Freight document extraction
Also called bill of lading data capture, shipping document OCR, customs document automation, trade document processing
AI extraction of bills of lading, commercial invoices and customs documents: what benchmarks show, what it costs and where it fails.
AI can read shipping documents into your systems, but we found no independent study of its accuracy or the time it saves at freight forwarders. Trade still runs on paper and PDFs: the ICC estimates that over 90% of the trading world holds trade data in paper forms, physical or PDF. Only about 11% of bills of lading were electronic in 2025. Academic benchmarks show extraction does well on familiar layouts but drops on unseen ones and on line-item tables. Headline accuracy figures are vendor claims.
At a glance
| Status | In pilots |
|---|---|
| What it does | Reads bills of lading, commercial invoices, packing lists and customs documents from email or scans, and fills the fields in your freight, customs or accounting system. |
| Typical cost | From US$33 per month, billed annually, for 100 pages with one general document-parsing tool (vendor list price, checked Oct 2026); otherwise varies by vendor. |
| Data it needs | A steady flow of shipping documents, a field list per document type, and bookings, purchase orders and supplier records to check the output against. |
| Setup effort | Medium: fields must be mapped per document type, and a review step is needed for low-confidence values. |
| Evidence | Thin. Academic extraction benchmarks on business documents and industry digitisation reports; no independent freight-specific outcome study found. |
What is freight document extraction?
A single shipment generates a stack of paperwork. The ICC's Digital Standards Initiative lists the bill of lading, commercial invoice, packing list, certificate of origin, insurance certificate, warehouse receipt and customs declaration as key trade documents. One shipment can involve up to 30 parties and up to 50 sheets of paper.
The usual way to handle this is by hand. The ICC describes staff extracting, checking and re-entering trade data into IT systems that are mostly not connected to each other.
Freight document extraction uses AI to read those documents, whether scanned paper, PDFs or email attachments, and pull out the fields: shipper, consignee, container numbers, tariff (HS) codes, weights, values and line items. A person then checks the result instead of typing it. For a freight forwarder, clearing agent or importer, the target is fewer keying hours and fewer errors on customs entries.
How does it work?
Documents arrive by email or scan. The tool reads the text, including text in images, and identifies each field from its label and position on the page. Older systems relied on a template set up for each supplier or carrier layout. Newer ones use large language models that can attempt layouts they have not seen, with less setup.
A sound set-up attaches a confidence score to each extracted field. Fields above a threshold flow straight through; the rest go to a person. The extracted data is then matched against what you already hold: the booking, the purchase order and the supplier's earlier invoices. Mismatches, such as a container number that does not match the booking or an invoice total that differs from the order, are flagged before the customs entry is lodged or the supplier is paid.
What it needs to work
Extraction needs a predictable flow of documents and reference data to check the output against. Without that, errors go straight into customs entries and supplier payments.
- One inbox or folder where shipping documents arrive, separate from general correspondence
- A written field list per document type (bill of lading, commercial invoice, packing list, customs declaration), each mapped to its field in your freight, customs or accounting system
- Reference data to match against: bookings, purchase orders, supplier and consignee master records, and the HS codes you have used per product
- A labelled sample of past documents for each major supplier or carrier layout to test accuracy before going live; in one study, ten labelled training documents lifted accuracy sharply
- A named reviewer and a confidence threshold below which a person checks the field before anything is lodged or paid
Most use cases in this guide succeed or fail here, on the data underneath them, before any AI is involved.
What does the evidence show?
We found no independent study of extraction accuracy on freight documents specifically, and none measuring time or errors saved at forwarders or importers. The nearest evidence is academic benchmarks on business documents such as invoices and forms.
The DocILE benchmark, built from 6,680 real business documents, found extraction still challenging. The best qualifying baseline scored about 0.66 out of 1 on key fields (F1, a combined score for finding fields and reading them correctly), and accuracy fell on layouts the model had not seen in training. The VRDU benchmark found that handling new templates and nested fields such as line items were the hardest problems. A 2023 study of a language-model approach reported large gains. On registration forms it scored about 87 to 91 out of 100 with ten training examples. On the line items in advertising order forms it scored only 21 to 23 with no examples, and 72 after 200 training documents.
One vendor advertises 99% extraction accuracy, qualified as applying to clearly written documents. Treat that as a vendor claim.
AI extraction vs manual capture of freight documents
| Compared on | AI extraction | Manual capture |
|---|---|---|
| Speed | Fast once set up; review time is the main labour cost | Slow, and grows with document volume and headcount |
| Error profile | Systematic: misses on line items and unfamiliar layouts | Random: typos and transposed digits |
| Cost model | Subscription per page or document | Staff time |
| Liability for customs errors | Stays with you or your agent | Stays with you or your agent |
| Best for | High volumes from repeat suppliers and carriers | Low volumes, unusual or handwritten documents |
Status and evidence
Academic benchmarks show document extraction works well on familiar layouts and struggles on new layouts and line items, but none tests freight documents, and we found no independent study of time or error savings at forwarders or importers. Off-the-shelf tools exist at small-business prices, yet adoption is unmeasured, so we rate it pilots with thin evidence.
Where does it go wrong?
- Line items. Multi-line tables on commercial invoices and packing lists are where models do worst. A missed or merged line changes quantities, values and duty.
- New layouts. Every new supplier or carrier format is a fresh test, and accuracy drops on layouts the model has not seen.
- Poor scans. Faxed, photographed or stamped documents degrade the text before extraction starts. Even one vendor's accuracy claim depends on how clear the source is.
- Liability does not move. HMRC guidance says that if you appoint someone to deal with customs, you are still responsible for due diligence on your declarations. Software does not change that: a wrong HS code or value is still your error.
