Product classification may sound like an obscure, back-office task that only concerns customs officials or tax accountants. But in reality, it is a cornerstone of tax and customs compliance for businesses of all shapes and sizes, whether they sell goods, services, or both. Accurate classification ensures that the right tax rates, duties, and exemptions are applied, helping companies avoid costly errors, audits, and penalties.
When we think about product classification, we often picture long spreadsheets filled with codes like “HS 8471.30” or “HTSUS 0101.21.” These codes come from global systems such as the Harmonized System (HS) and its regional versions, like the Harmonized Tariff Schedule of the United States (HTSUS) and the European Union’s Combined Nomenclature (CN). They create a common language for classifying goods in international trade and applying the correct import taxes and duties. But product classification is not only about international trade. Even domestic sales require assigning the right tax rate to products and services. Businesses that rely on tax engines or accounting systems often use tax codes—alphanumeric identifiers that tell the system whether a product is taxable, exempt, or qualifies for a reduced rate. In other words, classification is everywhere, touching every invoice and tax return, often without anyone outside the finance team noticing.
The hidden dangers of misclassification
Getting product classification wrong is not just a small technical mistake. It’s more like planting a tiny bug in your company’s software that quietly replicates itself until it’s everywhere. A single misclassified product can flow undetected into invoicing, accounting, financial reporting, and tax filing systems. Each platform, trusting the information it receives, passes the error along until one day, the mistake is discovered—usually by a tax auditor, and often with a hefty bill attached.
Errors in product classification can result in underpayment or overpayment of taxes, incorrect financial statements, and reputational damage. It can also mean years of retroactive corrections and fines. In short, it’s a nightmare scenario that every CFO wants to avoid.
From manual labor to machine learning: a new era in classification
Historically, product classification was done manually. Tax professionals would comb through product descriptions, technical specifications, and usage details, then use their knowledge of tax laws to assign the correct codes. This method required deep expertise, meticulous attention to detail, and endless patience. And unsurprisingly, it was slow and prone to human error.
Enter artificial intelligence. AI systems today can analyze vast amounts of product data—including descriptions, specifications, and images—to suggest accurate tax classifications. Hybrid systems that combine text and image analysis have become especially effective, as pictures can help clarify ambiguities that plain text struggles to resolve. By learning from historical data and classification patterns, AI can help reduce human error, speed up the classification process, and handle enormous product catalogs with ease.
It sounds like a dream, doesn’t it? But before you envision a future where AI bots run the entire tax department, it’s important to ask: Can AI truly master the complex, nuanced world of tax classification?
The gray areas: where AI may still struggle
Not every product fits neatly into a predefined category. Products that have multiple uses or complex components often fall into tax gray areas that require subjective judgment.
Take smartwatches, for example. Should they be classified as wristwatches or as communication devices? If the primary function is telling time, they belong in one category. If it’s making calls or sending messages, they belong to another. Similar dilemmas arise with multifunction printers, which could be classified either as printers or photocopiers depending on their main function.
Even seemingly simple products can turn into legal puzzles. Different countries and regions have their own classification quirks, often leading to results that defy common sense. The “Subway” case in Ireland is a famous example: the Irish Supreme Court ruled that Subway’s bread contained too much sugar to legally qualify as “bread” for VAT purposes.
Meanwhile, across the Irish Sea in the United Kingdom, there’s a £470,000 tax battle over a surprisingly squishy question: Are Mega Marshmallows sweets? This question matters because most food in the UK is zero-rated for VAT, but confectionery—sweets, chocolates, and the like—is taxed at 20%. According to the law, anything “sweetened and normally eaten with the fingers” counts as confectionery. Initially, the First-Tier Tribunal sided with the marshmallow company, arguing that Mega Marshmallows are so large that they are more of a barbecue ingredient than a snack you casually pop into your mouth. However, HMRC was not satisfied and continued to appeal up the court ladder. Eventually, the Court of Appeal weighed in, stating that the lower court had overlooked a crucial point: how people actually eat Mega Marshmallows. If most consumers simply eat them with their fingers straight out of the bag, then they qualify as sweets—and yes, the 20% VAT applies. Now, the case is heading back to the tribunal (once again) to resolve the big question: Are Mega Marshmallows normally eaten with fingers, or are they typically roasted first?
These examples highlight a crucial point: product classification is not purely technical. It is a legal process that often depends on interpretation, usage, perception, and even cultural habits. While AI can process millions of data points faster than any human, it may struggle with the subtle, context-dependent reasoning needed to resolve such cases.
Recent scientific research backs up these concerns. Studies have shown that zero-shot product classification—where an LLM tries to categorize without seeing examples beforehand—works reasonably well, but still struggles with ambiguous, or domain-specific product categories.
Why human expertise remains irreplaceable
Despite its impressive capabilities, AI is not yet capable of fully replacing human expertise when it comes to product classification. Complex legal interpretations and the need for nuanced judgment on intended use and function mean that humans are still needed at the controls.
For instance, AI can easily classify a chair as a chair. But can it determine whether a reclining massage chair equipped with heat sensors should be taxed as furniture, medical equipment, or luxury electronics? That requires understanding the product’s design, intended use, marketing claims, technical specifications, and often, the applicable case law.
In short, AI can automate the routine—scanning descriptions, suggesting matches, flagging inconsistencies—but it cannot (yet) automate the judgment, interpretation, and creativity that human tax professionals bring to the table. Leveraging AI in VAT determination is much like using a navigation system during a storm. Technology offers vital assistance, but experience and intuition guide critical decisions.
A future of collaboration: AI and humans together
The future of product classification is not about choosing between humans and machines. It’s about collaboration. AI can and should handle the heavy lifting: processing millions of product descriptions, highlighting likely matches, and spotting potential errors. This frees up human experts to focus on the challenging, high-value tasks that require experience, judgment, and an understanding of legal context. Let AI handle the volume, and humans handle the nuance.
One promising development from recent research is the idea of blending AI models with external sources of information, such as knowledge graphs or retrieval-augmented generation (RAG) systems. Instead of expecting AI to “know everything,” we help it access richer, curated domain knowledge.
As AI continues to evolve, it will be fascinating to see just how far we can push the limits. But for now, when it comes to navigating the fiscal theme park that is modern tax law, it’s still wise to keep a few experienced humans on hand—just in case the machines need a little help reading the menu.
At the same time, it’s worth asking a more fundamental question: before we rush to deploy increasingly complex AI systems to manage tax rules, are we addressing the root of the problem? Building layers of technology to manage an already overwhelming web of legal distinctions is, at best, a reactive strategy. It is like constructing a labyrinth and then inventing smarter and smarter tools to find the way out. Perhaps, instead, we should ask whether the labyrinth needs to be so complicated in the first place. If tax classification systems were simplified, standardized, and made more intuitive, we could dramatically reduce the need for technological aids—and maybe save a few Mega Marshmallows along the way.
The opinions expressed in this article are those of the author and do not necessarily reflect the views of any organizations with which the author is affiliated.
