# Logistics documents and AI: what to automate first.

> In this field note, veridive explains what to automate first in logistics documents. It covers CMR waybills, bills of lading, customs declarations, proofs of delivery and carrier invoices in several languages, starting where a document triggers a decision, checking freight invoices against agreed rates, routing exceptions to the right desk, and keeping customs decisions with authorized people.

Logistics runs on documents from many parties, in many formats and languages. Start where a document triggers a decision, such as a proof of delivery that releases an invoice, extract with evidence and route exceptions to the right desk.

## Key takeaways

- Automate first where a document triggers a decision, such as a proof of delivery that releases or holds an invoice.
- Extract into one structure per document type, with the evidence attached, across languages, stamps and handwriting.
- Check carrier invoices against agreed rates by rule, and route each exception to its owner desk with the evidence.
- Customs classifications and declarations stay with the people authorized to make them; AI prepares the documents and checks them for consistency.

A shipment to a customer abroad leaves a trail of paper: a waybill signed at loading, an export declaration, a delivery note stamped at the customer’s door, a carrier invoice a few weeks later. Each comes from a different party, in its own format and often its own language. Most of it is filed and never read again.

A few of those documents decide something, though: whether an invoice is paid, whether a shipment moves. That is where logistics document automation should start. Extract the fields with evidence, and route every exception to the desk that owns it.

## Which logistics documents cause the most manual work?

| Document | What it carries | Manual work today |
|---|---|---|
| **CMR waybill** | Parties, goods, carrier, signatures and remarks for international road transport | Filing, and remarks nobody reads |
| **Bill of lading** | Shipment, containers, parties and terms for sea freight | Re-keying into the transport or ERP system |
| **Customs declaration** | Goods, values and references to supporting documents | Cross-checking against the commercial invoice and packing list |
| **Proof of delivery** | Signature, stamp, date and remarks at delivery | Matching to the shipment and releasing invoices |
| **Carrier invoice** | Freight charges, surcharges and extras | Checking against agreed rates |

The re-keying is visible. The costly part is often invisible: a remark that was never read, or a surcharge that was never checked.

## Where does a document trigger a decision?

> Automate the documents that decide something, not the ones that are only filed.

- **A proof of delivery** can release the customer invoice and the carrier’s. A remark on it should hold the carrier invoice and open a claim.
- **A customs discrepancy**, such as a quantity on the declaration that differs from the packing list, can hold a shipment until someone resolves it.
- **A carrier invoice line** outside the rate agreement should be queried before it is paid.
- **A bill of lading** that doesn’t match the booking needs correcting before the goods travel on wrong papers.

Take an illustrative example: a proof of delivery comes back from a customer abroad with a stamp, a signature and a handwritten remark: “two pallets damaged, accepted under reserve”. The system extracts the delivery date, confirms that a signature and stamp are present, reads the remark and recognizes it as damage. It links the document to the shipment, puts a hold on the matching carrier invoice and sends the case to the claims desk with the scan, the shipment record and the invoice attached. The claims handler decides whether to file a claim. The hold makes sure the invoice isn’t paid before that decision.

The same logic sets the order of work. Flows where a document releases or holds money come first, flows where it only updates a status come next, and documents that are only archived may never need AI at all.

## How do you handle many formats and languages?

Use structured data wherever it exists. If a carrier sends electronic messages or an API feed, read that; keep AI for the scans, PDFs and phone photos.

Along export routes, documents arrive in Turkish, English, German and other languages, often in the same shipment, with stamps over the text and remarks handwritten in whatever language the driver or receiver spoke. Three rules keep extraction reliable:

1. **One structure per document type.** Every proof of delivery becomes the same fields, whoever issued it and in whatever language.
2. **Evidence attached.** Each value links to the region of the page it came from, so a reviewer checks the crop, not the whole scan.
3. **Uncertain readings go to a person.** A smudged date or an unreadable remark is flagged, never filled in.

Build the evaluation set with samples from each carrier and country, because layouts differ more between issuers than between document types.

Remarks are translated for the reader, with the original kept beside the translation. Our notes on [reading Turkish documents with AI](https://veridive.com/insights/turkish-ocr-documents/) and on [structured output](https://veridive.com/insights/llm-structured-output/) cover the techniques.

## How are exceptions routed?

Each exception type has an owner desk, written down before launch:

- damage and shortage remarks go to the claims desk;
- customs mismatches go to the trade compliance team or the customs broker;
- rate discrepancies go to freight audit or accounts payable;
- missing proofs of delivery go to transport planning.

Every routed case carries the same packet: the document, the extracted fields with their crops, the linked shipment and invoice, the rule that fired and a proposed next step. The desk should be able to decide without opening another system.

Freight invoices deserve their own mechanism. The agreed rates, surcharge rules and extras such as waiting time are stored as data, and rules, not a language model, calculate what each shipment should cost. AI reads the invoice, matches each line to the shipment and its proof of delivery, and explains any difference. For example: waiting time billed for three hours, while the delivery note shows arrival and unloading forty minutes apart. The proposed action, pay, hold or query the carrier, waits for a person.

## What stays with customs brokers and operations staff?

Customs decisions and declarations stay with the people authorized to make them: a licensed customs broker or your own authorized staff. AI prepares and checks: it compares the declaration with the commercial invoice and packing list, flags inconsistencies and drafts the question to the supplier. It doesn’t choose tariff classifications, and it doesn’t submit anything. Which documents are required, and how long to keep them, are questions for your customs broker and counsel.

Operations staff keep the decisions that cost money or touch customers: releasing holds, filing claims, disputing carrier charges, approving payments, and telling a customer about damage or delay.

## How do you measure the gain?

Record a baseline first, then track:

- time from delivery to invoice release;
- carrier invoices checked against rates, all of them rather than a sample;
- discrepancies found and credits recovered;
- damage remarks caught before the carrier invoice was paid;
- shipments held because documents didn’t match;
- hours of re-keying, and extraction accuracy by document type and language.

## Follow the decisions

Pick one decision-triggering flow, such as proof of delivery to invoice release, and one owner desk. The illustrative [supplier document desk](https://veridive.com/work/#supplier-desk) shows the same pattern for supplier confirmations, [supply chain intelligence](https://veridive.com/solutions/supply-chain-intelligence/) covers the wider exception queue, and the [ERP and enterprise workflows](https://veridive.com/solutions/erp-enterprise-workflows/) page shows how holds and releases reach the ERP with approval. For quotes and supplier documents before the shipment, see our note on [AI for procurement](https://veridive.com/insights/ai-for-procurement-teams/).

## Frequently asked questions

### Which logistics documents can AI process?

AI can read most documents a shipment produces, such as CMR waybills, bills of lading, customs declarations and their supporting papers, proofs of delivery and carrier invoices, including scans with stamps and handwritten remarks in several languages. It extracts the fields into one structure per document type, attaches the evidence and sends uncertain readings to a person.

### How does AI help with freight invoice audit?

The agreed rates, surcharge rules and extras are stored as data, and rules calculate what each shipment should cost. AI reads the carrier invoice, matches each line to the shipment and its proof of delivery, and explains any difference, such as waiting time billed that the delivery times don’t support. A person decides whether to pay, hold or query.