AI for Customs Brokers: Draft Client Updates and Entry Checks

A small customs brokerage can use AI to turn file notes into clear shipment status emails and to run a consistency check across commercial invoices, packing lists and transport documents before filing, catching mismatched names, units, weights and descriptions. The licensed broker still decides classification, value and origin, and still signs the entry.

Two jobs, two different risks. Status updates are low-risk drafting where tone and speed matter. Entry checks are higher-risk, so the model is used as a second pair of eyes on document consistency, never as the source of any declared fact. Keeping the two jobs apart, with different prompts and different review rules, is what stops a drafting tool drifting into declared facts.

Follow me on Instagram@sagnikteaches

A brokerage and a file, to make this concrete

Say a four-person brokerage handles about 60 entries a week for 35 clients, mostly small importers of goods such as lighting, tools and furniture. Each file has a client email chain, a supplier's commercial invoice, a packing list, a transport document and the brokerage's own notes. Two staff spend roughly two hours a day writing status emails and answering "where is my shipment?" messages, and one rejected or amended entry a week costs an hour or more of rework. Every number here is illustrative.

Connect on LinkedInSagnik Bhattacharya

The aim is modest: cut the writing time, and catch the plain inconsistencies that cause avoidable queries from the authority or the client. Nothing in this approach changes who is responsible.

Subscribe on YouTube@sagnikteaches

Part one: client status updates from file notes

Clients ask the same questions: has it arrived, has it cleared, what do you need from me, what will it cost. The answers are in your system and your staff's notes, so the job is turning notes into a message in the client's language.

The pattern: facts in, draft out, person sends

You draft shipment status emails for a customs brokerage.
Use ONLY the file notes below. Write 80 words or fewer.
Structure: one line on status, one line on what happens next and
roughly when (only if the notes give a date or range), one line on
anything we need from the client, a sign-off from [staff name].
Do not state duties, taxes, fees or release times unless they are in
the notes. Do not speculate about delays or inspections.
If the notes mention a hold, query or inspection, write
"PERSON NEEDED:" and one line, and no client wording.

FILE NOTES:
[paste, without bank or personal data]
CLIENT FIRST NAME:
[name]

Say the notes read: "Documents received 9:10. Commercial invoice and packing list match PO. Entry prepared, awaiting client confirmation of the invoice value. Goods arrive port Thursday per carrier." An illustrative draft: "Hi [first name], we've received the documents for your shipment and the entry is prepared. Before we file, could you confirm the invoice value is correct? The carrier shows arrival on Thursday, and we'll update you once the entry is filed. [staff name]". It states only what the notes support, asks the one question that blocks progress and promises one further update. The staff member checks the dates and sends.

Message types worth templating

• Documents received and what is missing. The model lists missing items from a checklist you provide, never from its own idea of what is required.
• Entry filed. Short, with the reference from your system.
• Query from the authority or hold. PERSON NEEDED. The first message on a hold is written by a broker.
• Released and next steps. What the client does next: collection, onward delivery.
• Invoice for brokerage fees. Amounts from your accounting system, never retyped.

What to fix in the first drafts

Early drafts tend to over-reassure ("everything is on track") and to guess timings ("should clear by tomorrow"). Add to the prompt: "never promise a time unless it is in the notes", and put "on track", "no issues" and "should clear" on a banned-phrases list. A realistic slip: a draft said "your goods have cleared" when the notes said "entry accepted", which is an earlier step. Clients then arranged a collection that could not happen. The fix was a status vocabulary in the prompt: a short list of the exact stage names your system uses, with the instruction to use only those words.

For how the same idea works for parcels and couriers, cutting tracking calls with AI delivery updates shows a low-risk version of the same idea.

Part two: a checklist pass on the documents before filing

The entry's declared facts come from the documents, so a mismatch between documents is the earliest warning of a problem. A broker doing this by eye on the sixtieth entry of the week will miss things. A model reading the three documents side by side, with a written checklist, will not tire, and can list candidate problems in seconds. It is not deciding anything; it is listing what a person should look at.

The checklist

The prompt

You are checking three shipping documents for consistency before a
customs entry is prepared. You are NOT classifying goods, valuing
them or deciding origin.
Compare the documents below against this checklist: party names,
invoice number and date, line items and quantities, units, weights
and package counts, description detail, currency and terms of sale,
origin statements, missing documents.
Output a table: finding number, what differs, where (document and
line), and a suggested question for the client or supplier.
Quote the exact text from each document. If everything matches for
a check, write "no difference found". Do NOT do arithmetic; I will
total the lines in a spreadsheet.

