Skip to content

AI on the desk

What AI can and cannot do on a small export desk

A language model is a fast reader and a tireless drafter. It is a poor authority on HS codes, duty rates and the law. Here is where we would use one on an export desk, where we would not, and what the evidence says.

Clay model of a small robot reading a document with a magnifying glass beside a desk lamp, a tray of envelopes and a rubber stamp.

The short version

  • AI is useful for reading purchase orders and letters of credit, drafting replies, and comparing documents, with a person checking.
  • It is not a source of truth. In published tests the best AI systems got the full tariff code right less than half the time; human experts scored 95 percent.
  • You make the declaration on the shipping bill. That responsibility stays with you when a broker files it, and it does not pass to software.
  • Extraction errors cluster in numbers, so quantities, prices and dates need a second look.
  • Check the data settings before putting a buyer’s documents into a consumer chatbot.

Sources last checked on . Rules like these change, so confirm the current position before you act.

Earlier this year two partners at EY India argued that AI-led compliance can power India’s export boom (opens in a new tab). Their case is written for companies with trade and tax departments: automated classification, origin checks and shipment visibility on one platform.

If your export desk is four people and a customs broker, the question is smaller and more useful. Which hours of the week can a language model take off your team, and which jobs should it never be left alone with? We sell this kind of system, so we have an interest in the answer. We have tried to give the evidence against as well as for.

Reading: good, with one habit to watch

The most valuable thing a model does on an export desk is read. A buyer’s purchase order arrives as a PDF, and somebody types its forty lines into your system. A letter of credit runs to three pages of conditions that somebody has to turn into a checklist. A forwarder’s email contains one new date buried in a thread.

Models are now good at this. In a 2025 benchmark (opens in a new tab) by Fraunhofer researchers, the best of eight models got 96.5 percent of fields exactly right on clean invoices and 92.7 percent on scanned ones. On a harder set of real business documents (opens in a new tab) tested by researchers at SAP and Stanford, the best score was lower, around 77 to 79 on the measure they used.

Look at what the errors were, though. The SAP paper gives examples: “002” returned for a quantity of 2, and “1” returned where the document said 8.00. The mistakes are in numbers, and a wrong number on an order is the expensive kind of wrong.

So 96.5 percent is not “nearly perfect”. It is about one field in thirty wrong on a clean document, and more on a scan or a photograph. The way to use this is to have the model fill the order record and show each value beside the place it came from, so a person checks quantities, prices and dates in a minute instead of typing them in fifteen.

Drafting: fast, and confidently wrong when it has nothing to go on

The second job is writing. Replies to product questions, follow-ups on quiet quotations, a polite chase for an overdue balance, a message in Spanish to a buyer whose English is as limited as your Spanish. A model produces a decent first draft in seconds, in the right tone.

The risk has a name. The US standards body NIST calls it confabulation (opens in a new tab): “the production of confidently stated but erroneous or false content”. It adds that people tend to believe such output because it sounds so sure. On an export desk that looks like a draft that promises a 20-day lead time you cannot meet, or mentions a certification you do not hold, because a plausible reply would contain one.

Two rules deal with most of this. Give the model your price list, your catalogue and your terms, and tell it to answer only from them. And have a person read every message before it goes. A buyer cannot tell which sentences a machine wrote. They will hold you to all of them.

Cross-checking: the dull job it does well

The third job is comparison. Does the packing list agree with the invoice? Does the invoice description match the letter of credit word for word? Is the notify party the one on the purchase order?

This is dull, it is done at the end of a long day, and it matters. The ICC estimates (opens in a new tab) that between 65 and 80 percent of documents presented under letters of credit are refused the first time. Banks examine a presentation on the basis of the documents alone (opens in a new tab), so a difference between two pieces of paper is all it takes. We went through the common ones in our article on document mismatches.

The banks have started using the same technology on their side. HSBC’s document-checking system is live in Hong Kong, the UAE and the UK, according to Global Trade Review (opens in a new tab), and several other large banks use tools from specialist vendors. HSBC published no accuracy figures, and said that “human judgement is essential to understanding exceptions and making the decisions that matter”. That is the right model for an exporter too. The software flags the differences; a person decides which ones matter.

HS codes and duty rates: do not ask it

Here the evidence points the other way, and clearly.

Researchers at Alibaba built a test (opens in a new tab) of 632 real products and asked AI systems for the full 10-digit tariff code. The best one was right 46.8 percent of the time. Human classification experts scored 95 percent. A second study (opens in a new tab), using 200 US customs rulings, found a leading model correct at 10 digits in 25 percent of cases and at 6 digits in 55.5 percent. That paper’s authors sell a tariff tool and their test was small, so do not lean on the exact figures. Both studies say the same thing: on classification, today’s models are wrong about as often as they are right.

That would be a curiosity if a wrong code were harmless. It is not. The code decides the duty your buyer pays, whether a trade agreement applies, and which incentive rate your shipment earns. And the declaration is yours. Under Section 50 of the Customs Act (opens in a new tab), the exporter subscribes to a declaration of the truth of the shipping bill’s contents and must ensure the information is accurate and complete. When a broker files for you, the Act treats what the agent does as done with your knowledge and consent (opens in a new tab). Knowingly making a false declaration can bring a penalty of up to five times the value of the goods, as a Government of India order (opens in a new tab) quoting the section sets out. Nothing in that chain moves responsibility to a piece of software.

Use a model to suggest two or three candidate headings and the questions that separate them. Then go to the sources that count:

The same caution applies to anything that is a matter of law or of a published rate: duty rates, scheme eligibility, rules of origin, sanctions. A model’s answer is a starting point for looking it up.

