Export Declaration: What AI Can Pre-Check Before You File

AI can pre-check that the invoice, packing list and draft declaration agree: names, values, weights, package counts, units, currency, the delivery term and the tariff code shown on each. It cannot classify the goods, decide whether a licence is needed or file for you. The exporter or the broker confirms those and answers for what is filed.

That split matters because errors on export papers are mostly boring ones: a transposed digit in the value, a gross weight that no longer matches the cartons, a consignee spelt two ways. Software is better than a tired person at catching these, and worse at deciding what the goods legally are.

Follow me on Instagram@sagnikteaches

What "pre-check" means in practice

A pre-check is a reading job done before the declaration goes to whoever files it. The input is the set of documents you already have: the commercial invoice, the packing list, the draft declaration (or the data sheet your broker asks you to complete) and sometimes the booking confirmation. The output is a list of disagreements and gaps, each pointing at the two places that clash.

Connect on LinkedInSagnik Bhattacharya

It is not a legal opinion and it is not a filing. If you want the cross-document comparison in more depth, the tutorial on cross-checking a commercial invoice against a packing list with AI goes line by line. This tutorial sits one step later: the declaration itself, who does what with it, and where AI's job stops.

Subscribe on YouTube@sagnikteaches

The fields worth comparing, and who confirms each

Different filing systems ask for different data, so this table is a general map rather than a form. The "who confirms" column is the point: AI finds the disagreement, a named person resolves it.

The last row is the one people most want AI to cover and the one it should not. Whether a licence is needed depends on the goods, the destination, the end user and current rules. A chatbot's answer on that is a prompt to go and check, nothing more.

Three product lines, four mismatches caught before filing

Illustration, with invented figures. A small exporter of stoneware tableware has a shipment of three product lines. The papers are correct in the sales order; the errors crept in while copying into the declaration data sheet.

The invoice (shortened):

Invoice INV-2041. Seller: Exporter Ltd. Buyer: Consignee Co. Delivery term: FCA, seller's warehouse. Currency: USD
1. Stoneware dinner plates 26 cm   240 pcs   @ 4.80 = 1,152.00
2. Stoneware bowls 15 cm           360 pcs   @ 3.10 = 1,116.00
3. Stoneware mugs 300 ml           480 pcs   @ 2.25 = 1,080.00
Total: 3,348.00

The packing list (shortened):

Packing list for INV-2041
Plates: 20 cartons x 12 pcs, gross 14.5 kg/carton, net 13.2 kg/carton
Bowls:  15 cartons x 24 pcs, gross 13.2 kg/carton, net 12.0 kg/carton
Mugs:   10 cartons x 48 pcs, gross 11.8 kg/carton, net 10.6 kg/carton
Total: 45 cartons, gross 606 kg, net 550 kg

The draft declaration data sheet:

Goods lines: plates (6912.00), bowls (6911.10), mugs (6912.00)
Packages: 45 cartons. Gross weight: 616 kg. Net weight: 550 kg
Declared value: 3,384.00 USD. Delivery term: FCA, buyer's warehouse

Now the prompt. Paste the documents as text (after removing bank details and contact data) and ask for a table, not a verdict.

You are checking export papers for consistency. Below are three documents for one shipment.
Compare them and list every disagreement or gap. For each, give: the field, what each document says,
the size of the difference, and which document is probably wrong and why.
Do not guess tariff codes. If codes differ between documents, say so and stop there.
Recompute line totals, carton totals and weights yourself and show the sums.
Finish with a list of things you could not check from this text.

