October 16, 2019
strategy | Global Trade Management

AI and ML come to GTM

AI and ML come to GTM

Some vendors are incorporating artificial intelligence and machine learning into their global trade management software. Here's how these technologies could improve trade compliance and classification.

By Toby Gooley

If you work for a business that imports products into the U.S., you know how important it is to correctly describe and declare them to U.S. Customs and Border Protection (CBP) when they enter the country. For customs purposes, the product description consists of two elements. The first is an English-language description of the product, and sometimes its function or purpose. The second is the product's tariff classification in the form of a numerical identifier.

The correct tariff classification is critical: As its name suggests, it's the primary factor in determining which tariffs apply and the amount of duty you'll pay. Get it wrong, and you could end up paying more than you should ... or less than you should and have to make up the difference later (and possibly pay a fine to boot). This is no small matter; for a large importer, a mistake could potentially add up to millions of dollars in unnecessary expense.

But tariff classification is a complex and difficult process, and human experts are hard to come by. Classification errors are common, and tales of epic disagreements between customs authorities and importers abound. And there's a new complication: President Trump's tit-for-tat trade war with China has led, often with little advance notice, to repeated rounds of tariff increases on both sides. It's more important than ever to verify that a commodity really is—or isn't—subject to those higher duties.

Importers have long turned to global trade management (GTM) software and stand-alone tariff-classification solutions to help them classify their products correctly. But the complexity of the task makes it hard to fully automate. Now, software providers are betting that incorporating artificial intelligence (AI) and machine learning (ML) into their products to make the software "think" and act more like a human expert will take classification technology to the next level. Proponents believe these technologies could significantly improve accuracy and prevent costly errors—a welcome development at a time when so much is riding on getting things right.


Like most countries, the U.S. bases its classifications on the global Harmonized Tariff Schedule (HTS), a numerical product-identification system administered by the World Customs Organization. The U.S. version, known as the Harmonized Tariff Schedule of the United States (HTSUS), takes the global version's descriptions and classifications and adds more fine-grained detail.

It's complicated, to say the least. The HTSUS assigns each imported item—which may be a finished product, part, component, or raw material—a unique 10-digit number. To find that number, importers drill down through chapters, headings, subheadings, and individual item numbers, progressing from the vague and general (vegetable products) to the excruciatingly specific (cut flowers and flower buds of a kind suitable for bouquets or for commercial purposes, fresh, dried, dyed, bleached, impregnated, or otherwise prepared: roses). The HTSUS currently includes more than 17,000 classification-code numbers. (U.S. exports require a similar product classification and an export control number, or ECN, to be filed with the U.S. Department of Commerce.)

Identifying the correct number can be a tedious, time-consuming process. It's also complex and rife with ambiguity, even for customs-compliance professionals. A product might fit several HTSUS descriptions, or it might fit none of them exactly.

Furthermore, the descriptions, which were developed for duty-assessment purposes, may bear little resemblance to those used by manufacturers or end-users. Just one example: electric toothbrushes, which are classified not with other toothbrushes but as electronics. Sometimes creativity is called for, such as when classifying sets and kits, which requires establishing the product's "essential character." To top it off, the General Rules of Interpretation (GRI) that govern the classification process are difficult to learn, says Beth Pride, president of BPE Global, a provider of customs-compliance consulting services.

It's a challenge for anyone—including the government officials who judge whether a classification is correct—to always get it right. Now, the hope is that bringing today's digital tools—namely, ML and AI—into the process will help cut through the complexity.


Artificial intelligence is an umbrella term for any type of technology that mimics human thought patterns or behavior, explains William McNeill, a Gartner analyst who writes an annual market report on GTM software. The algorithm-based technology conducts analyses, makes decisions, and responds or takes action much as humans do. For example, AI applies "natural language processing" to recognize, understand, and respond to or act on written or verbal communication. Machine learning is a subset of AI. This technology uses iterative processes to access and correlate data, recognize patterns, and use what it has "learned" or "experienced" to improve its predictive or decision-making model.

Most GTM software vendors have "put AI and related technologies on their product road map," Pride says, "but AI is still developing, and vendors are trying to find the best places to use it." She and others consulted for this article believe there's a strong use case for AI and ML in global trade, particularly for classification accuracy (the main focus of this article), denied-party screening, calculating estimated time of arrival, and risk prediction and avoidance.

AI and ML tools can process enormous amounts of information from multiple sources, including the tariff classification schedules, a user's own product data, customs rulings, and historical classification data. What's particularly valuable about machine learning is that it considers the same data sources that are available to humans but "looks for correlations that you can't get to with the human mind," McNeill says. For example, when human experts classify an imported product, they use their own country's version of the Harmonized Tariff Schedule and maybe review some examples of what other companies have done. But machine learning could go further, he says. A hypothetical example: analyzing historical submissions, cross-referencing them to dispute rulings and other relevant data, and detecting a recurring problem with imports associated with the classification number in question. "I would argue that it's not possible for classifications and denied-party screening to be accurate enough without automation," McNeill says. "You couldn't make the correlations that enable you to put a new lens on the data you have and find new value from it."

Automating routine tasks and freeing up experts to focus on problem-solving makes sense, especially since it's hard to find experts in this field. Furthermore, machine learning could force proper application of the GRI. And when confronted with new products that aren't explicitly provided for in the tariff schedules, Pride says, ML could "learn" from similar cases and recommend new classifications—not just to importers but also to governments. In fact, some governments are already using AI to identify classification errors and related violations.