Before you paste a buyer’s document into a chatbot

Free and personal versions of AI assistants are not the same product as the business versions, and the difference is what happens to what you type.

A purchase order has your buyer’s prices on it. An LC has their bank details. Check the setting, or use a business plan or an API-based tool, before those go anywhere.

There is also Indian law to keep in view. The Digital Personal Data Protection Act (opens in a new tab) defines personal data as any data about an identifiable individual, with no exception for business contacts, and its main obligations are being phased in through 2027. How it applies to a list of overseas buyers is a question for a lawyer, not for us.

How we would set it up

If we were starting on your desk tomorrow, we would do less than you might expect.

  1. One job first. Reading purchase orders into the order record. The output is easy to check and the time saved is easy to see.
  2. Source beside value. Every figure the model extracts sits next to the line it came from.
  3. A person approves. Nothing is sent to a buyer, a bank or a broker on the model’s say-so.
  4. Numbers get a second look. Quantities, prices, weights and dates are checked every time, because that is where the errors are.
  5. Codes and rates come from the official source, with the date you looked.
  6. Count the mistakes for a month. If the model’s error rate on your documents is not lower than your team’s typing errors, stop.

That is how our AI Export Desk is designed: it works from the information you provide, and a person approves every send. The same six rules apply if you are doing this yourself with a subscription and a spreadsheet.

Questions exporters ask

Can ChatGPT or another AI tool find the right HS code?

It can suggest candidates, but it should not be the final answer. In one published test the best AI system got the full 10-digit code right 46.8 percent of the time against 95 percent for human experts. Confirm the code against the official tariff and with your customs broker, and consider an advance ruling for a product that is hard to classify.

Is it safe to upload a buyer’s purchase order to an AI chatbot?

It depends on the product and its settings. Consumer versions of several assistants can use conversations to train their models unless you switch that off, while business plans and API access generally do not. Check before uploading documents that contain a buyer’s prices or bank details.

What should a small export team automate first?

Reading purchase orders into the order record. The output is easy to check against the original, the time saved is visible, and a mistake is caught before it reaches a buyer or a bank.

Who is responsible if an AI tool gets a customs declaration wrong?

The exporter. Under the Customs Act the exporter subscribes to a declaration of the truth of the shipping bill’s contents and must ensure the information is accurate, including when a customs broker files it.

Sources

  1. How AI-led compliance can power India’s export boom (opens in a new tab)EY India, 24 March 2026
  2. Benchmark of multimodal models on invoice extraction (opens in a new tab)Fraunhofer IAIS and Lamarr Institute, preprint, 2025
  3. Extraction from real business documents (opens in a new tab)SAP and Stanford researchers, preprint, 2026
  4. AI Risk Management Framework: Generative AI Profile (NIST AI 600-1) (opens in a new tab)US National Institute of Standards and Technology, 2024
  5. Technical Advisory Briefing No. 3: reducing discrepancy rates (opens in a new tab)ICC Banking Commission, June 2022
  6. Technical Advisory Briefing No. 12: reading data in context (opens in a new tab)ICC Banking Commission, November 2024
  7. HSBC goes proprietary on AI trade document checking (opens in a new tab)Global Trade Review, 29 September 2026
  8. HSCodeComp: a benchmark for tariff classification (opens in a new tab)Alibaba International Digital Commerce researchers, preprint, 2025
  9. ATLAS: tariff classification on US customs rulings (opens in a new tab)Flexify.AI researchers, preprint, 2025. Small test set; the authors sell a tariff tool
  10. Customs Act 1962, Section 50 (opens in a new tab)LawX. A reproduction of the section
  11. Customs Act 1962 (consolidated text) (opens in a new tab)WIPO Lex. An older consolidation; see Section 147
  12. Revision order of 30 September 2019 quoting Section 114AA (opens in a new tab)Department of Revenue, Government of India
  13. ITC(HS) import and export schedules (opens in a new tab)Directorate General of Foreign Trade
  14. Access2Markets (opens in a new tab)European Commission
  15. Customs Act 1962, Section 28I (opens in a new tab)LawX. A reproduction of the section
  16. Study of advance rulings in India, 2021 to 2024 (opens in a new tab)Global Trade and Customs Journal, Kluwer
  17. Gemini Apps Privacy Hub (opens in a new tab)Google
  18. Privacy FAQ for Microsoft Copilot (opens in a new tab)Microsoft
  19. Is my data used for model training? (opens in a new tab)Anthropic
  20. Your data: API data usage (opens in a new tab)OpenAI
  21. Digital Personal Data Protection Act, 2023 (opens in a new tab)Gazette copy hosted by PRS Legislative Research

This article explains how a rule or process works so you can ask better questions. It is not legal, tax, customs or banking advice. Export-Saathi is not a customs broker, bank or compliance adviser; confirm anything that affects a live shipment with your CHA, AD bank or adviser. Spotted an error? Tell us and we will correct it.

Keep reading

Export documents

The same shipment, typed five times: where export documents disagree

An invoice number typed two ways can hold up an IGST refund. A missed declaration on a shipping bill can cost an incentive. Here is where the documents for one shipment disagree, what it costs, and what can be repaired afterwards.

9 min read

Trade agreements

India–EU FTA: what an exporter must be able to prove

The agreement is concluded but not yet signed or in force. When it starts, the lower duty will go to exporters who can show where their goods were made, on a document customs can check online.

8 min read

What would make your next order easier?

Let’s walk through one enquiry or shipment and find a useful place to begin.

Book a discovery call