With 99 chapters, 1250 HS headings and more than 17,000 TARIC codes, the classification landscape is a wide-ranging world of commodity codes. With that being said, AI may offer the opportunity to reduce the risk of an incorrect commodity code for your business. Work that is normally done by humans and is very labour-intensive can be done a lot faster by AI. However, the question remains whether AI will do it as accurately as humans. From our experience, we know nevertheless that in very many cases, classification hasn't been seriously examined 'at all' and therefore no human intervention has taken place. Examples of hundreds of very different items - with just one commodity code being used - occur regularly. The use of AI will then quickly lead to improvement.
A problem in this regard is that we also often hear that the commodity code is not modified by Customs. Is this then 'approved' or 'condoned' by Customs? No! Often, checks are hardly ever carried out, or not at all, because of the lack of capacity to check every consignment to ensure the correct commodity code. However, the fact is that by using AI, Customs will be able to identify which commodity codes are wrong more often and more efficiently.
An increasing number of digital tools based on AI are available in the classification landscape market. These digital tools operate based on machine learning and deep learning. Machine learning is a form of artificial intelligence aimed at building systems that can learn from processed data or use data to perform better. Deep learning is a form of machine learning in which artificial neural networks (algorithms) learn from large amounts of data. We believe that AI certainly adds value in the world of classification.
This article explains one provider's AI tool called ChatGPT. ChatGPT is an online tool that suggests several CN codes based on the description of your product. You can choose the most appropriate commodity code from the suggestions. This seems like a foolproof system. After all, if all importers can match the correct CN codes to their goods, pick the right TARIC code while the system is self-learning, the tool should be able to display the correct CN code every time.
However, is this the case?
Let's
put it to the test and ask ChatGPT the following:
What is
the commodity code of an oxygen mask?
The
Harmonized System (HS) code for an oxygen mask can vary depending on the
specific type of mask, its use, and the materials from which it is made.
Generally, oxygen masks for medical use may fall under HS code 9020.00.0000,
which is the code for "other breathing appliances and gas masks, excluding
protective masks having neither mechanical parts nor replaceable filters."
However,
it's important to note that HS codes can be subject to change and may have
variations based on the specific characteristics of the product. For accurate
and up-to-date information, it's recommended to check with the customs
authorities of the specific country you are dealing with or refer to their
customs tariff database.
That seems about right!
But with 99 chapters consisting of more than 17,000 Taric codes, a mistake is easily made. In addition, commodity codes can obviously change in the meantime, jurisprudence also has an impact and classification regulations still exist. These circumstances are not included in the free version at all. Also of note is that AI learns through machine learning. In other words, ChatGPT learns based on approved user responses. If a commodity code changes and this is not noticed, this could - perhaps due to the use of AI notwithstanding - lead to an incorrect declaration.
That
ChatGTP does not consistently get it right is also evident from the following
example where the query had been asked to determine the commodity code of a
mouth mask. The result was as follows:
Face
masks can be classified under different HS codes depending on their materials,
design, and intended use.
Here are
a few possible HS codes that may apply to face masks:
6307.90.98:
Other made up articles, including dress patterns, of all types of textile
materials, including face masks and respirators
9020.00.00:
Other breathing appliances and gas masks, excluding protective masks having
neither mechanical parts nor replaceable filters
4818.90.90: Other articles of paper and paperboard, including face masks and respirators
3926.90.90: Other articles of plastics and articles of other materials of headings 3901 to 3914, including face masks and respirators
Plenty
of choice, so it seems. But the correct option is still missing, being 6307
9090. Protective mouth and nose masks - in fact, if they fall under Chapter 63
- have their own subheading. While the import rate for both 6307 9090 and 6307
9098 is 6.3%, 6307 9098 is not the correct code. ChatGPT's suggested GN Code
9020 0000 could possibly also be correct, but of course this should be
critically examined. In fact, most mouth masks (which we know from the COVID
period) are not covered by Chapter 90.
Critical or excited?
As
experts on classification, we are quite excited about ChatGPT's capabilities.
But we also remain very critical. We still have many questions, after all. As a
user, how do you know that the code is correct? How is ChatGPT 'filled' in
these very specific cases? Does ChatGPT know all the notes, explanations, case
law, classification regulations, national measures etcetera. The answer to this
question is 'No'! This is because the free version does not possess up-to-date
data.
We
conclude that AI can be a useful tool in determining the commodity code of a
large range of goods. Nevertheless, it is still necessary that human
intervention is required when classifying goods. After all, AI is still not
capable of error-free classification anno 2024. The future will tell whether
human intervention will eventually be displaced by AI.