How can you automate the processing of shipping documents?

A customs declaration arrives by email. Included with it are an invoice, a packing list, and a bill of lading.

The first level of automation is now well established: identifying documents, extracting data, and feeding that data into a TMS or ERP system.

But let's say the invoice shows 10,000 kg, the packing list shows 9,800 kg, and the B/L shows 10,200 kg.

The three data points were successfully extracted.

However, the case poses a problem.

What information should we focus on? Why do these values differ? And should we let the case proceed?

That's where document automation takes on a whole new dimension.

It's no longer just a matter of reading a document or extracting data from it. We need to cross-reference information, verify its consistency, detect anomalies, and know when to involve a human operator.

Automation isn't just about processing documents faster. It's about getting reliable information to the right system at the right time.

What exactly needs to be automated?

Automating a document workflow involves more than just extracting data from a PDF. You need to consider the entire process, from the moment the document arrives until it is used in the business system.

Capture

Retrieve documents, whether they arrive via email, PDF, a portal, or an existing system.

Understanding

Identify the type of document and the relevant information: invoice, packing list, bill of lading (B/L), shipping order, certificate, etc.

Improve reliability

Compare data across documents, repositories, and systems. This is the stage at which inconsistencies can be detected before they turn into operational errors.

Take Action

Transmit reliable data to the TMS, ERP, WMS, or customs system and trigger the planned actions.

Supervise

When information is uncertain or an anomaly requires a business decision, the operator takes over.

Automation must follow the flow of information, not just that of the document.

Automating data before verifying its reliability: the pitfall

Using the example of shipping.

‍

The extraction is working.

Integration is working.

Automation works.

But the process is flawed.

The system did retrieve three pieces of data. It doesn't yet know which one to use, or why they are different.

This is the limitation of automation that stops at data extraction: a piece of data may be extracted perfectly, yet still be inconsistent with the rest of the file.

Before transmitting the information to the business system, it must be reconciled, verified, and any discrepancies flagged. This is the purpose of document reconciliation.

Document reconciliation changes the problem

Reconciliation no longer involves processing each document separately, but rather matching the information that describes the same transaction:

• references;
• quantities;
• weight;
• values;
• origin.

The audit may cover several levels: between documents, against operational data, against regulatory standards, or against the company's internal data.

Here are a few examples:

• difference in weight between the invoice and the B/L;
• Quantity discrepancy between the invoice and the packing list;
• HS code does not match the description of the goods;
• a shipping invoice that differs from the negotiated rate;
• Required document missing.

Simply extracting data is not enough; you must verify that it is consistent with other sources before continuing the process.

Reconciliation thus becomes an essential component of document automation:

extract → compare → verify → detect discrepancies → trigger the appropriate action.

Automate an end-to-end document workflow

In the case of an international shipment, not all documents arrive at the same time, and each one provides a portion of the information needed for processing.

From the Customs Declaration to the TMS

A customs declaration arrives via email along with its attachments.

The system identifies the documents, extracts the relevant information, and sends the data to the TMS.

The operator no longer has to re-enter the same data. The operator steps in when a case requires verification or a decision.

New documents under review

The invoice, packing list, and bill of lading arrive next.

Their data is extracted and then cross-checked to verify references, quantities, weights, or values.

If the information matches, the case will continue to be processed.

If a discrepancy arises, the system identifies the relevant data and flags the item for verification.

From Anomaly to Decision

A discrepancy is detected: the system identifies the relevant data and the documents causing the discrepancy, and alerts the operator.

The user can then analyze, correct, or validate the text, depending on the context.

The operator then reviews the case and either approves it, corrects it, or requests further action.

The goal is not to eliminate human involvement, but rather to focus it on situations that truly require a decision. 

From Reliable Data to Business Systems

Once the checks have been completed, the data can continue on to the TMS, ERP, WMS, or the relevant business application.

Automation doesn't stop at extraction; it guides the document all the way to verified, actionable data.

‍

Which checks can be automated?

Once the data has been extracted, automation can go a step further: verifying that the information is complete, consistent, and compliant with the business process.

Consistency Check

Compare information that is common to multiple documents:

Invoice ↔ packing list ↔ B/L

References, quantities, weights, or values can be compared to identify discrepancies.