COMMERCIAL INVOICE:
[paste]
PACKING LIST:
[paste]
TRANSPORT DOCUMENT:
[paste]

The instruction "do not do arithmetic" is deliberate. Language models make slips with sums, and the awkward part is that the wrong total reads fluently. Why AI is bad at maths, and what to check in quotes and invoices explains the problem. Do the sums in a spreadsheet: line value is quantity times unit price, the invoice total is the sum of lines, and the weights should add up across packages.

A worked check on a sample file

The file (illustrative): a client imports 12 pallets of LED fittings. The invoice lists three lines: 2,400 units of model A at $6.10, 1,200 units of model B at $9.80 and 600 units of model C at $14.50. The invoice total printed is $35,160. The packing list shows 12 pallets, 4,200 units, gross weight 3,740 kg. The transport document shows 12 pallets, gross weight 3,470 kg, consignee a named freight agent.

The assistant's table flags: (1) invoice quantities total 4,200 units and the packing list agrees, no difference; (2) gross weight 3,740 kg on the packing list against 3,470 kg on the transport document, a possible transposition; (3) the consignee on the transport document differs from the buyer on the invoice, which may be normal if a freight agent is named; (4) the description on all three says "LED fittings" with no wattage or type, which may be too general for the classification work. It suggests questions for each.

The spreadsheet check does the arithmetic: 2,400 x 6.10 = 14,640; 1,200 x 9.80 = 11,760; 600 x 14.50 = 8,700; total 35,100. The invoice prints $35,160, so there is a $60 difference. That is the kind of error a tired reader misses and the one the authority may query. The broker contacts the supplier for a corrected invoice, and the client's question is a single email: "the invoice total is $60 higher than its lines add up to; can you ask the supplier to confirm which is right?" The assistant made three of the four finds (weight, consignee, description), the spreadsheet made the fourth, and the broker decided what to do with all of them.

What still belongs to the licensed broker

• Classification. Based on the tariff and your reading of the goods. AI can suggest candidates and quote the wording to compare; you decide, and you keep your reasoning on file.
• Value. The declared value, additions and adjustments, and the relationship between buyer and seller.
• Origin. Including any preferential treatment claimed and the proof you hold.
• Licences, permits and restrictions on the goods.
• Signing and filing the entry. The responsible person's name goes on it, so that person reads the final entry against the documents, not against the AI's summary.

For an importer's own view of these checks, can AI check customs compliance and what a person must still verify sets out the same boundary from the client's side, and an importer product compliance checklist shows the product-document side.

A second worked check: vague descriptions and mixed terms

Not every finding is a typo. Some are gaps that cause queries later. Take a second illustrative file: a client imports 800 pairs of work gloves and 300 safety helmets. The invoice describes the lines as "Safety goods A" and "Safety goods B", the packing list as "gloves" and "helmets", and the terms of sale on the invoice as "FOB" while the freight charge appears on the transport document as prepaid.

The assistant's table reports two findings. First, the invoice descriptions do not match the packing list wording and neither gives the material, type or intended use, so the line cannot be matched to a precise description without a question. Second, the terms of sale and the freight charge may not agree: under FOB the buyer contracts and pays for the main carriage, so a prepaid freight charge needs an explanation. It does not say which is right, because it cannot know. The broker asks the supplier for a corrected invoice with full descriptions and asks the client who paid the freight, and then decides whether the freight forms part of the declared value under the rules that apply.

Before and after on the description, illustrative: before "Safety goods A", after "Work gloves, nitrile-coated, knitted nylon liner, men's size 9 to 11, 800 pairs". The after version came from the supplier, not from the model. The model's job was spotting that the before version would not do.

Document sets by client: a checklist that stops chasing

Many delays are plain waiting for a missing paper. Keep, for each client, a short list of the documents their shipments normally need, and let the assistant compare each new file against it. The list lives in your system or a shared sheet, so the model does not guess what is required.

When a file arrives, the check is a one-line prompt: "Here is the client's usual document list and the documents received. List what is missing and write a three-line client request." The reply is checked by a person for the document names, which must match your own terms, and sent. Staff who used to read every file for gaps now read only the exceptions.

Onboarding a new client with a plain-language guide

New importers often send half-complete papers because nobody told them what is needed. A one-page guide helps, and AI can produce a first draft from your own notes. Give it your checklist and ask for a guide in plain words that explains what documents to ask the supplier for, which details must match across them, and how long the brokerage needs before arrival. A broker reviews it against what you actually require and adds the points only you know, such as your cut-off times. The guide goes out with the engagement letter. It is not legal advice, and it says so, with a line that requirements depend on the goods and the destination.