One example of how this works can be found in the classification solution offered by 3CE Technologies. President and CEO Randy Rotchin describes the software as "an expert system designed to emulate how an expert would tackle this problem." The software "reads" the commercial goods description, and if all the details needed for correct classification are included in that description, it suggests a classification code. If not, it "interacts" with the user until all the required details have been provided and then delivers a code. AI comes into play via the use of natural language processing to read, analyze, and understand product descriptions, Rotchin says. Currently, machine learning is being used to reduce the number of questions and/or choices presented to users by, for example, eliminating theoretically possible but unlikely options and offering only those that are known to be relevant.

Some other GTM software providers are developing their own AI and ML tools for classification, while some incorporate 3CE's solution into their products. Thomson Reuters, for example, does the latter, giving its software the ability to understand plain-language product descriptions and identify the correct HTS or ECN code (for imports or exports, respectively), says Mary Breede, a customer insight leader for Thomson Reuters' Onesource Global Trade Management solution.

Another example of how AI and ML are solving problems in global trade comes from Pawan Joshi, executive vice president of product management and strategy at E2open, the parent of GTM software provider Amber Road. He notes that many data sources include inaccuracies and inconsistencies. E2open uses machine learning to correct such errors and improve the data quality based on the source, he says. "For example, if we keep getting addresses with the city spelled incorrectly, we can correct that across multiple languages." And because each of the tens of thousands of entities on E2open's network platform has a distinct "signature," the system can identify the source of the incorrect information. "Machine learning has the ability to self-correct it without human intervention," he explains.

Joshi notes that an added advantage of using ML to improve data quality in global trade is that the number of errors constantly declines as the machines learn more, which may free up customers to focus more on preventing problems.


While the benefits of applying AI and ML to classification are clear, that doesn't mean they can—or should—completely replace human experts. One reason, says Pride, is that some of the historical data the technology is learning from may be incomplete, outdated, or simply incorrect.

What would happen if an AI- or ML-based solution recommended an incorrect classification? By law, the importer is ultimately responsible for customs compliance, and McNeill notes that most software vendors have clauses indemnifying them in case of errors. However, the technology could itself be a mitigating factor because it imposes a consistent process with proper controls in place, something CBP likes to see, he adds.

Experts say there's a line that shouldn't be crossed when it comes to automating classification. "Of course you want the machines to do as much as possible," says Joshi. "But the moment they reach a certain threshold or limitation on the degree of confidence, then you want to stop and flag the task for a human expert to assess."

There's another reason to think carefully about the use of AI, Breede says. "AI systems will be powered by algorithms analyzing an organization's vast volume of global trade data, and, if left unchecked, these tools have the potential to amplify any human biases the organization inadvertently has perpetuated in its supply chain operations."

All of the experts we consulted expect that AI and ML will become widely used for classification, in large part because human expertise is in short supply. The growth of cross-border e-commerce will provide further incentive to adopt these technologies. When shippers book an international shipment with a parcel carrier, they're forced by the booking software to provide a classification. Large shippers know what to do, but small companies and individuals are unlikely to learn the rules of classification, Pride says. Instead, they quickly "pick what they think is OK" and move on with the transaction. Until AI and ML are built into couriers' systems, classifications for many e-commerce shipments will continue to be incomplete and incorrect, she says.

Breede says that, while AI and ML are "still in an infancy stage" relative to trade compliance, there are other, more immediate opportunities in GTM. These include searching for patterns or anomalies to identify fraudulent transactions; learning from past mistakes and steering the next iteration of operational methods; and reducing supply chain risk by finding correlations in historical data to help forecast customer demand and predict supply chain disruptions.

Joshi of E2open believes that AI and ML could convert tactical functions like classification and denied-party screening into strategic tools. For example, a company could use them to gather data that would be relevant as an order progresses through the supply chain, identify important correlations, and attach the information early on, such as when a purchase order or booking is issued, he says. Such information might include why the product is being sourced from a particular region and an advisory noting that if it is trans-shipped through a certain port, that cargo will require denied-party screening. "Even though such considerations may seem to be tactical, when you pull them further upstream, they can be more strategic," he says.

In this scenario, the roles of customs brokers and third-party logistics service providers (3PLs) will change. "Rather than paper pushing, they will focus more on moving product, on logistics and warehousing, on being the relationship liaison with carriers and ports," Joshi predicts. "Some may go away, while others will change the work content and value they provide."

Breede sums it all up this way: "As AI and ML are introduced into an industry, they reshape that industry's practices. ... What's clear is that artificial intelligence and its associated technologies will continue to transform how trade experts interact with information and machines."

About the Author

Toby Gooley
Contributing Editor
Contributing Editor Toby Gooley is a freelance writer and editor specializing in supply chain, logistics, material handling, and international trade. She previously was Senior Editor at DC VELOCITY and Editor of DCV's sister publication, CSCMP's Supply Chain Quarterly. Prior to joining AGiLE Business Media in 2007, she spent 20 years at Logistics Management magazine as Managing Editor and Senior Editor covering international trade and transportation. Prior to that she was an export traffic manager for 10 years. She holds a B.A. in Asian Studies from Cornell University.

More articles by Toby Gooley

Resources Mentioned In This Article

Technology Videos

Join the Discussion

After you comment, click Post. If you're not already logged in, you will be asked to log in or register.

Subscribe to DC Velocity

Feedback: What did you think of this article? We'd like to hear from you. DC VELOCITY is committed to accuracy and clarity in the delivery of important and useful logistics and supply chain news and information. If you find anything in DC VELOCITY you feel is inaccurate or warrants further explanation, please ?Subject=Feedback - : AI and ML come to GTM">contact Chief Editor David Maloney. All comments are eligible for publication in the letters section of DC VELOCITY magazine. Please include you name and the name of the company or organization your work for.