It's been almost two years since I joined Customaite as a product manager, straight out of university and thrown into a business that tries to use the world's most rapidly evolving technology to optimize one of its most complex and slowest-moving industries.
Nothing breaks down the naivety of a junior faster than trying to understand Customs. And for those who aren't in the business: understanding Customs isn't really a thing. It's international trade, supply chains, geopolitical interests, taxes, economic and ecological considerations that have created decades of discussions, legislation, and countless attempts at harmonization, only for the Germans to decide they'd rather classify their goods with 11-digit codes instead of 10 as the rest of the EU does.
It all starts from the Harmonized System: a global framework of six-digit codes that sorts every product that exists, every product that has ever existed, and every product that might exist into a single, definitive category. The entire world agrees on those six digits . . . and then each country expands on them as it sees fit.
This is all done in the name of protecting their own interests, enforcing taxation, and imposing various restrictions.
Norway uses eight digits; most of the EU and the UK go as far as 10; China stretches all the way to thirteen. And between all these customs territories, only the first six digits will likely match.
These codes are not just a formality. They are the basis of everything. In 2003, Marvel won a court case against the United States government because Wolverine, the clawed superhero mutant, is technically not human. The dispute was whether the X-men figurines they import, should be classified as “Dolls representing only human beings and parts and accessories thereof …”, or as “Toys representing animals or other non-human creatures …” taxed at 12% and 6.8% respectively. The court agreed with Marvel and thus Wolverine became, legally, not human, and more importantly, cheaper to import.
Again and again, semantics and interpretation have triggered expensive and lengthy disputes. Botanists can scream all they want. According to US customs, tomatoes are vegetables.
So, when a shipment arrives and the question arises, “what is actually in this container?”. The answer is not a best guess, not a feeling, or a description, but a legal definition, with potentially enormous consequences.
Tons of goods get stuck at the border daily, because the correct local classification did not reach the declarant in time. You would be forgiven for trying to solve this costly inefficiency by feeding the invoice and any product information you may have to an AI to speed things up.
But in a world where a court has ruled that Wolverine is a mutant and not a human, and where us customs insists tomatoes are vegetables, letting an AI be confidently wrong about what is inside your container is a mistake you cannot afford.
As a product manager, I often ask: what is the user trying to accomplish? The answer sounds simple "declare a shipment of Wolverine figurines,” but then the questions start. Are the figurines staying in the country, or just passing through? Is it only figurines, or do they come with accessories? Were they produced in China, or did they only ship from China? And of course: are mutants human or not? Each of these questions has a multitude of possible answers, and each answer will have a relevant impact. Different codes, different procedures, different pieces of a larger, incomplete puzzle. Pieces a customs officer needs before they will let your shipment through.
Completing that puzzle correctly but efficiently is the challenge. Anyone can set up an account with an off-the-shelf AI, spend days or weeks perfecting their all-purpose prompt to make their shiny new AI agent complete a customs declaration, only for it to fail on the very first real-world example, because it timed out on page 29 of a Korean invoice. A couple of retries and a few tweaks later, the declaration is "ready". You never noticed it quietly guess the country of origin, because it couldn't find the information and there was nothing in the file to contradict it.
You now spend more money on completing a single declaration than anyone is willing to pay for it. And the lack of explainability and accountability has produced a string of subtle mistakes and best guesses, the kind a customs officer won't appreciate, and that you will be held responsible for.
In a normal product, edge cases are the weird scenarios that 1% of your users run into. In customs, the edge cases are your entire product.
Speeding up the paperwork with AI is not that hard. I can think, on the spot, of numerous methods to make a declarant’s workflow more efficient. But I cannot guarantee for any of them that the results will be compliant. I am no longer that naïve. The question we try to answer daily is not; “how can we speed up this process?”. But rather, “how do we increase compliance whilst gaining efficiency?” It’s a much tougher puzzle than you might expect. But getting it right is worth far more than any reckless AI implementation will cost you in the long run.