Illustrative guide lines: "Ask your supplier for a commercial invoice that names you as the buyer, lists each product on its own line with a clear description, quantity, unit price and currency." "Make sure the weights on the packing list and the transport document agree." "Tell us in advance if the goods need a licence or a certificate." They take the AI seconds to draft and save the first week of back-and-forth with every new client.

Data handling: what you may paste

Commercial documents hold client names, prices and sometimes bank details. Before using any AI tool, check four things: your client agreements and any confidentiality terms; whether the tool is a business plan that does not train on your content by default; whether you can keep a log of which files went through it; and whether you can strip data the task does not need. For the consistency check you do not need bank details, so black them out or leave them off the paste. Check what your data-protection obligations are with a qualified adviser, and tell clients in your terms that you use AI to help with drafting and document checks while a licensed person remains responsible.

Three more drafts that save real time

A "what we need from you" email

Give the assistant the findings table and ask for a client email listing only the questions that need the client's answer, each on a separate line, in plain words, with the document reference. Illustrative: "1. The invoice total is $35,160 but the three lines add to $35,100. Please ask the supplier which is correct. 2. The gross weight is 3,740 kg on the packing list and 3,470 kg on the transport document. Please check with the supplier. 3. Can you send a product description with wattage and type for each model?" A broker reads it and sends it.

A summary of a long email chain

A file with thirty emails is slow to pick up when a colleague is away. Ask the assistant for a ten-line summary of decisions made, open questions and dates. It saves minutes per handover. Spot-check it against the chain, because summaries drop exceptions.

A weekly client digest

From a list of each client's open shipments with stage and next step, the assistant writes a short digest per client. It is useful for clients with many shipments who want one message, not twelve. The facts come from your system export, so a mistake can only be a transcription slip that you can catch by sampling.

How to measure whether it is working

1. Time: minutes spent on status emails per day, before and after, for two weeks each.
2. Rework: entries amended or queried per week, tagged by cause (document mismatch, wrong value, classification, other). Document mismatch should fall.
3. Misses: any error the checklist pass did not catch. Add it to the checklist and the prompt.
4. Corrections: how often staff change a drafted email, and why. If they rewrite more than a third of drafts, the prompt or the notes are the problem, not the staff.

Run it alongside your normal process for a month before you rely on it. For the illustrative brokerage above, the aim is to win back part of the two hours a day spent writing, plus fewer simple document slips; the value is the second, and it only counts if you track it.

For related ground on bringing goods in, a landed cost calculator in Excel is a sheet your clients may ask you for, and a pre-shipment inspection checklist covers an upstream stage. If you would like help deciding which of these steps to automate and how to keep a licensed person in control, that is what my AI implementation consultation covers, and you can reach me through the contact page.

Questions brokers ask about AI on entries

Can AI classify goods for me?

It can suggest candidate classifications and the wording to compare against the tariff, which is useful for a first look. The classification you file is a licensed broker's judgement, based on the tariff text, notes and any rulings. Treat AI suggestions as leads to verify, never as the answer.

Is it safe to paste client invoices into an AI tool?

Only into a business plan where content is not used for training by default, and only with your client agreements in mind. Remove what the task does not need, such as bank details, and keep a note of which tool handled which file.

Should AI send status updates directly to clients?

Start with drafts that a person reads and sends. Once the drafts are consistently right for a month, you can automate the simplest messages, such as 'documents received', and keep exceptions with a person.

Does AI check arithmetic on invoices reliably?

Not reliably. Language models make arithmetic slips. Use a spreadsheet formula for totals and have the model flag mismatches in descriptions, parties and units, which is where it is strong.

Further reads

• Can AI Read Emailed Orders Into a Wholesaler's System? — Turning emailed requests into structured records.
• Supply Chain AI Software: What a Small Importer Needs First — What a small importer needs before buying software.
• How to Track Contract Renewals and Deadlines With AI — Track client agreements and renewal dates.
• AI Supplier Negotiation Emails for Importers: Price, MOQ, Terms — Wording for supplier emails from the importer's side.
• How to Automate Purchase Orders and Supplier Emails With AI — Purchase-order paperwork that feeds the shipment file.
• Can an AI Chatbot Handle Order Tracking and Returns? — The customer-service side of tracking questions.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: the author's guidance on document checks and client communication; no vendor pricing is quoted. Entry requirements, document rules and broker licensing differ between customs territories, so check the customs authority's published guidance for each. ICC Academy note on FOB under Incoterms 2020 (read 5 November 2026).