[invoice]
[packing list]
[declaration data sheet]

A realistic answer (illustrative) comes back like this:

1. Declared value: invoice total 3,348.00 (1,152 + 1,116 + 1,080 = 3,348). Declaration shows 3,384.00.
   Difference 36.00. Looks like transposed digits. Invoice is probably right because the line sums agree with it.
2. Gross weight: packing list 606 kg (20 x 14.5 = 290; 15 x 13.2 = 198; 10 x 11.8 = 118).
   Declaration shows 616 kg. Difference 10 kg. Packing list probably right.
3. Delivery term place: invoice says FCA seller's warehouse; declaration says FCA buyer's warehouse.
   One of them is wrong and I cannot tell which from these documents.
4. Tariff codes: plates and mugs show 6912.00, bowls show 6911.10. Same material, different codes.
   I am not deciding which is right.
Checked and consistent: carton count 45, net weight 550 kg, currency USD, names.
Not checked: licences, origin, whether the codes match the goods.

The sums are the strength here. A human skimming the declaration sees 3,384 and 3,348 as the same number. The model has to compute and compare, and it shows its working so you can verify it. Three of the four findings are fixed in two minutes by you; the fourth, the code, goes to whoever classifies the goods. The place named with the delivery term is a contract question, so you ask the buyer. Nothing was filed, and nothing had to be, to find all four.

A 15-minute routine for each shipment

1. Collect the three documents in final draft form and strip out bank details, tax numbers and personal contact data. Replace names only if you do not need the model to compare spelling.
2. Paste them into one chat with the prompt above. Work in a fresh conversation per shipment so nothing from the last one leaks in.
3. Check the sums yourself on the two biggest findings. If the model's arithmetic is wrong, discard its output and try again with the figures in a table.
4. Fix the documents at source (the sales order or warehouse record), not just in the declaration, or next month's shipment repeats the error.
5. Send the corrected set to whoever files, with a one-line note of what changed.

On a plan costing about $20 a month, such as ChatGPT Plus or Claude Pro (Claude Pro is $17 a month on annual billing), a shipper with 30 consignments a month spends about seven and a half hours a month on this routine (30 times 15 minutes), against the time a single held shipment can absorb in calls and corrections. Those minutes are my estimate, not a measured figure: your own number comes from timing five shipments.

Where the tariff code stays human

The World Customs Organization describes its Harmonized System as more than 5,000 commodity groups, each identified by a six-digit code, used by more than 200 countries and economies as the basis for their customs tariffs. Your market's tariff builds on that base, so the code your filer needs may run longer than six digits; take it from the official tariff, not from memory. Classification is a legal decision with consequences for duty, restrictions and statistics. A chatbot may offer a plausible code with a confident explanation, and it can be wrong without looking wrong.

What AI is good for here:

• Spotting that two documents show different codes for the same line, as in the example.
• Suggesting two or three candidate headings you then look up in the official tariff.
• Drafting the plain-language description you give your broker (material, use, how it is made), which is what a classifier needs.

What it should not do is settle the matter. The tutorial on whether ChatGPT can find the right tariff code covers how to check a suggestion before filing. Keep a note beside each product in your catalogue recording the code used, who confirmed it and the reason, so the question is answered once rather than every shipment.

Who files, and what that does to your duty to check

Filing is usually done in one of two ways. Either a broker or forwarder files on your behalf using the papers and data you give them, or you file yourself through the national filing system for your market. Which applies depends on the shipment, the goods and how you ship, and your forwarder will tell you which route they take. I am deliberately not naming systems, since they differ by market.

The same principle applies to both. Using a broker moves the keystrokes, not the facts: the values, weights and descriptions still come from you. How far responsibility for an error is shared depends on your market's rules and your agreement with the broker, so I would not state a general rule. Treat the accuracy of the data as yours as a working habit, check with your broker what their terms say about errors in data you supply, and ask whether they need your authority to act in writing.

This has a practical effect on your checking:

• If a broker files: send them the cleaned set and ask them to confirm any field they changed. A broker who "fixes" a weight silently has just created a fourth document that disagrees with your three.
• If you file yourself: your pre-check is the last line of defence, so run it twice: once on the drafts, once on the final data as you see it on screen before you submit.

For the broker's side of the relationship, how customs brokers use AI to draft client updates and entry checks shows what a well-run filing desk sends back.