Completeness Check

Verify that a file contains all the necessary documents and information before proceeding with its processing.

Regulatory Oversight

Compare the information in the file with the applicable rules and requirements, particularly those related to customs procedures.

Financial Control

Compare a shipping invoice with the quote or the agreed-upon pricing terms to identify any discrepancies.

Business Control

Compare the data extracted from the documents with the data already in the TMS or ERP.

In all these cases, the goal remains the same: not only to retrieve data, but also to verify that it is reliable within its business context.

What should you do when an anomaly is detected?

Detecting an anomaly does not mean it must automatically be corrected. In sensitive logistics processes, automation must identify deviations and guide the appropriate action.

What Can Be Automated

The system can:

detect → monitor → alert → recommend

He identifies the discrepancy, locates the relevant information, and indicates which point needs to be verified.

What Must Remain Under Supervision

Certain situations require a business decision, particularly when they involve:

• financial data;
• customs procedures;
• compliance requirements;
• an operational decision.

The process then becomes:

analyze → decide → approve

The goal is not to replace the expert, but to allow the expert to focus on exceptions rather than on systematically reviewing every document.

Keep a record of the decision

Reliable automation must also make it possible to understand why an anomaly was flagged and what decision was made.

You need to be able to find:

• the source document;
• the data subject;
• the detected anomaly;
• the control applied;
• the validation or correction that was performed.

This traceability is essential when processes are subject to control or audit requirements.

Where does automation create the most value?

Not all document-based processes have the same potential.

So the best place to start isn't necessarily the process that contains the most documents.

It is the one where volume, repetition, manual checks, and cross-referenced data create a workload or risk significant enough to justify automation.

How can you tell if a process is ready to be automated?

Before choosing a technology, you need to look at the process itself.

5 Questions to Ask Yourself

1. How much time do the teams spend handling this process?
2. How many times is the same data re-entered?
3. Does the same information appear in multiple documents?
4. Which anomalies still require human validation?
5. In which system will the data ultimately be used?

A process that is very large-scale, repetitive, and based on recurring document reviews is generally a good candidate.

Conversely, a low volume, highly variable documents, or a high proportion of business decisions may justify retaining a greater degree of human involvement.

We shouldn't automate just for the sake of automating; instead, we should focus on the steps where automation provides real benefits without losing control of the process.

How can you measure the success of document automation?

Automation cannot be measured solely by the number of documents processed. It is important to assess what it actually changes in the process.

Four indicators are used to track results:

• Processing time per case: How long does it take to process a transaction?
• Re-entry rate: What percentage of the data still needs to be entered manually?
• Rate of detected anomalies: How many discrepancies are identified before the process continues?
• Percentage of cases processed without human intervention: What portion of the workload can actually be automated?

The last indicator should be interpreted with caution: greater automation is not necessarily better if it increases the number of undetected errors.

The right KPI isn't just the extraction rate. It's the rate of cases that are actually processed all the way through.

Performance is thus measured at the business process level, not just in terms of the technology used.

This approach aligns with trends in the IDP market: classification, extraction, structuring, and even human-in-the-loop validation are now considered core capabilities. Differentiation is shifting toward process automation, workflows, and agents.

How far can automation go?

Automation goes far beyond classification and extraction. The level to aim for depends primarily on the complexity of the process and the decisions that need to be made.

Level 1: Automate Tasks

Document classification, data extraction, and data transfer to business systems.

Level 2: Automate Checks

Document reconciliation, consistency and completeness checks, and compliance checks.

Level 3: Automate Orchestration

The system can perform a sequence of actions, make recommendations, and handle certain decisions under human supervision.

This is where a document-centric, agentic AI approach can add a new dimension.

When the process requires cross-referencing multiple documents, preserving their context, applying business rules, and determining the next steps, the agent can analyze the case, recommend an action, and orchestrate the process, while allowing a human to intervene when the decision calls for it.

Not all decisions should be automated across the board, but we should automate every step that can be automated while retaining control where necessary.

Automating document processing isn't just about extracting data faster. It's about delivering reliable information to the right system with as little human intervention as possible.

‍

‍

‍