Stop rekeying commercial invoice data on the export desk

Copy-pasting line items from an £80,000 supplier invoice into a customs declaration is the single highest-risk action on your export desk today.

Every time a logistics operator moves their eyes from a PDF on their left monitor to a submission portal on their right, you introduce a margin for error. A misplaced decimal point on a unit price, a transposed Harmonised System (HS) code, or mixing up USD with AUD can trap a container at Botany Bay or trigger an immediate HMRC audit at Felixstowe.

For a long time, the only way to process commercial invoices at scale was to throw more people at the problem. As order volumes grew, so did the number of operators doing data entry.

Today, you no longer need humans to read and retype shipping documents. But you do need a system to control how that data is extracted, verified, and approved. This is how you replace manual data entry with an AI extraction workflow that keeps your operators firmly in control.

The shared inbox bottleneck

Walk into any mid-sized export or freight forwarding business, and you will find a shared logistics@ or shipping@ inbox holding the operation together.

Suppliers email commercial invoices, packing lists, and certificates of origin in a chaotic mix of formats. Some invoices are multi-page PDFs; others are Excel spreadsheets exported poorly. Some list the Incoterms clearly at the top; others bury them in the footer notes.

Your operators spend their days acting as expensive data routers. They open an email, identify the document type, open the corresponding file in their ERP or customs software, and start typing.

At 20 shipments a week, this is manageable. At 100 shipments a week, it becomes a severe bottleneck. Why manual export document sorting breaks at 100 shipments is a mathematical certainty: human attention depletes, backlogs form, and urgency forces operators to rush the data entry. When you rush customs data, you pay for it in demurrage charges and delayed sailings.

Why template OCR fails on messy supplier formats

Ten years ago, the proposed solution to this problem was Optical Character Recognition (OCR). Software vendors promised you could scan an invoice and automatically populate your database.

The reality for operators was a nightmare of template management. Traditional OCR relies on coordinate mapping or "zonal" reading. You tell the software: Look exactly three inches down and two inches across from the top left corner to find the Total Invoice Value.

This works perfectly if you control the document format. But exporters do not control their suppliers' templates. The moment a manufacturer in Shenzhen updates their accounting software and moves the 'Total Value' box a centimetre to the right, the OCR system breaks. It reads a blank space, or worse, it reads a nearby discount figure and enters that as the total value.

Ops teams end up spending more time fixing OCR mapping errors than they would have spent just typing the data themselves.

How an AI extraction workflow actually operates

Modern AI does not read coordinates; it reads context. Large Language Models (LLMs) understand what a commercial invoice is trying to communicate, regardless of where the data sits on the page.

When you build an AI-driven extraction workflow, you are not building a fully autonomous, touchless system. You are building a digital assembly line that prepares the data for a human to review.

Here is how the workflow functions in practice.

Phase 1: Ingestion and triage

The workflow connects directly to your shared inbox. When an email arrives, the system strips out the attachments and reads them.

It performs the first critical job: classification. It determines whether the PDF is a commercial invoice, a packing list, or a bill of lading. If a supplier has merged three different documents into a single 15-page PDF, the system splits them apart.

Moving your export desk from shared inboxes to AI document triage is the foundational step. You cannot extract data reliably if you do not know what document you are looking at.

Phase 2: Contextual data extraction

Once the system identifies a commercial invoice, it uses an LLM to pull the required fields. You configure the system to look for specific data points required by the Customs Declaration Service (CDS) in the UK or the Integrated Cargo System in Australia:

• Shipper and Consignee details (Name, Address, EORI or ABN)
• Invoice Date and Number
• Incoterms (e.g., FOB, CIF, EXW) and named places
• Currency (distinguishing between £, $USD, and $AUD based on context)
• Line items (Description, HS Code, Quantity, Unit Price, Total Price)
• Total Gross and Net Weights

The model finds this information whether it is in a neat table or written as a continuous paragraph in the document footer.

Phase 3: The arithmetic check

This is where an extraction workflow proves its value over a standalone AI agent. LLMs are notoriously bad at unprompted maths, so you do not rely on them to calculate totals.

Instead, you use traditional software rules to verify the AI's work. The system takes the extracted unit price and quantity for each line item and multiplies them. It adds up all the line totals and checks if the sum matches the 'Total Invoice Value' extracted from the bottom of the page.

If the numbers match, the system gains confidence. If they do not match—perhaps the AI misread a smeared digit or the supplier made a calculation error—the system flags the discrepancy immediately.

Phase 4: Human-in-the-loop handoff

This is the most critical phase. The workflow does not push the data straight into your customs portal.

Instead, it presents the extracted data to your operator on a split-screen dashboard. The original PDF sits on one side, and the extracted data fields sit on the other.

Crucially, the system highlights fields based on confidence. A verified line item might be green, but a missing HS code or a failed arithmetic check is highlighted in red.

The operator's job changes entirely. They stop typing and start QAing. They click the red field, look at the original document, correct the error, and hit 'Approve'.

If you want to understand the exact financial return of moving from manual typing to this review model, the Human-in-the-Loop AI Cost Model shows how the unit cost per document drops when operators shift to exception handling.

Shipping a pilot in weeks, not months

Historically, commissioning custom operations software meant committing to a six-month build before seeing any real utility. The shift to AI-assisted product delivery changes that timeline entirely.

Because modern extraction relies on off-the-shelf foundation models rather than rigid template rules, AI product studios can shorten the discovery-to-release cycle dramatically. Instead of spending months hard-coding coordinate maps for fifty different suppliers, an engineering team builds the core routing logic, plugs in the LLM for the extraction layer, and tests it against a batch of your historical invoices.

This means you can put a working triage and extraction prototype into your operators' hands in a matter of weeks. You test real data flows, measure the exact time saved during the human review phase, and refine the validation rules based on actual operator feedback—all without promising impossible, fully autonomous outcomes upfront.

Designing for operator control

The temptation with new technology is to try and automate the entire process end-to-end. In customs and logistics, full autonomy is a trap. The regulatory penalties for incorrect declarations are too high, and supplier data is too messy.

How to build an AI agent without breaking operations comes down to drawing clear boundaries. The machine excels at reading 50 pages of PDFs in seconds, extracting the data, and checking the arithmetic. The human excels at holding the commercial context, spotting a suspicious valuation, and signing off on the legal declaration.

Stop paying your operators to be typists. Build an extraction workflow that handles the reading, and let your team do the reviewing.

Useful tool

Try Samvara's Incoterms Chooser — Pick a practical Incoterm.