- No link to the freight system. If extracted data is not pushed into your freight or customs software, staff copy it across by hand and the saving shrinks.
- Ignoring the shift to digital. The industry target is 100% electronic bills of lading by 2030, so plan for structured data feeds as well as PDFs.
Numbers it should move
Ask the demo business
Northgate Supply is a fictional wholesale distributor of catering and packaging supplies: 28 staff, R46.6m in sales over the last year. Its data is generated for this demo, to 30 September 2026.
Payroll took 12.6% of sales over the year. December's payroll was nearly double a normal month because of 13th cheques, but record sales held that month's ratio to 14.9%, about the same as February 2026, the quietest month for sales, at 14.7%. March's increases lifted monthly payroll by about 5.5%.
How this was worked out
Monthly payroll cost, including 13th cheques, divided by monthly sales.
Where the data usually lives
Where the records this use case needs are usually kept in a small business. It is not a list of recommended tools.
Frequently asked questions
Can AI read a bill of lading accurately?
Often, for clear documents in layouts it has seen. Academic benchmarks show accuracy drops on unfamiliar layouts and is weakest on line-item tables, where one language-model approach scored 21 to 23 out of 100 with no training examples. One vendor advertises 99% accuracy on clearly written documents; that is a vendor claim. Test any tool on a sample of your own documents before trusting a figure.
How much does freight document extraction software cost?
One general document-parsing tool lists plans from US$33 per month for 100 pages, billed annually, with each PDF page using one credit (checked Oct 2026). Pricing for freight and customs platforms varies by vendor. Budget for the staff time spent reviewing low-confidence fields, which remains a real cost after automation.
Do I need a developer to automate shipping document capture?
Not to start: some parsing tools process emails as well as PDFs. You may need help to push extracted data into your freight, customs or accounting system automatically. The ICC notes that trade data is typically entered by hand into IT systems, most of which are not connected to each other, and without that link staff still copy data across by hand.
Who is liable if AI puts the wrong value on a customs declaration?
You are, or your appointed agent, depending on your arrangement. HMRC guidance says that if you appoint someone to deal with customs, you are still responsible for due diligence on your declarations. Software does not change that. Keep a human review step for duty-relevant fields such as HS codes, customs values and country of origin.
Does POPIA or UK GDPR apply to freight documents?
Often, yes. Shipping documents carry names, addresses and contact details of consignees and contact people, which can be personal information. POPIA requires security safeguards and a written contract with any operator that processes personal information for you, such as a software provider. UK GDPR requires a contract with processors covering your instructions, security and sub-processors. This is not legal advice.
Will electronic bills of lading make extraction pointless?
Not soon. About 11% of bills of lading were issued electronically in 2025, against an industry target of 100% by 2030. Other documents, such as commercial invoices and certificates from smaller suppliers, may keep arriving as PDFs. Once a document arrives as structured data, you no longer need to extract it.
More in stock, supply & fulfilment
Sources
- 1Digital trade for MSMEs: momentum and accelerationICC Digital Standards Initiative (presented to the WTO), 2023Survey
- 2ICC DSI: 7 key trade documents for digitalisationTrade Finance Global, 2023News
- 3The multi-billion-dollar paper jam: Unlocking trade by digitalizing documentationMcKinsey & Company, 2022Survey
- 4Soft barriers to eBL adoption: Addressing the hidden frictions that slow digital progressDigital Container Shipping Association (DCSA), 2025Survey
- 5DocILE Benchmark for Document Information Localization and ExtractionarXiv (Šimsa et al.), 2023Study
- 6VRDU: A Benchmark for Visually-rich Document UnderstandingarXiv (Wang et al.), 2023Study
- 7LMDX: Language Model-based Document Information Extraction and LocalizationarXiv (Perot et al.), 2023Study
- 8Appoint someone to deal with customs on your behalfHM Revenue & Customs (GOV.UK), 2025Government
- 9AI Logistics Document Parser - Extract Shipping Data AutomaticallyAirparser, 2026Vendor
- 10Pricing and Free Trial | AirparserAirparser, 2026Vendor
- 11Protection of Personal Information Act 4 of 2013Government of South Africa, 2013Government
- 12What needs to be included in the contract?Information Commissioner's Office (ICO), 2026Government
This entry is educational. It is not legal, financial or professional advice, and it does not recommend any vendor. Figures attributed to vendors are their own claims.
- 1Digital trade for MSMEs: momentum and accelerationICC Digital Standards Initiative (presented to the WTO), 2023Survey
- 2ICC DSI: 7 key trade documents for digitalisationTrade Finance Global, 2023News
- 3The multi-billion-dollar paper jam: Unlocking trade by digitalizing documentationMcKinsey & Company, 2022Survey
- 4Soft barriers to eBL adoption: Addressing the hidden frictions that slow digital progressDigital Container Shipping Association (DCSA), 2025Survey
- 5DocILE Benchmark for Document Information Localization and ExtractionarXiv (Šimsa et al.), 2023Study
- 6VRDU: A Benchmark for Visually-rich Document UnderstandingarXiv (Wang et al.), 2023Study
- 7LMDX: Language Model-based Document Information Extraction and LocalizationarXiv (Perot et al.), 2023Study
- 8Appoint someone to deal with customs on your behalfHM Revenue & Customs (GOV.UK), 2025Government
- 9AI Logistics Document Parser - Extract Shipping Data AutomaticallyAirparser, 2026Vendor
- 10Pricing and Free Trial | AirparserAirparser, 2026Vendor
- 11Protection of Personal Information Act 4 of 2013Government of South Africa, 2013Government
- 12What needs to be included in the contract?Information Commissioner's Office (ICO), 2026Government