Trade Documents, Tariff Codes, and Where the Liability Sits

International shipments generate documents that look like bureaucracy and are mostly doing specific jobs. Knowing which job each does tells you which ones tolerate automation.

The commercial invoice states what was sold, to whom, for how much. It is the basis for customs valuation and therefore for duty.

The packing list states what is physically in each carton and container, and it is what a customs officer or warehouse checks against.

The bill of lading is the important one. For sea freight it is a receipt for the goods, evidence of the contract of carriage, and, when issued in negotiable form, a document of title: whoever holds it can claim the goods. That last property is why it is handled carefully and why an air waybill, which is not a document of title, behaves differently.

The certificate of origin states where goods were produced, which determines whether preferential tariff rates under a trade agreement apply.

Customs declarations are the formal legal statement to the authority, carrying the classification, valuation and origin.

And where payment runs through a letter of credit, the documents themselves become the payment trigger, examined against the credit terms with notorious strictness.

The pattern worth noticing. Some of these are operational records, and some are legal statements with penalties attached. Automation is straightforward for the first group and needs care with the second, which is the rest of this lesson.

The largest genuine saving in this field is unglamorous: turning documents into data.

The current state in most operations. A supplier sends a commercial invoice as a PDF, or a scan, or an image taken on a phone. A person reads it and types the values into a system. The same data is re-keyed from the packing list, then again into the customs declaration, then again into the payment record. Every re-keying is an opportunity for a transposition error, and a wrong figure in a customs declaration is a legal problem rather than a typo.

Document extraction handles this well, and it is the mature end of the technology. Invoices, packing lists and bills of lading are semi-structured: the same fields appear, in different layouts, in different languages, from thousands of counterparties. That variation is exactly what defeated older template-based systems and what modern extraction handles.

The control that makes it safe is arithmetic rather than trust. Line items sum to the invoice total. Carton counts on the packing list match the bill of lading. Weights are consistent across documents. Values match the purchase order. Where those reconcile, confidence is high and the record can flow. Where they do not, a person looks.

That pattern, automate the flow and route the mismatches, is the shape of nearly every good automation in this field.

And note what it also buys: the cross-document consistency check that nobody was doing manually. Catching a discrepancy before submission is far cheaper than after.

Of everything in this lesson, tariff classification is where the exposure concentrates, and it is routinely treated as a data entry task.

Goods entering a country are classified under the Harmonized System, maintained by the World Customs Organization, which provides six digits used internationally. Countries extend it: the United States uses ten-digit codes in the Harmonized Tariff Schedule, the European Union eight in the Combined Nomenclature.

That code determines the duty rate, the eligibility for preferential treatment, and whether import restrictions, quotas or licensing apply. Getting it wrong means underpaying duty, which is recoverable years later with penalties, or overpaying, which nobody refunds unless you notice.

The responsibility sits with the importer, not the broker and not the software. Under the United States Customs Modernization Act of 1993 the importer of record must exercise reasonable care in classification, valuation and declaration, and penalties for negligence, gross negligence and fraud are set out in the customs statute. Other jurisdictions place equivalent duties on the declarant.

Why classification is hard, and hard in a way that defeats confident automation. It turns on legal interpretive rules rather than on what a product looks like. The General Rules of Interpretation govern how to classify, essential character determines the outcome for composite goods, and there is a body of rulings and case law where near-identical products classify differently for reasons that are not intuitive.

So a system that returns a code with high confidence is answering a question of legal interpretation using pattern similarity. Sometimes it is right. It cannot tell you when it is not.

The workflow that uses assistance without transferring the decision.

Start with the product, and specifically with facts the classifier needs: what it is made of, what it does, how it is presented for sale, and how it is packaged. Most classification errors trace back to an inadequate product description rather than to a bad decision.

Assisted search then proposes candidate headings, and this is where it genuinely helps. Finding the plausible headings in a schedule of thousands is a real cost, especially for a product type you have not classified before.

A person then applies the General Rules of Interpretation to choose between candidates, reads the section and chapter notes, which frequently exclude what the heading appears to include, and checks the published rulings for similar goods.

Where the answer is uncertain and the volume is material, the correct move is not more confidence. It is a binding ruling from the authority: a Binding Tariff Information in the European Union, a ruling from Customs and Border Protection in the United States. That converts an interpretation into certainty you can rely on.

And the decision is recorded with its reasoning, because reasonable care is demonstrated by showing how you decided, not by asserting that you were careful.

The recurring shape: the tool narrows the search space, the human applies the rules, the authority resolves what remains.

The second high-exposure decision, and the one people most often get wrong through honest confusion.