What the importer sees from the same papers

Your documents travel. The importer, or their broker, uses the same invoice and packing list to make the import declaration at the other end. If your papers disagree with each other, their declaration inherits the disagreement, and the importer is the party standing in front of the officer. Typical consequences, none of them certain: a hold while someone asks which number is right, extra storage charges, a corrected declaration, and in the worse cases penalties for misdeclaration that the importer may try to recover from you under the contract.

You can lower the odds by sending the importer the same cleaned set you gave your own filer, and by flagging anything unusual before the vessel or truck arrives: a split shipment, a partial invoice, a change of consignee. A short message works:

Hi [first name], the papers for INV-2041 are attached. Two things to know before arrival:
the declared gross weight is 606 kg (an earlier draft said 616; the packing list is correct),
and the delivery term is FCA at our warehouse. If your broker needs anything different, tell me today.
[your name]

If you want a second view on the importer's side, can AI check customs compliance and what a person must still verify is the mirror image of this tutorial.

Four more checks AI handles well

Names that look the same and are not

A consignee written "Consignee Co" on the invoice, "Consignee Co. Ltd" on the packing list and "Consignee Company Limited" on the declaration is one party to a person and three to a system that matches text. Ask the model to list every name variant per party and say which matches the contract.

Units that quietly change

Plates sold in sets of four on the sales order and pieces on the invoice, with the packing list counting boxes of 12, are a classic. Ask for a table that converts every line into pieces, cartons and sets side by side.

Dates that cannot be true

An invoice dated after the booking, or a packing list dated before the order, is easy for a model to flag and easy for a person to miss. Ask for a timeline of every date in the set.

Missing attachments

Give the model the list of documents the filer expects (invoice, packing list, certificate, licence copy) and the files you have, and ask what is absent. This is a simple matching job and it saves the Friday-afternoon scramble.

A vague goods description, before and after

Filers dislike descriptions such as "kitchen items" or "samples", and a pre-check can flag them in seconds. Ask the model to mark any line whose description would not let a stranger picture the goods, and to propose a clearer version using only facts already in your papers. Here is an illustrative pair.

Before (declaration line 2): "Kitchen items, ceramic"
After (suggested, to be confirmed by you): "Stoneware bowls, 15 cm diameter, glazed, for table use, 360 pieces"

The suggestion uses only what the invoice already says. If the model invents a feature that is not in your papers, such as "microwave safe", strike it. A description is a statement you make to an authority, so every word in it must be one you can support.

Mixed currencies and split invoices

A second case that trips people up is a shipment with two invoices, one in dollars for the goods and one in another currency for freight or insurance. The declared value then depends on the delivery term and on what the filing system wants included. A model can lay out each charge, its currency and the document it came from, and show which total you would get under each reading. It cannot tell you which reading your filing system requires. Take that table to your broker and ask for the answer in writing, then record it in the product notes so the next shipment uses the same rule.

Keep the arithmetic visible. If the model converts a currency, make it state the rate and its source, because a rate pulled from memory is a guess. Use the rate your filer or the official source for your filing gives, and write down the date it applies to.

When the model said everything matched and it did not

Here is how a clean bill of health goes wrong. An exporter pasted a scanned packing list through a free tool that read only the first page. The second page, with six extra pallets, never reached the model. The model compared what it saw and reported all consistent. The shipment arrived with 6 pallets more than the declaration said.

The fix is a habit rather than a tool. Ask the model to state, at the top of its answer, the number of lines and pages it read from each document, and compare that to what you pasted. If your package count is 45, the model should say it saw 45. A reply that cannot give those counts is one you do not trust.

The other common failure is false reassurance on arithmetic. A model that adds a five-line invoice correctly can still slip on a forty-line one. For any document with more than ten lines, do the totals in a spreadsheet and let the model compare the spreadsheet's totals to the declared ones. The tutorial on AI tools for shipping documents: commercial invoices and packing lists shows ways to produce a clean line table first.

