AI Tools for Shipping Documents: Commercial Invoices and Packing Lists

A small exporter can draft commercial invoices and packing lists with a chat assistant such as ChatGPT or Claude fed from an order spreadsheet, and add extraction software such as Nanonets if supplier paperwork piles up. A person must still verify tariff codes, origin, values, weights and the buyer's document requirements.

The split that matters is between layout and judgement. AI is good at turning forty order lines into a tidy, correctly added-up invoice and a matching packing list. It is poor at knowing what your buyer's border authority will accept, so every field that touches duty or compliance stays with you and your customs broker.

Follow me on Instagram@sagnikteaches

What the two documents contain and who reads them

A commercial invoice is the sales document customs uses to value the goods. It normally carries seller and buyer details, invoice number and date, a line for each product with description, quantity, unit price and currency, the total, the agreed trade terms (for example FOB or DAP), the country of origin and usually the tariff code. A packing list does a different job: it tells the carrier, the warehouse and the receiving customs officer what is physically in each carton or pallet, with counts, weights and dimensions.

Connect on LinkedInSagnik Bhattacharya

Because they describe the same goods from two angles, they must agree. If the invoice says 480 units and the packing list adds up to 468, the shipment can be held while someone explains the gap. That agreement check is where most of the value of AI sits, and it is covered in the companion tutorial on cross-checking both documents with AI. This tutorial is about choosing the tool that produces the drafts in the first place.

Subscribe on YouTube@sagnikteaches

Four ways to produce the drafts, compared

There are four realistic routes for a business shipping a few consignments a week. They differ in cost, in how much setup they need and in where the data lives.

Prices are list prices read from the vendors' own pages in November 2026 and change often, so confirm before you commit. The Nanonets figures come from its pricing page: a free starter with $50 of credit, then $100 a month for 100 credits, with each workflow step charged by type. A typical invoice workflow runs four to six steps and the page puts that at under $2 per document. For a business reading ten supplier invoices a week, that is cheap. For one that writes its own documents and reads none, it buys nothing.

The spreadsheet-plus-assistant route, step by step

Most small exporters should start here because the data already exists and the cost is almost nothing. The trick is to separate fixed facts from per-order facts.

Build a product master sheet once

Create one tab with a row per product and these columns: your SKU, a plain-English description a stranger could understand, material or composition, unit of measure, units per carton, net weight per unit, gross weight per carton, carton dimensions, country of origin, and a column marked "tariff code (broker-confirmed)" that stays blank until your broker has filled it in. Only the last column is slow, and it is also the one that protects you.

Descriptions deserve care. "Mugs" invites questions; "Ceramic mug, stoneware, 350 ml, glazed, with handle, decorated" gives a classifier and a customs officer something to work with. This is a good job for the assistant: paste your rough names and ask it to write precise descriptions in a fixed pattern, then read every one yourself.

Add a per-order tab

For each order, one tab holds the buyer, delivery address, trade terms, currency, invoice number and one row per line with SKU and quantity. Pull in prices and weights from the master with lookups so nobody retypes them. The assistant should receive the finished sheet, not do the lookups, because lookups in a spreadsheet are exact and a language model is not.

Ask for both documents in one go

Paste the order tab and the master rows it uses and give a prompt like this.

You are drafting export paperwork for a small seller. Using only the data below, produce:
1. A commercial invoice table: line number, SKU, description, quantity, unit price, line total, country of origin, tariff code (copy exactly from the data; write "BROKER TO CONFIRM" if blank).
2. A packing list: carton number, SKU, units in carton, net weight, gross weight, dimensions.
3. A reconciliation block: total units on invoice, total units on packing list, total net weight, total gross weight, invoice total.
Do not invent any value. If a field is missing, write MISSING and list it at the end.
Seller: [seller name and address]
Buyer: [buyer name and address]
Terms: [FOB / DAP / etc.]  Currency: [USD]
DATA:
[paste order rows and master rows]

The "write MISSING" instruction matters more than anything else in the prompt. Without it, an assistant will cheerfully fill a blank weight with a plausible figure, and a plausible figure is the worst kind of error on a customs document.

A worked order with the numbers

Take a six-person homeware seller shipping one order to a retail buyer abroad. Illustrative figures throughout. The order has five product lines.

The invoice total is 1,536 + 984 + 990 + 768 + 594 = 4,872.00 USD, units total 1,260, and the packing list has 50 cartons. The assistant's reconciliation block should say exactly that. If it says 1,250 units or 4,827.00, you have caught an arithmetic slip, which language models still make on long tables, and the fix is to recompute in the spreadsheet and paste the result back.

