Classification accuracy is a function of input quality, not model quality. A vague product description produces a vague classification, no matter how capable the underlying system is. This is the first thing to understand before setting up an AI-assisted classification workflow, and it shapes every decision in the process below.

This guide walks through creating a new product classification in Trade Insight AI (TIA), using a recorded product walkthrough as the anchor. The video covers the setup stage only: opening a new product record, entering a description, attaching supporting material, and selecting jurisdictions. It stops before the classification runs. No result, code, or reasoning output is shown, and none is described here. What follows is a step-by-step breakdown of that setup process, plus a direct answer to the question that actually determines classification quality: what information should go into these fields, and why.

Watch: Creating a New Product Classification in Trade Insight AI

The walkthrough shows a new product classification being created in Trade Insight AI: product information is entered, supporting material is attached, jurisdictions are selected, and the classification is prepared to run. It ends before any result is displayed.

Step-by-Step: Setting Up a Product Classification

1. Select the correct organization and workspace

Confirm the organization and workspace before creating anything. Workspaces typically separate business units, brands, clients, or environments. Classification records, jurisdiction settings, and reporting are scoped to the workspace they're created in. Start in the wrong one and the classification either has to be recreated or ends up disconnected from the rest of the team's records. This is a five-second check that prevents an hour of cleanup later.

2. Open Classification

From the workspace dashboard, navigate to the Classification section, where existing classifications are managed and new ones are created.

3. Select New → Product

The New menu separates single-item classification from batch and organizational actions. Product is the right choice for a single item. Import package is built for batches. Folder is for organizing classifications that already exist. This walkthrough covers Product only.

Selecting Product opens a classification modal introduced with a short instruction: describe the product, add files, pick jurisdictions, then save for later or run now. That order (description, files, jurisdiction, run) is also the order that matters for classification quality. Description does the most work. Files and images add context. Jurisdiction determines which tariff schedule the classification is actually evaluated against.

4. Add product information, image, and description

A product record accepts an image and a written description. The description field is not limited to a short label. It can hold detailed product information, references to specific legal rules, or anything else that helps a classification engine reason about the product with precision.

This is where most classification inputs fall short. A one-line description gets a general suggestion. A description that also states material composition, construction, and intended use gives the system, and any human reviewer, something they can actually check against tariff language.

5. Name the classification

The classification name is separate from the description and determines how the product appears in the classification listing. It's an organizational label, not an input to the classification logic itself. A consistent naming convention matters once a workspace holds hundreds of records, not because it improves any single classification.

6. Select the relevant jurisdiction

Jurisdiction selection determines which tariff schedule and nomenclature the classification is evaluated against. More on why that's not a minor step below.

7. Run the classification

With description, image, name, and jurisdiction in place, the last step is running the classification. The result, the reasoning behind it, and any confidence indicators the platform surfaces are outside what this video shows, so they're outside what this article claims.

What Information Actually Helps an AI Classification System

More information does not mean better classification. More relevant information does. This distinction gets skipped constantly in how classification tools are explained, and it's worth being blunt about it: padding a description with irrelevant detail does not raise accuracy, and it can bury the details that matter under ones that don't.

What tends to matter, depending on the product:

• Material or composition. What the product is made of, including blends or percentages where the blend affects classification.
• Function and intended use. What the product does. This is frequently the deciding factor between two visually similar products classified under different headings.
• Physical characteristics. Form, structure, or attributes that a tariff heading specifically references.
• Dimensions. Where size or capacity determines which subheading applies.
• Manufacturing method. How the product is produced, when the method itself is part of the tariff distinction (woven versus knit, cast versus forged).
• Product components. For composite or multi-part goods, what each component is and how they combine.
• How the product is presented or sold. Retail packaging, marketed use, or set composition can shift classification for certain goods.
• Technical specifications. Relevant for machinery, electronics, and similarly regulated categories.
• Legal or regulatory information. Known rulings, section notes, or exclusions that apply to the product.
• Relevant classification rules or notes. Section and chapter notes the classifier should weigh.
• Supporting documentation. Spec sheets or certificates that substantiate claims made in the description.
• Product images. Visual reference material, covered in detail below.