Origin for customs purposes is not the address of the factory. It is a legal determination under rules of origin, and it decides whether goods qualify for preferential duty rates under a trade agreement.

The complexity comes from goods that are made from inputs from several countries, which is most manufactured goods. The rules then ask whether sufficient transformation occurred in the claimed country, and they express that in several ways: a change in tariff classification between inputs and output, a regional value content threshold, or a specific process requirement. Which test applies depends on the product and on the agreement, and different agreements set different rules for the same goods.

So the same product can originate in one country under one trade agreement and not under another. That is not an anomaly; it is how the system works.

Why this resists automation more than classification does. The determination depends on your bill of materials, the origin of each input, and the values involved, which are commercial facts held across your suppliers rather than properties of the product. A model that has never seen your bill of materials cannot determine origin, and if it produces an answer it has assumed one.

Where assistance helps. Structuring the analysis, extracting supplier declarations, tracking which inputs lack documented origin, and flagging when a bill of materials change might break a preference claim.

And the standing rule: a preference claim is a statement to a customs authority. Claiming preference you cannot substantiate on audit means repaying the duty with penalties, years later, on every shipment.

A short but high-value topic, because the errors here are common, expensive and entirely avoidable.

Incoterms, published by the International Chamber of Commerce with the current set issued as Incoterms 2020, are three-letter rules allocating responsibilities between seller and buyer: who arranges carriage, who pays which costs, and crucially where risk transfers.

Three misunderstandings recur.

That the cost transfer point and the risk transfer point are the same. In several terms they are not, and the gap is exactly where uninsured loss happens.

That an Incoterm determines ownership. It does not. Title passes according to the sale contract, which is a separate matter.

And using a term inconsistent with the mode of transport. Some terms are intended for any mode and some only for sea and inland waterway, and applying a maritime term to a container moved door to door creates ambiguity about where risk actually transferred.

Where language tooling helps genuinely. Checking whether the term stated on a purchase order, the sales contract, the invoice and the transport document are consistent with each other, which is a cross-document comparison and a common source of disputes. And explaining the practical consequences of a proposed term to a commercial team who are agreeing to it without knowing what it means.

What it should not do. Choose the term. That is a commercial negotiation about who bears cost and risk, informed by your insurance, your freight capability and your leverage, and it belongs to the people who will live with it.

One compliance area deserves separate treatment because the consequences are categorically more serious than duty errors.

Trade sanctions and export controls restrict who you may deal with, what may be shipped where, and for what end use. Screening obligations cover counterparties against restricted party lists, the goods against control classifications, and the destination and end use against prohibitions.

Why this is different from the rest of the lesson. Sanctions breaches are frequently strict liability in the sense that intent is not required, penalties are severe, and in serious cases they are criminal. A duty misclassification is money. A sanctions breach can be a prosecution.

What automation does well here, and it is genuinely necessary at volume. Screening names against lists, which no human can do across thousands of transactions. Flagging near matches, since evasion uses spelling variants and transliteration. And re-screening continuously, because lists change and a counterparty compliant last month may not be today.

What it does badly. Resolving whether a near match is the same entity, which is a judgement about identity with serious consequences either way. Determining beneficial ownership, since restrictions can extend to entities owned by listed parties and ownership is deliberately obscured. And assessing end use, which turns on what a customer actually intends.

So the design is automated screening with mandatory human adjudication of every hit, no exceptions for volume. And unlike everywhere else in this cursus, the correct response to uncertainty is to stop the transaction and take advice, not to proceed with a caveat.

Collecting the lesson into a single division.

Automate freely, with arithmetic reconciliation as the control. Extracting fields from invoices, packing lists and transport documents. Cross-checking totals, counts and weights between documents. Populating systems from extracted data. Flagging mismatches. Matching documents to purchase orders. Chasing missing paperwork.

Assist, with a person deciding. Tariff classification, where the tool proposes candidates and a person applies the interpretive rules. Origin determination, where the tool structures the analysis against your bill of materials. Incoterm consistency checking, where the tool compares documents and a person owns the commercial choice.

Screen automatically and adjudicate manually. Sanctions and restricted party hits, every one, regardless of volume.

And never generate. A classification code entered into a declaration without a person deciding it. A preference claim nobody can substantiate. A value on a customs declaration that came from anywhere other than the commercial reality.

The organising principle for the whole lesson. Documents that record what happened tolerate automation, because errors are visible and correctable. Documents that state something to an authority are legal declarations, and the person making them carries a duty of care that no software vendor assumes on their behalf.

The extraction saving is large and real. It is available without touching the decisions, which is why it is where to start.