Now the realistic failure. Suppose the order sheet lists the jug at 120 units, but the warehouse packed only 11 jugs into the last carton because one was chipped. The packing list the warehouse sends says 119. A good prompt run on both documents flags: "Invoice line 4 shows 120 units; packing list shows 119; difference 1." The decision is yours: amend the invoice to 119 and credit the buyer, or hold the shipment for a replacement. The assistant spotted it; it cannot decide it.

Sample output and what to fix in it

Here is a trimmed, illustrative response to the prompt above, followed by what a careful person would change.

COMMERCIAL INVOICE (draft)
Line 1 | MUG-350-W | Ceramic mug, stoneware, 350 ml, glazed | 480 | 3.20 | 1,536.00 | Origin: [seller country] | Tariff: BROKER TO CONFIRM
Line 2 | BWL-600-G | Ceramic bowl, stoneware, 600 ml | 240 | 4.10 | 984.00 | Origin: [seller country] | Tariff: BROKER TO CONFIRM
...
RECONCILIATION
Units on invoice: 1,260 | Units on packing list: 1,260
Net weight: 612.4 kg | Gross weight: 701.0 kg
Invoice total: USD 4,872.00
MISSING: gross weight for TRY-ACA-S

Three edits. First, the missing gross weight stops the document: weigh a carton, update the master and rerun. Second, "Origin" should say the actual country of manufacture in the form your buyer's customs expects, which is your fact, not the model's. Third, the tariff column stays "broker to confirm" until the broker replies. A draft with visible gaps is better than a finished-looking one with invented values.

When extraction software earns its place

Suppose instead you are a distributor receiving supplier invoices and delivery notes as PDFs, then re-keying them into your own export documents. Extraction software reads the PDF into fields (supplier, date, lines, quantities, prices) so that nothing is retyped. Nanonets' pricing page describes a block-based model: simple routing or formatting steps cost $0.02 a run, classification or validation $0.10, and extraction or generative steps $0.30. A workflow of five steps with one extraction and one validation costs around $0.30 + $0.10 + three simple steps at $0.02, which is $0.46 per document. At 200 documents a month that is about $92, close to the $100 credit pack, against roughly ten hours of retyping you stop doing.

Two cautions. Extraction accuracy falls on scanned, skewed or handwritten pages, so test twenty of your worst real documents before you trust it with the rest. And a workflow that writes straight into a customs document with no human review is exactly the arrangement to avoid. Keep a review step where a person sees the extracted fields beside the original page.

Inventory systems sometimes bundle the same ability. Cin7's pricing page lists its Intelligent Document Recognition as an add-on on several Core plans rather than a standard feature, so if you are shopping for that, ask what the add-on costs on top of the $349 to $1,199 monthly plans. If you are weighing a system purchase mainly for stock reasons, the tutorial on AI for ecommerce inventory and reorder points covers that side.

Tariff codes: where AI helps and where it must stop

The code that classifies a product for duty is the field most exporters worry about, and rightly. Classification depends on the product's material, function and sometimes the way it is sold, and the correct heading can differ between importing countries after the first six digits. A language model can read your product description and propose candidate headings with its reasoning, which is a useful starting list for a conversation. It cannot take legal responsibility, and it will occasionally state a wrong code with complete confidence.

A workable routine looks like this. Use the assistant to tighten each description. Ask it for two or three candidate codes with a one-line reason each. Send the product list, descriptions and candidates to your customs broker. Paste the confirmed codes into the master sheet, record the date and who confirmed, and never let the assistant overwrite that column. A classification service such as Zonos Classify automates the candidate step for large catalogues. Its product page says it returns a six-digit code by default and a longer, destination-specific code (typically ten digits) when you give the ship-to country, along with a confidence score and alternative codes. That confidence score is useful: send every low-confidence suggestion to the broker first. The page gives no price, so treat it as a quote conversation, and the broker still signs off.

Example: a bad description and the better one

Before: "Cutlery set". After: "Cutlery set, 24 pieces, stainless steel, dishwasher safe, packed in a retail gift box". The second tells a classifier the material, the count and the presentation, all of which can affect the heading. Run each product through this kind of rewrite before it goes near a broker.

Eight checks that stay with a person

However the drafts are produced, run this checklist before anything leaves the building. The first four are about agreement between documents, the rest about facts only you can know.

1. Units per line match between invoice and packing list, and the totals match.
2. Invoice total equals the sum of the lines, and equals the order confirmation or proforma the buyer agreed.
3. Gross weight is at least net weight on every line and in total, and the carton count matches what was physically packed.
4. Buyer name and delivery address are identical on every document and on the carrier booking.
5. Trade terms and currency are the ones in the sales contract, because they decide who pays freight and duty.
6. Country of origin reflects where the goods were made, in the wording the buyer's customs expects.
7. Every tariff code was confirmed by your broker, and the descriptions on the invoice match the descriptions you gave them.
8. Any certificate, licence or declaration the buyer's side requires is attached, which your broker or the buyer can list for you.