None of this needs to be exhaustive. A description that nails material, function, and construction for a garment does more work than one that also lists five irrelevant technical specifications. Specificity beats volume.

Can a Product Image Actually Help With Classification?

Yes, but less than most people assume, and it should never be the only input. An image confirms form, styling, and general construction. For products where visual presentation has direct bearing on classification (retail sets, certain apparel categories, packaged goods), that confirmation has real value.

What an image cannot do is tell the system what a product is made of, how it's built internally, what it's rated for, or what legal notes might apply to it. A photograph of a garment shows cut, color, and print. It does not show fiber content, knit versus woven construction, or how the item is marketed, and any of those three can change the classification outcome. The weight of evidence here points one direction: image plus description outperforms either alone, and description carries more of that weight than the image does. Treat the image as confirmation, not substance.

Why Jurisdiction Selection Is Not a Minor Step

The Harmonized System standardizes classification at six digits internationally. Past that point, countries diverge. The United States extends the six-digit HS code into its own Harmonized Tariff Schedule (HTS), typically to eight or ten digits. The European Union maintains a separate extended nomenclature, the Combined Nomenclature. A product classified correctly under one jurisdiction's extension is not automatically correct under another's, even when both share the same six-digit root.

What is the difference between HS and HTS classification? HS classification is the internationally standardized six-digit code most trading countries use as a common baseline. HTS classification is the United States' extension of that code to eight or ten digits, which sets the applicable U.S. duty rate and import requirements. Other countries maintain comparable extensions under different names.

A classification tool that defaults to a single schedule is only useful for one trade lane. Teams working multiple markets need to select jurisdiction deliberately, every time, because the answer changes with it.

Where AI Fits, and Where It Doesn't

AI accelerates the research and reasoning steps of classification. It does not replace the judgment call at the end of them. That's not a hedge, it's the actual shape of the work: classification frequently comes down to interpreting a section note, a prior ruling, or a fact pattern that reasonable classifiers could read two ways. A system that organizes the relevant rules and prior reasoning quickly is genuinely useful for that process. A system that outputs a code with no way to check the reasoning behind it is not something a compliance team should rely on unreviewed, regardless of how the tool is marketed.

The honest position: AI-assisted tools are strong at organizing and speeding up the research a classifier already has to do. They are not a substitute for review on classifications that carry real compliance or duty exposure. Teams that treat the AI output as a first draft, not a final answer, get the most out of this kind of tool.

Frequently Asked Questions

What information is needed for AI-assisted HS classification?
A description covering material, function, and intended use, at minimum. Additional detail (physical characteristics, manufacturing method, technical specifications, relevant legal notes) improves the reasoning when it's actually relevant to the product. Supporting files and images add context but work best alongside a written description, not instead of one.

Can an image be used for HS or HTS classification?
It can support classification but shouldn't carry it alone. An image confirms general form and construction. It cannot show material composition, technical specifications, or intended use, which are frequently the deciding factors in a classification. Pair it with a detailed description.

How detailed should a product description be for tariff classification?
As detailed as the product's classification actually requires, no more. A description that nails material, function, and construction for the relevant tariff heading outperforms a longer one padded with irrelevant detail. Specificity on the facts that matter, not length, is the target.

Why does jurisdiction matter for HTS classification?
The same product can carry different extended codes, duty rates, and requirements depending on the destination country's tariff schedule. HS is standardized internationally at six digits, but each country extends it differently, so a classification run against the wrong jurisdiction is not valid for the trade lane it's actually needed for.

Can AI classify products without a product image?
Yes. A detailed, accurate description is the primary input most classification systems reason from. An image adds context but isn't required when the text already covers the classification-relevant facts.

What is the difference between HS and HTS classification?
HS is the internationally standardized six-digit code most trading countries share. HTS is the United States' extension of that code to eight or ten digits, setting U.S.-specific duty rates and import requirements. Other countries maintain their own extensions in similar fashion.

Should AI-generated classifications be reviewed by a trade professional?
Yes, especially for anything ambiguous, high-value, or carrying compliance risk. AI-assisted tools speed up research and reasoning. They don't replace the judgment call a classification professional makes on a genuinely contested fact pattern.

Teams that classify products regularly and want to see how an AI-assisted workflow fits their process can start here: Try Trade Insight AI!