One of the less obvious problems that comes with scaling an import business is figuring out how to keep product classification under control.
At 100 products, manually reviewing HS classifications may be perfectly manageable.
At 10,000 products, it becomes a different problem.
The challenge isn't just the number of classifications. It's the amount of product information that has to be reviewed, the consistency of decisions, the changes to products and tariffs over time, and the amount of specialist time required to keep everything current.
When companies think about scaling operations, they usually plan for more orders, customers, suppliers, inventory, and employees.
Classification often sits somewhere in the background.
Then the catalog grows.
A new supplier provides incomplete product descriptions. Another product uses a different material. A seemingly similar SKU has a different intended use. Someone copies the HS code from an older product because it looks close enough.
Individually, these decisions may seem harmless.
Across thousands of products, they can create a serious data-quality and compliance problem.
Adding people solves the immediate workload problem, but it doesn't necessarily solve consistency.
Different specialists can interpret complicated products differently. The same specialist may also reach a different conclusion later if the original reasoning wasn't recorded properly.
And every additional person adds another operational dependency.
That doesn't mean manual classification is bad.
Quite the opposite.
For difficult products, human expertise is extremely valuable.
The question is whether specialists should spend their time manually processing every straightforward product or concentrating on the products that actually require judgement.
This is where automated HS classification becomes interesting from a SaaS and operations perspective.
Instead of treating every SKU as an identical task, a system can evaluate product attributes such as:
Material
Function
Intended use
Product form
Degree of processing
Technical specifications
Packaging and presentation
Straightforward products can move through an automated workflow, while uncertain cases can be separated for human review.
That distinction matters.
The goal isn't necessarily to automate 100% of classifications.
The goal is to reduce the amount of repetitive work that experts have to perform.
For a deeper breakdown of the trade-offs, including cost, accuracy, consistency, auditability, and maintenance, this manual vs. automated HS classification comparison provides a useful framework.
There's a tendency to frame AI automation as:
Human does the work → AI replaces human
For classification, that isn't necessarily the most useful model.
A better workflow can look more like:
Product data → automated analysis → confidence/risk assessment → human review when required → approved classification
That changes what the specialist spends time doing.
Instead of reviewing thousands of routine items, they can focus on unusual products, incomplete data, difficult tariff boundaries, and decisions with significant financial or compliance consequences.
That's a much more realistic use of AI.
If you're evaluating an AI system for a compliance-related workflow, I'd be more interested in how it handles uncertainty than how impressive its best-case demo looks.
A system that confidently produces an answer for everything isn't necessarily better.
What happens when the product description is incomplete?
What happens when two tariff provisions appear plausible?
What happens with a genuinely novel product?
What happens when the system isn't confident?
A useful system should have a mechanism for identifying those situations rather than forcing every product into an automated answer.
There is another scaling issue that is easy to overlook.
A classification isn't necessarily "set it and forget it."
Tariff nomenclature changes. Products change. Regulations and rulings can change. A business may also expand into new markets with different tariff structures.
So the technology needs to deal with versioning and reassessment, not just initial classification.
Otherwise, a company can end up with a beautifully automated workflow producing classifications based on outdated information.
Automation without maintenance is just technical debt wearing a compliance badge.
If I were evaluating an automated classification workflow, I wouldn't start with:
"How many products can it classify per minute?"
I'd look at:
1. Review rate
What percentage of products require human intervention?
2. Exception quality
Does the system send the genuinely difficult cases to specialists?
3. Classification consistency
Are equivalent products being handled consistently?
4. Reasoning capture
Can someone understand why a classification was assigned?
5. Reclassification effort
How easily can affected products be reassessed when tariff information changes?
6. Data quality
How does the system react when product information is incomplete?
Those metrics tell you much more about whether automation is actually solving the operational problem.
Reducing manual work is an obvious benefit.
But there can be a bigger benefit in building a more controlled classification process.
A centralized workflow can make it easier to understand which products have been classified, who reviewed them, what information was used, and which decisions need attention.
That becomes increasingly valuable as the business grows across suppliers, products, employees, and markets.
The interesting business question isn't:
"Can AI classify products?"
It's:
"Can AI help a growing company manage classification at scale without losing control over the decisions?"
That's a much harder problem, but also a much more useful one to solve.
Further reading: Manual vs. Automated HS Classification: Comparing Cost and Accuracy