Classifying products using AI is something that still needs to be approached with a degree of caution and knowledge on the limits and capabilities of AI today. AI can support classification but get a tariff code wrong and the stakes can be high. It can mean you’re overpaying duty rates, risking fines and penalties and don’t have a true and accurate landed cost, all of which affect profit margins. It’s why we’re constantly evaluating where AI can help with tariff codes, and where it is currently lacking.
Key takeaways
- TariffTel tested ChatGPT three times on the same product and got three different tariff codes
- The example was presented at the 2026 Customs Compliance Conference
- AI tools are only as accurate as the product description fed into them
- TTExplore and TTVerified were built to bridge this gap with guided, expert-backed questions
One example that hits home is where we took a single product, a knitted polyester baby blanket with a stuffed toy elephant head stitched to one corner, and asked ChatGPT to classify it three separate times. Three different search terms. Three different codes. Three different duty rates.
Same product. Completely different outcomes. The only variable was how we described it.
That example, which we presented at the 2026 Customs Compliance Conference, reveals something most businesses probably know, but don’t pay close enough attention to. Tariff codes are only as accurate as the information you put in to get them. AI tools like ChatGPT are fast and accessible, but unreliable, especially when you don’t put in an accurate description. It’s why we built TTExplore and TTVerified to bridge the gap, prompting you with questions to answer about your product to get to the correct tariff code. Fast and accurate but with the knowledge that codes are expert-backed, not guessed by an algorithm.
How AI tariff code tools work
Most AI classification tools work by pattern-matching. You describe a product, and the tool compares that description against its training data and known tariff codes to suggest the codes it thinks fit. The better tools go further, pulling product data from your systems and suppliers and flagging where the information looks inconsistent.
That makes data the deciding factor. An AI tool is only ever as good as the data it works from. Details like composition, use, origin and dimensions directly change the code, so accurate, complete product data is what gets you to the right answer. Fragmented legacy systems and incomplete supplier data are where errors, delays and penalties creep in. Wrong data in, wrong code out.
Used well, AI genuinely helps. It can aggregate data from internal and external sources, narrow thousands of options to a shortlist, and flag inconsistencies it has learned to spot over time. Regulations like the EU Deforestation Regulation (EUDR), which demand detailed origin data on raw materials, are exactly the kind of task where pulling clean data together at speed is valuable.
What AI cannot do is the judgement: interpreting the rules, reaching paywalled references like the WCO Explanatory Notes, or leaving the audit trail customs authorities expect. We look at why in detail below. AI on its own is not a reliable substitute for verified classification. The right approach pairs AI speed with expert verification, which is exactly how our tools are built.
What our tests using AI show
We’ve been testing AI classification tools across a range of products, from t-shirts to plant pots to machinery. The results were consistent.
- AI suggested the correct code outright in only 34% of cases.
- In a further 60%, it narrowed options to a shortlist that still required expert verification to reach the right answer.
- In 6% of cases, AI hallucinated entirely, generating codes that didn’t exist or applying reasoning that was completely wrong.
- For straightforward, single-material products, AI can be a useful starting point. For complex, composite or context-dependent goods, it falls short.
Why AI gets classification wrong
- It doesn’t understand the rules
Tariff classification requires interpretation. It isn’t always a straight yes or no answer. The General Rules of Interpretation (GRIs), Explanatory Notes and binding rulings define how products are classified and there are nuances that need to be understood to reach the correct decision. AI isn’t there yet. Where AI does support classification is in pattern-matching against training data. But the minute that data is out of date, incomplete or wrong, the outcomes reflect this. Wrong data in, wrong code out. - AI can’t access critical sources
Key classification references, like the WCO Explanatory Notes, sit behind paywalls. AI tools can’t reach them. - There’s no audit trail
HMRC, EU Customs and US CBP all require documented reasoning behind every tariff code. An AI-generated answer with no methodology and no documentation, won’t satisfy an auditor. HMRC’s own guidance now warns against relying on generative AI for classification decisions. - Bad description in, wrong code out
AI is only as good as what you type in. A commercial product name like “premium kitchen tool set” tells a classification engine almost nothing. It needs exact materials, individual components, intended function, and whether any part is electrical. Without that detail, every answer is a guess.
What an accurate tariff code actually requires
When we review codes, we typically find that 2 out of every 5 are wrong. Those errors compound across thousands of declarations, sometimes for years before anyone notices. The duty you’re paying, the margins you’re forecasting, the landed costs you’re expecting, are all built on codes that may not hold up.
Every tariff code in your system should be able to answer three questions.
- What data was used? Complete product specifications from suppliers including composition, function, dimensions.
- What rules were applied? GRIs, Explanatory Notes, binding rulings, chapter notes. A documented methodology that holds up to scrutiny.
- Who validated it? Expert review from experienced classification specialists.
TariffTel’s approach: automation with expert accountability
AI has a genuine role in classification. It can accelerate initial searches and narrow options. But the final decision needs tariff knowledge, documented reasoning and human accountability. That’s why we built TTExplore and TTVerified. TTExplore gives you fast AI code suggestions with the reasoning and WCO notes so you can check them yourself, and TTVerified adds expert verification and a full audit trail for codes that stand up to scrutiny.
Start classifying in minutes with TTExplore, or get expert-backed, audit-ready codes with TTVerified. Platforms like TTVerified catch exactly this kind of nuance, combining automation with expert review.
Find out more at www.tarifftel.com