Testing the pre-check before you trust it

Before using the prompt on live work, test it the way you would test a new employee. Take ten finished shipments whose papers you know were correct. Copy each set and plant errors: change one digit in a value, shift a weight, rename a consignee, switch a code. Record which errors you planted in a sheet with columns for shipment, planted error and found yes or no.

Run the prompt on each copy. Score two numbers: how many planted errors it found, and how many false alarms it raised on the clean originals. If it misses more than one in ten planted errors, tighten the prompt or add structure (a table of fields) and test again. If false alarms exceed what you are willing to read, tell it to report only differences it can show with figures from the documents.

Checks that no AI should be trusted with

A short list of what stays with you or your broker, whatever tool you use.

• Licences and restrictions. Whether the goods or destination need permission is a legal question that changes. Ask your broker or the official authority.
• Sanctions and restricted parties. Screening names against official lists is a process with records; a chatbot has no authoritative live list.
• Valuation rules. What goes into the declared value, and whether related-party prices are acceptable, is a rules question with documents behind it.
• Origin claims. A certificate of origin states a fact you must be able to prove.
• Anything the broker says they will check. Do not assume you and they both verified something; agree who does.

If you would rather map which of these checks you can safely hand to software and which to keep with a person, that is the kind of job my AI implementation consultation covers, and you can reach it through the contact page.

Keeping the papers private while you check them

Shipping documents carry names, addresses, prices and sometimes bank details. Before pasting anything, look at the plan's data terms: whether content is used to train models, how long conversations are kept and who in your business can see them. The tutorial on what to check in an AI tool's privacy policy and terms lists the questions. As a working habit, remove what the comparison does not need: bank account lines, tax numbers, phone numbers and any buyer prices you would not want a competitor to see. Keep the originals, and store the model's output beside the shipment file so you can show later what was checked and when.

Questions exporters ask about AI and export declarations

Can AI file my export declaration for me?

No. Filing goes through your broker, forwarder or the official filing system, using your authority and your data. AI can compare documents and draft a clean list of fields beforehand, but a person submits and stands behind what is filed.

Is it safe to upload shipping documents to a chatbot?

Only if the plan's data terms suit the content. Remove bank details, tax numbers and contact details you do not need to compare, and check whether the plan trains on your data. OpenAI and Anthropic both say their business plans do not train on your content by default; consumer plans have their own settings, so read the terms for the plan you hold.

Can AI choose the tariff code for my goods?

It can suggest candidates and flag where two documents show different codes. It cannot take responsibility for the classification. The exporter or broker confirms the code against the official tariff and keeps the reasoning.

How do I test whether the pre-check works?

Take ten past shipments, plant three known errors in copies, and see whether the prompt finds them. Count misses and false alarms before you rely on it for live shipments.

What if the importer finds a mismatch after the goods arrive?

A mismatch can hold the goods at the destination while the importer's broker queries it, and corrections may carry costs or penalties. Fix the papers on your side first, and tell the importer and your broker the same day if you spot an error late.

Further reads

• AI Freight Quoting: Turn Customer Emails Into Draft Quotes — See how forwarders turn customer emails into draft quotes.
• Carrier Packet Onboarding: AI Collects and Checks the Documents — Collect and check the documents a carrier must supply.
• A Simple AI Risk Register for Small Businesses (With Template) — Record the document-checking risks and who owns each.
• AI for Freight Business Analytics: Margin by Lane From a Spreadsheet — Look at which shipping lanes earn their keep.
• AI Security Risks for Small Businesses and How to Close Them — Close the gaps that matter when documents leave your computer.
• Letter of Credit Checklist: Check Documents With AI Before the Bank — Turn the credit into a clause table, let AI compare each document against it, then fix mismatches before the bank sees them.
• AI for Exporters: Six Jobs From Finding Buyers to Paperwork — Six export jobs AI can take on, from buyer research to shipping documents, each with a tool, a prompt to copy and the check that catches errors.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: the World Customs Organization's page on what the Harmonized System is, the ChatGPT and Claude pricing pages, and OpenAI's and Anthropic's pages on business data and model training. Document examples and timings are illustrative.