Since the introduction of the Harmonized Tariff System (HTS or HS) in January of 1988 and its global implementation in following years (for example, in the U.S. on January 1, 1989), classifying products (i.e., associating the tangible product to its related HS code) has been a global party. Used on import (and export) declarations, HS codes identify the duty rates applicable to the specific goods, relate to statistics, give regulators an opportunity to link Anti-Dumping Duties (ADD) and Countervailing Duties (CVD) to products, dictate how to qualify for preferential treatment, and can govern document and license requirements. Quite a laundry list—and that makes the correct HS code classification an important piece of information, especially when using an incorrect classification can lead to penalties and delays upon import.

In the U.S., with roughly 16,000 HS codes to choose from, a customs ruling database for U.S. classifications only (CROSS) that is 206K classifications strong, ongoing changes to import tariffs, and a massive World Customs Organisation-initiated overhaul every five years (2022 here we come), it is no wonder classification is evergreen on trade compliance professionals’ list of concerns that demand a significant amount of attention.

To make it more complicated—although harmonized globally at the six-digit level, local authorities are allowed to differentiate down to a local n^th digit (usually eight or 10) and have not been hesitant to do so. For example, the U.S. and European Union both support HS codes that have 10 digits, but few are the same or represent the same products. As import declarations are filed locally, this implies that, for each importing country, a different HS code must be identified and then maintained for any product shipped into that country. Do the math: a product catalog of 50,000 parts that ship to 50 different countries adds up to a solid 2.5 million classifications. Not something to maintain on the back of an envelope—unless it’s a really, really big one, erasers are cheap, and pencils are free.

With widely diverse needs for classification (e.g., from a B2C ecommerce shipment of two cotton T-shirts that need an HS code for a quick landed costs calculation, to raw materials and semi-finished products for manufacturers, to a single unique import of a $10 million factory engine), it is no surprise that any self-respecting Global Trade Management (GTM) solution or consultant is happy to assist companies in desperate need for those classifications. And no wonder that, since around 2000, numerous software companies have been trying to solve the mystery of auto-classification.

The diversity in the initial reason for classification comes with different parameters for success. For an ecommerce retailer, an autoclassification tool can solve many challenges (e.g., quick returns, high volume of items are immediately classified), but accuracy can be a challenge. A lack of accuracy is not something importers can afford when, for example, the classification determines whether the import is subject to ADD, is heavily restricted from a license perspective, or is subject to quotas. Basically, (auto-) classification is like a freeway and, depending on the exact needs, companies take a different exit.

There are three key components to a successful (auto-) classification project—other than, of course, the hopefully not superfluous statement that a decent amount of classification expertise comes in handy when either classifying or building a tool.

First, the quality of the product description. ‘Garbage in, garbage out’ also applies to classifying. Poor descriptions, lack of product detail, or even incorrect specifications will likely lead to an incorrect HS code with all related consequences. For quality descriptions, product managers or developers may get involved to provide the necessary technical detail as some classification decisions are made based on those elements.

Second, the classification logic. Whether the classification is assigned by a person or a tool, classification logic cannot lack, well, logic. This means many things: rules that decide to classify a piece of clothing that is not gender-specific as textiles for female or male (and the U.S. handles it differently from the EU); rules-based classification that guides the correct classification in a decision matrix fashion; the ability to ignore information not relevant to the classification (e.g., color); or the ability to observe characteristics that may be needed in one case but not for another (e.g., weight), including material compositions that are usually very important. The logic must also account for a way to ‘smart search’, or search across different references to generate results from, such as synonyms, natural language, industry jargon, and even from images. In addition, classification logic means integrating Artificial Intelligence (AI) and Machine Learning (ML) into the application so results can automatically improve, which enhances both the number of items classified and the quality of the classifications without human intervention.

Third, the classification reference database. The classification logic must look to match a description with an HS code not only by matching it with a ‘word in the tariff’ but also with the explanatory notes and, preferably, for broader context a natural language reference. This might include a shipping manifest reference or information gained via access to previous imports and classification repositories of identical products. Regardless, all types of references need to be reviewed before the final classification is determined. The logic is only as sound as the foundation on which it is built.

It’s important to keep in mind that references are also where, as an industry, companies should actually assist one another. Data privacy concerns notwithstanding, there must be a way to ‘crowd source’ references, which could reduce the efforts made and resources spent on classification in sensational fashion—engineering a classification freeway that is even more well-marked and efficient to traverse.