailbox.
2. Reading. OCR and a language model pull out the fields and map item descriptions to the item codes in your WMS.
3. Checking. The system sets the packing slip against the advance shipping notice or the purchase order, and the CMR against the trip in your TMS.
4. Pass or hold. If everything matches, the receipt is booked. If something differs, the document goes to a review queue with the reason attached.

On a CMR, reading the free text is no extra. Under the CMR Convention the carrier enters his reservations, with reasons, in the consignment note at loading. If the consignee takes delivery without making a reservation straight away for visible damage, he is presumed, unless proven otherwise, to have received the goods as the consignment note describes them. That one handwritten line next to the stamp is often the most important field on the paper.

Where reading consignment notes and packing slips goes wrong

Most of it goes wrong on paper that has been through several hands: handwriting, stamps, poor scans and documents in another language.

Handwriting

A driver who writes ‘2 pallets wet’ in ballpoint on the bonnet, or a warehouse operative who crosses out a quantity and writes 46 above it. Language models read handwriting better than classic OCR, but a crossed out number stays a doubt. That line belongs in the queue.

Stamps and signatures

The CMR Convention allows the signatures of sender and carrier to be replaced by stamps, if the law of the country where the consignment note is made out permits it. So a stamp often sits across printed text. The system then has to be able to say there is a stamp without reading the text underneath as a value.

Scans and photos

A copy of a copy, a photo taken at an angle in a dark cab, the second carbon copy where the ink barely shows. A photo that cannot be read should go straight back to the driver while he is still at the dock, and not only when the trip is closed in the TMS.

Several languages

A German Lieferschein, a Polish CMR, an English delivery note. A language model reads them side by side without a separate template per language. The units are harder: ‘Stk’, ‘szt’ and ‘pcs’ are the same, ‘Karton’ is not. You set that translation into your own units once.

How accurate is AI document recognition?

It depends so much on your own documents that a general percentage tells you nothing, so you measure it on your own pile. A packing slip from a large supplier’s system is a different thing from a CMR that has been written over three times.

What happens to an error matters more than the percentage. A system that books everything also books its mistakes, and you only find them at the next stock count. So nothing goes into the WMS without a check against something you already know: the advance shipping notice, the purchase order, the trip. Whatever does not match, or whatever the system could not read with confidence, lands in a review queue. Your colleague sees the document with the field highlighted and the value from the WMS next to it, and decides.

A proof of delivery often carries the name of whoever signed and an address. According to the Dutch Data Protection Authority, a name and an address are personal data, so reading and storing them falls under the GDPR. What that means for AI in your warehouse is in GDPR and AI in logistics.

How do you measure whether document processing in logistics pays off?

You measure it against a baseline of how it goes now, on the same documents, before anything is built.

• How many documents per week, per type, and how many minutes someone spends on each.
• How many errors surface afterwards now: wrong stock, a claim you could not back up, a receipt booked on the wrong order.
• A test set of a few hundred real documents for which a colleague has recorded the correct values. You measure the system on that, field by field.

After go live you count which share passes without a person, how long a document waits in the queue and how many errors still slip through. That last number matters most. It has to be lower than with retyping, or you have only moved the work.

What do you ask a document processing supplier?

1. Does the system show, for each field, where on the page it read that value? Without that you cannot see whether a language model filled in something that is not there.
2. What does it check a document against before anything goes into the WMS, and what does your colleague see when it is in doubt?
3. Can you measure it on a test set of your own documents, field by field, and does the supplier also count the errors that slip through without a person?

Then you pick the first document. That is the one where an error costs most and where your WMS already holds something to check it against, and that is usually goods in: the packing slip against the advance shipping notice. The CMR follows only once that first step works and has been measured, as our way of working describes. If part of your trips already runs on an e-CMR, read e-CMR and eFTI first, because what arrives digitally does not need reading. To see which of your own packing slips or CMRs the system would hold back, bring a week of scans to an introductory call.