How to classify products
at scale, without losing accuracy.
Classifying a handful of products by hand is a manageable afternoon. Classifying ten thousand SKUs across a dozen destination markets is an entirely different problem, and the habits that work at small volume tend to break quietly, long before anyone notices. This guide covers how to build a classification process that actually holds up as the catalog grows.
9 min read · Trade compliance operations
Why scale changes the problem
Classifying a single product correctly is a research task. Classifying thousands of products correctly, on an ongoing basis, is an operations problem, and the two require almost entirely different approaches. A company with fifty SKUs can afford to have one knowledgeable person sit down and work through each one carefully. A company with fifty thousand SKUs, adding new products every week and shipping into a dozen destination countries, cannot run the same process and just do it more times. The math does not work.
The volume problem is compounded by a second one: variety. Large catalogs are rarely uniform. A single company might sell electronics components, packaging materials, and finished apparel under one roof, each of which sits in a completely different part of the tariff schedule and calls on different expertise to classify correctly. Add multiple destination countries, each with its own extensions past the shared six-digit code, and the number of correct answers a team needs to track grows far faster than the number of products.
None of this means scale makes accurate classification impossible. It means the process has to be designed for volume and variety from the outset, rather than being a manual workflow that a company hopes will keep holding up as the catalog grows.
Why manual processes break down first
Manual classification does not fail all at once. It degrades gradually, and the early signs are easy to miss because each individual classification still looks reasonable in isolation.
| Symptom | What is actually happening |
|---|---|
| Backlog keeps growing | New products are arriving faster than any reviewer can research and classify them. |
| Same product, different codes | Different reviewers, or the same reviewer on different days, reach different answers for similar items. |
| Codes go unreviewed for years | Nobody owns the job of rechecking codes after a product spec or a tariff schedule changes. |
| Reasoning lives in someone's head | The person who made the original call is the only one who can explain it, and eventually leaves. |
| New markets get rushed | Codes assigned quickly for a new destination country skip the second half of the schedule. |
Each of these is survivable on its own. Together, they compound. A backlog pushes reviewers to work faster, which increases inconsistency, which makes the eventual audit trail harder to reconstruct. The failure mode of manual classification at scale is rarely one dramatic error. It is a slow accumulation of small ones that surface all together, usually during an audit or a customs inquiry.
The building blocks of a scalable process
Before touching any tooling, a classification process that can handle volume needs four things in place. These apply whether the classification itself is done by people, software, or some mix of the two.
A structured product description, not a product name
At volume, "product name" is not a usable input. What holds up is a consistent record of material, function, and how the item is sold, captured the same way for every product so it can be compared and reasoned about systematically.
A single source of truth for each code
Every product needs one current, authoritative code per destination market, along with the reasoning behind it, stored somewhere the whole team can see, rather than scattered across spreadsheets, emails, and one reviewer's notes.
A trigger for rechecking, not just assigning
Codes need to be revisited when a product's materials or function change, and separately, on a periodic cycle, because tariff schedules themselves are revised. Neither of these happens automatically unless something is built to catch them.
A defined role for human review
Even a heavily automated process needs a clear answer to the question of which products get a person's eyes before the code is finalized, and which do not. Deciding this in advance, based on ambiguity or risk, is far better than deciding it case by case under time pressure.
A workflow that holds up at volume
In practice, most teams that classify successfully at scale converge on a similar shape, even if the tools underneath differ. It separates the parts of the job that scale well from the parts that genuinely need a person.
The part most teams underestimate is step three. Confidence is not a nice-to-have; it is what makes the whole workflow scale, because it decides which products a person actually needs to look at. Without some way to separate a routine, high-confidence match from a genuinely borderline one, teams either review everything, which does not scale, or review nothing, which is how errors accumulate unnoticed.
Quality control at volume
Volume also changes what quality control looks like. Checking every code individually is not realistic once a catalog reaches a few thousand products, so the check has to shift from individual verification to sampling and pattern detection.
- Sample a percentage of high-confidence, automatically classified products for spot review, rather than assuming they are all correct.
- Review 100% of anything flagged as low-confidence or genuinely ambiguous before it ships.
- Track which product categories generate the most reviewer overrides, since that pattern usually points to a gap in the underlying product data, not a one-off mistake.
- Re-run the full catalog, or at least the affected chapters, whenever a tariff schedule update lands, rather than waiting for individual products to be touched again.
The goal of quality control at scale is not zero errors on the first pass. It is catching the errors that do occur before they turn into a pattern across thousands of shipments.
Common pitfalls when scaling up
Automating before the product data is clean
Bulk classification run against vague or inconsistent product descriptions just produces bulk errors, faster. Cleaning up how products are described is a precondition for scaling, not an optional step to revisit later.
Treating every market's extension as an afterthought
Getting the shared six-digit code right and then guessing at each country's extra digits works fine for one destination. It falls apart fast once a company is shipping into ten markets with ten different national schedules.
No process for catalog changes
A workflow built for the current catalog quietly stops working the moment a new product line is added, unless there is a defined step for bringing new SKUs into the same pipeline as everything else.
Losing the reasoning during the handoff to automation
When a manual process moves to a more automated one, it is easy to keep the codes but lose the documented reasoning behind them. That reasoning is exactly what gets asked for during an audit, so it needs to travel with the code, not get left behind.
Getting started
Teams that scale classification successfully tend to sequence the work the same way, regardless of catalog size:
- Audit the current state first: how many products, how many destination markets, and how consistent the existing codes actually are.
- Standardize product descriptions before attempting to automate anything on top of them.
- Decide, in advance, what counts as low-confidence or ambiguous enough to route to a person.
- Build in a recheck trigger for both product changes and periodic tariff schedule updates, rather than relying on someone to remember.
None of this requires solving every edge case on day one. It requires a process that treats classification as ongoing infrastructure rather than a one-time project, since the catalog, the destination markets, and the underlying schedules will all keep changing after the first pass is done.
See classification
run across your entire catalog.
Enthron classifies at volume against national tariff schedules across dozens of countries, flags low-confidence results for review, and keeps every determination up to date as the underlying rules change.