Example: the filled-in check on one shipment

For the homeware order, a person ticks checks 1 and 2 in a minute because the reconciliation block shows 1,260 units and 4,872.00 USD on both sides. Check 3 fails: gross weight is blank for the trays, so a carton goes on the scale and reads 9.4 kg, which the master sheet now stores. Check 7 is held because the new tray design has no confirmed code yet, and the broker's reply the next morning sets the shipping date. The documents were drafted in minutes; the shipment waited a day because the facts were not there, which is the right outcome.

Two smaller jobs the same setup can take on

Once the master sheet exists, the same prompt style handles two chores that usually eat an afternoon. The first is the proforma invoice a buyer asks for before paying: it is the same layout with the status marked as a quotation, so you paste the quote lines and ask for a proforma with a validity date and the payment terms copied from your sales conditions. The second is the shipper's instruction email to the carrier or forwarder. Paste the packing list totals and ask for a short, plain message stating cartons, gross weight, dimensions, pickup address and the goods description, then read it against the booking.

A third use is a pre-flight read of the finished pair. Paste the invoice and the packing list and ask, "List every difference in quantity, weight, description, name or address between these two documents, and say nothing else if there are none." The last clause stops the assistant from inventing discrepancies to look helpful. A shipment of 1,260 units that returns "no differences" still gets the person's eight checks, because the assistant can only compare what is on the page; it cannot see that the carton on the dock has a different label.

Data handling when documents contain customer details

Commercial invoices carry names, addresses and sometimes tax or bank details. Before pasting an order into any assistant, check the plan's data controls. OpenAI and Anthropic both say their business plans do not train on your content by default, whereas their consumer plans have a model-training switch ("Improve the model for everyone" in ChatGPT's data controls) that you should turn off if you use them. Strip bank account numbers and anything the draft does not need. If you automate the flow with a workflow tool, keep the review step so that a person reads the document before it is emailed to a buyer or a carrier.

Choosing for your volume

A rough rule from the numbers above. Under about twenty shipments a month with order data you already hold in a sheet, use the spreadsheet and assistant route and spend nothing extra. If supplier PDFs are the bottleneck and you read more than a hundred a month, price an extraction workflow against the hours it saves. If you hold stock for many channels and want documents generated inside the same system as orders, look at an inventory platform, but cost the whole plan, not the AI add-on alone. If you ship a wide, changing catalogue, ask for classification-service quotes while keeping the broker as the final word.

Whichever route you take, treat the first ten shipments as a trial. Produce the documents the old way and the new way side by side, compare every field, and log each difference and its cause. Most differences will be missing master data rather than AI mistakes, and fixing them once makes every later shipment cleaner. Before you rely on the setup for a new buyer, check their document requirements, since some insist on specific wording, extra declarations or particular document formats. For wider uses of AI around buyers and paperwork, see AI for exporters: six jobs from finding buyers to paperwork.

Questions exporters ask before trusting a draft

Can ChatGPT or Claude produce a commercial invoice?

Yes, as a draft. Give it your order lines, buyer details and terms and it will lay out a clean invoice and arithmetic. It cannot know your buyer's import rules, so tariff codes, origin statements and declared values need a person, and the codes need a customs broker.

Is an AI-chosen tariff code safe to put on an invoice?

Treat it as a suggestion to take to your broker. A wrong code can change duty, delay clearance or trigger questions later. Use the AI to write a clear product description, then let the broker confirm the code.

What is the cheapest setup that works?

A spreadsheet holding your product master data, plus a chat assistant on a plan you have to draft the documents from each order. Add paid extraction software only when you are retyping supplier paperwork every day.

Should I upload customer details to a chat tool?

Check the plan's data settings first. Business plans from OpenAI and Anthropic do not train on your content by default, while their consumer plans have a model-training switch in settings. Remove bank details you do not need to paste.

Further reads

• 3PL Pricing Calculator: AI Quotes Storage, Pick and Pack Rates — Price storage and pick-and-pack before quoting delivered terms.
• 3PL Quote: What 3PLs Charge and How to Compare Every Fee With AI — Compare what a warehouse charges to pack and ship the order.
• Best AI for Ecommerce Operations: Orders, Stock, Returns, Suppliers — Orders, stock and suppliers alongside the paperwork.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Nanonets pricing page, Cin7 pricing page, Zonos Classify product page, ChatGPT pricing page, OpenAI enterprise privacy page and Anthropic privacy centre, as read in November 2026.