Stop paying freight operators to split mixed PDF attachments

A client sends an email with the subject line "docs attached". Inside is a single, 34-page scanned PDF. The first three pages are a commercial invoice. Page four is a packing list. Pages five through twelve are a certificate of origin and various fumigation certificates. They are out of order, one page is upside down, and the shipment reference is missing from the email body.

Right now, a freight operator earning £35,000 a year downloads that file, opens Adobe Acrobat, and acts as an expensive PDF splitter. They manually extract the pages, rename the new files, and drag them into your customs software or forward them to the clearance desk.

When a forwarder processes twenty shipments a day, this is a minor annoyance. When that scales to two hundred shipments a day ahead of the peak retail season at Felixstowe or Port Botany, manual triage becomes a hard bottleneck. You miss customs clearance windows, containers sit on the quay, and you swallow the demurrage charges.

Operators often treat this as an unavoidable cost of doing business, assuming that messy client data simply requires human intervention. But paying skilled logistics staff to drag and drop files is a waste of resource. Here is how you replace manual splitting with an automated triage workflow, and why a clear human review screen is the only way to make it work in production.

The manual reality of the clearance desk

If you want to understand why your operations team feels stretched, watch them process a consolidated shipment. The process usually looks exactly like this:

1. Monitor the shared inbox: Operators watch an inbox flooded with replies, automated alerts, and client emails.
2. Download and interpret: They open a massive, unstructured PDF and read it visually to figure out where the commercial invoice ends and the packing list begins.
3. Extract and rename: They manually split the file and type out standard naming conventions (e.g., INV_AWB123456789.pdf).
4. Data entry: They rekey the core shipment identifiers into a transport management system (TMS) like CargoWise so the clearance team knows the documents have arrived.

This takes between four and seven minutes per shipment. If a team of five operators handles 200 shipments daily, they lose over 16 hours every single day just preparing documents to be read. As volume grows, you are forced to choose between freezing headcount and commissioning a custom workflow or hiring another junior administrator just to manage the queue.

Why traditional IT rules fail on freight data

When operations managers try to fix this, they usually ask their IT team for help. The IT team typically attempts to solve it using two legacy methods, and both fail.

First, they try email routing rules. They set up a system that reads the subject line. If the subject contains "Invoice", route it to the billing queue. But freight clients are notoriously lazy. They will reply to a three-week-old email thread titled "Urgent: Missing Invoice" to send the packing list for a completely new shipment. Email rules break immediately on messy human behaviour.

Second, they try template-based OCR (Optical Character Recognition). They draw extraction boxes where the total value and document title should be. But you do not control the templates of the hundreds of different shippers you deal with globally. The total value might be in the bottom right for a supplier in Shenzhen, and on page two for a supplier in Manchester. Zonal OCR shatters the moment a supplier updates their accounting software.

How an AI document triage pipeline actually works

Modern AI does not rely on fixed templates or strict email rules. It processes documents much closer to how a human operator does: by looking at the visual layout and reading the context.

When you build a custom AI triage workflow, the system acts as a buffer between the messy inbox and your core freight management system. The pipeline executes four specific steps:

1. Ingestion and Splitting

The system watches the inbox and intercepts the 34-page PDF. Instead of asking "where is the invoice box?", a visual language model scans the document to find the natural boundaries. It recognises that page three has a total and a signature, and page four starts with a list of weights and dimensions. It digitally splits the PDF at that exact boundary.

2. Classification

Once split, the system classifies the individual files. It labels pages 1-3 as Commercial Invoice, page 4 as Packing List, and flags page 5 as an Unknown Document.

3. Key Entity Extraction

The AI reads the classified documents to pull out the critical routing data—usually the Air Waybill (AWB), Bill of Lading (BOL), or container number. It does not matter where this number sits on the page; the system understands the context of the text.

4. The Handoff

This is where mixed export attachments stop choking your clearance desk. If the system is highly confident in its splits and classifications, it automatically renames the files, attaches the AWB, and pushes them directly into your TMS via API. The clearance operator simply sees the files appear against the correct shipment record, ready to work.

The human review screen: Why black boxes fail in ops

If you try to fully automate freight document processing on day one, you will fail. The risk of misrouting is too high. If an AI system incorrectly classifies a dangerous goods certificate as a standard packing list and automatically files it away, the container gets held at the border.

Practical AI requires a safety net. You need to design the system so it knows when it is confused, and flags those exceptions for an operator.

When you build this software, you must build a side-by-side triage interface. If the AI calculates a confidence score below a set threshold (say, 85%), it pauses the automation. The operator gets an alert. On their screen, they see the original PDF on the left, and the AI's proposed document boundaries on the right.

The operator can instantly see where the machine went wrong, drag the boundary line down one page, and click 'Approve'. The AI learns from the correction, the document is routed correctly, and the operator spent 15 seconds reviewing instead of five minutes manually splitting. Deciding where to put the human in the workflow is the single most important decision in an AI build.

The economics of a custom build vs adding headcount

Many forwarders hesitate to commission custom operations software because of the upfront capital expense. It is easier to sign off on a £35,000 salary for a new document sorter than to budget for a software build.

But a human operator maxes out at a fixed number of shipments per day. When a vessel arrives late and 400 documents flood the inbox simultaneously, your human capacity breaks. Backlogs form, clearance is delayed, and clients call your account managers demanding updates.

A custom AI classification pipeline handles infinite volume spikes. Processing 50 documents or 5,000 documents takes the same amount of time. You fix the cost of processing at a fraction of a pence per page, rather than paying full-time salaries for manual data entry. If you are modelling this decision for your own team, you can map the exact financial crossover point using an AI human-in-the-loop cost model.

Getting from discovery to release in weeks

Operations leaders often fear that custom software means a twelve-month waterfall project that disrupts the whole business. That is how legacy IT operates, but it is not how modern AI product delivery works.

By taking a product studio approach to AI, you shorten the cycle from discovery to release. Engineers do not need to map out every single business rule for six months. Instead, they take a sample of 1,000 historical freight emails from your actual inbox. They use that messy, real-world data to train and test the classification model.

Within weeks, you can have a prototype triage interface sitting on a staging server. You put one senior operator in front of it and let them test the system against yesterday's inbound emails. You see exactly how the AI handles your specific client quirks before you commit to rolling it out across the entire clearance floor.

Stop paying your most knowledgeable logistics staff to act as file routers. The technology to intercept, split, and classify unstructured freight documents is already here. You just have to decide to build the system.

Useful tool

Try Samvara's Document Readiness Checklist — Export/import docs by mode.